Wind power forecasting using hybrid ICEEMDAN-KPCA and IWOA-BiLSTM models

• ICEEMDAN+KPCA extracts compact nonlinear features from noisy meteorological sequences. • An improved WOA (chaotic map + adaptive weights + mutation) enhances global search and convergence. • IWOA-tuned BiLSTM captures nonlinear temporal dependencies and boosts forecasting accuracy. • Real wind farm tests outperform benchmarks in RMSE, MAE and R² across multiple horizons. • The framework is practical for deployment and extensible to other energy forecasting tasks. This paper addresses inadequate noise suppression, weak feature representation, and unstable model training and hyperparameter optimization in traditional short-term wind power forecasting under complex wind-farm conditions. We propose a forecasting framework that integrates improved complete ensemble empirical mode decomposition with Adaptive noise (ICEEMDAN) for signal decomposition, kernel principal component analysis (KPCA) for feature construction, and an improved whale optimization algorithm (IWOA) to tune a bidirectional long short-term memory (BiLSTM) predictor. First, ICEEMDAN decomposes wind-power and meteorological series into multiscale intrinsic mode functions, thereby reducing nonstationarity and mode mixing. Next, KPCA extracts salient cross-scale features while removing redundancy and reducing dimensionality. Finally, IWOA improves the balance between global exploration and local refinement to fine-tune BiLSTM hyperparameters, thereby mitigating errors arising from heuristic settings. Validation on real wind-farm data shows that the proposed method outperforms classical models and strong contemporary baselines in accuracy, demonstrating its effectiveness for short-term wind power forecasting.