A regional distributed photovoltaic power generation forecasting method based on grid division and TCN-Bilstm

Accurate power forecasting for distributed photovoltaic (DPV) sites is crucial for optimizing power system dispatch and operation, especially under high penetration of DPV. Traditional power forecasting methods, especially designed for single large-scale photovoltaic stations, are clearly unsuitable for the unique characteristic of DPV sites. Moreover, the severe data loss in DPV sites further complicates the power forecasting process. To address these issues, this paper proposes a novel regional DPV power forecasting method based on grid division and a hybrid neural network architecture of Temporal Convolutional Network-Bidirectional Long Short-Term Memory (TCN-BiLSTM). The proposed method begins with a DPV site power reconstruction method based on an adaptive weighting algorithm. And the algorithm leverages weighted information fusion of power data from neighboring sites to effectively reconstruct power data for sites with poor power data quality. Then, the method integrates the sum of power of DPV sites within grid and the numerical weather predictions (NWP) at the corresponding grid intersection as input information for the neural network. This enables short-term power forecasting for grid divided DPV site clusters. The main innovation of the paper lies in the systematic development of a DPV grid division power forecasting framework. By enhancing DPV sites data quality, the proposed method achieves efficient and accurate power forecasting. Verification using actual site data demonstrates the superior performance of the proposed method, achieving the highest accuracy with an R2 value of 0.997. Compared to other existing methods, the proposed method exhibits improvements of 1.14 %, 1.15 %, 0.69 %, 1.15 %, and 1.01 %, respectively.