Cross-Modal Contrastive Learning for Net Load Prediction Based on Public Weather Forecasts

As the penetration rate of distributed photovoltaics (PVs) increases, behind-the-meter (PVs) amplify the impact of meteorological factors on net load predictions. Public weather forecasts (PWFs) can serve as an alternative meteorological covariate to ensure prediction accuracy when numerical weather prediction (NWP) data are unavailable owing to high costs or data delays. PWFs are freely accessible to the public and significantly influence resident behavior. However, they provide limited information and have inherent heterogeneity with historical net load data in terms of data structure and features, posing challenges in feature extraction and fusion. To address these challenges, this study proposes a day-ahead net load prediction model cross-modal contrastive learning network (CCL-Net) that incorporates PWFs. The CCL module employs contrastive learning to minimize the gap between the representations of temporal meteorological data and PWF features, enriching the semantic features of PWFs. Considering the heterogeneity in features and data structure between future weather forecasts and historical net load data, CCL-Net incorporates a cross-modal fusion block to fuse both features, enhancing prediction accuracy. The proposed model is compared with benchmark methods, demonstrating superior performance. In addition, the analysis and ablation experiments of cross-modal technologies validate the rationality and superiority of the proposed approach.