An improved hybrid model for wind power forecasting through fusion of deep learning and adaptive online learning

Abstract

Accurate and effective wind power forecasting is crucial for wind power dispatch and wind energy development. However, existing methods often lack adaptive updating capabilities and struggle to handle real-time changing data. This paper proposes a new hybrid wind power forecasting model that integrates the Maximal Information Coefficient (MIC), Density-Based Spatial Clustering of Applications with Noise (DBSCAN), an improved Harris Hawks Optimization (IHHO) algorithm, and an Adaptive Deep Learning model with Online Learning and Forgetting mechanisms (ADL-OLF). First, MIC is used to reconstruct input features, enhancing their correlation with the target variable, and DBSCAN is employed to handle outliers in the dataset. The ADL-OLF model enables continuous updating with new data through online learning and forgetting mechanisms. Its deep learning component consists of Bidirectional Long Short-Term Memory (BiLSTM) networks and self-attention mechanisms, which improve the prediction accuracy for sequential data. Finally, IHHO optimizes the parameters of the ADL-OLF model, achieving strong predictive performance and adaptability to real-time changing data. Experimental simulations based on actual wind power data over four seasons from a U.S. wind farm show that the proposed model achieves a coefficient of determination exceeding 0.99. Compared with 12 benchmark models (taking IHHO-ADL-OLF as an example), the Root Mean Square Error (RMSE) is reduced by more than 20%. These results indicate that the model significantly improves the accuracy and robustness of wind power forecasting, providing valuable references for the development and optimization of wind power systems.