Day-ahead photovoltaic power forecasting based on corrected numeric weather prediction and domain generalization

Day-ahead photovoltaic (PV) power forecasting is usually built upon numeric weather prediction (NWP) data. However, NWP data could be significantly different from locally measured data (LMD), seriously degrading the PV power forecasting performance. Moreover, season changes may yield big PV power variation, challenging PV power forecasting. To resolve these issues, we propose a novel PV power forecasting method based on NWP correction and domain generalization. Specifically, we first propose decoupled seasonal representation to learn seasonal as well as local weather changes. Then we perform seasonal encoding and fluctuating encoding for both LMD and NWP data. Their seasonal encoding contains rich information of the current season while their fluctuating encoding can well characterize local weather changes. Afterwards we design an NWP loss function to measure the consistency between the local fluctuations of NWP and LMD, and construct an NWP correction method to efficiently mitigate the negative effects of NWP errors on PV power forecasting. Furthermore, we implement domain generalization to deal with meteorological feature distribution shifts caused by season changes. During the domain generalization process, we split historical training data into multiple domains and design a cross-domain loss function to ensure the consistency of these domains. Therefore season-invariant representation is learned, our forecasting method's generalization capability is well improved, and the adverse effects caused by season changes is greatly attenuated. We validate the superiority of our proposed PV power forecasting method on a publicly available dataset.