Improving Isolated Microgrid Inverter Lifetime with Reinforcement Learning-based Energy Management

This paper proposes a novel method for including the inverter reliability as an objective in the isolated microgrid scheduling. Inverter switches degradation modelling is highly non-linear. With the unlimited scenarios of solar irradiance, ambient temperature and load, building an optimal controller or policy becomes sophisticated. This work proposes using reinforcement learning (RL) as the energy management system (EMS) of an isolated microgrid. The isolated microgrid is successfully formulated as a Markov decision process, with 4-step ahead scheduling. The resulting formulation included the accumulated IGBT degradation in the reward function by converting it to a cost based on expected failure limit. Soft Actor-Critic is then trained on summer period from actual environmental and loads datasets. Having achieved the required training criteria, the proposed RL-based EMS is then tested on 10 days. Testing shows the proposed EMS reduces the IGBT damage by 21% compared to base EMS, while still extracting more energy from the Photovoltaic plant.