A multivariate ultra-short-term wind speed forecasting model by employing multistage signal decomposition approaches and a deep learning network

Abstract

Wind speed forecasting (WSF) accuracy is vital for exploiting renewable and environment-friendly wind energy. Therefore, a hybrid WSF model, namely VIL (VMD-ICEEMDAN-LSTM), was constructed in this study. VIL is a three-layered structure utilizing variational mode decomposition (VMD), improved complete ensemble empirical mode decomposition with additive noise (ICEEMDAN), and a long short-term memory (LSTM) neural network optimized using a differential evolutionary algorithm (DEA). The VIL was trained on multivariate meteorological data collected from Sujawal station, Pakistan. The partial autocorrelation function (PACF) analysis and an ensemble of VMD and ICEEMDAN techniques were employed to determine the appropriate input variables. The forecasted results obtained through LSTM were combined to create an ensemble forecasting model for the WSF. Seven other models including radial basis function (RFB), support vector regression (SVR), random forest regression (RFR), extreme learning machine (ELM), extreme gradient boosting (XGB), gated recurrent unit (GRU), and LSTM were developed as benchmarks for comparison purposes. As per the results, the VIL model revealed the lowest statistical errors, the best goodness-of-fit, the consistency of variation trend following the pattern of the observed wind speed, and the nearly matched probability distribution with the observed wind speed data. VIL reduced MAPE compared to RFB, SVR, RFR, ELM, XGB, GRU, and LSTM models by 7.297%, 5.704%, 4.993%, 5.233%, 6.514%, 5.013%, 5.204%, respectively, during testing. Similarly, VIL revealed highest R2 = 0.984 value during testing compared to the counterpart models including RFB (R2 = 0.911), SVR (R2 = 0.927), RFR (R2 = 0.930), ELM (R2 = 0.932), XGB (R2 = 0.931), GRU (R2 = 0.934), and LSTM (R2 = 0.935). The results validated the efficacy and statistical significance of the VIL model over the seven other constructed standalone models. The superior results of the VIL model compared to the counter developed models highlight the viability of integration of VMD and ICEEMDAN decomposition approaches and LSTM deep learning network for multivariate WSF. Therefore, the VIL model is performant for WSF and has good prospects for the energy-environment-transportation-health nexus.