A two stage feature extraction and synchronized feature–parameter learning framework for reliable multistep wind speed forecasting

To address the pronounced non-stationarity, heavy noise, and mixed multi-scale characteristics inherent in wind speed series, this study proposes a novel multi-step forecasting framework (IESF-ICEEMDAN-AFS-XGBABiGRU-HO) integrating two-stage decomposition, synchronous feature–parameter optimization, and differentiated modeling. The framework first employs Improved Exponential Smoothing Filtering (IESF) with adaptive smoothing parameter α to extract global trends and mitigate non-stationarity. Subsequently, the residual series undergoes Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN), where Laplacian noise injection enhances stochastic perturbations to suppress mode mixing, complemented by a Bayesian adaptive β strategy to enable fine-grained multi-scale separation. Phase space reconstruction generates dynamic inputs, while extreme gradient boosting (XGBoost) and bidirectional gated recurrent units (BiGRU) differentially model trend components and intrinsic mode functions (IMFs) respectively. The Hippopotamus Optimization (HO) algorithm, integrated with adaptive feature selection (AFS), concurrently optimizes lag orders and model hyperparameters in a unified search space, effectively resolving the feature–parameter decoupling problem prevalent in conventional methods. Experimental results demonstrate that: (1) the IESF-ICEEMDAN two-stage decomposition with Laplacian noise lowers forecasting errors and improves decomposition accuracy relative to single-stage methods; (2) Diebold–Mariano tests at the 1% significance level confirm the framework’s superior and robust predictive performance over benchmark models. The proposed method therefore provides a precise and reliable tool for wind speed forecasting in power systems with high wind-power penetration.