Posters

Abstract

The inherent variability and stochastic nature of wind power generation presents significant challenges for grid integration and system stability. This paper proposes a novel hybrid deep learning model that combines Grey Relational Analysis (GRA), Wavelet Packet Decomposition (WPD), and Gated Recurrent Unit (GRU) networks for ultra-short-term wind power forecasting. The methodology introduces three key innovations: (i) optimal meteorological feature selection using Pearson correlation analysis combined with GRA-based similar day pattern recognition, (ii) multi-resolution temporal decomposition through WPD to handle non-stationary characteristics, and (iii) parallel GRU networks optimized for different frequency components. The proposed framework first employs Pearson correlation to identify crucial meteorological parameters (wind speed, direction, temperature, and pressure) affecting wind power generation. The GRA module then constructs an optimal training dataset by analyzing historical similar day patterns using grey correlation degrees (γ). The original wind power time series P(t) is decomposed into n frequency sub-bands through L-level WPD, producing detailed coefficients (D₁, D₂, ..., Dₙ) and approximation coefficients (A₁, A₂, ..., Aₙ). Each sub-sequence is processed by an independent GRU network with optimized hyperparameters, capturing temporal dependencies at different scales. The final prediction P`(t) is obtained through wavelet reconstruction of individual GRU outputs. The model's performance was validated using 6-month operational data from a real wind farm, with 15-minute resolution for ultra-short-term (4-hour ahead) predictions. Experimental results demonstrate significant improvements over benchmark models, with an average improvement of 12.5% in forecasting accuracy compared to traditional deep learning methods, such like GRU and LSTM. The model exhibits robust performance across different prediction horizons and weather conditions, evaluated through Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and coefficient of determination (R²). The proposed methodology maintains consistent accuracy across varying weather conditions and seasonal changes, showing particular strength in handling extreme weather scenarios where traditional methods exhibit significant performance degradation. Cross-validation tests across seasonal variations demonstrate the model's stability and generalization capability. The experimental results establish that the proposed GRA-WPD-GRU framework effectively addresses the challenges of ultra-short-term wind power prediction through its multi-scale analysis capabilities and optimized feature selection. The demonstrated superior performance under varying weather conditions indicates significant potential for real-time grid operations. Future research directions include extending the framework to incorporate spatial correlations among multiple wind farms and exploring adaptive decomposition levels based on weather patterns.