Sim-to-Real Deep Reinforcement Learning for Maximum Power Point Tracking of Photovoltaic Systems

The maximum power points (MPPs) of photovoltaic (PV) panels vary with atmospheric conditions (solar irradiance, ambient temperature, shading conditions, etc.). Simulation platforms are widely used by engineers and scientists to create models, analyze data, and develop algorithms for the maximum power point tracking (MPPT) of PV systems. However, the behavior, which the algorithm develops in simulations, is usually specific to the characteristics of the models. Strategies that succeed in simulations may not be victoriously transferred to the real world because the modeling errors and sensor noise. In this paper, we create a dynamic model of the PV system only based on the data-sheet from manufacturers. A sim-to-real transfer strategy is proposed for the maximum power point tracking of PV systems. By dynamically randomizing the environments for the agent during the training, the proposed policy can adapt to very different atmospheric conditions. Simulations show that the proposed strategy can maintain a similar level of performance when deployed on the real PV panels.