A correction framework for day-ahead NWP solar irradiance forecast based on sparsely activated multivariate-shapelets information aggregation

Abstract

Numerical Weather Prediction (NWP) is widely used in day-ahead solar irradiance forecast, which is of great significance to the optimization of power systems. Due to unescapable inherent errors of numerical techniques, NWP need correction. However, most correction schemes lack of error analysis, making the correction insufficiently efficient. Meanwhile, obtaining sufficient historical data in practical applications is challenging. Therefore, it is important to utilize the limited historical data to provide more meaningful information and increase the utilization of data information. To solve these problems, this paper proposes a day-ahead NWP solar irradiance correction framework. NWP global horizontal irradiance (GHI) error analysis is first conducted to determine the correction parts. Then, multivariate-shapelets analysis (MSA) is performed to select samples with high correlation to the correction parts. Mixture-of-experts (MoE) is adopted to sparsely activate high correlation samples contributing more to enhancing accuracy. Finally, a sequence-level information aggregation named SAIA is employed to obtain the corrected NWP forecasts. The proposed MSA-SAIA is evaluated with publicly available dataset and actual field dataset. The results demonstrate that MSA-SAIA yields the highest improvement, with increases of 16.15 % and 19.65 %. Additionally, we analyzed the performance of MSA-SAIA across different weather conditions, indicating its superior weather robustness.