Dual-path ultra-short-term wind power forecasting based on numerical weather prediction and multi-order temporal dynamic gating fusion

As the proportion of wind power in the power grid increasing, accurate forecasting of wind power generation has become a critical requirement for grid scheduling. However, due to the influence of unstable weather factors such as wind speed and wind direction, the accuracy of ultra-short-term wind power forecasting faces substantial challenges. To address this issue, a novel comprehensive wind power forecasting framework is established by integrating numerical weather prediction (NWP) with a dual-path model fusion strategy. First, maximal information coefficient is employed to select weather factors highly correlated with wind power, and complete ensemble empirical mode decomposition with adaptive noise is applied to decompose the power sequence into multiple intrinsic mode functions. Whereafter, a dual-path forecasting framework is then constructed, where one path uses extended long short-term memory (xLSTM) to forecast the wind power, and the other path employs extreme gradient boosting (XGBoost) to forecast the wind power combined with key meteorological features from NWP. The dual-path forecasting results are obtained by superposition. Furthermore, an innovative multi-order temporal dynamic gating fusion module is designed to dynamically fuse the dual-path forecast results through the enhanced attention mechanism and the gating network to obtain the final forecast results. The proposed method is validated using datasets from a wind farm in China in June and December. The results show that the average normalized mean absolute error of the proposed method reach 0.0078 and the average normalized root mean square error reach 0.0093, which leads to a reduction in forecast error by 55.65 % and 59.74 %, respectively. Compared to the xLSTM and NWP-XGBoost models, the proposed method significantly improves the accuracy and stability of wind power forecasting.