Ultra-short-term wind power forecasting based on the MIFCformer model and a critical low wind speed region power revision strategy

The inherent intermittency, randomness, and volatility of wind power generation pose significant challenges to the stable operation of large-scale power grids. Accurate ultra-short-term forecasting is crucial for maintaining grid safety. Most existing ultra-short-term wind power forecasting methods have limited effectiveness in modeling multi-scale temporal dependencies and often fail to address the inadequate modeling capability under critical low wind speed conditions. To address this issue, this paper proposes a method for ultra-short-term wind power prediction using Multiscale and Interactive Fusion Convolution Transformer and Critical Low Wind Speed Region Power Revision Strategy (MIFCformer-CRS), to improve power prediction accuracy. The MIFCformer model is designed for initial power prediction. Leveraging multi-scale inputs, ProbSparse self-attention from Informer model, and temporal convolutional networks, the model enhances its capacity to capture complex patterns, while interactive top-down fusion convolution ensures deep fusion of multi-scale features. The revision strategy addresses overestimation and lag in initial predictions within critical low wind speed regions, enabling precise power adjustments. The revision strategy begins with constructing a Decomposition and Mixing of Factors (DMF) model for wind speed prediction. The wind speed sequence is decomposed into subsequences, processed individually using Multilayer Perceptron (MLP), and reconstructed to ensure prediction accuracy while effectively incorporating the influence of external factors. Next, predicted wind speed and preliminary power outputs are input into the multi-factor dynamic revision function to adjust the preliminary power prediction results. The proposed model was validated using wind farm data from Yunnan Province, China. Our method significantly outperformed 12 benchmark models, including LSTM, Informer, FEDformer, DLinear, and the hybrid method WSTD-Autoformer. Compared with other models, the proposed approach achieved reductions in Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) ranging from 7.08% to 34.27%, and attained improvements in the highest coefficient of determination (R²) by 2.48% to 17.91%. These results demonstrate the superior prediction performance of the proposed model.