High-Precision Photovoltaic Potential Prediction Using a Multi-Factor Deep Residual Network

With the continued increase in global renewable energy threats, accurate identification and assessments of solar photovoltaic (PV) potential areas has become a fundamental undertaking in order to maximize solar deployment strategies. A novel multi-factor deep residual network (ResNet-DNN) prediction framework is proposed to provide high-accuracy prediction of PV potential areas. This model develops a multibranch input coding module to consolidate features from factors pertaining to terrain, albedo, solar radiation, temperature and land use, thus preserving the heterogeneous characteristics of different modal data. A residual hierarchical fusion structure is developed to generate multi-factor deep interaction while preserving gradient propagation stability. The global-local attention mechanism is proposed to adjust the importance weights of different factors based on the spatial context and time changes. Contrary to traditional ResNet or DNN methods, the proposed ResNet-DNN model can efficiently adaptively capture the nonlinear coupling among multiple factors and modulate the depth and connectivity of the network through the residual gating mechanism. The experiments show that majority of baseline methods in the existing literature have a significantly less prediction accuracy than the proposed model.