Graph Neural Network With Dual-Stream Attention for Global Maximum Power Point Tracking Under Partial Shading Conditions

Abstract

Due to the complex nonlinear coupling between circuit topology and environmental changes, precise maximum power point tracking under partial shading conditions remains a challenge. Existing data-driven methods typically view photovoltaic (PV) arrays as unstructured feature vectors or Euclidean grids, which fail to capture the fundamental series-parallel constraints that control the distribution of current and voltage. This study proposes a graph neural network with dual-stream attention (GNN-DSA), which is a physical perception framework that can strictly decouple static topologies from dynamic environmental signals. Static streams encode arrays into message-passed directed graphs with specific edge types that capture series-parallel circuit constraints, while dynamic streams capture irradiation patterns through the Transformer. A novel bidirectional cross-attention mechanism explicitly simulates the interaction between circuit constraints and environmental stress. Hardware verification indicates that the tracking efficiency of GNN-DSA on real PV panels is 99.1%–99.6%, on solar array simulators is 99.99%, and the response time is less than 100 ms. Comparative analysis shows that compared with the multilayer perceptron baseline, GNN-DSA reduces the mean squared error by 45.6%, and compared with the state-of-the-art Crossformer model, it reduces it by 22.3%.