Evaluation of High-Intensity Precipitation Prediction Using Convolutional Long Short-Term Memory with U-Net Structure Based on Clustering

Abstract

Recently, Asia has experienced significant damage from extreme precipitation events caused by climate change. Improving the accuracy of quantitative precipitation forecasts over wide regions is essential to mitigate the damage caused by precipitation-related natural disasters. This study compared the predictive performances of a global model trained on the entire dataset and a clustered model that clustered precipitation types. The precipitation prediction model was constructed by combining convolutional long short-term memory with a U-Net structure. Research data consisted of precipitation events recorded at 10 min intervals from 2017 to 2021, utilizing radar data covering the entire Korean Peninsula. The model was trained on radar precipitation data from 30 min before the current time (t − 30 min, t − 20 min, t − 10 min, and t − 0 min) to predict the precipitation after 10 min (t + 10 min). The prediction performance was assessed using the root mean squared error and mean absolute error for continuous precipitation data and precision, recall, F1 score, and accuracy for the presence or absence of precipitation. The research findings indicate that, with sufficient training data for each precipitation type, models trained on clustered precipitation types outperform those trained on the entire dataset, particularly for predicting high-intensity precipitation events.