A Hybrid Decomposition-Based Deep Learning Approach for Enhanced Wind Power Forecasting
Almansoori, Abdulrahman Khaled
Almansoori, Abdulrahman Khaled
Supervisor
Department
Machine Learning
Embargo End Date
2025-05-30
Type
Thesis
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
The massive demand for fossil fuels has negatively impacted our climate. Renewable energy sources, such as wind, solar, and hydraulic energy, have recently become an excellent choice for sustainable and climate-friendly energy. Wind turbine energy generators have become an excellent solution in this area. However, the non-stationary nature of the wind presents challenges in achieving high levels of efficiency. To address this issue, this study has proposed a method that considers a hybrid prediction model integrating three preprocessing and prediction modules, which are a Complete Ensemble Empirical Mode Decomposition with an Adaptive Noise (CEEMDAN) module, an Empirical Wavelet Transform (EWT) module, and a Gated Recurrent Unit (GRU) module. Those components have been utilized to process the data by reducing the noise and providing stability to improve the GRU model’s captivity in the prediction. Furthermore, the proposed integrated CEEMDAN-EWT-GRU model has been evaluated in three structured experiments, each with unique objectives to see the effectiveness of the proposed method. The first experiment consists of evaluating the proposed method against baselines of deep learning models, benchmark [1], and a hybrid model that integrates CEEMDAN and EWT across time horizons before selecting the optimal time horizons that will be selected for further investigation. The second experiment evaluates the performance of the proposed method by integrating more features, which is a novelty in this work. In the third experiment, we introduce a hybrid method, CEEMDAN-EWT-GRU-BiLSTM, which applies CEEMDAN and EWT to the original data. The Intrinsic Mode Functions (IMFs) will be divided into highfrequency and lowfrequency IMFs using the permutation entropy. The GRU will be trained on highfrequency IMFs, while Bi-LSTM will be trained on lowfrequency IMFs. The findings of this research confirm that the integration of the decomposition technique has provided superior accuracy and robustness compared to the standalone deep learning prediction models. The GRU has shown more suitability for integration with the CEEMDAN and EWT for realtime forecasting applications, especially in wind energy systems. This work has a number of contributions to wind power forecasting: (1) introducing a more efficient method CEEMDAN-EWT-GRU for shortterm forecasting, (2) demonstrating the effectiveness of the GRU against LSTM and Bi-LSTM, (3) evaluating the effect of integrating more features into the proposed method, (4) introducing the hybrid method of GRU and Bi-LSTM based on the frequency of IMFs, and (5) evaluating the performance of the methods for varying historical and forecasting horizons.
Citation
Abdulrahman Khaled Almansoori, “A Hybrid Decomposition-Based Deep Learning Approach for Enhanced Wind Power Forecasting,” Master of Science thesis, Machine Learning, MBZUAI, 2025.
Source
Conference
Keywords
Wind Power Forecasting, Deep learning, Decomposition Techniques, CEEMDAN, EWT