Prediction of the water quality of Ogouchi Reservoir, Japan, using machine learning and time-series analysis models based on continuously monitored indicators

The impact of climate change on water resources is a growing concern. As reservoirs are highly susceptible to climate and land-use changes in their catchments, it is necessary to quantify the actual impact of climate change using models and apply them to management. The Soil and Water Assessment Tool (SWAT) is a widely used watershed-scale hydrological model; however, it requires detailed and accurate boundary condition inputs. In this study, we focus on using machine learning and time-series analysis as an alternative to the SWAT model. By comparing long short-term memory (LSTM), seasonal autoregressive integrated moving average (SARIMA), and vector autoregression (VAR) models, we found that the LSTM model could predict total nitrogen with a high degree of accuracy, with R2 and Nash–Sutcliffe efficiency values of 0.92. Turbidity was difficult to predict using all models, suggesting that conventional hydrological models are better suited to turbidity prediction. The total inflow model achieved higher accuracy than the subwatershed models, suggesting that small-scale prediction is potentially challenging. Continuously monitored indicators, such as water temperature, pH, flow rate, precipitation, and sunshine duration, were useful explanatory variables for predicting nutrient fluctuations.