Reinforcement Learning Approaches for Load Forecasting in Microgrids: A Comprehensive Review

Abstract

Increased use of renewables makes good load forecasting crucial to the more efficient running of microgrids, and to the proper management of their energy. The conventional methods of prediction are not typically capable of dealing with the highly non-linear, stochastic and time-varying dynamics commonly observed in modern microgrid systems. During the last several years, reinforcement learning (RL) has started to be preferred due to its capabilities of assisting systems to progress independently and follow the best learning patterns depending on their real-world experiences. In this review, a large number of RL-based load forecasting approaches applied in microgrid environments are considered. The method is used to arrange past research according to forecasting horizon, the kind of algorithms employed, character of the information and judgment basis of outcome. Their efficiency and drawbacks in terms of real-time forecasting assignments are compared in terms of Q-learning, Deep Q-Networks (DQN), Proximal Policy Optimization (PPO) and Actor-Critic. Hybrid models, computation problems and challenges of merging IoT and edge computing layouts are also examined by the authors. It talks about the fields in which the recent research has been lacking and outlines how to proceed, naming federated learning, multi-agent reinforcement learning and the standardization of datasets as requirements. This work aims at demonstrating to the research and developer communities how they can deploy solid RL techniques to achieve smart, scalable and reliable microgrid load forecasting.