Wind power prediction based on meteorological forecast and convolution simplified long short-term memory network

Abstract

To enhance the accuracy of ultra-short-term wind power forecasting, this study proposes an efficient prediction model that integrates meteorological forecasts with a Convolutional Simplified Long Short-Term Memory Network(ConvSLSTM).Firstly, to address the common issue of outliers in wind power datasets, the boxplot method is applied for outlier detection, and the K-nearest neighbor imputation method is utilized to correct the detected outliers, thereby improving data quality and reliability. Secondly, to simplify model inputs and enhance feature relevance, the Maximal Information Coefficient(MIC) is employed to identify key meteorological factors strongly correlated with wind power, such as wind speed and hub-height wind speed. Subsequently, convolutional operations are incorporated into the Long Short-Term Memory(LSTM) network to strengthen feature extraction capabilities. A novel cross-coupled gating mechanism and peephole structure are further designed to reduce network parameters and improve training efficiency. Finally, leveraging both historical data and numerical weather prediction data, the ConvSLSTM model is used to achieve high-precision forecasting of wind power sequences.Experimental result demonstrate that, compared with ConvSLSTM without forecasts, Forecast-LSTM, and Forecast-Transformer models, the proposed method achieves reductions in mean absolute error(MAE) of 42.18%, 2.26%, and 32.66%, respectively, across multiple forecasting horizons.The ConvSLSTM model based on meteorological forecasts exhibits superior prediction accuracy, stronger robustness, and better generalization capability. These result validate its significant advantages in wind power forecasting tasks and indicate its broad application potential.

Keywords

convolutional simplified long short-term memory network / boxplot method / K-nearest neighbor complementation method / numerical weather prediction