Real-time water quality management in Lake Taihu, China: a machine learning application to high-frequency time series

ABSTRACT Graph illustrating the framework to collect data from six stations around Lake Taihu and apply machine learning algorithms to support the real-time management of water quality. Decades of water pollution and blue-green algae blooms had detrimental effects on the aquatic environment of Lake Taihu, necessitating real-time management. Methods with limited interpretability can be inadequate for the comprehensive study of high-frequency multivariate time series. To address this research gap, we leveraged a data-driven machine learning framework to investigate the data streams of six monitoring stations surrounding Lake Taihu. For water quality classification using eutrophication indicators, random forest was most accurate with a test root mean square error (RMSE) of 0.4041, outperforming gradient boosting machines (RMSE is 0.4443) and neural network (RMSE is 0.4887). For predicting the chlorophyll a concentration, total phosphorus and potassium permanganate were identified as the dominant drivers, and the test RMSE of random forest was 0.5061, the smallest of the three methods too. The convolutional neural network plus long-short-term memory models were well-suited for high-frequency time series forecasting, with the coefficient of determination ranging from 0.4856 to 0.8699, despite outliers and anomalous spikes. We then drew insights into the interaction mechanism of water quality and eutrophication dynamics. We recommended expanding open-access high-frequency monitoring coverage (especially in the lake's northwest region) to enhance spatial-temporal modeling, real-time risk identification, and algae bloom control.