Comparison of the performance of a hydrologic model and a deep learning technique for rainfall- runoff analysis

Abstract

Rainfall-runoff analysis is the most important and basic analysis in water resources management and planning. Conventional rainfall-runoff analysis methods generally have used hydrologic models. Rainfall-runoff analysis should consider complex interactions in the water cycle process, including precipitation and evapotranspiration. In this study, rainfall-runoff analysis was performed using a deep learning technique that can capture the relationship between a hydrological model used in the existing methodology and the data itself. The study was conducted in the Yeongsan River basin, which forms a large-scale agricultural area even after industrialization, as the study area. As the hydrology model, SWAT (Soil and Water Assessment Tool) was used, and for the deep learning method, a Long Short-Term Memory (LSTM) network was used among RNNs (Recurrent Neural Networks) mainly used in time series analysis. As a result of the analysis, the correlation coefficient and NSE (Nash-Sutcliffe Efficiency), which are performance indicators of the hydrological model, showed higher performance in the LSTM network. In general, the LSTM network performs better with a longer calibration period. In other words, it is worth considering that a data-based model such as an LSTM network will be more useful than a hydrological model that requires a variety of topographical and meteorological data in a watershed with sufficient historical hydrological data.

Keywords

• Rainfall-runoff analysis
• LSTM
• SWAT
• Deep learning