Deep Reinforcement Learning For Peak Load Reduction In Aggregated Residential Houses

Abstract

Demand response aggregators will create a new economic value allowing the control and the management of a large cluster of residential houses. The flexibility enabled by consumers demand combined with power system incentives may create more opportunities for economic efficiency. However, aggregators face privacy problems and unavailability of a clear description of each household dynamics. In this case, achieving optimal control problem must rely on observations from various system trajectories. This study proposes a deep RL-based energy management in a cluster of houses as a continuous-learning agent. The objective is to achieve a maximum peak load reduction, while scheduling a cluster of thermostatically controllable loads and respecting the occupant-chosen temperature limits. In order to overcome the uncertainties related to houses’ power consumption, a deep neural network uses the state of each house, which includes power consumption history, inner temperature, outer temperature and humidity, to predict the aggregated peak load. The considered RL technique shows good performances at reducing both the overall power consumption and the consumption during peak periods, while maintaining a given temperature.