Precipitation forecasting through numerical model correction considering the korean peninsula terrain

Abstract

In this paper, a study was conducted to improve precipitation accuracy by correcting the results of the numerical model used for precipitation prediction using a deep learning model. Precipitation predictions are typically classified into short-term prediction (nowcasting) and long-term prediction. Recently, there are several studies that perform nowcasting using deep learning model. However, the precipitation prediction model using deep learning has a weakness in long-term prediction of more than a day, so the existing numerical model continues to be used. This study aims to improve the accuracy of precipitation prediction by using deep learning for long-term prediction. For this, the U-Net model, which is a deep learning model previously used for nowcasting, was used, and based on the fact that the location information of the prediction data is fixed, a layer that can learn information by terrain was added to the model. Through this, we will show that the performance is improved by correcting the precipitation prediction result of the existing numerical model.