Open Access

ARTICLE

Wind Power Forecasting Utilizing Bidirectional Gated Recurrent Units in Conjunction with Empirical Mode Decomposition and Bayesian Neural Networks

Xiaolan Li^1,2, Yanting Wang^1,2,*

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

* Corresponding Author: Yanting Wang. Email:

(This article belongs to the Special Issue: Advances in Renewable Energy Systems: Integrating Machine Learning for Enhanced Efficiency and Optimization)

Energy Engineering 2026, 123(7), 7 https://doi.org/10.32604/ee.2026.076417

Received 20 November 2025; Accepted 19 January 2026; Issue published 18 June 2026

Abstract

To address the operational challenges of power systems with high renewable penetration, this research targets the non-stationarity and stochasticity of wind power. A novel hybrid framework for probabilistic forecasting and risk assessment is proposed. Initially, Empirical Mode Decomposition (EMD) adaptively decomposes the raw power signal into multi-scale Intrinsic Mode Functions (IMFs) and a residual trend, effectively segregating temporal features and reducing complexity. These components are then fused with historical data to form a comprehensive input. The core predictor is a Bidirectional Gated Recurrent Unit (BiGRU) network enhanced with a Temporal Attention (TA) mechanism. The BiGRU captures bidirectional long-term dependencies, while the TA mechanism dynamically focuses on the most influential historical time steps, enabling precise temporal pattern extraction. To quantify uncertainty, a Bayesian Neural Networks (BNNs) layer is integrated, transforming deterministic point forecasts into probabilistic outputs with prediction intervals. Finally, leveraging these probabilistic forecasts, the Value at Risk (VaR) metric is applied to assess potential operational risks under specified confidence levels, translating uncertainty into quantifiable reliability or financial risk. Simulation results confirm the framework’s superiority, achieving a normalized Root Mean Square Error (nRMSE) of 15.73% and a normalized Mean Absolute Error (nMAE) of 10.94%, significantly outperforming benchmarks. The innovative integration of signal processing, attentive deep learning, Bayesian inference, and risk theory within a unified model enhances forecasting accuracy, quantifies uncertainty, and enables proactive risk assessment, providing robust decision support for grid dispatch and renewable integration.

---

Keywords

---