Neural network modeling for rainfall prediction: observation extrapolation and numerical forecast correction

Abstract

Torrential rain and inundation are becoming more common because of global warming and extreme weather, leading to an increase in human and property damage. It is critical to improve the performance of the rainfall prediction model in order to minimize damage by identifying as early as possible when a natural disaster may occur. In this study, we propose a model that performs very short-term rainfall prediction based on past observation data and a model that performs short- and medium-term rainfall correction based on future forecast data, respectively, to improve the performance of rainfall prediction. In Study 1, we focused on the performance improvement of the nowcasting model, which predicts very short-term rainfall using radar extrapolation. We improved radar extrapolation performance by extending the architecture of the existing encoding-forecasting model to emphasize the phenomenon of the most recent data. However, the nowcasting method based on spatiotemporal analysis of radar data is limited in that the prediction frame blurs as the time step increases, even though it accurately predicts near future rainfall. Furthermore, there is a limitation in that it does not properly represent the sudden change in rainfall patterns caused by rapid changes in weather conditions. Therefore, Study 1 is a method suitable for very short-term rainfall prediction, and a new method needs to be developed for short- and medium-term rainfall prediction. In Study 2, we proposed a method to improve the accuracy of forecast data by correcting future simulation data produced by a numerical weather prediction (NWP) model close to the ground truth data observed, as a method for predicting short- and medium-term rainfall. The NWP model is the pinnacle of human knowledge about the Earth's atmospheric circulation, currently used for weather forecasting in the practical field, however, room for improvement remains because there are many differences from the actual ground truth. Therefore, Study 2 proposed a deep learning-based correction method that transforms the distribution of the forecast data of the NWP model into the distribution of the ground truth data measured by ground observation equipment such as radar by applying the GAN model, which exhibits excellent performance in data distribution transformation in different domains. Using this method, the quality of the forecast data is improved by correcting the error of the simulation result of the NWP model close to the ground truth. In Future Study, we will develop an integrated rainfall prediction model that combines the extrapolation result of radar data from Study 1 and the correction result of forecast data from Study 2 to produce the final rainfall prediction result. By combining the two results, the final model can respond to rapid changes in future weather conditions while mitigating the spatial smoothing problem. In addition, through the future study, we can improve the rainfall prediction model's performance by combining the accuracy of the prediction data obtained in Study 1 and the stability of the forecast data obtained in Study 2. Overall, this study dealt with rainfall prediction, which is a field of weather forecasting, but it also presents practical guidelines for the design of integrated deep learning models for various weather forecasting fields that fuse past observation data and future forecast data using spatiotemporal prediction modeling.