Prospective Experiment for Reinforcement Learning on Demand Response in a Social Game Framework

Abstract

Improving demand response can help optimize renewable energy use and might be possible using current tools in machine learning. We propose an experiment to test the development of Reinforcement Learning (RL) agents to learn to vary a daily grid price signal to optimize behavioral energy shift in office workers. We describe our application of Batch Constrained Q Learning and Soft Actor Critic (SAC) as RL agents and Social Cognitive Theory, LSTM networks, and linear regression as planning models. We report limited success within simulation with SAC and linear regression. Finally, we propose an experiment timeline for consideration.