Improving Multivariate Runoff Prediction Through Multistage Novel Hybrid Models

The exploitation of hydropower provides cleaner, more sustainable, and cheaper energy than fossil fuels. Therefore, hydropower offers prospects to meet the sustainable development goals of the United Nations. These benefits motivate this study to develop different models for efficient runoff prediction utilizing multivariate hydro-meteorological data. The techniques employed for this purpose include correlation analysis, time series decomposition, sample entropy (SE), and sequence2sequence (S2S) algorithm with spatio-temporal attention (STAtt). The decomposition techniques include improved complete ensemble empirical mode decomposition with additive noise (ICEEMDAN) and the maximal overlap discrete wavelet transform (MODWT). The ICEEMDAN-STAtt-S2S model reveals the best prediction results over the counterpart hybrid and standalone models in terms of statistical metrics and comparison plots. The ICEEMDAN-STAtt-S2S model decreases RMSE by 19.348 m3/s, 14.35 m3/s, 13.937 m3/s, 13.681 m3/s, 11.988 m3/s, 9.066 m3/s, 7.7 m3/s, 7.129 m3/s, 5.511 m3/s, 4.071 m3/s, 2.011 m3/s for SVR MLR, MLP, XGBoost, LSTM, S2S, SAtt-S2S, TAtt-S2S, STAtt-S2S, MODWT-STAtt-S2S, and ICEEMDAN-SE-STAtt-S2S models, respectively. In terms of NSE, the ICEEMDAN-STAtt-S2S model is 10%, 7.4%, 7.3%, 7.1%, 5.9%, 4.5%, 3.7%, 3.4%, 2.6%, 1.9%, and 0.9% more efficient compared to SVR MLR, MLP, XGBoost, LSTM, S2S, SAtt-S2S, TAtt-S2S, STAtt-S2S, MODWT-STAtt-S2S, and ICEEMDAN-SE-STAtt-S2S models, respectively. The surpassed prediction outcomes substantiate the merger of ICEEMDAN and S2S utilizing STAtt for runoff prediction. Moreover, ICEEMDAN-STAtt-S2S offers the potential for reliable prediction of similar applications, including renewable energy, environment monitoring, and energy resources management.