CFTL: A Future-Aware Asymmetric-Stacking Feature Fusion Network for Multi-Task Wind Power Forecasting

The significance of Wind Power Forecasting (WPF) becomes evident within the strategic planning of wind power integration and operation. Consequently, considerable attention has been directed towards this domain in recent times. However, accurate wind power forecasting faces challenges from the following two aspects: i) a substantial number of abnormal data, and ii) complex correlations between historical and future data. To address these challenges, we propose a hybrid anomaly correction algorithm and an asymmetric-stacking model (named CFTL) that can capture the history-future correlations for wind power forecasting. The hybrid anomaly correction algorithm consists of a learning-based detection module, an RF&KNN-based hybrid imputation module, and a curve-based adjustment module. The CFTL leverages Temporal Convolutional Networks (TCN), Long Short-Term Memory (LSTM), and Residual Networks (ResNet) for multi-scale fusion of historical and future features. In the Longyuan Wind Power Track of the “China Software Cup” University Student Software Design Competition
¹
(henceforth, referred to as “Competition”), the proposed algorithm and CFTL achieved 1st place in Phase 1 and 6th place in Phase 2, demonstrating its significant advantages in WPF.