Reinforcement Learning-based Control System of a Hybrid Power Supply

The control system of a hybrid power supply (HPS) is optimised using reinforcement learning. Hybrid power supplies combines two or more power sources. As more sources are integrated, the complexity of the control and energy exchange increases significantly. A model-free Q-learning RL controller is proposed to reduce the loss of power supply and to increase the cost optimisation of system components. The HPS consists of PV power, battery storage, a limited grid supply and a diesel generator. The use of renewable sources is a priority to reduce the amount of fuel usage. The results are compared to two baselines: random action controller and a rule-based controller. Different intervals are also assessed to find an optimal training interval to reduce computational power. The results showed that the RL-controller has the lowest LPS and optimises the use of the system components. This indicates that reinforcement learning-based control is viable and feasible for a HPS.