An ensemble of AI time-series models for reservoir storage rate in South Korea: Accuracy improvement using expert knowledge-based rainfall-runoff modeling

In regions such as South Korea, where rainfall is irregular over time, it is important to operate dams and reservoirs to regulate the supply and demand of water, ensure proper irrigation water discharge to downstream agricultural areas, and prevent flood damage. However, during periods of heavy rainfall, it becomes difficult to operate and manage reservoirs normally, and time-series AI models can be used to produce reservoir storage rate forecasts to help smoothly operate reservoirs and manage water balance. In this study, multivariate time series forecasting was performed for the top 46 reservoirs with effective storage in Korea, using historical storage, precipitation, Julian day, and estimated inflow and outflow calculated using an expert knowledge-based rainfall-runoff model as input variables to predict storage from 1 day up to 20 days-ahead, and quantitative evaluation of reservoir storage rate predictions was performed using MAE (mean absolute error), MBE (mean bias error), RMSE (root mean square error), and CC (correlation coefficient) metrics. Predicting reservoir storage rate by considering only seasonal and meteorological effects reduces accuracy during the rainy season. However, when the estimated inflow and outflow of reservoirs, derived using an expert knowledge-based rainfall-runoff model, were additionally incorporated as inputs into the time series model, the average MAE of the 46 reservoir storage forecasts improved from 0.088%p to 0.240%p, particularly enhancing the low forecast performance in the rainy season. Furthermore, an ensemble of time series models with recurrent neural network structure, which has strengths in short-term forecasting and transformer structure, which has strengths in medium-term forecasting produced better reservoir storage rate predictions than the single model in both short-term and medium-term. The MAE and RMSE averages of the ensemble model's 1-day-ahead reservoir storage rate predictions for 46 reservoirs were 1.384% and 2.496% respectively, an improvement of 0.573% and 0.644% over the single model, and the statistical superiority of the ensemble model increased as the number of days in the future was predicted. The importance of the input variables of the ensemble model was evaluated, and the historical reservoir storage rate was the most important with 67.24%, followed by the estimated inflow and estimated outflow with 12.36% and 8.73%, and the sum of the importance of the two variables calculated through rainfall-runoff modeling was 21.09%. By using variables that reflect the management practices of the reservoir, the model provided information that the model could not learn from the reservoir storage rate, meteorological data, and seasonal data alone. This study enables stable reservoir operation throughout the year, even in areas with irregular rainfall, and is expected to improve agricultural stability in downstream irrigated areas and prevent rainfall flooding.

This work is supported by the Korea Agency for Infrastructure Technology Advancement (KAIA) grant funded by the Ministry of Land, Infrastructure and Transport (Grant RS-2022-00155763).