Open Access

ARTICLE

Ultra-Short-Term Wind Power Forecasting Based on Hierarchical Signal Refinement and Intelligently Optimized Deep Learning

Xiaolan Li^1,2,*, Jinyu Shen^1,2, Jinhuang Liang^1,2, Yanting Wang^1,2

1 School of Electrical and Information Technology, Yunnan Minzu University, Kunming, China
2 Yunnan Key Laboratory of Unmanned Autonomous System, Kunming, China

* Corresponding Author: Xiaolan Li. Email:

Energy Engineering 2026, 123(7), 21 https://doi.org/10.32604/ee.2026.076521

Received 22 November 2025; Accepted 05 January 2026; Issue published 18 June 2026

Abstract

The intrinsic volatility and stochasticity of large-scale wind power generation pose significant challenges to grid stability. To address the limitations of conventional models in capturing strong non-stationarity, this study proposes a novel Multi-Stage Adaptive Forecasting Network (MSAF-Net). The framework features a hierarchical signal refinement strategy coupled with an intelligently optimized hybrid predictor. Initially, input redundancy is minimized via Pearson Correlation Coefficient (PCC) analysis to isolate significant meteorological variables. A two-phase decomposition-reconstruction mechanism is then implemented: the wind power series is first decomposed using Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN). To optimize the trade-off between signal complexity and computational cost, the resulting components are reconstructed based on Sample Entropy (SE), with the highest-complexity component specifically targeted for secondary denoising via Empirical Wavelet Transform (EWT). For the prediction stage, a hybrid architecture integrates Bidirectional Temporal Convolutional Networks (BiTCN) to extract multi-scale local features and Bidirectional Long Short-Term Memory (BiLSTM) networks to model long-term temporal dependencies. Crucially, an Attention Mechanism is embedded to weigh critical time steps, while the Sparrow Search Algorithm (SSA) automatically optimizes the network hyperparameters. Experimental results demonstrate that MSAF-Net achieves an RMSE of 41.59, MAE of 26.67, and MAPE of 1.36%. Notably, the proposed model achieves a 23.16% reduction in MAPE compared to the competitive CEEMDAN-EWT-LSTM benchmark, verifying its superior predictive accuracy and generalization capability.

---

Keywords

---