| Abstract |
| This study was conducted to compare the predictive accuracy of three popular machine learning algorithms, i.e., Bayesian Regularized Neural Networks (BRNN), Neural Network (NN), and Support Vector Machines with Radial Basis Function Kernel (SVM), for modeling river water quality and quantity. Input data with 11 parameters collected from three monitoring sites in the lower part of the Nakdong River between January 2008 to December 2018 were input into the three algorithms. The data were divided into two subsets, which included a training data set and test data set in a ratio of 70:30. The results showed that NN displayed better performance in prediction accuracy than BRNN and SVM when optimizing one or two tuning parameters. The prediction accuracy was, on average, higher for chlorophyll a than for discharge, regardless of which machine learning algorithm was used. Identical results were also observed in the training data set. The two variables biochemical oxygen demand (BOD) and chemical oxygen demand (COD) played an important role in predicting chlorophyll a, whereas suspended solids (SS) was designated as the largest contributor for the prediction of discharge in all three algorithms. In terms of error metrics and performance indices such as root mean squared error (RMSE), mean absolute error (MAE), and R2, the three machine learning algorithms exhibited improved performance in the test data set, as compared to the training data set. We believe that this preliminary study helps identify suitable methods for enhanced prediction of emerging water quality or quantity parameters such as harmful algal blooms. |
|
|
| Key Words |
| Machine learning, Predictive performance, Variable importance, Water quality, Streamflow, 기계학습, 예측 성능, 변수 중요도, 수질, 하천유량 |
|
|
|
|
|
PDF