UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Modelling short-range path loss in smart meter mesh networks Lancashire, Sol Joseph Lapierre 2016

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
24-ubc_2016_september_lancashire_sol.pdf [ 5.1MB ]
Metadata
JSON: 24-1.0300442.json
JSON-LD: 24-1.0300442-ld.json
RDF/XML (Pretty): 24-1.0300442-rdf.xml
RDF/JSON: 24-1.0300442-rdf.json
Turtle: 24-1.0300442-turtle.txt
N-Triples: 24-1.0300442-rdf-ntriples.txt
Original Record: 24-1.0300442-source.json
Full Text
24-1.0300442-fulltext.txt
Citation
24-1.0300442.ris

Full Text

  MODELLING SHORT-RANGE PATH LOSS IN SMART METER MESH NETWORKS  by   Sol Joseph Lapierre Lancashire B.Eng., University of Victoria, 2000    A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF   MASTER OF APPLIED SCIENCE in THE FACULTY OF GRADUATE AND POSTDOCTORAL STUDIES (Electrical and Computer Engineering)    THE UNIVERSITY OF BRITISH COLUMBIA (Vancouver)   May 2016   © Sol Joseph Lapierre Lancashire, 2016 ii  Abstract  Electrical power utilities have begun to deploy IPv6-based Low power and Lossy wireless mesh Networks (LLNs) in suburban areas to support smart metering, distribution automation, interaction with customers, and demand response. To date, most insights concerning network performance have been obtained through simulations based upon simplistic channel models and measurements conducted using testbeds of limited size and extent. Here, we seek to overcome the limitations of previous work and develop better insights into the factors that affect LLN performance by analyzing data collected from BC Hydro’s 1.9-million-node Multi-Service Grid Network (MSGN). The network includes both device-to-device (D2D) and device-to-infrastructure (D2I) links. First, we review the essential aspects of the MSGN and propose an incremental strategy based on using network performance data to: 1) develop measurement-based short-range path loss models that will make network simulations more accurate, 2) identify correlations between network layout and performance, and 3) develop schemes for optimizing network performance through infilling relay nodes when the node density is low and transmit power adjustment when the node density is high. Second, we conclude that because the links are obstructed, power law path loss models generally apply. We show how distance errors or finite distance spans may degrade estimates of the model parameters in typical environments and should be accounted for in the development of path loss models for LLNs. Third, after developing procedures for managing and preparing such data for analysis, we reduce path loss data collected in three representative areas with hilly terrain and light vegetation, hilly terrain and heavy foliage, and flat terrain and light foliage and compare the results to previous work. The results show that: 1) modelling path loss using data collected over short distance spans and with substantial distance errors is challenging but can be accomplished with appropriate care, 2) the path loss experienced over such short D2D and D2I links is fairly independent of terrain and foliage density, and 3) data analytics can greatly improve modelling efficiency and support network tuning and optimization. The results serve as an important foundation for the remainder of our proposed strategy. iii  Preface  This thesis presents original research conducted by me under the supervision of Prof. David G. Michelson.  Prof. Michelson and I collaborated in the design of both the research plan and the thesis. I managed the PostgreSQL database that served as a repository for the network performance data collected from BC Hydro’s Multi-Service Grid Network (MSGN), conducted all of the pre-analysis cleaning and processing of the data, and performed all of the Tableau-based statistical analysis of the data. Hamed Noori and Zahra Vali wrote the Matlab scripts used to generate the results presented in Chapter 3 concerning the effects of distance errors and finite distance spans on the estimation of power law path loss model parameters. BC Hydro provided access to the data, software tools, and the resources required to set up and manage the MSGN performance database. However, the research reported here was conducted independent of my regular employment responsibilities with the company.  iv  Table of Contents  Abstract .......................................................................................................................................... ii	Preface ........................................................................................................................................... iii	Table of Contents ......................................................................................................................... iv	List of Tables ................................................................................................................................ vi	List of Figures .............................................................................................................................. vii	List of Abbreviations ................................................................................................................... ix	Acknowledgements ........................................................................................................................x	Dedication ..................................................................................................................................... xi	Chapter 1: Introduction ................................................................................................................1	1.1	 Grid Modernization ......................................................................................................... 1	1.2	 Wireless Connectivity for the Distribution Grid ............................................................. 2	1.3	 BC Hydro Multi Service Grid Network .......................................................................... 4	1.4	 Objectives and Outline .................................................................................................... 5	Chapter 2: Optimizing and Tuning Wireless Mesh Networks in Urban Areas .......................7	2.1	 Introduction ..................................................................................................................... 7	2.2	 Physical vs. Network Layer Optimization ...................................................................... 8	2.3	 BC Hydro’s Multi-Service Grid Network ....................................................................... 9	2.3.1	 Network Layout ...................................................................................................... 9	2.3.2	 Network Performance Data ................................................................................... 10	2.3.2.1	 Metrics Available from the NMS ..................................................................... 10	2.3.2.2	 Metrics Available from the Collector but Not Collected by the NMS ............. 11	2.3.2.3	 Metrics from the Meter Data Management System .......................................... 11	2.3.2.4	 Metrics That can be Calculated from Available Data Sources ......................... 11	2.3.3	 Network Data Management .................................................................................. 12	2.3.4	 Database and Visualization Tools ......................................................................... 12	2.4	 A Strategy for Development of U-LLN Tuning and Optimization Techniques ........... 14	2.4.1	 Path Loss Modelling ............................................................................................. 14	2.4.2	 Correlation of Performance Metrics with Design Parameters .............................. 15	2.4.3	 Alternative Tuning and Optimization Strategies .................................................. 15	v  2.4.3.1	 Conventional Relays ......................................................................................... 16	2.4.3.2	 Street Light Relays ............................................................................................ 16	2.4.3.3	 Power Control or Adjustment ........................................................................... 16	2.4.3.4	 Modulation ........................................................................................................ 17	2.5	 Discussion ..................................................................................................................... 17	Chapter 3: Distance Effects in the Estimation of Power Law Path Loss Models ..................18	3.1	 Introduction ................................................................................................................... 18	3.2	 Model Formulation ....................................................................................................... 19	3.3	 Effect of Limited Distance Spans on Model Parameter Estimation ............................. 20	3.4	 Effect of Distance Errors on Model Parameter Estimation .......................................... 25	3.5	 Discussion ..................................................................................................................... 27	Chapter 4: Short-Range Power-Law Path Loss Models in Suburban Environments ..........28	4.1	 Introduction ................................................................................................................... 28	4.2	 Concepts ........................................................................................................................ 29	4.3	 Methodology ................................................................................................................. 30	4.4	 Results ........................................................................................................................... 32	4.4.1	 Measurement Database ......................................................................................... 32	4.4.2	 Device to Infrastructure Path Loss Models ........................................................... 41	4.4.3	 Device to Device Path Loss Models ..................................................................... 47	4.5	 Comparison with other Path Loss Models .................................................................... 53	4.6	 Discussion ..................................................................................................................... 56	Chapter 5: Conclusions and Recommendations .......................................................................57	5.1	 Conclusions ................................................................................................................... 57	5.2	 Recommendations ......................................................................................................... 59	References .....................................................................................................................................61	   vi  List of Tables    Table 4-1: Quantities of meters and collectors in each region. ..................................................... 34	Table 4-2: Quantities of D2I and D2D links and measurements for each region. ........................ 34	Table 4-3: Node distribution statistics for each region. ................................................................ 34	Table 4-4: Device-to-Infrastructure path length statistics for each region. .................................. 34	Table 4-5: Device-to-Device path length statistics for each region. ............................................. 34	Table 4-6: Device to Infrastructure model parameters for different regions. ............................... 42	Table 4-7: Device-to-Device model parameters for different regions. ......................................... 48	Table 4-8: Parameters for several short-range path loss models. ................................................. 55	   vii  List of Figures  Figure 1-1: Elements of the electrical power distribution system. ................................................. 1	Figure 1-2: BC Hydro’s Smart Grid applications framework. ....................................................... 2	Figure 2-1: BC Hydro Multi Service Grid Network Coverage Map. ........................................... 10	Figure 3-1: Mean value of correlation coefficient vs. distance span, d2/d1. ................................ 21	Figure 3-2: Mean value of error in path loss exponent, n vs. distance span, d2/d1. ..................... 21	Figure 3-3: Error in path loss intercept, PL0 vs. distance span, d2/d1. ......................................... 22	Figure 3-4: Error in shadow fading coefficient, σ vs. distance span, d2/d1. ................................ 22	Figure 3-5: Distribution of the correlation coefficient vs. distance span, d2/d1. .......................... 23	Figure 3-6: Distribution of errors in n vs. distance span, d2/d1. .................................................. 24	Figure 3-7: Mean value of correlation coefficient vs. distance error, σd. ..................................... 25	Figure 3-8: Mean value of error in n vs. distance error, σd. ......................................................... 26	Figure 3-9: Mean value of error in PL0 vs. distance error, σd. ..................................................... 26	Figure 4-1: Gray scale overlay showing the density of smart meters in Richmond/Delta. .......... 35	Figure 4-2: Map showing device-to-infrastructure links in Richmond/Delta that are used 2 or more times. .................................................................................................................................... 36	Figure 4-3: Distribution of measured RSSI values in dBm for D2I links. ................................... 37	Figure 4-4: Distribution of measured RSSI values in dBm for D2D links. .................................. 37	Figure 4-5: Distribution showing the number of neighbouring nodes within a 50 metre radius. . 38	Figure 4-6: Distribution showing the number of neighbouring nodes within 1 square kilometre.38	Figure 4-7: Distribution of path lengths for D2I scenarios. .......................................................... 39	Figure 4-8: Distribution of path lengths for D2D scenarios. ........................................................ 39	Figure 4-9: D2D Path length vs. node density. ............................................................................. 40	Figure 4-10: Distribution showing the re-use of D2I links over the 119 day analysis window. .. 40	Figure 4-11: Distribution showing the re-use of D2D links over the 119 day analysis window. . 41	Figure 4-12: Example of the RSSI vs. path length for D2I link to the ACAD collector in Kamloops. ..................................................................................................................................... 43	Figure 4-13: Example of the Residual vs Path Length for D2I links to the ACAD collector. ..... 43	viii  Figure 4-14: Regression lines for Coquitlam terrain type A - hilly with moderate to heavy tree density. .......................................................................................................................................... 44	Figure 4-15: Regression lines for Kamloops terrain type B - hilly with light tree density. .......... 44	Figure 4-16: Regression lines for Richmond and Delta terrain type C - flat with light tree density........................................................................................................................................................ 45	Figure 4-17: Regression lines for RSSI vs. Path Length for D2I links in all regions. .................. 45	Figure 4-18: Distribution of path loss exponents for D2I scenarios. ............................................ 46	Figure 4-19: Distribution of the residuals for D2I scenarios. ....................................................... 46	Figure 4-20: Example of the RSSI vs. path length for D2D links below the ACAD collector in Kamloops. ..................................................................................................................................... 49	Figure 4-21: Example of the Residual vs. Path Length for D2D links below the ACAD collector........................................................................................................................................................ 49	Figure 4-22: Regression lines for Coquitlam terrain type A - hilly with moderate to heavy tree density. .......................................................................................................................................... 50	Figure 4-23: Regression lines for Kamloops terrain type B - hilly with light tree density. .......... 51	Figure 4-24: Regression lines for Richmond and Delta terrain type C - flat with light tree density........................................................................................................................................................ 51	Figure 4-25: Regression lines for RSSI vs. Path Length for D2D links in all regions. ................ 51	Figure 4-26: Distribution of path loss exponents for D2D scenarios. .......................................... 52	Figure 4-27: Distribution of the residuals for D2D scenarios. ...................................................... 52	Figure 4-28: Comparison of mean path loss vs. distance for several short-range path loss models........................................................................................................................................................ 53	 ix  List of Abbreviations  D2D - Device to Device D2I - Device to Infrastructure ETX - Expected Transmission Count GIS - Geographic Information System I2I – Infrastructure to Infrastructure IETF – Internet Engineering Task Force IoT - Internet of Things LLN – Low Power and Lossy Networks LOS - line-of-sight NMS - Network Management System  MDMS - Meter Data Management System NLOS - Non-line-of-sight OF0 - Objective Function Zero OF1 - Objective Function One ORS - Openway Reporting System RDBMS - Relational Database Management System RFC – Request for Comment ROLL - Routing Over Low power and Lossy networks RPL - Routing Protocol for Low-power and lossy networks RSSI – Received Signal Strength Indication V2V - vehicle to vehicle WSN - Wireless Sensor Network x  Acknowledgements  I am grateful to my professional colleagues, especially Dr. Vidya Vankayala, Director of Smart Utilities at Powertech Labs; Wayne Cross, Principal Engineer at BC Hydro; David DeYagher, Senior Manager for Smart Metering Operations and Deployment at BC Hydro; and Anthony Blake of CaZaTech Consulting for their support and assistance as I have pursued this project.  I owe particular thanks to my research advisor, Prof. David G. Michelson, who contributed so much to the research strategy, showed me how academic rigour and discipline can be so usefully applied in an industrial setting, and shared so much of his encyclopedic knowledge and real world experience with me during the course of this work. I thank Stephen Johnson, Director Solutions Architecture and Michael Belanger, Senior Product Line Manager, both at Itron; Rob Barton, Systems Engineer, Dr. JP Vasseur, Cisco Fellow & Chief Architect of Self Learning Networks, Eruch Kapadia, Chief Architect Internet of Everything, Connected Energy, all at Cisco Systems; Ted Thomas, Director, Smart Grid Emerging Technologies at Duke Energy Corporation, and Phil Beecher, Chair of the WiSUN Alliance and IEEE 802.15.4 Task Group 4g for helpful discussions at various times during the course of this work. I also thank the Pacific Institute for Climate Solutions (PICS) and the Clean Energy Research Centre (CERC) at UBC for their support and encouragement during the latter stages of this work. I am, of course, most grateful to my family for allowing me to devote so much of my time after hours to my thesis project.   xi  Dedication  To my lovely wife, Dorothy, and wonderful children, Matthew and Ella.  1 Chapter 1: Introduction 1.1 Grid Modernization Electrical power distribution systems are traditionally divided into four functional blocks: Generation, Transmission, Distribution, and Customer, as depicted in Figure 1-1. Since the early 2000s, there has been a worldwide effort to modernize the architecture of electrical power systems and prepare for a new era in which distributed generation (DG), demand response (DR), voltage and VAR control, automated fault location, isolation, and service restoration (FLISR) and other advanced features and services will be fully supported [1][2][3][4][5]. The North American effort has been led by EPRI, NIST, and the IEEE. In 2004, EPRI released the IntelligridSM Architecture. In 2009, NIST established the Smart Grid Interoperability Panel. In 2011, IEEE Standards Association released the IEEE 2030 standard which provides guidelines for smart grid interoperability [6]. In their contributions, each organization has emphasized the need to develop and deploy information communication technologies (ICT) that provide efficient, reliable and secure connectivity between the geographically dispersed functional blocks that comprise the electrical power system in order to ensure that the system can quickly and effectively respond to changing needs and conditions. Until the advent of grid modernization, the bulk of the ICT infrastructure in the electrical power system was deployed within the generation and transmission infrastructure and ended at the distribution substations. Smart meter networks are fundamental to grid modernization because they extend the electrical power utility’s ICT infrastructure past the distribution   Figure 1-1: Elements of the electrical power distribution system [7].  2 substations into the distribution grid and customer premises. The original role of such networks was to provide connectivity to the smart meters used to: 1) automatically record energy used by a customer, 2) provide a real-time indication of service interruption to the customer, and, eventually, 3) allow interaction with the customer and control their discretionary loads. However, current plans and strategies acknowledge that such networks can also provide connectivity to other devices in the distribution grid including: 1) sensors that provide near real-time measurements of load, voltage, power factor, 2) devices used to control switches, voltage regulators, and other distribution automation equipment, and 3) faulted circuit indicators that are key to automated fault location, isolation, and service restoration (as noted above) [7]. BC Hydro’s Smart Grid applications framework is depicted in Figure 1-2.    Figure 1-2: BC Hydro’s Smart Grid applications framework [8]. 1.2 Wireless Connectivity for the Distribution Grid The distribution grid includes several orders of magnitude more nodes than the generation and transmission portions of the electric power network and covers a far larger geographic extent. The architects of grid modernization realized that this precluded the re-use of the relatively expensive ICT technology used in generation and transmission grid and sought lower cost  3 alternatives. The need for suitable radio transmission technology was addressed by IEEE 802.15 Smart Utility Networks (SUN) Task Group 4g [9]. Their goal was to create a PHY amendment to IEEE 802.15.4 that would provide a global standard that will facilitate very large scale process control applications. This includes smart-grid networks capable of supporting large, geographically diverse networks with minimal infrastructure, with potentially millions of fixed endpoints. The principal attributes of the IEEE 802.15.4g include: 1) operation in both licenced and licence-exempt bands between 700 MHz and 2.5 GHz, 2) data rates between 40 and 1000 kbps, 3) PHY frame sizes that permit an IP packet to be transmitted without fragmentation and 4) incorporation of mechanisms for coexistence with other systems operating in the same licence-exmpt bands. Because many vendors had already released proprietary solutions, the standardization effort faced many challenges but was ultimately successful. The details of this effort has been described in detail in [10]. At about the same time, members of the Internet Engineering Task Force (IETF) realized that the capabilities of existing wireless sensor network (WSN) technology were severely limited and began to develop an enhanced architecture that would allow WSNs to directly connect to the Internet and incorporate other advanced features. The IETF chartered the 6LoWPAN (IPv6 in Low-Power Wireless Personal Area Network) and ROLL (Routing over Low-Power and Lossy Links) working groups to specify standards at various layers of the protocol stack that would allow Low-power and Lossy Networks (LLNs) to connect to the Internet. The standardization work conducted by these groups has been described in depth in [10]. The combination of IEEE 802.15.4g radio transmission technology with IPv6 based LLNs provides a powerful platform for implementing Smart Grid, Smart Utilities and Smart City applications. Such networks are often referred to as Urban LLNs or U-LLNs to distinguish them from LLNs used in healthcare, industrial automation, building automation and home automation. Much effort has been devoted to understanding the factors that affect the performance of U-LLNs. The principal approaches have included software-based simulations and testbed-based experimentation. The latter involves a multiplicity of network nodes that have been suitably deployed or interconnected. Issues of interest to researchers and developers have included: neighbour discovery, route selection, convergence and repair mechanisms, throughput and delay performance analysis. Kermanjani [11] analyzed the impact of RPL (Routing Protocol for Low power and lossy networks) and 6LoWPAN neighbour discovery parameters on link availability  4 and end-to-end path availability, and showed that careful tuning of the relevant parameters is critical for obtaining good network performance. Lee et al. [12] used the Contiki test-bed to analyze 6LoWPAN performance with different RSSI and payload sizes and proposes RSSI based routing metrics. Gaddour et al. [13] used ContikiRPL to simulate and analyze network formation with RPL, and compares OF1 (Objective Function 1) which minimizes ETX (Expected Transmission count) and OF0 (Objective Function 0) which minimizes the number of hops [14]. Gaddour et al. also evaluated stability in terms of throughput and number of parents available to each node. The principal limitation of these works has been lack of realism in the airlink model or implementation which make the results obtained less useful when applied to operational networks.  Like all wireless networks, the performance of U-LLNs is highly affected by the wireless propagation environment. However, short-range wireless links in urban and suburban environments have not been well characterized to date. Of the few studies that have been reported, some concern path lengths that are far shorter than those typically found in smart meter networks, others focus on links that are far less obstructed than those typically found in smart meter networks, or were conducted in frequency bands far removed from those mandated for use by IEEE 802.15.4g. None characterize the variability of the links day to day or the reliability of the links over the long term. We conclude that there is a pressing need to develop short range path loss, short-term variability and long-term reliability model for use in the design and simulation of U-LLNs used in smart grid applications. While characterization based upon data collected via a large-scale, operational network is likely the most cost-effective option, we have not been able to locate any results of this sort in the literature to date.  1.3 BC Hydro Multi Service Grid Network Since 2010, BC Hydro has deployed one of the largest U-LLNs in the world. Its Multi Service Grid Network (MSGN) provides connectivity to 1.9 million Smart Meters across its entire service territory which covers 95% of the population of British Columbia. The MSGN covers a broad range of environments: 1) Terrain ranges from flat to rolling, 2) Foliage ranges from very light to dense, 3) Building cover includes urban, light urban, suburban and rural, 4) Meter densities range from 20,000 smart meters per square kilometre in Metro Vancouver to only a few smart meters per square kilometre in certain locations in the BC Interior. The scale and range of  5 physical conditions offers an exceptional opportunity to analyze the physical operating environment (path loss) and network behaviours experienced by a large low-power and lossy wireless sensor network.  1.4 Objectives and Outline The objective of this study is to take the first steps toward using data analytics (the science of examining raw data with the purpose of drawing conclusions and making effective decisions) to examine raw performance data collected from BC Hydro’s Multi-Service Grid Network and reveal the environmental, design and configuration factors that affect the performance of U-LLNs. In this manner, we hope to overcome the limitations of previous efforts based upon software simulations and testbed experimentation and contribute to the development of effective and efficient techniques for tuning and optimizing operational U-LLNs. Such techniques will be critical as the smart grid networks grow and evolve and in order to meet the needs of increasingly demanding applications. The specific goals of this work are to: 1) propose a strategy for developing U-LLN tuning and optimization methods based upon analysis of network data collected from BC Hydro’s Multi-Service Grid Network (MSGN), 2) assess the impacts of errors in transmitter-receiver distance and limited distance spans on the estimation of power law path loss models from measured data and 3) to demonstrate both the operational and technical feasibility of formulating short-range power law path loss models applicable to device-to-device and device-to-infrastructure links based upon analysis of network data collected from BC Hydro’s MSGN. Such path loss models will: 1) assist in the interpretation of higher level network performance data, 2) improve the accuracy of U-LNN simulations and testbeds that represent similar environments, and 3) inform the design of tuning and optimization strategies. The remainder of this thesis is organized as follows: In Chapter 2, we: 1) review the essential aspects of the BC Hydro Multi-Service Grid Network including the physical layout and network structure, 2) identify key performance parameters that are generated by the network and review the process for collecting and managing the performance data, and 3) propose a progressive strategy for using MSGN performance data to develop effective and efficient U-LLN tuning and optimization methods.  6 In Chapter 3, we: 1) review the essential aspects of linear regression and correlation as they apply to estimation of power law path loss models, 2) consider the effect of limited distance span on the estimation of power law path loss model parameters over the range of typical values of path loss exponent and shadow fading and 3) consider the effect of distance errors on the estimation of power law path loss model parameters over the range of typical values of path loss exponent and shadow fading. In Chapter 4, we: 1) review the limitations of past efforts to characterize path loss over short-range wireless links in suburban environments and summarize our approach to overcoming these limitations using performance data collected from BC Hydro’s MSGN, 2) describe our method for collecting, managing, processing and reducing the network performance data, 3) present the statistics of our data set, our estimates of the power law path loss models that best fit the device-to-device and device-to-infrastructure links in our data set, and compare our results to previous work.  In Chapter 5, we present our conclusions and offer recommendations for further work, including: 1) continued development of short-range D2I,  D2D, and I2I path loss models and 2) pursuit of the second and third steps of our progressive strategy for using MSGN performance data to develop LLN tuning and optimization methods that incorporate conventional relays, street light controllers and transmit power adjustment as appropriate.    7 Chapter 2: Optimizing and Tuning Wireless Mesh Networks in Urban Areas  2.1 Introduction Wireless mesh networks are an alternative to conventional point-to-multipoint wireless networks that allows a multiplicity of low power (and low cost) transceivers to provide wide area coverage. The essential aspects of wireless mesh networks have been described in several review papers [15][16][17]. Wireless mesh networks are particularly associated with low power wireless sensor networks and wireless personal area networking standards developed by the IEEE 802.15.4 working group and promoted by the ZigBee and SUN alliances. However, mesh architectures may be used by other short range wireless technologies such as wireless local area networking standards developed by IEEE 802.11 and promoted by the WiFi Alliance as well. The advantages of mesh architectures come at the expense of increased latency due to the need for data packets to traverse multiple nodes on their way from the source to the destination and the relatively complicated protocols used to determine the route that the packet should take.  The performance of wireless mesh networks will suffer if packets are directed to traverse too many hops or if packets are directed to traverse unreliable links with the result that transmissions fail and retransmissions are required. Transmissions may also fail if the path loss between nodes is too large and the received signal strength is too low. Alternatively, they may fail if too many nodes in the same general vicinity attempt to transmit and packets collide. When IETF’s RoLL (Routing over Low-power and Lossy networks) working group proposed to develop RPL (Routing Protocol for Low-power and lossy IPv6 networks), they sought to build on previous wireless mesh routing protocols and meet the requirements of four different target environments: 1) urban environments [18], 2) industrial low-power-networks [19], 3) home automation [20], and 4) building automation [21] . Low-power and Lossy Networks designed to operate in urban environments are referred to as U-LLNs. The essential aspects of RPL development have been reported by Ko et al. [10].  In 2015, the network protocol of BC Hydro’s Multi Service Grid Network (MSGN) that supports smart metering and other applications was converted from an ANSI C12.22-based protocol to a U-LLN based on RPL/IPv6. It was the first full scale production RPL/IPv6 meter system deployment by Itron and Cisco and one of the first by any Smart Meter network  8 manufacturer. The conversion from RFLAN to IPv6 took place on a live operating network and required a substantial amount of performance analysis and tuning to ensure success of the conversion. Any stranded meters would have to be physically visited so their firmware could be upgraded manually. At the time, performance analysis was based primarily on “meter daily read rate” and tuning relied heavily on highly manual and costly field surveys. With almost two million nodes in the network, continued reliance on manual tuning and optimization was deemed impractical. BC Hydro and its suppliers were motivated to investigate opportunities to automate performance analysis and develop efficient techniques for network tuning. Since then, inspection of performance data has revealed that further tuning and optimization are required in order to meet future applications such as demand response. However, more detailed analysis is required to determine the appropriate steps. In this chapter, we propose an incremental strategy for developing physical layer optimization techniques that will meet the needs of BC Hydro’s MSGN and similar U-LLNs. In Section 2.2, we distinguish between physical and network layer optimization. In Section 2.3, we summarize the essential characteristics of BC Hydro’s Multi Service Grid Network. In Section 2.4, we present our incremental strategy and review the tuning options available to network developers.  2.2 Physical vs. Network Layer Optimization Tuning and optimization of a U-LLN may take place on two levels: the physical layer and network layer. We define physical layer optimization as adjustment of the network to ensure that each node has reliable links to between, say, three and five of its nearest neighbours. If the number of links is less than this, the node runs the risk of being isolated or stranded if one or more of the links becomes unreliable or drops completely. If the number of links is greater than this, the node runs the risk of interfering with or being interfered by other nodes. We define network layer optimization as adjustment of media access, routing and other protocols and settings in a manner that accounts for shortcomings in node connectivity and the impact of link variability and reliability. As a general rule, physical layer optimization should always be the first step in efforts to improve network performance. Ultimately, a physical tuning adjustment seeks to either increase the connectivity of a node that has too few reliable links to its nearest neighbours or reduce the connectivity of a node that  9 has too many links to its nearest neighbours. There are a limited set of physical layer optimization options available.  Some of these options are more difficult to implement that others. The number and location of the meter nodes themselves are pre-determined by the customer base and cannot be changed. The sheer number of customers makes it cost prohibitive to return to replace or make any other adjustments once the meter nodes are installed. Over-the-air adjustments are permissible but the risk of misconfiguration that strands or isolates a meter and requires a site visit to correct is a significant concern.  2.3 BC Hydro’s Multi-Service Grid Network 2.3.1 Network Layout BC Hydro’s Multi Service Grid Network (MSGN) provides connectivity to over 1.9 million smart meters across its entire service territory. One hundred load control devices were recently added on a trial basis. The Itron Centron meters deployed by BC Hydro implement the air interface using a TI CC1101 transceiver and an ARM Cortex M3 Core based custom ASIC MCU [21]. New nodes (e.g., Distribution Automation Gateway Modem, or Demand Response Load Control Modules) are based on the TI CC1200 transceiver. The RF output power can be adjusted from 200 to 630 mW. Cisco’s implementation uses 802.15.4g mode 2. which is 2FSK. Symbol duration is 20 µs. Channel Dwell time is 145  µs. The MSGN has a coverage area of ~95,000 square kilometres. While this represents just 10% of the total area of the province, it covers 95% of the population. Notable exceptions include the cities of Kelowna and of Penticton among other cities and towns in the south eastern part of the province which are part of the Fortis BC’s service territory. The node density varies between 1 node per square kilometre to almost 20,000 nodes per square kilometre. with an average of almost 1000 nodes per square kilometre. A map depicting the MSGN service areas (billing regions) is presented in Figure 2.1.  The majority of the smart meters and load control devices in the network are elements of a wireless mesh. A small fraction of the meters link to almost 2400 collectors that connect via WiMAX, cellular and satellite links to the MSGN Network Management System (NMS). Over 5000 conventional relay nodes or range extenders (0.25% of the total number of nodes) are installed throughout the mesh to enhance the connectivity of meters that are isolated by excessive distance or path loss. Meters and relays may be installed on the sides of single detached homes,  10 commercial buildings, or industrial buildings, or in meter closets within multiple dwelling units. Collectors and relays may be mounted on pads or on utility poles.  2.3.2 Network Performance Data Cisco’s mesh network implementation is based on IEEE 802.15.4g, 6LoWPAN, and RPL. Since Cisco uses storing mode, all the fundamental performance metrics are managed by and stored in the Collector Router. Many of the metrics gathered by the router are collected by the NMS and available for analysis.   2.3.2.1 Metrics Available from the NMS 1. RSSI. RSSI samples are recorded in both the forward and reverse directions. The measurement is by either by the TI CC1101 (for Centron Meters) or CC1200 (for DAGW and DR devices). In Cisco’s SUN implementation a node will not join the network until it   Figure 2-1: BC Hydro Multi Service Grid Network coverage map.   11 finds a parent with an RSSI of -95 dBm or better. Once connected it will remain on the network to the node receiver threshold at -105 dBm. The RSSI sample is taken from the RF SoC. (TI CC1101 or CC1200). The RSSI samples are of machine quality and subject to error from external interference.  We may be able to identify external or mutual interference by comparing forward and reverse RSSI and observing changes over time.  2. Expected Transmission Count (ETX). A key parameter in the RPL protocol is the expected transmission count (ETX). The Cisco NMS records both the expected transmission count for the nodes link to its parent (ETXL), as well as the expected transmission count to reach the root (ETXP).  3. Hop depth. Hop depth is available from the NMS. Hop depth has a significant impact on the performance of a SUN network. 2.3.2.2 Metrics Available from the Collector but Not Collected by the NMS 1. Link Quality Indicator (LQI). Measurement is required in 802.15.4g [9]. The LQI is the error vector magnitude when comparing a received signal to the ideal constellation. 2. PHY layer counters. There are a number of PHY layer message and error counters available in aggregate for the mesh interface on the collector node. These are not automatically collected by the NMS.  2.3.2.3 Metrics from the Meter Data Management System 1. Read rate. Each day the meter system data collection system reports if a meter was read by the 6AM deadline. This is a pass fail business process metric to confirm if data is available for billing and presentation to the customer portal but it doesn’t account for the multiple attempts that may have been made to read each meter the previous day. 2. Communications rate. Each day, the meter system data collection system reports the number of successful contacts from the previous 10 attempts. This metric is a better indicator of message performance than the pass / fail read rate. 2.3.2.4 Metrics That can be Calculated from Available Data Sources 1. Node Density. Meter density is calculated by using the GIS location data and geospatial nearest neighbor tools. We speculate that there may be a high level of mutual interference in high density installations. Ancillotti [23] has identified by simulation and using the SensLAB test bed with 90 nodes that RPL can have issues selecting the best parents in high density scenarios. Ancollotti determined that this was due to the constrained design  12 of RPL to use a relatively short neighbor table. To investigate the impacts of density we want to test the correlation of performance metrics under varying node density. E.g., compare RSSI to ETXL, and Communications Rate to ETXP at different node densities. 2. Path Length. Path length is calculated by using the NMS statistics to determine the node and parent ids. The installed locations are then determined from GIS data. Finally, the path length is calculated. There is error in the location measurement as BC Hydro’s GIS data is based on legacy property parcel maps that as of 2015 are not aligned with a true GIS projection. A conflation process is underway that will allow GPS measurements to be directly used with BC Hydro GIS data.  3. Join Time. Join time can be calculated by comparing the timestamps in the systems event records. There are restoration events indicating when power is restored to a meter as well as registration events indicating when a meter has obtained an IP address, exchanged security credentials, and communicated with the head end system. 2.3.3 Network Data Management Network performance data statistics are gathered by the NMS. The NMS maintains only a limited history as required for plotting trends for the benefit of the operators. In order to build an archive of network statistics, an SQL Extraction, Transformation and Load process is performed to move network statistics to a reporting data warehouse. The data warehouse retrieves a snapshot of network statistics daily and stores them for 30 days for reporting purposes. A further “Data Pond” infrastructure is used for long term storage. The Data Pond retrieves data from the warehouse and appends that data to previously collected data. The Data Pond has been used for long term storage of the NMS statistic since November 5th, 2015.  The location data for individual meters is gathered from BC Hydro customer records. Note a key limitation of the data set is that the meter location is not captured with GPS accuracy. Typically, the meter is assumed to be placed near the front of the land parcel. The location data for conventional relay nodes or range extenders and collectors is gathered from BC Hydro’s enterprise GIS. Read performance statistics are gathered from the Collection Engine. 2.3.4 Database and Visualization Tools A variety of Big Data and software tools are available that can handle large volumes of data. They are distinguished by their features, cost, and user community. In our case the selection of a storage tool is particularly influenced by the amount of data we have. In the case of the network  13 statistics, the tool used for storing and providing access to the data need to be capable of storing 2,000,000 records for each day in the archive. The statistical value of the data increases when many months’ worth of data is stored. Having multiple months of data provides multiple samples for each unique node to parent link allowing for averaging. On a longer scale, multiple years of data will allow seasonal variations to be analyzed. Thirty days of NMS statistics (approximately 60 Million records) has a data volume of 20 GB. Indexing the data to improve query performance increases the storage needs by 50%. Unlike other Big Data problems, the velocity (the rate of arrival of new data) of the data is not an issue. New data is available each day but the most current data is not needed to conduct the analysis. Additionally, there is limited variety in the data used as it is well structured from the machine source. Given a large volume but low velocity and limited variety the problem is best classified as a data warehousing problem. A data warehousing tool with an inherent spatial model to support queries is desirable because location, distance, and geographic properties such as land cover are relevant to the analysis. The warehousing tool must be able to interface easily with analysis tools as well as the other source systems. There is a cost to moving data in and out of the warehouse or Big Data system. As data warehousing matures a growing trend is to enable analytics within the system which avoids issues that result from extraction and transport. Ideally, the tool should will allow migration from data warehousing to Big Data as the volume of our data grows or if we start to apply analysis on a more real-time basis.  For the purposes of this study, we selected PostgreSQL, an open source DBMS, as our primary database. Due to the availability of a PostGIS geospatial extension, it is well suited to applications which required geospatial analysis. In the latest version, additional Big Data features sponsored by the European FP7 project were added. An example of a new and useful Big Data feature is the implementation of SQL TABLESAMPLE, a function that return a random sample of a table. This can be very useful in limiting the runtime of large statistical queries.  The principal limitation of PostgreSQL is that each query uses a single thread so large queries are not optimized across multiple cores. This is not an issue in transactional DBMS applications where there are many smaller queries. However, it becomes an issue in data warehousing and Big Data applications where there are few very large queries. PostgreSQL does not provide data analysis and visualization tools. Several alternatives exist, including R (an open  14 source version of S that is very popular among statisticians), the ubiquitous Matlab (from The Mathworks), and Tableau, an interactive data visualization products focused on business intelligence (from Tableau Software). We chose the latter. 2.4 A Strategy for Development of U-LLN Tuning and Optimization Techniques The vast amount of data available permits comprehensive study of U-LLN performance on a scale that was not previously possible. Recent experience at BC Hydro and elsewhere has shown that tuning and optimization of a U-LLN is more challenging than had been expected. We propose a phased, incremental strategy for using MSGN performance data based on: 1) development of measurement-based short-range path loss models that will make network simulations more accurate, 2) identification of correlations between network layout and performance, and 3) development of schemes for optimizing network performance through infilling relay nodes through various means when the node density is low and transmit power control when the node density is high. 2.4.1 Path Loss Modelling Path loss between network nodes and between network nodes and infrastructure is an important parameter for predicting link reliability and mutual interference. Realistic path loss models are required by standards bodies, manufacturers and operators to facilitate both simulation and design. Observations of the manner in which path loss varies with distance, environment, and time will be crucial in interpreting higher level network behaviour and devising appropriate physical layer tuning and optimization techniques. As of early 2016, BC Hydro has collected over 260 million RSSI data samples for meter-to-meter, meter-to-collector, and relay-to-collector links from most regions of British Columbia. The opportunity to develop highly reliable path loss models that apply across the terrain, building and foliage environments within BC Hydro’s service area is immense. Compared to data measured using survey or lab grade instruments, the amplitude of the data collected from an operational network has a slightly higher uncertainty. The location of the meter nodes is known only to the lot level but collector and pole mounted relay locations are surveyed with conventional GPS accuracy. The meter antenna patterns are somewhat less regular than those employed in purpose-built instruments. The measured RSSI samples available in the data set are from links chosen by the RPL algorithm and are biased toward the best available  15 links. The range of distances found in the data set is relatively small with most of the links between 20 and 250 metres. However, the vast amount of data available, including both the number of links measured and the number of times each link is measured during the observation period, more than compensates for the minor limitations of individual measurements. The first steps towards developing short-range path loss models based upon network performance data are reported in Chapter Four of this thesis.  2.4.2 Correlation of Performance Metrics with Design Parameters After analyzing path loss, the next logical step is to determine the relationship between the network performance metrics described in Section 2.3.2 and various aspects of the network layout and configuration. Such observations of higher level network behaviour will play a particularly crucial role in devising appropriate physical layer tuning and optimization techniques. Relationships of particular interest include the manner in which Read Rate varies with ETXP (expected transmission count to reach the root), Hop Depth, RSSI (received signal strength indication), ETXL (expected transmission count for the nodes link to its parent) and the difference between the forward and reverse RSSI. 2.4.3 Alternative Tuning and Optimization Strategies Ideally, we would perform physical layer optimization by creating a list of the nodes in the network, determine the quality of each node’s links to its neighbours according to some metric, e.g., signal strength, link quality or the number of retries required to complete a transmission, identify those that were above or below a specified threshold, and simply add relay nodes where required. For example, Cheng et al. [26] have presented algorithms that place the minimum number of relay nodes and maintain the connectivity of a single-tiered WSN, under the assumption that all relays and ordinary sensors have identical transmission ranges. Experience with BC Hydro’s MSGN suggests that situation is complicated by: 1) variability in link characteristics over time, 2) the multiplicity of network performance parameters that one may wish to optimize, and 3) the cost-performance tradeoffs between different solutions. There are also some indications that mutual interference is degrading performance in locations where the node density is too high. In such cases, transmit power adjustment may be used to reduce node coverage. Development of an efficient and effective physical layer optimization technique depends critically on thorough understanding of the network and its behaviour. In this section,  16 we briefly review the alternative means for implementing physical layer optimization and consider the tradeoffs as a precursor to development of such a technique.  2.4.3.1 Conventional Relays A conventional relay is a node very similar to a meter. It is used when there is inadequate node density and street light relays or other useful nodes with relay functionality are not available. It may be installed at a location where the electric utility has a right-of-way (e.g. on a utility pole), it may have an external antenna to improve performance, or it may be DC powered to facilitate battery backup. BC Hydro has already installed approximately 5,000 conventional relays.  Its principal strengths include the following: The hardware is similar to meters and it can easily be deployed on utility poles or at a customer’s premise. Its main limitations: It serves no purpose beyond improving mesh performance but must be maintained. It requires additional agreements if it is to be installed on customer property. To avoid placing unnecessary conventional relays an engineering study or site survey is often conducted. 2.4.3.2 Street Light Relays A street light relay is a communicating street light control module that replaces the photocell controller on individually controlled street luminaires. It may be used where street lights are available and ideally owned by the electric utility.  Its principal strengths: A street light relay can both improve the mesh network and also provide control and monitoring for the street lighting. Luminaires that have monitoring capability can be used to detect failures allowing for proactive replacement of lights. Its principal limitations: It may not be used where the street light is not property of the electric utility or an agreement cannot be effectively reached with the luminaire owner. 2.4.3.3 Power Control or Adjustment IEEE 802.15.4g [9] states that, “Devices should transmit lower power when possible in order to reduce interference to other devices and systems.” However, the current version of BC Hydro’s MSGN does not have any provision for transmit power adjustment; all nodes transmit at the same power level. We distinguish between power control, which is performed by each node in response to current inputs, and power adjustment, which is performed occasionally in response to a system wide optimization scheme.    17 Its principal strengths: Power adjustment is easy to implement but determining the best value for any given meter is challenging. Its principal limitations: Determining the best transmit power level for any given node is challenging.  2.4.3.4 Modulation BC Hydro currently operates its network in at 150 kbps 2FSK. 802.15.4g allows for other modes including 50 kbps 2FSK, 200 kbps 2FSK, 50 kbps OFDM, and 800 kbps OFDM. Understanding the path loss seen by the network will help the value of introducing higher or lower order modulations. The RX sensitivity of the CC1101 -95 for 500 kbps GFSK. The RX sensitivity of the CC1200 increases from -107 dBm at 100 kbps 2FSK with a channel filter bandwidth of 208 kHz to -122 dBm at 1.2 kbps 2FSK with a channel filter bandwidth of 11kHz. 2.5 Discussion Tuning and optimization of U-LLNs such as BC Hydro’s MSGN represents a major challenge. Both physical and network layer approaches to tuning and optimization are possible. Although the underlying issues are generally well understood, detailed understanding of higher level network behaviour based upon performance data collected over live networks will play a particularly crucial role in devising appropriate schemes.  18 Chapter 3:  Distance Effects in the Estimation of Power Law Path Loss Models 3.1 Introduction When path loss over obstructed paths such as urban, in-building and foliated environments is measured, it has been found that: 1) path loss generally exceeds that experienced in free space, 2) the mean path loss tends to follow a power law relationship with distance where the path loss exponent is generally greater than 2, and 3) the residual tends to follow a log normal distribution that is independent of distance. Many standardized measurement-based path loss models of this type have been developed over the years [27][29] [29][30][31]. The conventional method for estimating the power law path loss model that best fits measurement data are to apply ordinary least squares to path loss and distance data that is expressed on a log scale. The slope of the resulting regression line gives the path loss exponent and the intercept at a specified distance becomes the path loss intercept. The residual is characterized by subtracting the mean path loss from the measured data and then determining the zero-mean distribution that fits best. As noted above, a log normal distribution that is uniform with distance fits the vast majority of cases. The ordinary least squares technique is based on the tacit assumptions that: 1) the span is long enough that variations in the log of dependent variable about the regression line do not obscure the linear relationship and 2) the independent variable is known with complete accuracy. As the span decreases, however, the correlation between the log of the dependent and independent variables will rapidly decrease. As the correlation coefficient drops, our ability to resolve the power law relationship decreases. Increasing errors in the independent variable will result in a phenomenon referred to as attenuation bias or regression dilution in which the slope of the regression line and, possibly the value of the intercept, will decrease. Errors in variables models are widely used in other fields such as economics, chemistry and ecology that deal with large amounts of data [32][33][34]. They could be used to mitigate the effect of distance errors on model parameter estimates. Finite distance span effects can only be mitigated by increasing the distance span which must be addressed at the data collection stage.     When path loss is measured in macrocell or even microcell environments, the ranges covered extend for hundreds or even thousands of metres. Modern GPS equipment is capable of  19 routinely measuring locations with accuracies of a few metres unless severe shadowing and multipath, e.g., due to tall buildings in urban settings, degrades accuracy. As a result, most previous researchers have assumed that this renders any concerns about the span and accuracy of the range to be moot. In wireless sensor network and similar short range scenarios, however, the spans may be very short and the positions of individual nodes may be difficult to measure. Accordingly, we are motivated to determine the specific conditions under which these effects become significant.      The remainder of this chapter is organized as follows: In Section 3.2, we formulate the problem. In Section 3.3, we consider the effect of limited distance spans on estimation of power law path loss parameters over the range of typical values of path loss exponent and shadow fading. In Section 3.4, we consider the effect of distance errors in estimation on power law path loss parameters over the range of typical values of path loss exponent and shadow fading. In Section 3.5, we consider the implications of these results for engineering practice and future research. 3.2 Model Formulation A power law path loss model at a given frequency takes the form  𝑃𝐿 𝑑 = 	𝑃𝐿 𝑑& + 10𝑛 log.& //0 +𝑋2 , (3.1) where PL is expressed in dB,  𝑃𝐿 𝑑&  is the path loss intercept, d0 is the intercept distance, n is the path loss exponent, X is a zero-mean Gaussian random variable and  σ is the shadow fading coefficient (or depth of shadow fading), also expressed in dB. As the span decreases and the shadow fading coefficient increases, the correlation coefficient will decrease and the linear relationship will be increasingly difficult to discern. A critical distance is the span over which ∆𝑃𝐿 = 	𝜎,  i.e., the span over which the variation due to the linear relationship is comparable to the variation due to shadow fading. If the change in the mean path loss over the span from d2 to d1 is given by  ∆𝑃𝐿 = 10𝑛 log.& /6/0 −10𝑛 log.& /8/0 	,  (3.2) then it can be shown that the span over which ∆𝑃𝐿 = 	𝜎 is given by  𝑑9 𝑑. = 	102/.&;	.  (3.3)  20 The ratio of the maximum and minimum distances, not their difference, is the determining factor. For ratios of σ/n = 3, 2, and 1, this corresponds to d2/d1 = 2, 1.6 and 1.2, respectively. 3.3 Effect of Limited Distance Spans on Model Parameter Estimation  In this section, we investigate the manner in which limited distance span degrades the correlation between path loss and distance and leads to errors in estimating the model parameters in the following manner. For given values of path loss exponent and shadow fading coefficient, and across a specified distance span, we estimate PL(d) from (3.1) then attempt to estimate 𝑃𝐿 𝑑&  , n and σ. We repeat this many times, then estimate the mean value and distribution of our estimates and the relative error. We repeat the above over a range of distance spans and plot the results. In Figure 3-1, the degradation in the mean value of the correlation coefficient as d2/d1 decreases is shown. It can be seen that the correlation coefficient decreases as the ratio σ/n increases. In Figure 3-2, the mean value of the error in the path loss exponent as d2/d1 decreases is shown. It can be seen that the error increases as the ratio σ/n increases and increases markedly after d2/d1 drops below 3. The corresponding errors in the path loss exponent and shadow fading are given in Figures 3-3 and 3-4. The distribution of the correlation coefficient and mean value of the error in the path loss exponent over the 5000 instances as a function of d2/d1 are depicted in the box plots given in Figures 3-5 and 3-6.   21  Figure 3-1: Mean value of correlation coefficient vs. distance span, d2/d1.   Figure 3-2: Mean value of error in path loss exponent, n vs. distance span, d2/d1.  1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4d2/d100.10.20.30.40.50.60.70.80.91Correlation Coefficientn= -2.00 < = 4.00n= -2.00 < = 6.00n= -2.00 < = 8.00n= -4.00 < = 4.00n= -4.00 < = 6.00n= -4.00 < = 8.001.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4d2/d100.511.52Error in nn= -2.00 and < = 4.00n= -2.00 and < = 6.00n= -2.00 and < = 8.00n= -4.00 and < = 4.00n= -4.00 and < = 6.00n= -4.00 and < = 8.00 22  Figure 3-3: Error in path loss intercept, PL0 vs. distance span, d2/d1.   Figure 3-4: Error in shadow fading coefficient, σ vs. distance span, d2/d1.    1.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4d2/d10.60.811.21.41.6Error in PL0n= -2.00 and < = 4.00n= -2.00 and < = 6.00n= -2.00 and < = 8.00n= -4.00 and < = 4.00n= -4.00 and < = 6.00n= -4.00 and < = 8.001.5 1.75 2 2.25 2.5 2.75 3 3.25 3.5 3.75 4D2/D101234567Error in < (%)n= -2.00 and < = 4.00n= -2.00 and < = 6.00n= -2.00 and < = 8.00n= -4.00 and < = 4.00n= -4.00 and < = 6.00n= -4.00 and < = 8.00 23      Figure 3-5: Distribution of the correlation coefficient vs. distance span, d2/d1.    1.5 1.75 2   2.25 2.5 2.75 3   3.25 3.5 3.7500.20.40.60.81.5 1.75 2   2.25 2.5 2.75 3   3.25 3.5 3.7500.20.40.61.5 1.75 2   2.25 2.5 2.75 3   3.25 3.5 3.7500.20.40.61.5 1.75 2   2.25 2.5 2.75 3   3.25 3.5 3.750.20.40.60.81.5 1.75 2   2.25 2.5 2.75 3   3.25 3.5 3.7500.20.40.60.81.5 1.75 2   2.25 2.5 2.75 3   3.25 3.5 3.75D2/D100.20.40.60.8Correlation Coefficientn = 2< = 4n = 2< = 6n = 2< = 8n = 4< = 4n = 4< = 6n = 4< = 8 24      Figure 3-6: Distribution of errors in n vs. distance span, d2/d1.   1.5 1.75 2   2.25 2.5 2.75 3   3.25 3.5 3.750501001.5 1.75 2   2.25 2.5 2.75 3   3.25 3.5 3.750501001.5 1.75 2   2.25 2.5 2.75 3   3.25 3.5 3.750501001.5 1.75 2   2.25 2.5 2.75 3   3.25 3.5 3.750501001.5 1.75 2   2.25 2.5 2.75 3   3.25 3.5 3.750501001.5 1.75 2   2.25 2.5 2.75 3   3.25 3.5 3.75D2/D1050100Error in n (%)n = 2< = 4n = 2< = 6n = 2< = 8n = 4< = 4n = 4< = 8n = 4< = 6 25 3.4 Effect of Distance Errors on Model Parameter Estimation In this section, we investigate the manner in which distance errors degrade the correlation between path loss and distance and leads to attenuation bias. For given values of path loss exponent and shadow fading coefficient, and across a specified distance span, we estimate PL(d) from (3.1), degrade our estimates of d by incorporating a Gaussian distributed error with specified standard deviation then attempt to estimate 𝑃𝐿 𝑑& , n and σ. We repeat this 5000 times, then estimate the mean value and distribution of our estimates and the relative error. We repeat the above over a range of distance errors and plot the results. In Figure 3-7, the degradation in the mean value of the correlation coefficient as d2/d1 decreases is shown. It can be seen that the correlation coefficient decreases as the ratio σ/n increases. In Figure 3-8, the mean value of the error in the path loss exponent as d2/d1 decreases is shown. It can be seen that the error increases as the ratio σ/n increases and increases markedly after d2/d1 drops below 3. The corresponding errors in the path loss exponent is given in Figures 3-9.   Figure 3-7: Mean value of correlation coefficient vs. distance error, σd. 0 0.2 0.4 0.6 0.8 1Distance Error (<d)00.20.40.60.81Correlation Coefficientn= -2.00  < = 2.00n= -2.00  < = 4.00n= -2.00  < = 6.00n= -2.00  < = 8.00n= -4.00  < = 2.00n= -4.00  < = 4.00n= -4.00  < = 6.00n= -4.00  < = 8.00 26  Figure 3-8: Mean value of error in n vs. distance error, σd.   Figure 3-9: Mean value of error in PL0 vs. distance error, σd.    0 0.2 0.4 0.6 0.8 1Distance Error (<d)01020304050607080Error in n (%)n= -2.00 and < = 2.00n= -2.00 and < = 4.00n= -2.00 and < = 6.00n= -2.00 and < = 8.00n= -4.00 and < = 2.00n= -4.00 and < = 4.00n= -4.00 and < = 6.00n= -4.00 and < = 8.000 0.2 0.4 0.6 0.8 1Distance Error (<d)00.511.522.533.544.555.566.5Error in PL0n= -2.00 and < = 2.00n= -2.00 and < = 4.00n= -2.00 and < = 6.00n= -2.00 and < = 8.00n= -4.00 and < = 2.00n= -4.00 and < = 4.00n= -4.00 and < = 6.00n= -4.00 and < = 8.00 27 3.5 Discussion Our results reveal that a path loss model with a shallow path loss exponent and high degree of shadow fading is more susceptible to error than a model with a steeper path loss exponent and less shadow fading. Errors in distance can almost certainly be addressed using errors-in-variables models. Finite distance span effects can only be mitigated by increasing the distance span which must be addressed at the data collection stage. Our results provide useful practical guidance to others engaged in fitting power law path loss models to measured data which contain distance errors or which were collected over a short distance span.  28 Chapter 4:  Short-Range Power-Law Path Loss Models in Suburban Environments 4.1 Introduction Path loss between network nodes and between network nodes and infrastructure is an important parameter for predicting link reliability and mutual interference. Realistic path loss models are required by standards bodies, manufacturers and operators to facilitate both simulation and design. Observations of the manner in which path loss varies with distance, environment, and time will be crucial in interpreting higher level network behaviour and devising appropriate physical layer tuning and optimization techniques.  Most previous path loss modelling efforts have been concerned with P2MP macrocell links; examples include Hata [27] and Erceg [28]. Recently there has been growing interest in wireless sensor networks and peer-to-peer communications resulting in an emergence of D2D models; such as Matolak [31], Ghassemzadeh [35], Tsuchiya [36], Alsayyari [37][38][39], and Chandrasekharan [40]. However, these new D2D models are difficult to apply to our smart meter application because of the unique characteristics of our use case. Most D2D models include low height antennas that are positioned in relatively open space. Some consider frequency ranges much higher than our use case [35]. Others consider links shorter than our scenario and in LOS or in building use cases [36][37][38][39]. All of the models surveyed were based on small datasets collected over a very short period of time. New models based on larger data sets and realistic scenarios are required to support industrial WSN in suburban environments. As of early 2016, BC Hydro has collected over 260 million RSSI data samples for meter-to-meter, meter-to-collector, and relay-to-collector links across most regions of British Columbia. The opportunity to develop highly reliable path loss models that apply across the terrain, building and foliage environments within BC Hydro’s service area is immense. Compared to data measured using survey or lab grade instruments, the amplitude of the data from this live network has a slightly higher uncertainty. The location of the meter nodes is known only to the lot level but collector and pole mounted relay locations are surveyed with conventional GPS accuracy. The meter antenna patterns are somewhat less regular than those employed in purpose built instruments. The measured RSSI samples available in the data set are from links chosen by the RPL algorithm and are biased toward the best available links. The range of distances found in  29 the data set is relatively small with most of the links between 20 and 250 metres. However, the vast amount of data available, including both the number of links measured and the number of times each link is measured during the study period, more than compensates for the minor limitations of individual measurements.  The remainder of this chapter is organized as follows: In Section 4.2, we review the concepts that underlie our data collection, reduction and modelling approach. In Section 4.3, we describe our method for collecting, managing, processing and reducing the network performance data, In Section 4.4, we present the statistics of our data set, our estimates of the power law path loss models that best fit the device-to-device and device-to-infrastructure links in our data set, and compare our results to previous work. In Section 4.5, we briefly discuss the implications of our results.  4.2 Concepts We follow a commonly used empirical modelling approach in which path loss is expressed as  𝑃𝐿 𝑑 = 	𝑃𝐿 𝑑& + 10𝑛 log.& //0 +𝑋2 , (3.1) where PL(d0) and n are calculated by fitting a sum of least squares linear regression line to the collected data.  σ is calculated from the standard deviation of the residuals (difference between our measurements and the predicted value along the regression line). The shadow fading is inevitably normally distributed. Several D2D modellers have used this approach and we will compare our results to theirs in Section 4.4.  Due to our unique circumstances, our approach is different than previous efforts to model path loss for short range links. We have unique advantages: 1) our data has been collected over a period of several months, 2) we have substantially more data. The time span will allow us to be able to see long duration or seasonal trends if they occur. The volume of data will help us overcome our unique challenges: 1) we don’t have precision location information for our nodes, 2) we have embedded device SoC grade RSSI measurements. With our short links, the position error will results in path length errors in the range of 5 to 10% (10-metre position error for 100-200-metre links).  Because our data is collected from an in-service network, we anticipate some atypical challenges. The data set include only ‘best’ paths selected by the routing algorithm. The network will normally choose a path that will get it close to the root. As shown in Figures 4-3 and 4-4,  30 under this condition, the majority of the measurements will be near the designed RSSI floor of -95 dBm. This will truncate measurements that would be gathered in a deliberate measurement campaigns. We expect that our use of OLS, with truncated measurements, may underestimate the pathloss exponent and large scale fading [41]. In order to form our power law path loss models, we will need position data from the geographic information system, as well as RSSI data from the network management system.  We are able to generate a more useful model if we can further classify our nodes by their environment. Most meters are installed on the sides of buildings. We desire to known the building type, lot size, local land cover, and node density characteristics. We will need to join the network statistics with the node locations. This may seem trivial but the two system have different inventory approaches. The GIS tracks meter sockets not meters, and the NMS tracks meters not meter sockets. Accurate meter inventory data will be needed to tie these two sources together. We will segment our data for three important reasons: 1) to manage the volume of data, 2) to generate a model representative of the diversity of environments seen in BC, 3) to permit comparison between our regions as well as other models. Processing all our data is possible, but we can develop our processes and likely a suitable model with a small fraction of the data available, greatly reducing our data handling effort. Well known power law path loss models differentiate between terrain types [28]; we can take advantage of the diversity of environments found in BC to evaluate the effects of terrain on smart meter LLNs.  While we do not expect terrain to be a significant factor as paths are very short, we can test our terrain assumption by choosing three environments with different terrain characteristics. 4.3  Methodology We processed the raw data into an analyzable data set by following these steps: 1) Collected the necessary data from source systems (locations from GIS, network statistics from NMS, and inventory from MDMS), 2) Joined the GIS and NMS data using the MDMS data. 3) Created node-to-parent pairs for each of D2I and D2D measurement, 4) Created aggregate for all links used more than once, 5) Calculated path lengths for the node-to-parent pairs, 6) Mapped the links and removed any obviously incorrect data (e.g., links longer than 50km), 7) generate node and links statistics (path length histograms, link re-use counts, RSSI averages, variances, etc.), 8)  31 calculate regression lines by collector/neighbourhood for both D2I and D2D, 9) check for goodness of fit (e.g., R-Squared values, mean of residuals = 0 across the full span of path length), 10) remove collectors and neighbourhoods with poor R-Squared values or small sample populations, 11) calculate predictions and residuals for each measurement, 12) calculate the intersection of regression lines for each area, 13) calculate statistics for the regressions (e.g., standard deviation of the residuals), 14) create indexes and bins for data as required, 15) produce tables and plots for analysis. Typical data mining issues include: lack of long term vision, not all files up to date, struggle between departments, poor cooperation for the electronic data processing department, legal and privacy issues, files are hard to connect for technical reasons, timing problems, and interpretation problems [42]. These issues are not unfamiliar to the BC Hydro environment. The data needed to conduct our study was not collected for this purpose. Network statistics are available but extracting them requires significant effort on the part of the operations department. Data from multiple system needs to be joined and there are reporting delays between these systems. A large amount of data need to be stored, indexed, and queried to produce the desired outcome. There are also data quality issues with the location of nodes. Our data cleaning was a two phase process. In the first pass, we carefully removed data that we believed to be erroneous. We removed any links with a path length greater than 50km. This resulted in the removal of 527 nodes or 5,237 links out of 2,931,840 links. Quick analysis suggested that these nodes were not installed in the locations indicated by the MDMS. There are a few plausible explanations: 1) data entry error, 2) after a meter is replaced, it can take a few weeks before the MDMS is updated (within the billing cycle), 3) after an infrastructure node is replaced, it can take a long time before the GIS is updated, 4) occasionally a customer will move and take their meter with them. We likely could have corrected the data but that effort did not seem warranted given the volume of good measurements available.  As a second pass. we excluded collectors (neighbourhoods) if the regression results did not meet a quality threshold defined as: 1) the regression correlation coefficient R-Squared was greater than 0.25, 2) the sample size was greater than 4000 measurements, and 3) the span of the path lengths was greater than 100 metres. These criteria were selected based on inspection of the data. Table 4-1 shows the number of collector in each region and Table 4-6, and 4-7 show the  32 number of collectors used in the models. Further review is recommended as these criteria removed a large portion of the regression lines. We maintained our data set in a ProgreSQL 9.5 RDBMS. Due to the size of the data set we tried to do as much of the analysis as possible within the database. We made use of the built-in ‘aggregate functions for statistics’ for our regression analysis. In particular, we the built-in ‘ordered-set aggregate functions’ to calculate medians and percentiles. When developing and testing scripts, we used the TABLESAMPLE function to query a random sample of data (often 0.1% is enough to confirm the expected result). The TABLESAMPLE function is also useful for producing scatter plots. We use PostGIS extensions to enable geospatial features (calculating path lengths, counting nearest neighbours, etc). We used QGIS for viewing the data set in map form. QGIS is an open source mapping tool able to directly connect to PostgreSQL PostGIS databases. We used Tableau to produce plots. With the exception of Tableau, we have used community open source tools freely available to any researcher. For a fully open source approach we could have used R, Orange3, or Zeppelin as alternatives to Tableau.  4.4 Results In this section, we present the results of analyzing and reducing the network performance data collected in Coquitlam between mid-November 2015 and mid-February 2016. The remainder of the section is divided into four parts: In Section 4.4., we review the essential aspects of the measurement database. In Section 4.4.2, we present the results of our efforts to model device to infrastructure (meter-to-collector) links. In Section 4.4.3, we present the results of our efforts to model device to device (meter-to-collector) links. In Section 4.4.4, we compare the results that we obtained for D2I and D2D links. In Section 4.4.5, we compare our results to those reported previously and put our contributions in context.  4.4.1 Measurement Database  As of January 31st, 2016, BC Hydro’s MSGN consists of 1,939,746 meters, 5260 relay nodes, and 2395 collectors. BC Hydro smart meters were fully converted to IPv6 in August 2015. Automated collection and storage of daily network statistics was enabled November 5th, 2015. The data set used for this study spans 119 days from November 16th, 2015 to March 7th, 2016 and includes 231 collectors (aka neighbourhoods) and 264,772 meters from three representative regions of BC, as detailed in Table 4-1. Figure 4-1 shows the density of smart meters in  33 Richmond/Delta. White areas represent the densest deployment of nodes and dark areas represent the least dense deployments. Figure 4-2 shows device-to-infrastructure links in Richmond/Delta coloured by collector. The quantity of D2I and D2D links in each regions is shown in Table 4-3. There are plenty of links in each area for analysis.  The node density in Kamloops is much lower than in the the other two regions. Coquitlam, and Richmond/Delta have very similar meter density. On average the collectors in Kamloops have 30% the number of meters as in the other two regions, as shown in  Table 4-3. In Kamloops, 50% of the meters have less than 8 neighbours within 50 metres; whereas, in Coquitlam and Richmond/Delta, 50% of the meters have less than 14 neighbours within 50 metres, as shown in Figure 4-5. In Kamloops, 47% of nodes have a neighbour density less than 500 metres per square kilometre; whereas, only 10% of the nodes in Coquitlam, Richmond/Delta have a neighbour density of less than 500 meters per square kilometer, as shown in Figure 4-6.  Table 4-4 shows device-to-infrastructure path length statistics, and Table 4-5 shows device-to-device path length statistics. There are clear trends in the path lengths between regions. D2I links are twice as long as device-to-device in all three regions. Both device-to-device and device-to-infrastructure links are 50% longer in Kamloops than the other two regions. Figure 4-7 shows the distribution of path lengths for D2I scenarios. Figure 4-8 shows the distribution of path lengths for D2D scenarios. Overall Kamloops has lower node density and longer paths. Figure 4-9 shows a scatter plot comparing D2D path length vs. node density for all regions studied. This plot is generated using the TABLESAMPLE function to randomly sample 0.1% of the data set. The use of a sample reduces the clutter on the plot while still showing the presence of long links with lower density nodes, and the grouping around 100m path length for very high densities.  We expected the routing protocol to select links with a RSSI better than -95dBm, and due to the high packet loss and ETX routing metric, select longer links if needed to reach the collector in fewer hops. Figure 4-3 shows the distribution of RSSI values for D2I links. For D2I links, 50% of the measurements are between -80 and -95 dBm, and 15% of links with an RSSI worse the -95dBm suggesting that this condition is more favourable than 2 hops. Figure 4-4 shows the distribution of RSSI values for D2D links. For D2D links, 32.5% of the measurements are between -80dBm and -95dBm.   34  Table 4-1: Quantities of meters and collectors in each region. Region Type A  Coquitlam Type B Kamloops Type C   Richmond/Delta Number of Neighbourhoods (Collectors) 71 61 99 Number of Nodes (Meters) 89,916 51,321 123,535 Median Number of Nodes per Neighbourhood 1645 500 1543  Table 4-2: Quantities of D2I and D2D links and measurements for each region. Region Type A  Coquitlam Type B   Kamloops Type C  Richmond/Delta D2D links 847,252 658,647 1,097,481 D2D measurements  7,193,939 4,638,114 10,185,843 D2D median number of links per device 10.8 13.8 10.1 D2D standard deviation of the number of links per device 6.4 7.9 6.3 D2I links 23,154 19,742 24,191 D2I measurements 1,104,993 775,925 1,344,256 D2I median # of links per collector 315 300 205  Table 4-3: Node distribution statistics for each region. Node Density \ Region Coquitlam Kamloops Richmond/Delta Median (nodes per sqkm) 1164 602 1300 95th Percentile (nodes per sqkm) 5081 1630 5699 standard deviation (nodes per sqkm) 1350 517 1578  Table 4-4: Device-to-Infrastructure path length statistics for each region. Path Length \ Region Coquitlam  Kamloops  Richmond/Delta Median 189m 308m 175m 95th Percentile  850m 2217m 509m Standard Deviation 498m 1353m 310m  Table 4-5: Device-to-Device path length statistics for each region. D2D Path Length \ Region Coquitlam  Kamloops  Richmond/Delta Median 102m 154m 96m 95th Percentile  384m 1902m 297m Standard Deviation 321m 1519m 191m  35   Figure 4-1: Gray scale overlay showing the density of smart meters in Richmond/Delta.  Lighter shades indicate higher density. Darker shades indicate lower density.   36  Figure 4-2: Map showing device-to-infrastructure links in Richmond/Delta that are used 2 or more times.    37  Figure 4-3: Distribution of measured RSSI values in dBm for D2I links.    Figure 4-4: Distribution of measured RSSI values in dBm for D2D links.   38  Figure 4-5: Distribution showing the number of neighbouring nodes within a 50 metre radius.   Figure 4-6: Distribution showing the number of neighbouring nodes within 1 square kilometre.    39  Figure 4-7: Distribution of path lengths for D2I scenarios.   Figure 4-8: Distribution of path lengths for D2D scenarios.    40  Figure 4-9: D2D Path length vs. node density.   Figure 4-10: Distribution showing the re-use of D2I links over the 119 day analysis window.   41  Figure 4-11: Distribution showing the re-use of D2D links over the 119 day analysis window.  4.4.2 Device to Infrastructure Path Loss Models Figure 4-12 show an example of the RSSI vs. path length for D2I link to the collector called “ACAD” in Kamloops fit with an OLS regression line. Most of the data is in the range from 100m to 300m. Within this range there is a good distribution above and below the zero mean as seen in Figure 4-13. The longer links have a worse fit with most the sample having a positive residual. This regression line has an R-Squared of 0.44 and is based on 22679 samples. We generated regression lines for each collector in all regions. The results are shown in Figures 4-15 to 4-17. Figure 4-14 shows the similarities between regions. We see that the results are fairly well grouped. We note that the R-squared correlation coefficient is low for most of the regression lines indicating that there is a high degree of shadow fading. However, we expect the model can be applied to similar smart meter use cases. We choose to remove from the model development regression lines with R-Squared values lower than 0.25, distance span less than 100m, and sample sizes less than 4000 measurements. As a result, we used 116 of 231 regression lines. The final D2I model parameters are shown in Table 4-6. Figure 4-18 shows the distribution of path loss exponents for each region. There is a distinct mode for each region as well as a fairly high degree of variation. This result is consistent with  42 the results of other D2D modellers who have observed a high degree of variation between sites when conducting measurement campaigns [31][43].  There are many approaches to finding the intercept distance d0. Some modellers use a point at which their regression line intersects FSPL [44]. Another approach is to select a small value relative to the model (e.g., 1 m for D2D) [36]. Because we have regression lines from many neighbourhoods, we can select the intercept by determining the mean of their intersections. We believe that this results in a more robust model.  Similar to other D2D modellers, we see a high degree of shadow fading [31]. The example for collector “ACAD” shown in Figure 4-13 is representative of the typical case where we see a reduction in fading for longer links; however, this is in part due to the soft floor at -95 dBm, and the fact that the routing protocol will only select longer links if they perform well. Figure 4-19 shows that overall distribution, from a random 1% sample, of the RSSI vs distance regression residuals fits within a Gaussian distribution.  Over the 119-day span of the data set, 67,087 D2I links and 3,225,174 measurements were observed. This allowed the long term variability of path conditions to be observed and analyzed.  Figure 4-10 shows a distribution of D2I link re-use over the analysis window. Over this period only 26% of the links were used more than 100 times and 40.8% were used less than 10 times. The mean of the standard deviation of RSSI for those links used more than 100 times is 3.58 dB, the maximum standard deviation is 14.13 dB, standard deviation of the standard deviation is 1.0 dB. This suggests the path loss for any given path is relatively stable, and also shows that the network is very dynamic in its route selection.  Table 4-6: Device to Infrastructure model parameters for different regions. Model  d0 n σS PL(d0) # BS used Coquitlam  (Type A) 394 m 0.94 to 2.87, median 2.00 1.31 to 14.92 median 6.13 90.19 50 Kamloops  (Type B) 385 m 0.74 to 3.09, median 1.80 1.41 to 12.12 median 6.02 85.07 34 Richmond / Delta (Type C) 342 m 1.13 to 3.16, median 2.32 1.25 to 14.11 median 6.47 89.43 32 Overall 366 m 0.74 to 3.16 median 2.00 1.25 to 14.92 median 6.16 88.66 116  43   Figure 4-12: Example of the RSSI vs. path length for D2I link to the ACAD collector in Kamloops.   Figure 4-13: Example of the Residual vs Path Length for D2I links to the ACAD collector.   44   Figure 4-14: Regression lines for Coquitlam terrain type A - hilly with moderate to heavy tree density.   Figure 4-15: Regression lines for Kamloops terrain type B - hilly with light tree density.   45  Figure 4-16: Regression lines for Richmond and Delta terrain type C - flat with light tree density.   Figure 4-17: Regression lines for RSSI vs. Path Length for D2I links in all regions.    46   Figure 4-18: Distribution of path loss exponents for D2I scenarios.   Figure 4-19: Distribution of the residuals for D2I scenarios.    47 4.4.3 Device to Device Path Loss Models Figure 4-20 shows an example of the RSSI vs. path length for D2D link to the collector called “ACAD” in Kamloops fit with an OLS regression line. Most of the data is in the range from 50m to 600m. Within this range, there is a good distribution above and below the zero mean as seen in Figure 4-21. The very long links have a worse fit with most the samples having a positive residual. This regression line has an R-Squared of 0.28 and is based on 169,416 samples. We generated regression lines for each collector in all regions. The results are shown in Figures 4-22 to 4-25. Figure 4-22 shows the similarities between regions. We see that the results are fairly well grouped.  We note that the R-squared correlation coefficient is low for most of the regression lines indicating that there is a high degree of shadow fading. However, we expect the model can be applied to similar smart meter use cases. We choose to remove from the model development regression lines with R-Squared values lower than 0.25, distance span less than 100m, and sample sizes less than 4000 measurements. As a result, we used 68 of 231 regression lines. The final D2I model parameters are shown in Table 4-6. Figure 4-26 shows the distribution of path loss exponents for each region. There is a distinct mode for each region as well as a fairly high degree of variation. This result is consistent with the results of other D2D modellers who have observed a high degree of variation between sites when conducting measurement campaigns [31][43].  We choose to use the intersection of all regression lines to select a value for d0. Other modellers use a point at which their regression line intersects FSPL. However, that approach is not suitable in our scenario as such a point does not occur. Another approach is to select a small value relative to the model (e.g., 1 m for D2D). However, given the tight grouping of our regression lines at the 109m point we choose that point for our d0 reference. Similar to other D2D modellers [31] we see a high degree of shadow fading. The example for collector “ACAD”, shown in Figure 4-21, is representative of the typical case where we see a reduction in fading for longer links; however, this is in part due to the soft floor at -95 dBm, and the fact that the routing protocol will only select longer links if they perform well. Figure 4-27 shows that overall distribution, from a random 1% sample, of the RSSI vs distance regression residuals fits within a Gaussian distribution.   48 We observed 2,503,380 D2D links and collected 22,017,896 measurements over the 119-day span of the data set. This allowed for the analysis of long term variability of path conditions.  Figure 4-11 shows a distribution of D2D link re-use over the 119-day analysis window. Over this period only 0.4% of the links were used more than 100 times. 79.2% were used less than 10 times. The mean of the standard deviation of RSSI for those D2D links used more than 100 times is 3.62 dB, max is 7.9 dB, standard deviation is 0.9 dB. This suggests the path loss for any given path is relatively stable, and also shows that the network is very dynamic in its route selection. In both D2I and D2D, we saw a lot of attenuation. This is likely due to the obstructed locations in which our smart meters are mounted. The relative ranking of path loss exponent is similar between D2I and D2D. We see the lowest path loss exponents in Kamloops and the highest in Richmond/Delta in both cases. The relative ranking of PL0 is also similar between D2I and D2D. We see less attenuation in Kamloops and the highest in Coquitlam in both cases.  It has been generally believed that the device-to-device links in WNSs will allow them to cover large areas more easily by providing improved link budgets; However in our measure data, we see slightly higher levels of shadow fading on D2D links compared to D2I links. However, in our case with tens of thousands of links, across 3 different regions.  In both the D2I and the D2D scenarios the standard deviation of RSSI over reused links is ~3.6dB, and the standard deviation of the standard deviations is below 1.0. This suggests the path loss for any given path is relatively stable in both the D2I or D2D scenarios. It is more likely that a link will be re-used in the D2I case.  Table 4-7: Device-to-Device model parameters for different regions.  Model  d0 n σS PL(d0) # BS used Coquitlam (Type A) 136 m 1.31 to 1.70, median 1.46 1.04 to 17.32 median 6.98 83.22 18 Richmond / Delta (Type C) 115 m 1.27 to 1.92, median 1.56 1.39 to 18.53 median 7.47 82.19 20 Kamloops  (Type B) 59 m 0.87 to 1.65, median 1.06 1.39 to 18.27 median 6.23 76.68 30 Overall 109 m 0.87 to 1.92 median 1.42 1.04 to 18.53 median 6.83 81.39 68   49   Figure 4-20: Example of the RSSI vs. path length for D2D links below the ACAD collector in Kamloops.    Figure 4-21: Example of the Residual vs. Path Length for D2D links below the ACAD collector.  50    Figure 4-22: Regression lines for Coquitlam terrain type A - hilly with moderate to heavy tree density.     51 Figure 4-23: Regression lines for Kamloops terrain type B - hilly with light tree density.   Figure 4-24: Regression lines for Richmond and Delta terrain type C - flat with light tree density.   Figure 4-25: Regression lines for RSSI vs. Path Length for D2D links in all regions.  52    Figure 4-26: Distribution of path loss exponents for D2D scenarios.   Figure 4-27: Distribution of the residuals for D2D scenarios.  53 4.5 Comparison with Other Path Loss Models It is useful to compare the results of different short-range path loss models in similar environments but it is also difficult to do so due to different frequency ranges, distance spans, and antenna heights or mounts. In Figure 4-28 we show a comparison of D2D and D2I mean path loss models found in literature review. The two marked D2D and D2I are ours. The models depicted have been scaled to a reference frequency of 915 MHz. The different use cases modeled result in three clusters: 1) very short range links modelled by Tsuchiya [36] and Alsayyari [37][38][39], 2) line-of-sight links modelled by Ghassemzadeh [35], and Chandrasekharan [40], and 3) short-range obstructed D2D links modeled by Matolak, Ghassemzadeh. Our work fits well with the other short range obstructed D2D models. To simplify the presentation, we select a single line per use case where modellers have reported several similar scenarios with a high variation in results at each site (e.g., Matolak, Chandrasekharan). This figure shows the variation in slope, attenuation, and the distance span between the models. Parameters for each of the models compared are shown in Table 4-8. Despite modelling a smart meter use case, the work by Tsuchiya is not comparable to our case since it is modelling a smart meter mounted outdoors that is communicating with indoor devices. The LOS models by Ghassemzadeh and near-line-of-sight models by Chandrasekharan and Walden are not representative of our use case either.   Figure 4-28: Comparison of mean path loss vs. distance for several short-range path loss models.   54  The most similar use case is found in the work by Ghassemzadeh who has studied small cell deployments at 5.7 GHz in two neighbourhoods. He conducted four measurement campaigns in each neighbourhood and reports results for both LOS and NLOS scenarios. His results are based on more than 3000 measurement points using diversity antennas at both ends of the link collected using corrected GPS (50cm resolution).  The antenna heights for his test bed are similar to our D2I case with collectors at 6.1m and remote stations at ~1.9m. We scale the results from 5.7 GHz to 915 MHz for the comparison of mean path loss exponent. We observe that the regression lines for Ghassemzadeh’s NLOS results lay between the results of our model and Matolak’s with all the regression lines are clustered in a similar range. Matolak has studied peer-to-peer wireless channel in urban environments in public safety bands including 905 MHz. Measurements were taken in downtown Denver around large buildings (over 20-stories). There were five stationary receivers and a pedestrian transmitter. Antennas for four of the collectors and the remote station were at ~1.6 metres.  Samples were continuously recorded at a rate of 2 samples/second allowing for small-scale multipath effects to be averaged out. We compare our results to site 1 where both transmitter and receiver are at 1.6m, which is similar to our D2D model. He reports NLOS path loss parameters ranging from 4 to 6 and notes these are higher than most other published results.  While the path loss exponents observed by Matolak are greater than our models and Ghassemzadeh’s NLOS model, Matolak notes a high degree of variation between his 5 sites. We wonder if more sites would lead to greater variability as is seen in our modelling which is based on several million links. Surveying D2D literature reveals that our dataset has unique advantages and disadvantages. Advantages include having: 1) substantially more data available, 2) a long measurement time span allowing for the opportunity to observe long term temporal/seasonal variations, 3) applicability to the smart meter, smart street lighting use cases which are seeing significant investment worldwide. Disadvantages include having: 1) slightly less accurate location information, 2) slightly less accurate RSSI measurements, 3) somewhat irregular antenna patterns. On the whole, the advantages far outweigh the disadvantages.     55 Table 4-8: Parameters for several short-range path loss models. Model Path Type f (MHz) d0 (m) n S (dB) PL(d0) (dB) Equivalent PL(d0) at 915 MHz (n1) D2I (Meter to pole top) NLOS (n2) 915  366  1.97 1.25 to 14.92 median 6.16 114.66  D2D (Meter to meter) NLOS (n2) 915 109  1.21 1.04 to 18.53 median 6.83 107.39  Matolak (Urban mobile to mobile) NLOS 905 50  3.95 to 6.06 5.73 to 9.03 80 80 Walden (Street lights) Near LOS 868 54.9  3.3 to 4.3 and 1.68 to 2.08 (n3) 6 to 11 66.01  66.53  Ghassemzadeh  (LOS NAN small cell) LOS 5700 5.3  1.77 4.5 60.6  47.8  Ghassemzadeh (NLOS NAN small cell) NLOS 5700 10.3 3.57 6.23 80 63 Chandrasekharan (Hewson structures on both sides) LOS 922  20 3.55 4.32 46.81 46.76 Chandrasekharan Follet structures on one side) LOS 922  20 2.97  5.39 52.96 52.90  Alsayyari LG (Long Grass) LOS 1900 1 2.56 3.84 59.42  53.67 Alsayyari T (Turf) LOS 1900 1 2.75 4.2 67.68 61.13  Alsayyari (Concrete Surface) LOS 1900  1  3.21 2.19 64.84 58.65 Alsayyari ST (Sparse Trees) LOS 1900 1  3.34 7.30 60.84  54.95 Tsuchiya A (meter to indoor device) NLOS  950  1  2.85 6.91 44.35  44.11  Tsuchiya B (meter to indoor device) NLOS 950 1 1.34 5.69 65.05  64.69  n1 - scaled by the ratio of the square of the frequencies.  n2 - the data set is assumed to be primarily obstructed NLOS but there may be LOS paths included in this data set. n3 - Walden rejected 4 of the 10 regression results indicating that the OLS solution was not meaningful because node density was too low at certain ranges.  56 4.6 Discussion Although path loss is key to understanding the performance of wireless mesh networks deployed in suburban environments, remarkably little work relevant to U-LLNs has been published in the literature. While the network performance data available to us suffers from certain limitations compared to data collected using survey or lab grade instruments, the vast amount of data available, including both the number of links measured and the number of times each link is measured during the study period, more than compensates for the minor limitations of individual measurements. While our results are broadly consistent with previous work, they are based on a far larger and richer database of measurements and are statistically more reliable. Unlike previous work, we have been able to directly compare device-to-device and device-to-infrastructure path loss in the same environments. Our results contradict the traditional belief that wireless mesh networks will experience less shadow fading than point-to-multipoint networks because they tend to select only the best paths. Because we collect measurements over a long term, we have the option of incorporating both link reliability and path loss variability into our model.       57 Chapter 5:  Conclusions and Recommendations 5.1 Conclusions Urban Low power Lossy wireless mesh Networks (U-LLNs) based upon IEEE 802.15.4g, IEEE 802.15.4e, 6LoWPAN and RPL will play a key role in enabling future smart grids, smart utilities, and smart cities. However, their performance is highly affected by the manner in which the wireless propagation environment, the network layout and configuration, and the network protocols combine to determine both the reliability of individual links and the degree of mutual interference between links. Effective and efficient techniques for tuning and optimizing operational U-LLNs to increase link reliability, reduce mutual interference and improve overall performance will be required as: 1) U-LLNs grow and evolve and 2) increasingly demanding applications are introduced. Given the size and geographic extent of most U-LLNs, manual techniques are inadequate and automated methods are required. Although previous work has shown that software-based simulation and testbed-based simulation can provide useful insights concerning tuning and optimization, they are not sufficient and may even be misleading in certain aspects because they do not fully account for the nature of the wireless propagation environment.  Analysis and reduction of performance data from live U-LLNs almost certainly offer the most reliable and cost-effective method for characterizing the wireless propagation environment in which U-LNNs operate and the manner in which it affects U-LLN performance. The objective of this study has been to take the first steps toward using data analytics to examine raw performance data collected from a large-scale operational network and reveal the environmental, design and configuration factors that affect the performance of U-LLNs. The source of our data has been three areas within BC Hydro’s Multi-Service Grid Network (MSGN), an IEEE 802.15.4e/g and IPv6-based LLN provided by Itron and Cisco. The MSGN includes over 1.9 million smart meters, over 5000 range extenders/relays and over 2000 collectors deployed across a service territory that covers 95% of the population of British Columbia. The MSGN network performance data archive offers an unparalleled opportunity to observe network behaviour over time across a wide range of terrain, building and foliage environments.   58 The first major contribution of this work has been setting up the software infrastructure required to manage and prepare the data from the MSGN so that they can be analyzed and reduced to models useful in simulation and design. The effort required is considerable and should not be underestimated. The preliminary results presented in this thesis confirm the value and potential of the database and sets the stage for using it to pursue further studies.  The second major contribution of this work has been to achieve all three of our specific research goals, namely: 1) to develop a strategy for developing U-LLN tuning and optimization methods based upon analysis of network data collected from BC Hydro’s MSGN, 2) to assess the impacts of errors in transmitter-receiver distance and limited distance spans on the estimation of power law path loss models from measured data and 3) to demonstrate both the operational and technical feasibility of formulating short-range power law path loss models applicable to device-to-device, device-to-infrastructure, relay-to-infrastructure links based upon analysis of network data collected from BC Hydro’s MSGN.  Development of a Tuning and Optimization Strategy for U-LLNs. Recent efforts at BC Hydro and elsewhere have revealed that it is much more difficult to tune an optimize an operational U-LLN than many had originally expected. Our proposed strategy for developing U-LLN tuning and optimization methods acknowledges the many factors that affect network performance and proceeds in a logical and measured way. The specific stages that we propose include: 1) development of short-range path loss models that capture our knowledge and understanding of the relevant airlink impairments in a form useful in design and simulation, 2) identification of correlations between network layout and configuration, airlink impairments and observed network performance, and 3) assessment of the merits of alternative schemes for infilling relay nodes when the node density is low and using transmit power adjustment to reduce mutual interference when the node density is high, based upon insights developed during the first two stages. Effect of Distance Errors or Finite Distance Span on Model Estimation. Although power law path loss models are widely used to represent the manner in which path loss varies with distance in urban, in-building, and foliated environments, the impact of errors in transmitter-receiver distance and limited distance spans on estimation of the relevant model parameters from measured data has not been previously considered. We generated ideal power law path loss data then either introduced errors in the distance or restricted the distance span over which the path  59 loss was observed. We assessed the error in estimating model parameters as a function of either distance error or distance span over a range of parameters representative of typical propagation environments. Our results reveal that a path loss model with a shallow path loss exponent and high degree of shadow fading is more susceptible to error than a model with a steeper path loss exponent and less shadow fading. Errors in distance can almost certainly be addressed using errors-in-variables models. Finite distance span effects can only be mitigated by increasing the distance span which must be addressed at the data collection stage. Our results provide useful practical guidance to others engaged in fitting power law path loss models to measured data which contain distance errors or which were collected over a short distance span. Short-Range Path Loss Models for Suburban Environments. Although path loss is key to understanding the performance of wireless mesh networks deployed in suburban environments, remarkably little work relevant to U-LLNs has been published in the literature. While our results are broadly consistent with previous work, they are based on a far larger and richer database of measurements and are statistically more reliable. Unlike previous work, we have been able to directly compare device-to-device and device-to-infrastructure path loss in the same environments. Our results contradict the traditional belief that wireless mesh networks will experience less shadow fading than point-to-multipoint networks because they tend to select only the best paths. Because we collect measurements over a long term, we have the option of incorporating both link reliability and path loss variability into our model.  5.2 Recommendations Our success in achieving all three of our specific research goals bodes well for further work and execution of the progressive strategy for using MSGN performance data to develop U-LLN tuning and optimization methods.  Short-Range Path Loss Models. We anticipate that short-range path loss models based upon MSGN network performance data will: 1) assist in the interpretation of higher level network performance data, 2) improve the accuracy of software simulations of LLNs deployed in similar environments, and 3) inform the design of more realistic testbeds. We recommend that the modelling effort described in Chapter 4 be continued and efforts made to model relay-to-infrastructure (pole-top to pole-top) links, and to incorporate reliability, variability and, possibly, correlated shadow fading into the model. Issues such as measurement uncertainty, selection bias  60 and factors that degrade the correlation between path loss and distance over short ranges should be further studied. Correlation between Network Performance and Network Layout & Configuration. We anticipate that knowledge of the correlation between network performance and network layout & configuration will assist in the interpretation of higher level network performance data. We recommend that the strategy described in Chapter 2 be pursued.  U-LLN Tuning and Optimization Methods. The current technique for tuning and optimizing the BC Hydro MSGN by infilling conventional relay nodes has proven to be problematic. We recommend that the proposal to assess the merits of alternative schemes for infilling relay nodes, including use of street light controllers as relays, when the node density is low and using transmit power adjustment to reduce mutual interference when the node density is high, based upon insights developed during the first two stages of the progressive strategy that we described in Chapter 2, be pursued. The ultimate goal will be to develop a system that will analyze network performance data and then determine: 1) where to place infill nodes and 2) which meter and relay nodes should have their transmit powers reduced in order to help the network achieve optimal performance.       61 References  [1] R. Davies, “Hydro One’s smart meter initiative paves way for defining the smart grid of the future,” in Proc. IEEE PES, 2009, pp. 1–2. [2] “Smart Metering and Infrastructure Program Business Case,” 2011. [Online]. Available: https://www.bchydro.com/content/dam/BCHydro/customer-portal/documents/projects/smart-metering/smi-program-business-case.pdf. [Accessed: 01-Apr-2016]. [3] V. Dabic and D. Atanackovic, “Voltage VAR optimization real time closed loop deployment - BC Hydro challenges and opportunities,” in Proc. IEEE PES, 2015, pp. 1–5. [4]  “The Canadian Smart Grid Standards Roadmap: A strategic planning document.” Standards Council of Canada, Oct. 2012, 38 pp. [5] Clean Energy Act: Smart Meters and Smart Grid Regulation. 2010. [6] “IEEE Guide for Smart Grid Interoperability of Energy Technology and Information Technology Operation with the Electric Power System (EPS), End-Use Applications, and Loads,” IEEE Std 2030-2011, pp. 1–126, Sep. 2011. [7] J. Messerly, “Simple diagram of electricity grids in North America,” U.S. Department of Energy, 2008. [8] “The Challenges in Achieving a Smart Grid: The BC Hydro Journey,” 05-Mar-2010. [Online]. Available: http://icics.ubc.ca/comm-workshop/files/Zucker%20-%20Achieving%20a...Journey.pdf. [Accessed: 02-Apr-2016]. [9]  “IEEE Standard for Local and metropolitan area networks–Part 15.4: Low-Rate Wireless Personal Area Networks (LR-WPANs) Amendment 3: Physical Layer (PHY) Specifications for Low-Data-Rate, Wireless, Smart Metering Utility Networks,” IEEE Std 802.15.4g-2012 (Amendment to IEEE Std 802.15.4-2011), pp. 1–252, Apr. 2012. [10] K-H Chang and B. Mason, "The IEEE 802.15.4g standard for smart metering utility networks," in Proc. IEEE SmartGridComm 2012, pp. 476-480.J. Ko, A. Terzis, S. Dawson-Haggerty, D. E. Culler, J. W. Hui, and P. Levis, “Connecting low-power and lossy networks to the internet,” IEEE Commun. Mag., vol. 49, no. 4, pp. 96–101, Apr. 2011. [11] H. R. Kermajani and C. Gomez, “Route change latency in low-power and lossy wireless networks using RPL and 6LoWPAN Neighbor Discovery,” in Proc. IEEE ISCC, 2011, pp. 937–942. [12] T. H. Lee, X. S. Xie, and L. H. Chang, “RSSI-based IPv6 routing metrics for RPL in low-power and lossy networks,” in Proc. IEEE SMC, 2014, pp. 1714–1719. [13] O. Gaddour, A. Koubâa, S. Chaudhry, M. Tezeghdanti, R. Chaari, and M. Abid, “Simulation and performance evaluation of DAG construction with RPL,” in Proc. ComNet, 2012, pp. 1–8.  62 [14] P. Thubert, “Objective Function Zero for the Routing Protocol for Low-Power and Lossy Networks (RPL).” [Online]. Available: https://tools.ietf.org/html/rfc6552. [Accessed: 02-Apr-2016]. [15] I. F. Akyildiz, X. Wang, and W. Wang, “Wireless mesh networks: A survey,” Comput. Netw., vol. 47, pp. 445-487, 2005. [16] I. F. Akyildiz and X. Wang, “A survey on wireless mesh networks,” IEEE Commun. Mag., vol. 43, no. 9, pp. S23-S30, Sep. 2005. [17] S. Vural, D. Wei, and K. Moessner, “Survey of experimental evaluation studies for wireless mesh network deployments in urban areas towards ubiquitous Internet,” Commun. Surveys Tuts., vol. 15, no. 1, pp. 223-239, 2013. [18] M. Dohler, D. Barthel, T. Watteyne, and T. Winter, “Routing Requirements for Urban Low-Power and Lossy Networks.” [Online]. Available: https://tools.ietf.org/html/rfc5548. [Accessed: 02-Apr-2016]. [19] S. Dwars, T. Phinney, and P. Thubert, “Industrial Routing Requirements in Low-Power and Lossy Networks.” [Online]. Available: https://tools.ietf.org/html/rfc5673. [Accessed: 02-Apr-2016]. [20] A. Brandt and G. Porcu, “Home Automation Routing Requirements in Low Power and Lossy Networks.” [Online]. Available: https://tools.ietf.org/html/rfc5826. [Accessed: 02-Apr-2016]. [21] “Itron Selects Accent Single Chip Solution for Smart Metering Applications.” [Online]. Available: https://www.itron.com/na/newsAndEvents/Pages/Itron-Selects-Accent-Single-Chip-Solution-for-Smart-Metering-Applications.aspx. [Accessed: 29-Apr-2016]. [22] J. Martocci, W. Vermeylen, N. Riou, and P. D. Mil, “Building Automation Routing Requirements in Low Power and Lossy Networks.” [Online]. Available: https://tools.ietf.org/html/rfc5867. [Accessed: 02-Apr-2016]. [23] E. Ancillotti, R. Bruno, and M. Conti, “Reliable data delivery with the IETF routing protocol for low-power and lossy networks,” IEEE Trans. Ind. Informat., vol. 10, no. 3, pp. 1864–1877, Aug. 2014. [24] J. Li, L. L. H. Andrew, C-H Foh, M. Zukerman and H-H Chen, "Connectivity, coverage and placement in wireless sensor networks,” Sensors, vol. 9, pp. 7664-7693, 2009. [25] P. Agrawal and N. Patwari, “Correlated link shadow fading in multi-hop wireless networks,” IEEE Trans. Wireless. Commun., vol. 8, no. 8, pp. 4024-4036, Aug. 2009. [26] X. Cheng, D-Z Du, L. Wang, B. Xu, "Relay sensor placement in wireless sensor networks,” Wireless. Netw. vol. 14, pp. 347–355, 2008. [27] M. Hata, “Empirical formula for propagation loss in land mobile radio services,” IEEE Trans. Veh. Tech., vol. 29, no. 3, pp. 317–325, Aug. 1980. [28] V. Erceg, L. J. Greenstein, S. Y. Tjandra, S. R. Parkoff, A. Gupta, B. Kulic, A. A. Julius, and R. Bianchi, “An empirically based path loss model for wireless channels in suburban environments,” IEEE J. Sel. Areas Commun., vol. 17, no. 7, pp. 1205–1211, Jul. 1999.  63 [29] COST Action 231, “Digital mobile radio towards future generation systems, final report,” tech. rep., European Communities, EUR 18957, 1999. [30] “ITU-R Rec. P.1238 : Propagation data and prediction methods for the planning of indoor radiocommunication systems and radio local area networks in the frequency range 300 MHz to 100 GHz.” [Online]. Available: https://www.itu.int/rec/R-REC-P.1238-8-201507-I/en. [Accessed: 26-Apr-2016]. [31] D. W. Matolak, Q. Zhang, and Q. Wu, “Path loss in an urban peer-to-peer channel for six public-safety frequency bands,” IEEE Wireless Commun. Lett., vol. 2, no. 3, pp. 263–266, Jun. 2013. [32] S. van Huffel and P. Lemmerling, Total Least Squares and Errors-in-Variables Modeling: Analysis, Algorithms and Applications. Springer Science & Business Media, 2013. [33] T. Söderström, “Errors-in-variables methods in system identification,” Automatica, vol. 43, no. 6, pp. 939-958, Jun. 2007. [34] D. M. Hawkins and C. Weckwerth, “Errors in variables regression with value-censored data,” J. Chemometrics, 2016. [35] S. S. Ghassemzadeh, H. R. Worstell, and R. R. Miller, “Wireless neighborhood area network path loss characterization at 5.7 GHz,” in Proc. IEEE VTC 2010-Fall, 2010, pp. 1–6. [36] H. Tsuchiya, “Characterization of the radio propagation channel in residential environment for smart meter communications,” in Proc. APMC, 2011, pp. 713–716. [37] A. AlSayyari, I. Kostanic, and C. E. Otero, “An empirical path loss model for wireless sensor network deployment in an artificial turf environment,” in Proc. IEEE ICNSC, 2014, pp. 637–642. [38] A. Alsayyari, I. Kostanic, and C. E. Otero, “An empirical path loss model for wireless sensor network deployment in a concrete surface environment,” in Proc. IEEE WAMICON, 2015, pp. 1–6. [39] A. Alsayyari, I. Kostanic, C. Otero, M. Almeer, and K. Rukieh, “An empirical path loss model for wireless sensor network deployment in a sand terrain environment,” in Proc. IEEE WF-IoT, 2014, pp. 218–223. [40] S. Chandrasekharan, A. Al-Hourani, K. Magowe, L. Reynaud, and S. Kandeepan, “Propagation measurements for D2D in rural areas,” in Proc. IEEE ICCW, 2015, pp. 639–645. [41] C. Gustafson, T. Abbas, D. Bolin, and F. Tufvesson, “Statistical modeling and estimation of censored pathloss data,” IEEE Wireless Commun. Lett., vol. 4, no. 5, pp. 569 - 572, Oct. 2015. [42] P. Adriaans and D. Zantinge, Data Mining. Boston, MA: Addison-Wesley Longman, 1997. [43] J. Turkka and M. Renfors, “Path loss measurements for a non-line-of-sight mobile-to-mobile environment,” in Proc. ITST 2008, pp. 274–278.  64 [44] M. C. Walden, T. Jackson, and W. H. Gibson, “Development of an empirical path-loss model for street-light telemetry at 868 and 915 MHz,” in Proc. IEEE APS/URSI), 2011, pp. 3389–3392. 	

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.24.1-0300442/manifest

Comment

Related Items