Short-term wind speed forecasting based on a novel KANInformer model and improved dual decomposition

Accurate short-term wind speed forecasting is crucial for optimizing wind power generation plans and ensuring the quality of power supply. However, the inherent nonlinearity and frequent fluctuations of wind speed render the task exceedingly challenging. This study proposes a hybrid model based on KANInformer and VMD-CA-EWT, with enhanced predictive and generalization capabilities. KANInformer, a novel predictor, integrates the multidimensional spatial expression of Kolmogorov-Arnold Networks (KAN) with the effective feature extraction of Informer, attaining robust adaptability and deep nonlinear mapping. VMD-CA-EWT consists of two decomposition processes and one aggregation process. Initially, Variational Mode Decomposition (VMD) decomposes the original wind speed into appropriate components. The Component Aggregation (CA) method is then introduced to aggregate highly unpredictable components into a new component. Finally, Empirical Wavelet Transform (EWT) further decomposes the fused component into multiple sub-modes. The improved dual decomposition effectively mitigates the random fluctuations and prevents incomplete decomposition. Four comparative experiments are conducted in Brentwood to assess the superior performance of the hybrid model. The MAPE for the 3-step prediction results across the four datasets reach 3.3 %, 5.8 %, 5.4 %, and 7.8 %, respectively. Results indicate that the proposed model adapts well to the nonlinear characteristics of wind speed, achieving reliable and stable predictive accuracy.