Short Term Wind Power Prediction Base on STL Decomposition and LSTM-Attention-Linear Model

This paper introduces a method for short-term wind power prediction that utilizes neural networks trained on historical wind power data and numerical weather prediction data. The neural network is based on a combination of temporal attention and Long Short-Term Memory (LSTM) networks. Initially, the data is preprocessed through time series decomposition to isolate features. Subsequently, these distinct features are fed into the prediction model to complete the wind power prediction. The paper concludes with the presentation and analysis of test results. The results indicate that the model has small errors and exhibits strong robustness. In cross-validation across different datasets, the average prediction accuracy can reach 90%. Compared to a single LSTM model, the accuracy has improved by over 30%, demonstrating a significant advantage in wind power prediction.