A Machine Learning-Based Global Maximum Power Point Tracking Technique for a Photovoltaic Generation System Under Complicated Partially Shaded Conditions

When the photovoltaic generation system (PVGS) operates under partially shaded conditions (PSC), its output power versus voltage (P-V) characteristic curve becomes multimodal, which complicates the search for the global maximum power point (GMPP). This paper proposes a GMPP tracking (GMPPT) method based on machine learning (ML). In the first stage, the regression tree (RT) is used to predict the approximate location of the GMPP. In the second stage, the α-perturb and observe (α-P&O) method is used to obtain the precise GMPP. This study first establishes a PVGS simulation platform and generates the training data required for RT, then optimizes the obtained RT and integrates it into the simulation platform. Finally, this paper compares the proposed method with the state-of-the-art approaches. It can be seen from the results that the proposed method has an average tracking power loss of 2.13 W and an average tracking time of 0.11 seconds under 252 different shading patterns (SPs). It can correctly identify 244 intervals where the exact GMPP is located among the 252 test SPs. The experimental results show that the proposed method outperforms 5 state-of-the-art approaches in terms of tracking accuracy and tracking time under three shading patterns, thus confirming its excellence.