Predicted Irradiance Correction and Contrastive Learning Feature Extraction–Based Day‐Ahead Photovoltaic Power Forecast Method

Existing research has enhanced short‐term photovoltaic (PV) power forecasting through weather‐type classification frameworks; yet two limitations remain: (1) Numerical weather prediction data are overly smoothed, and errors are amplified under complex weather conditions, causing multi‐stage error accumulation; and (2) weather type labelling methods are constrained – manual labelling is labour‐intensive and subjective, while clustering based on statistical features struggles to capture nonlinear patterns, making accurate labelling difficult for complex weather. To address these issues, we propose a short‐term PV power classification forecasting method leveraging spatio‐temporal irradiance correlations and contrastive learning. First, a temporal convolutional network corrects predicted irradiance at the target site using correlations with neighbouring sites. Next, high‐dimensional features extracted from the corrected irradiance via a TS2Vec contrastive learning model enable robust weather type classification. A Timemixer++‐based model then predicts weather types for the target day. Finally, a classification‐based forecasting framework uses the predicted weather types to generate PV power forecasts. Simulation results demonstrate that the proposed framework significantly improves the accuracy of short‐term PV power forecasting.