Efficient Wind Power Forecasting Based on Data Reconstruction and the PSO-CNN-LSTM Hybrid Model

Accurate wind power forecasting contributes to the safe and stable operation of the power system. Influenced by climate factors, wind power fluctuates greatly, which affects the accuracy of model predictions. To extract deeper hidden information from existing data to mitigate the impact of this uncertainty on the prediction model. This paper proposes a Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) model that integrates feature transformation, data reconstruction, and particle swarm optimization (PSO) algorithms. First, Principal Component Analysis (PCA) is applied to the original feature data to perform feature transformation, effectively eliminating redundant feature information in the data. Then, Using K-Means to optimize the initial parameters of the Gaussian Mixture Model (GMM), followed by clustering the transformed data. This process reconstructs the data to create a dataset with similar feature characteristics. Finally, PSO is employed to refine the parameters of the CNN-LSTM model, which is then used for wind power forecasting. To validate its effectiveness, the proposed model is tested using real data from a wind power plant, followed by experiments comparing it with other models. The experimental results indicate that the K-Means-GMM-PSO-CNN-LSTM model delivers more accurate predictions and exhibits greater stability in performance compared to other models, especially at points with significant data fluctuations.