Downscaling Rainfall in ERA5 Using Deep Learning for Rainfall Prediction in the Chi River Basin

Abstract

In analyzing and assessing the impacts of hydro-climatic disasters, such as floods and droughts, hydrological models require high-resolution data to accurately simulate possible scenarios. However, the insufficient data resolution limits analysis and forecasting. Therefore, this study aims to apply the artificial intelligence technology to improve the resolution of rainfall data at the sub-basin scale. The climate variables from the European Centre for Medium-Range Weather Forecasts (ECMWF) and measured rainfall data from the Thai Meteorological Department's monitoring stations in the Chi River Basin would be downscaled by using the Recurrent Neural Network (RNN) model which is developed to learn and analyze the relationship between climate variables and measured rainfall data. Moreover, from the model training and testing results during 1995–2024, it was found that the developed model can accurately estimate rainfall, especially during the rainy season, which is a period of high weather variability. From model testing results indicated the range of R² was 0.77 to 0.93 and the RMSE was 2.4 to 5.8 mm. This knowledge can be applied to support water management policy formulation and disaster risk reduction in various areas.