Open Collections

Energy management in wireless communications : from convex optimization to machine learning

University of British Columbia

Ever-increasing energy consumption of network infrastructures motivates wireless operators to exploit renewable energy resources (e.g., sunlight and wind) to network infrastructure. When solely powered by weather-dependent renewable energy, a base station can experience power outages. Therefore, a smart-grid powered communication system (SGPCS) is proposed to avoid the power outage at base stations. To successfully apply renewable energy to an SGPCS, wireless operators need to develop energy management algorithms that can handle the unpredictable and intermittent arrival of renewable energy. Since machine learning algorithms are inherently designed for problems with random sources (i.e., stochastic optimization problems), we start by investigating machine learning algorithms for different stochastic optimization problems. Then, we adapt the potential machine learning algorithms to the long-term grid-energy expenditure minimization problem under various practical constraints. Using the finite-sample analytical methods, we quantify the convergence rates of the proposed offline learning and online learning algorithms. Based on the derived convergence rates, we have the following findings. When faulty users exist in the federated learning framework, our proposed fault-resilient proximal gradient and local fault-resilient proximal gradient algorithms require fewer communication rounds than the state-of-the-art benchmarks. Therefore, they are more energy-efficient than the benchmarks. The proposed linear function approximation based decentralized \emph{Q}-learning converges as fast as the tabular \emph{Q}-learning while retaining robustness to the large state and action spaces. Based on Lyapunov learning algorithms, we can successfully integrate the renewable energy in single-cell and multi-cell SGPCSs. Moreover, our proposed two time-scale resource allocation algorithm can trade the grid-energy expenditure for access delay of user equipments in single-cell SGPCS. Our proposed two time-scale resource allocation algorithm can trade grid-energy expenditure for the end-to-end delay of user equipments in multi-cell SGPCS.

Text

2020-08-28

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/75738

Electrical and Computer Engineering

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/