A Hybrid Hydrological Modelling Approach Combining Physically-Based and Deep-Learning-Based Models to Consider Dam Operations

Physically-based hydrologic models can accurately simulate streamflow in natural environment, but they cannot precisely consider the anthropogenic disturbance caused by the operation of large-scale dams. We tried to overcome this issue by developing a hybrid modeling framework, consisting of physically-based models for simulating upstream natural watersheds and deep-learning-based models for simulating dam operation. The model was developed for the Paldang Dam watershed, a major water source for Seoul metropolitan area, where the importance of stable water supply has increased due to the increase of population and water use per capita. The prediction performance of the hybrid model was compared with that of models built based only on the physically-based hydrologic model, namely the Variable Infiltration Capacity model, with single and cascaded structure. For the validation period, Nash-Sutcliffe Efficiency from the developed hybrid model, the single model, and the cascaded model were 0.6410, -0.1054, and 0.2564, respectively, suggesting that the consideration of dam operation aided by the machine learning algorithm is essential for accurate assessment of streamflow.

This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (No. NRF-2022R1A4A3032838).