Insulated Gate Bipolar Transistors (IGBTs) are widely deployed in power electronic systems due to their superior performance. However, at the same time, they are one of the most critical and fragile components in electronic systems. The failure prediction of IGBTs can precisely forecast the potential risk to guarantee system reliability. In this paper, Bayesian-optimized CEEMDAN is adopted to extract fault features efficiently, and a prognostic model named Performer-KAN is proposed for IGBT failure prediction. The proposed model combines the efficient FAVOR+ mechanism from the Performer with the flexible spline-based activation of the Kolmogorov–Arnold Network (KAN), enabling improved nonlinear approximation and predictive precision. Comprehensive experiments were conducted using the IMFS, which were decomposed by BO-CEEMDAN. The model’s performance was evaluated using key metrics such as MAE, RMSE, and R². The Performer-KAN demonstrates superior prediction accuracy while maintaining low computational overhead, compared to six representative deep learning models. The results demonstrate that the proposed method offers a practical and effective solution for real-time IGBT health monitoring and fault prediction in industrial applications.

Short-term wind power forecasting is challenging because power output exhibits strong non-stationarity, stochastic fluctuations, and multi-scale temporal characteristics. This study proposes an ICEEMDAN-EWT forecasting framework whose principal methodological contribution is a complexity-based network assignment strategy. ICEEMDAN decomposes the original power series, and EWT further refines the highest-frequency intrinsic mode function. Zero-crossing rate, spectral entropy, and volatility are integrated into a complexity score, according to which TCN-BiLSTM-Attention, BiLSTM-BiGRU-Attention, and GRU are assigned to high-, medium-, and low-complexity components, respectively. The component forecasts are reconstructed, and a BiLSTM-Attention model is further employed for residual error correction. Experiments are conducted separately on wind power datasets from Turkey and Inner Mongolia. Relative to TCN, the proposed model reduces MAE, RMSE, and NMAPE by 87.67%, 89.51%, and 87.67% on Dataset A, and by 79.92%, 81.11%, and 79.92% on Dataset B, respectively. The results demonstrate the effectiveness of combining complexity-adaptive network assignment with multi-scale decomposition and residual correction for short-term wind power forecasting.