A hybrid correction framework using disentangled seasonal-trend representations and MoE for NWP solar irradiance forecast

Day-ahead solar irradiance forecast holds important value for optimizing energy utilization within the power system and ensuring stable grid scheduling. The forecast outputs of numerical weather prediction (NWP) are widely acknowledged as one of the indispensable data sources for day-ahead solar irradiance forecast tasks. In previous studies, post-processing methods have generally been employed as correction models to enhance the accuracy of NWP solar irradiance forecasts. However, irradiance sequences contain complex mixed patterns and exhibit various seasonal periodic differences. Based on the analysis of NWP global horizontal irradiance (GHI) error characteristics in this study, errors in NWP GHI forecasts also show obvious seasonal variations. Given these issues, it is challenging for a single correction model to achieve good correction performance and strong seasonal robustness. Therefore, this paper proposes a hybrid model comprising representation learning module, feature sparse activation module, and encoder-decoder-based correction module to address the aforementioned problems. A contrastive-learning-based representation learning module named CoST is introduced to learn disentangled seasonal features and trend features of irradiance sequences. A learnable mixture-of-experts (MoE) layer is adopted to sparsely activate the seasonal-trend features that contribute more to improving correction accuracy. The encoder-decoder-based correction module takes the sparsely activated seasonal-trend features as inputs, achieving the final corrected NWP GHI forecasts. The correction performance of the proposed method was validated on both publicly available datasets and actual field dataset. The results for various datasets show that our proposed CoST-MoELSTM model achieves the highest improvement for NWP forecasts, with increases of 29.82 %, 36.54 %, and 26.58 %. Additionally, we conducted a detailed analysis of the correction performance of CoST-MoELSTM across different seasons, indicating its superior seasonal robustness.