Short-term wind power forecasting using the hybrid model of multivariate variational mode decomposition (MVMD) and long short-term memory (LSTM) neural networks

Abstract

This paper presents a novel hybrid forecasting procedure for wind power using meteorological and historical data. The introduced method consists of three parts: effective feature selection, time series decomposition, and forecasting each decomposed time series. The minimum redundancy and maximum relevance (mRMR) algorithm is first utilized to choose the most effective features. In this stage, those selected historical features whose values are needed at the prediction time will be decomposed by the variational mode decomposition (VMD) technique and then forecasted by the long short-term memory (LSTM) networks. Then, the multivariate variational mode decomposition (MVMD) algorithm is exploited to simultaneously decompose selected features to address frequency mismatches between different series and capture the correlation among them. Given that various series and variables are involved in wind power forecasting, considering the correlation among them significantly affects prediction results. Afterward, LSTM neural networks are utilized to forecast each decomposed time series. Finally, two cases and several evaluation criteria are elaborated to assess the performance of the presented method. Experimental results confirm that the developed hybrid model, compared to the VMD-LSTM model, results in a decrease of 9.97, 4.33, and 3.32% in root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE), respectively. The mean values of these criteria are, respectively, 4.6, 3.5, and 20.8 for the proposed model.