Ultra-Short-Term Wind Power Forecasting Integrating Multi-Scale Decomposition and Dual Dependency Interaction

Abstract

Accurate wind power forecasting is an essential prerequisite for ensuring the safe and stable operation of power systems and improving the scheduling and planning capability of power grids. To address the large prediction errors caused by the strong nonlinearity and non-stationary fluctuations of wind power time series, as well as the coupling effects among multiple meteorological variables, this paper proposes a wind power forecasting model integrating multi-scale decomposition, dual dependency interaction, and cross-variable linear mapping. The multi-scale decomposition module employs multi-scale average pooling to separate the trend and periodic components of the sequence, thereby effectively mitigating the non-stationary interference of the original series. The dual dependency interaction mechanism explores long-term temporal correlations and coupling relationships among meteorological factors from both temporal and variable dimensions. Finally, a cross-variable linear structure is adopted to accomplish prediction. Experiments are conducted using annual measured data collected from a wind farm in Inner Mongolia, China. The proposed model is compared with several mainstream forecasting models, including Informer, xLSTM-Informer, GRU, and CNN-LSTM. Experimental results demonstrate that the proposed model achieves MSE, RMSE, MAE, and R2 values of 10.32, 3.214, 1.966, and 0.961, respectively. Compared with the xLSTM-Informer model with the best overall baseline performance, the proposed model reduces MAE by 7.35% and cuts training time by 97.51%, thus conclusively demonstrating that the proposed method achieves much better forecasting accuracy without sacrificing training efficiency.