Latest revision as of 12:42, 25 August 2026

Abstract

This study presents a hybrid modelling framework for predicting river water temperature (Tw) by combining advanced boosting algorithms with signal decomposition techniques. Four models were compared: (i) categorical boosting (CatBoost), (ii) extreme gradient boosting (XGBoost), (iii) adaptive boosting (AdaBoost), and (iv) light gradient boosting machine (LightGBM). Each model was developed using air temperature (Ta) as input, both with and without signal decomposition. Three algorithms, empirical mode decomposition (EMD), ensemble EMD (EEMD), and complete EEMD with adaptive noise (CEEMDAN), were applied to decompose Ta into intrinsic mode functions (IMFs), which served as additional predictors. The models were trained and validated using data from four Polish river stations, and their performance was evaluated using root mean squared error (RMSE), mean absolute error (MAE), correlation coefficient (R), and Nash–Sutcliffe efficiency (NSE). Tw was modelled as an instantaneous function of Ta, without incorporating a forecasting horizon, to assess the ability of decomposition-boosting hybrids to reproduce observed thermal dynamics at the same time step rather than perform multi-step predictions. Unlike previous single-hybrid studies, this work systematically compares multiple boosting-decomposition combinations under identical settings, isolating the decomposition effect and assessing accuracy, efficiency, and robustness. Results show that integrating signal decomposition substantially improves prediction accuracy. The best performance was achieved by XGBoost_EEMD (RMSE = 0.804, MAE = 0.598, R = 0.993, NSE = 0.987), reflecting a ∼64% improvement over the standalone model. The proposed hybrid modelling framework offers methodological innovation and practical potential for real-time, datadriven river temperature prediction in support of climate-resilient water management.

Document