Ultra-short-term wind power forecasting method based on multi-variable joint extraction of spatial-temporal features.

The article focuses on the development and validation of a novel ultra-short-term wind power forecasting method named MDSLS, which integrates the maximum information coefficient (MIC), deep separable convolutional neural network (DSCNN), and long- and short-term time-series network (LSTNet) to jointly extract multi-variable spatial and temporal features. Using real wind farm data from China, the method demonstrates improved forecasting accuracy over benchmark models such as LSTM, GRU, BiLSTM, BPNN, and Seq2Seq, reducing mean absolute error by up to 4.66%. The approach incorporates MIC to select relevant input variables by capturing nonlinear correlations, DSCNN to efficiently extract spatial features with reduced model complexity, and LSTNet to capture both long-term and short-term temporal dependencies, enhanced by an autoregressive module. Comprehensive experiments, including ablation studies and interval predictions using kernel density estimation and Gaussian mixture models, confirm the method’s effectiveness, interpretability through causal analysis, and its potential to provide more precise and reliable wind power forecasts for power system operations.