一种端到端集成学习方法以提升风电功率预测精度

Accurate wind power forecasts are crucial for grid stability, thereby ensuring a reliable and efficient power supply. To characterize the complicated fluctuation of wind power, numerous ensemble models with multiple base forecasters have been developed. However, most existing ensemble forecasting models contain several modeling stages, which increases the risk of error accumulation and inefficiency in model training. Moreover, the limited number of base forecasters results in forecasts with reduced diversity, thereby diminishing the performance of ensemble models. To address these challenges, MG-DS, a simple but efficient end-to-end ensemble learning model based on the Dempster-Shafer (DS) evidence theory, is proposed to unify base model learning and ensemble learning into a single process. It comprises an all-MLP-based nonlinear feature extraction module, a GRU and cross attention-based base forecast generation module, and a DS-based self-ensemble forecasting module with a DS-based magnifying glass to enhance the diversity of base forecasts. Further, a DS-based self-ensemble (DSSE) plugin is proposed to integrate the trained RNN-type and non-RNN-type base forecasters. Experiments on five wind power datasets show that MG-DS outperforms popular wind power forecasting models and ensemble techniques, and the effectiveness of the DSSE plugin is also validated in enhancing the performance of ensemble wind power forecasting.