Meteorological information calculation and power forecasting for regional distributed photovoltaic power based on information fusion and deep learning

Abstract

Distributed photovoltaic (DPV) generation is characterized by multiple sites, wide geographic distribution, and relatively low capacity, which pose significant challenges for accurate power forecasting. These challenges include meteorological information gaps, the excessive number of models required for individual site modeling, and difficulties in self-calibration of forecasting models. To address these issues, this study proposes a novel approach that utilizes regional DPV grid centers as virtual representative DPV sites. And by integrating information fusion and deep learning techniques, this method aims to achieve regional DPV meteorological information fusion and power forecasting. The approach capitalizes on the superior parallel computing capabilities and computational precision of the eXtreme Gradient Boosting(XGBOOST) model. It employs direct irradiance, diffuse irradiance, wind speed, wind direction, humidity, and historical power data as inputs of information fusion model for DPV sites’ solar radiation and ambient temperature. Subsequently, the DPV power forecasting model is constructed based on the Temporal Convolutional Network-Bidirectional Gated Recurrent Unit-Attention (TCN-BiGRU-Attention) architecture. Finally, SHapley Additive exPlanations (SHAP) is utilized to optimize feature variables. Validation using real-world data demonstrates that the optimized TCN-BiGRU-Attention model achieves a forecasting accuracy of 99.3 %, representing a 0.90 % improvement over traditional methods. The introduction of virtual representative PV sites enables the efficient construction of self-iterative models for DPV meteorological information and power forecasting, providing an effective solution for regional DPV power forecasting.