Battery Energy Storage Control Using Reinforcement Learning

With the increasing adoption of solar PV installations in Australian households, the availability of cheap renewable power during the day has surged. However, the challenge lies in rising electricity prices during morning and evening peak-consumption times. This project assessed the feasibility and profitability of using a Reinforcement Learning (RL) controller in a Battery Energy Storage System (BESS) to make cost-effective decisions by purchasing power when it's inexpensive and selling when it's costly. MATLAB/Simulink is used to create a BESS simulation model integrated into the electricity market, with the RL agent trained using normalized observation data and a reward function. Benchmarking demonstrated the RL controller's consistent outperformance of the timer-based controller in various market scenarios, emphasizing its adaptability and profitability advantages, particularly in volatile markets.

I. Introduction

In recent years, with the rapidly growing awareness and adoption of solar PV installations on household roofs across Australia, it has led to having an increase of renewable power cheaply available throughout the daytime [1]. However, the major challenge that comes with the great adoption of solar generation is the surging electricity prices during the evening peak-consumption times. This is caused by the absence of solar generation in the evenings when people are typically coming home and turning on air-conditioning and cooking appliances, creating a 24 hour load graph phenomenon known as a “duck curve”.