A hybrid approach for wind power forecasting using successive variational mode decomposition and long short-term memory

This study presents novel hybrid model aimed at improving the accuracy of short-term wind power forecasting (WPF) by combining Successive Variational Mode Decomposition (SVMD) with Stacked Long Short-Term Memory (SLSTM) networks. The wind power time series exhibits non-linear and non-stationary characteristics. The SVMD has the ability to extract features from such complex data which makes it effective for identifying patterns. In the first stage, the SVMD is utilised to decompose the WP data into various modes, isolating different patterns in the time series without any need of predefining the number of modes. Subsequently, the second stage utilises the SLSTM to capture intricate temporal dependencies and hierarchical patterns within time series which is essential for predicting complex sequences. The stacking of multiple LSTM layers helps in modelling hierarchical patterns, making it possible to understand both short-term variations and long-term trends within WP. The proposed SVMD-SLSTM model achieved overall improvement of 54.67% in seasonal forecasting accuracy. Specifically, it demonstrates impressive performance during winter, summer and fall periods, with Mean Absolute Percentage Errors (MAPE) of 2.346%, 1.267% and 1.255%, respectively. It combines the strengths of SVMD’s feature extraction with SLSTM’s advanced temporal modelling, making it an effective model for short-term WPF.