Deep multivariable spatial attention CNN-based spatial grid NWP error correction for accurate one-day-ahead photovoltaic power forecast

Accurate day-ahead solar photovoltaic (PV) power forecasting is crucial for optimizing solar energy utilization. Traditional methods typically involve interpolating numerical weather prediction (NWP) data to a specific location, often overlooking spatial information from neighbouring areas. Entropy serves as an effective tool for quantifying forecasting uncertainty. In this study, we introduce transfer entropy to capture the information gain from incorporating NWP data from spatially adjacent regions, and propose the concept of spatial grid NWP correction. A deep spatial multivariable attention convolutional neural network-based framework for spatial grid NWP error correction is then proposed for day-ahead PV forecasting. Experiments conducted on three years (2017–2019) of actual data from two PV stations in Belgium demonstrate the method’s effectiveness. For the two stations, the proposed approach yields a normalized mean absolute error (nMAE) of 2.08% and 2.22%, and a normalized root mean squared error (nRMSE) of 4.45% and 4.81%, respectively. Compared to physical model chain without NWP correction, the proposed method reduces nMAE by 23.97%–24.79% and nRMSE by 15.54%–20.19%. Furthermore, comparison with eight conventional NWP correction methods further verifies its superior performance. Further experiment at a PV station in San Diego, USA, verifies that the proposed method can outperform conventional methods in different geographical regions and weather patterns.