A Short-Term Calgorithm Based on Improved LSTM Neural Network

[Introduction] The volatility and intermittency of wind energy pose significant challenges for large-scale wind power integration. An effective approach to address this issue is to provide accurate wind power forecasting. [Method] In response to this challenge, this study proposed a wind power forecasting model for deep learning neural networks based on an improved LSTM (Long Short-Term Memory) architecture. The model incorporated a wind power forecasting approach that included independently developed data anomaly detection and processing, wind speed feature extraction and hyperparameter tuning. To enhance the neural network model's ability to accurately learn the impact of wind speed features on wind power, a feature engineering method combining feature screening and feature augmentation was also defined. [Result] The simulation results demonstrate that the proposed data cleaning and data augmentation algorithm can enhance the accuracy of various machine learning algorithms by approximately 5%. Furthermore, the proposed improved LSTM neural network model, after data cleaning, outperforms traditional algorithms and state-of-the-art neural network algorithms in the industry, achieving a 2.5% increase in accuracy. [Conclusion] The improved approach not only exhibits robust capability in cleaning noisy data but also consistently outperforms other algorithms in terms of forecasting accuracy across all experiments. This model provides valuable guidance for practical applications in the field of wind power forecasting.