Modeling of Dissolved Oxygen Concentration and Its Hysteresis Behavior in Rivers Using Wavelet Transform-Based Hybrid Models

Concentration of dissolved oxygen (DO) is an indicator to evaluate environmental health in many riverine systems. In this study, two hybrid models, the wavelet-based regression model (WR) and the wavelet-based artificial neural network model (WANN), are proposed for short-interval (1 day ahead) and long-interval (31 days ahead) prediction of DO concentration in Clackamas River near Oregon City, OR, USA. It was found that both in the short-and long-prediction intervals, the hybrid models performed better than the multiple linear regression and artificial neural network models. Also, the performance of the models to predict seasonal DO variations, low (DO< 10 mg/L) and high (DO> 14 mg/L) DO ranges and simulating its hysteresis was evaluated. For the short-interval modeling, results of the WR model were better than those of the other three models. For the long-interval modeling during spring and summer, the WANN model presented better predictions of DO. And during autumn and winter, the WR model presented higher accuracy than the others. In addition, these hybrid models were able to simulate DO daily hysteresis behavior. The most accurate WR model predicted DO with the minimum root mean square error (RMSE) of 0.097 mg/ L and the maximum Nash-Sutcliffe coefficient of efficiency (E) of 0.996 for the short interval. For the long interval, the best performing model was the WANN model with RMSE and E-statistics of 0.744 mg/L and 0.766, respectively.