UBC Theses and Dissertations
Predictive analytics for building power demand : day-ahead forecasting and anomaly prediction Lin, Jing
The rising energy demand of buildings contributes to global resource consumption and greenhouse gas emissions. Understanding the energy consumption habits of a building is the first and most crucial step in order to achieve energy efficiency. Forecasting and anomaly prediction of power demand play an essential role in the electric industry, as it provides the basis for making decisions in power system planning and operation. However, there are several challenges in predictive analytics to improve energy efficiency. For instance, the information in the forecast needs to be interpreted by a person with domain expertise. Moreover, automated interpretation of upcoming abnormal behaviors needs ground-truth labeling, but labels are not always available from power meter data. To address those problems, a novel predictive power demand analytics methodology (PPDAM) is proposed in this thesis, based on long short-term memory neural networks and symbolic aggregate approximation. The main target of PPDAM is to identify upcoming normal and anomalous demand patterns, through mining historical power demand and temperature data, and day-ahead power demand forecast. The PPDAM consists of two modules: power demand forecasting and anomaly prediction. The power demand forecasting module is utilized to predict day-ahead power profiles (time sequences), as a baseline for pattern generation and anomaly prediction. In the module of anomaly prediction, the historical and predicted time-series profiles are first transformed into patterns and later retrieved from a system of labeled repositories. The patterns are classified as normal or anomalous based on their frequency of appearance in the pattern repositories. The power demand data cover the total electrical power consumption in seven buildings at the University of British Columbia. The experimental results indicate that a power forecast could be mapped as different foreseeable demand patterns, each with a specific probability of occurrence. Classification results indicate that our method is robust and fairly accurate in predicting anomalies despite the diversities of building types and behaviors. The outcomes of this work could provide building operators with a solution to derive latent information in power consumption data. The derived information could be used to improve the working conditions of the building's power system.
Item Citations and Data
Attribution-NonCommercial-NoDerivatives 4.0 International