Prediction of Cyanobacteria Using Decision Tree Algorithm and Sensor Monitoring Data

Abstract

A multifunctional weir was built on the Nakdong River. As a result, changes in the river environment occurred, such as an increase in river residence time. This causes changes in water quality, including green algae. The occurrence of green algae in the Nakdong River, which is used as a water source, also affects the purified water supply system. In particular, the mass spread of harmful algae is becoming a major problem as the frequency and intensity of occurrences increase. There are various causes of blue-green algae. We would like to examine the relationships between causal factors through a decision tree-based algorithm. Additionally, we would like to predict the occurrence of green algae based on the combination of these factors. For prediction, we studied categorical prediction based on the blue-green algae warning system used in Korea. RF, Catboost and XGBoost algorithms were used. Optimal hyperparameters were applied. We compared the prediction performance of each algorithm. In addition, the predictability of using sensor-based data was reviewed for a preemptive response to the occurrence of blue-green algae. By applying sensor-based data, the accuracy was over 80%. Prediction accuracy by category was also over 75%. It is believed that real-time prediction is possible through sensor-based factors. The optimal forecast period was analyzed to determine whether a preemptive response was possible and the possibility of improvement was examined through the segmentation of prediction categories. When there were three categories, 79% of predictions were possible by the 21st day. In seven categories, 75% prediction was possible up to 14 days. In this study, sensor-based categorical predictability was derived. In addition, real-time response and proactive response were determined. Such sensor-based algae prediction research is considered important for future blue-green algae management and river management.