Development of a short-term water quality prediction model for urban rivers using real-time water quality data

• Published in: Water science & technology. Water supply 2022-04, Vol.22 (4), p.4082-4097
• Main Authors: Lee, J. H., Lee, J. Y., Lee, M. H., Lee, M. Y., Kim, Y. W., Hyung, J. S., Kim, K. B., Cha, Y. K., Koo, J. Y.
• Format: Article
• Language: English
• Subjects: Accuracy; Algorithms; Artificial neural networks; Carbon; Classification; classification model; Climate change; Creeks & streams; Decision trees; Dependent variables; Dissolved oxygen; dissolved oxygen prediction model; Electrical conductivity; Electrical resistivity; Hydrocarbons; Long short-term memory; Machine learning; Neural networks; Nitrogen; Phenols; Phosphorus; Precipitation; Prediction models; Rain; Real time; real-time automatic measurement data; Rivers; Sensitivity; Static models; Stormwater; Time measurement; Time series; urban river; Water monitoring; Water pollution; Water quality; water quality accident; Water quality management; Water temperature; Waterfronts
• ISSN: 1606-9749; 1607-0798
• DOI: 10.2166/ws.2022.038

We developed a classification model and a real-time prediction model for short-term dissolved oxygen (DO) at the junction of the Han River in Anyangcheon, where water quality accidents occur frequently. The classification model is an analysis model that derives the main factors affecting DO changes in the Anyangcheon mobile water quality monitoring network using decision tree, random forest, and XGBoost. The model identified the key factors affecting DO changes to be electrical conductivity, cumulative precipitation, total nitrogen, and water temperature. Random forest (sensitivity, 0.9962; accuracy, 0.9981) and XGBoost (sensitivity, 1.0000; accuracy, 0.9822) showed excellent classification performance. The real-time prediction model for short-term DO that we developed adopted artificial neural network (ANN), long short-term memory (LSTM), and gated recurrent unit (GRU) algorithms. LSTM (R2 = 0.93 − 0.97, first half; R2 = 0.95 − 0.96, second half) and GRU (R2 = 0.94 − 0.98, first half; R2 = 0.96 − 0.98, second half) significantly outperformed ANN (R2 = 0.64 − 0.86). The LSTM and GRU models we developed used real-time automatic measurement data, targeting urban rivers that are sensitive to water quality changes and are waterfront areas for citizens. They can quickly reflect and simulate short-term, real-time changes in water quality compared with existing static models.