Control of Residential Battery Charge Scheduling using Machine Learning

Sammanfattning: This thesis proposes the use of a Reinforcement Learning (RL) agent to control the charge scheduling of a residential battery system. The system consists of a house located in Sweden equipped with a photo-voltaic array and grid-connection. Real residential load data is used while the PV output is simulated. The RL agent is trained using the Proximal Policy Optimization (PPO) algorithm to charge and discharge the battery within a continuous action space. The agent is trained and tested on three price contracts: fixed, monthly, and hourly. The perfor-mance of the agent is compared to a system without the battery, and to a Mixed Integer Linear Programming (MILP) optimizer controlling the battery. Results showed that while it was possible to train the agent to control the charge scheduling of the battery, the economic perfor-mance was only marginally better than a battery-less system and much poorer than MILP control. Of the price contracts, the agent had rela-tively better performance on the fixed price contract. The sensitivity of the RL algorithm to parameters and the reward function suggests that further investigation is needed in order to draw firm conclusions on the suitability of this method to the task in question.