Variable boundary reinforcement learning for maximum power point tracking of photovoltaic grid-connected systems

With clean energy development, photovoltaic power generation has become an important direction. The volt-ampere characteristic of photovoltaic modules in solar photovoltaic systems is nonlinear and changes with the environment. Therefore, tracking maximum power points in a changing environment is important for the research of photovoltaic power generation. This work proposes a variable boundary reinforcement learning (VBRL) algorithm to achieve maximum power point tracking, which is a novel reinforcement learning method that automatically divides the appropriate state range in the interaction with the environment. As a model-free control method, the VBRL algorithm can interact with the environment to learn and control without prior knowledge. The proposed algorithm realizes fast-tracking for maximum power point and has high tracking accuracy in the variable environment. Furthermore, the VBRL algorithm is a general solution to find the extremum of similar unimodal curve models. This paper compares the VBRL algorithm with some other methods in the complex variable environment. The VBRL algorithm achieves excellent comprehensive performance and has the largest average power generation when the irradiance changes rapidly.