Decomposition and Harris hawks optimized multivariate wind speed forecasting utilizing sequence2sequence-based spatiotemporal attention.

Accurate wind speed forecasting (WSF) is important for effectively harnessing wind energy with clean and sustainable energy benefits. Therefore, this study develops different models established through the use of correlation analysis (CA) and decomposition techniques, Harris hawks optimization algorithm (HHO), and S2S (sequence2sequence) based spatial and temporal attention (STAt-S2S) for effective WSF. First, the CA selects variables of significant correlation with the wind speed data. In the next stage, improved complete ensemble empirical mode decomposition with additive noise (ICEEMDAN) and discrete wavelet transform with maximum overlap (MODWT) techniques are employed to decompose the components having significant correlation. Afterwards, HHO selects suitable features from the decomposed data. Finally, STA-S2S extracts spatial, temporal features and performs forecasting. The CA-ICEEMDAN–HHO–STAt-S2S and CA-ICEEMDAN-STAt-S2S models reveal better forecasting outcomes over the other standalone and hybrid foresting models. The RMSE, MAE, and sMAPE values presented by CA-ICEEMDAN-STAt-S2S are 0.639 m/s, 0.474 m/s and 15.710 m/s with NSE of 0.922. The lowest error values with the highest efficiency values of ICEEMDAN, HHO, and STAt-S2S-based hybrid models corroborate the feasibility of these models for WSF with equal applicability for similar time series applications. [Display omitted] • Multivariate meteorological data is utilized for wind speed forecasting. • MODWT and ICEEMDAN decompose data to reduce nonlinearity and nonstationary. • Harris hawks optimization selects suitable decomposed subcomponents. • Spatiotemporal attention extracts spatial and temporal features and Sequence2Sequence framework performs forecasting. • CA-ICEEMDAN–HHO–STAt-S2S and CA-ICEEMDAN-STAt-S2S models reveal better results than the other developed models. [ABSTRACT FROM AUTHOR]