Abstract
Wind energy's stochasticity and volatility challenge grid stability and dispatch reliability. To overcome limitations of existing decomposition-based forecasting methods, this paper proposes an Improved Escape Algorithm (IESC) incorporating chaotic mapping to optimize the hyperparameters of Time-Varying Filter Empirical Mode Decomposition (TVF-EMD), effectively mitigating mode mixing and enhancing non-stationary wind signal separation. Departing from uniform modeling, we employ a frequency-adaptive strategy: XLSTM captures high-frequency volatility, LSTM models medium-frequency transitions, and ELM rapidly processes low-frequency trends. Evaluated on a large-scale dataset, IESC outperforms standard ESC, GWO, and DE by 6.2%, 11.3%, and 8.4%, respectively. The proposed hybrid model demonstrates superior robustness, achieving a 29.8% lower 1-step MAE (0.5109) and a 65.6% higher 15-step R² (0.7685) compared to XLSTM alone. Crucially, error growth (1-15 steps) is contained within 12% and R² degradation is 35% slower. These results confirm that the method significantly enhances forecasting precision and effectively bridges multi-step accuracy with real-time dispatch needs, ensuring dynamic grid-demand matching and improved operational stability.