Optimal real-time scheduling of battery operation using reinforcement learning

Adoption of battery energy storage systems working with solar photovoltaic distributed systems for residential household applications strongly depends on their return on investment. Battery energy storage system (BESS) technology costs have been strongly decreasing during the last decade. However, such a tendency has to be supported by optimal BESS real-time operation strategies that adapt to the stochastic operation conditions (residential load, solar generation, and electricity prices) and minimize the customer's electric bill. This work presents a real-time adaptive BESS controller that implements a load-shifting strategy under time-of-use and feed-in-tariff (microFlT) regulatory incentives. The optimization of the battery operating strategy is carried out by a Q-learning algorithm and later encoded in a neural network that implements the optimal strategy at a fraction of the computation cost. Real residential demand and solar generation profiles during the summer and winter seasons in Edmonton, Canada, are utilized to train and test the controller. Two battery technologies, lithium-ion and vanadium redox flow, are simulated; real charge-discharge experimental data from an installed system was used. The proposed adaptive controller outperforms the optimal strategy, both during the summer and winter testing periods.