River water temperature prediction using hybrid machine learning coupled signal decomposition: EWT versus MODWT

Accurate prediction of water temperature (Tw) will greatly help in improving our understanding of the overall thermal regime fluctuation, and it can help in making sound decisions. While great efforts have been devoted to the development of Tw models, further improvement in the prediction accuracy is challenging. Here, we propose a new hybrid machine learning (ML) models for predicting Tw from air temperature (Ta). First, two signal decomposition algorithms, i.e., the empirical wavelet transform (EWT) and maximum overlap discrete wavelet transform (MODWT) are used for decomposing the Ta into several subsequences. Second, the obtained subsequences are used as input variables for four ML models, i.e., the feedforward neural network (FFNN), the optimally pruned extreme learning machine (OPELM), the adaptive boosting (AdaBoost), and the bootstrap aggregating (Bagging). The development of the hybrid models is based on measured data form five measurement stations distributed over Poland. The experimental results show that, the new hybrid ML models achieved high predictive accuracies with Pearson correlation coefficient (R), Nash-Sutcliffe efficiency (NSE), root mean square error (RMSE) and mean absolute error (MAE) of approximately ≈0.99, ≈0.98, ≈0.718 °C, and ≈0.599 °C, respectively. Finally, the experimental results have demonstrated that, the MAE and RMSE of the single models were reduced by 66.12% and 65.30%, while the R and NSE values were improved by 4.85% and 10.02%, respectively, showing a very promising prediction performance. Our new modelling approach proposed in the present study clearly highlight the high contribution of the EWT and MODWT in improving the Tw estimation that is most strongly influenced by Ta. The significant difference between single and hybrid models can be explained by the ability of the EWT and MODWT to capture the high nonlinearity between Ta and Tw. Our new approach is of interest for water resources management and for assessing the variability of Tw over time and space. Finally yet importantly, this is the first study in the literature for which EWT and MODWT were used for modelling Tw, through which we have demonstrated that, an important amount of information is available and cannot be captured using single ML models, and signal decomposition have helped to overcome this challenge.