Extreme learning machine-based prediction of daily water temperature for rivers

Abstract

Water temperature impacts many processes in rivers, and it is determined by various environmental factors. This study proposed an extreme learning machine (ELM)-based model to predict daily water temperature for rivers. Air temperature (Ta), discharge (Q) and the day of the year (DOY) were used as predictors. Three rivers characterized by different hydrological conditions were investigated to test the modeling performances and the model results were compared with multilayer perceptron neural network (MLPNN) and simple multiple linear regression (MLR) models. Results showed that inclusion of three inputs as predictors (Ta, Q and the DOY) yielded the best modeling accuracy for all the developed models. It was also found that Q played a minor role and Ta and DOY are the most important explanatory variables for river water temperature predictions. Additionally, sigmoidal and radial basis activation functions within the ELM model performed the best for river water temperature forecasting. ELM and MLPNN models outperformed MLR model, and ELM model with sigmoidal and radial basis activation functions performed comparably to MLPNN model. Overall, results indicated that the ELM model developed in this study can be effectively used for river water temperature predictions.