A nature-inspired and noise-assisted feature extraction integrating spatiotemporal attention-based sequence2sequence for multivariate wind speed prediction

• Published in: Stochastic environmental research and risk assessment Vol. 39; no. 1; pp. 343 - 359
• Main Authors: Sibtain, Muhammad, Li, Xianshan, Saleem, Snoober, Qurat-ul-Ain, Shi, Qiang, Li, Fei, Apaydin, Halit
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.01.2025; Springer Nature B.V
• Subjects: Alternative energy sources; Aquatic Pollution; Artificial intelligence; Attention; Chemistry and Earth Sciences; Computational Intelligence; Computer Science; Correlation analysis; Deep learning; Earth and Environmental Science; Earth Sciences; Efficiency; Emissions; Energy industry; Environment; Environmental research; Feature extraction; Greenhouse gases; Irradiance; Machine learning; Math. Appl. in Environmental Science; Multivariate analysis; Neural networks; Noise prediction; Optimization algorithms; Original Paper; Performance measurement; Physics; Prediction models; Probability Theory and Stochastic Processes; Renewable resources; Risk assessment; Solar radiation; Spatiotemporal data; Statistics for Engineering; Support vector machines; Temporal variations; Time series; Waste Water Technology; Water Management; Water Pollution Control; Wavelet transforms; Wind power; Wind speed; Multivariate wind speed prediction; ICEEMDAN; Harris Hawks optimizer; Spatiotemporal attention; Sequence2sequence
• ISSN: 1436-3240; 1436-3259
• DOI: 10.1007/s00477-024-02866-1

An efficient and reliable multistage feature extraction facilitates the selection of useful features from multivariate time series data, resulting in the enhanced predicting ability of prediction models. For this reason, this study proposes a hybrid prediction model (HPM), namely, CA-HHO-ICEEMDAN-STA-Seq2Seq based on multistage feature extraction, to enhance the prediction outcomes of multivariate wind speed prediction (MWSP). The proposed HPM integrates correlation analysis (CA), Harris Hawks Optimizer (HHO), improved complete ensemble empirical mode decomposition with additive noise (ICEEMDAN), and spatiotemporal attention (STA) based sequence2sequence (Seq2Seq). First, the CA selects variables demonstrating a significant correlation with the wind speed data that undergo further selection through HHO. Afterward, noise-assisted ICEEMDAN decomposes the variables selected through HHO into subcomponents. The subsequent stage extracts spatial and temporal features through STA. The last stage employs Seq2Seq to finalize the prediction task. The superior prediction outcomes of the proposed model over eleven other counterpart models using different plots and performance metrics substantiate its viability for MWSP. The CA-HHO-ICEEMDAN-STA-Seq2Seq model reduced MAE by 0.674 m/s, 0.648 m/s, 0.618 m/s, 0.593 m/s, 0.297 m/s, 0.269 m/s, 0.168 m/s, 0.094 m/s, 0.089 m/s, 0.086 m/s, and 0.017 m/s compared with the MLR, SVM, ANN, EGB, GRU, LSTM, Seq2Seq, SA-Seq2Seq, TA-Seq2Seq, STA-Seq2Seq, and CA-ICEEMDAN-STA-Seq2Seq models, respectively. Similarly, in terms of NSE, the CA-HHO-ICEEMDAN-STA-Seq2Seq model was 0.340%, 0.332%, 0.308%, 0.287%, 0.124%, 0.116%, 0.064%, 0.032%, 0.035%, 0.027%, and 0.001% more efficient than the MLR, SVM, ANN, EGB, GRU, LSTM, Seq2Seq, SA-Seq2Seq, TA-Seq2Seq, STA-Seq2Seq, and CA-ICEEMDAN-STA-Seq2Seq models. Besides MWSP, the proposed model is feasible for similar tasks, including solar irradiance, hydro, GHG emissions, and load predicting.