Prediction of chlorophyll-a concentration using machine learning models trained by the higher resolution data combining the monitoring data and SWAT simulation data

Abstract

Harmful algal blooms (HABs), which result from the rapid growth of algae, destroy the aquatic ecosystem and increase the water treatment cost for managing drinking water quality. The number of HAB occurrences has increased due to abnormal temperatures caused by climate change, an increase in human activities such as industries, and stagnant water owing to artificial structure installations. Therefore, it is important to predict chlorophyll-a (chl-a) which represents the distribution of algae to reduce the damage caused by algal blooms in advance. Machine learning models, which have remarkable prediction performance and are widely used to manage algal blooms, are considered suitable tools in terms of time and cost. Therefore, this study aims to predict chl-a concentration using machine learning models and optimize the time lag between input variables and chl-a that affects prediction efficiency. Acquiring high-resolution data is important for using machine learning models, but it is difficult to directly use the monitoring data provided by Korean public water quality monitoring stations. In particular, as the data near the target site (Gwangju Metropolitan City) is low-resolution, countermeasures for the low-resolution problem of machine learning models are required. To obtain the high-resolution data, the monitoring data and simulation data of the soil and water assessment tool (SWAT) were combined. Artificial neural network (ANN), Support Vector Regression (SVR), and Extreme Gradient Boosting (XGBoost) were constructed to predict chl-a and the models were performed from 2015 to 2021. The results of machine learning models that used combined data indicated that the XGBoost model has the highest prediction accuracy among the three models. In the time lag examination from 1 to 8 days, the model using the data monitored 3 days ago showed the highest prediction performance. To efficiently manage the algal blooms, this study proposed new methods of higher resolution data production, presented a useful model to predict chl-a, and recommended considering the optimal time lag between input data and chl-a when making chl-a prediction models.