CE-FTSMixer: target-side CEEMDAN and a frequency-enhanced TSMixer for short-term wind power forecasting

Short-term wind power forecasting remains challenging because of high-frequency disturbances and multi-scale non-stationarity. We propose CE-FTSMixer, a lightweight TSMixer-based model combining target-side complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) and a frequency-enhanced block (FEB). CEEMDAN decomposes the target power series into scale-specific modes to reveal multi-resolution patterns, while FEB learns frequency-wise weights to suppress noise and retain informative spectral content. A parallel head generates multi-step forecasts. Experiments on a public SCADA dataset cover four prediction horizons and assess robustness to random missing values and additive Gaussian noise. Compared with TSMixer, PatchTST, and a temporal convolutional network (TCN), CE-FTSMixer achieves lower absolute errors and less degradation under corruption. Its robustness index (RI) is lower at every horizon, with average relative reductions of 35–52%. RI decreases as the forecast horizon increases. Ablation and sensitivity studies indicate that target-side decomposition primarily improves scale-sensitive losses, whereas FEB enhances robustness, especially at longer horizons. Combining both components delivers the best overall accuracy and stability, supporting multi-scale representation and frequency-domain suppression for ultra-short-term wind power forecasting.