Comparing Artificial Neural Networks and Regression-based Methods for Modeling Daily Dissolved Oxygen Concentration: A Study Based on Long-term Monitored Data

Abstract

In this study, the ability of regression-based methods, namely conventional regression analysis (CRA) and multivariate adaptive regression splines (MARS), and artificial neural networks (ANNs) method was investigated to model the river dissolved oxygen (DO) concentration. Daily average data for discharge and water-quality (WQ) indicators, which include DO concentration, temperature, specific conductance, and pH, were provided for the monitoring stations USGS 14210000 (upstream) and USGS 14211010 (downstream) in the Clackamas River, Oregon, USA. Eight models were established using different combinations of the input parameters and tested to determine the contribution of each parameter used in the modeling to the performance of the models. The results of the models and methods were compared with each other using several performance statistics. Although the performances of the methods were quite close to each other, the highest estimation performance was obtained from the ANNs method in the testing data sets. According to the performance statistics, Model 8, in which all WQ indicators were included as input parameters, was selected as the optimal model to estimate DO concentration of different periods of the same stations. However, when estimating the DO concentration from one station to another, the highest performance statistics were obtained from Model 8 for upstream and Model 1 for downstream station using the CRA method. For the ANNs method, Model 1 having the single input for both stations was the best model.