Hybrid multilayer perceptron and convolutional neural network model to predict extreme regional precipitation dominated by the large-scale atmospheric circulation

Recent advances in deep learning have provided tools that enable meteorologists to predict extreme precipitation using massive atmospheric data. However, individual models are constrained by imbalanced samples and prone to false alarms owing to the rarity of extreme precipitation events. In this study, a novel ensemble learning model, that is, hybrid multilayer perceptron and convolutional neural network (MLP-CNN) model is proposed for the binary prediction of extreme precipitation in Central-Eastern China (CEC), with a daily time horizon. The MLP-CNN model achieves an overall accuracy of 86% in predicting extreme and non-extreme precipitation days using the anomalous fields of two large-scale atmospheric predictors, i.e., geopotential height at 500 hPa and vertically integrated water vapor transport. Subsequently, we employ the MLP-CNN to predict extreme precipitation with a 1–15 day leadtime. The performance of MLP-CNN tends to decrease with increasing leading time of circulation anomalies. However, 1–2 days of advance forecasting can be considered a reference for predicting the occurrence probabilities of extreme precipitation. Finally, based on various evaluation metrics, MLP-CNN outperforms the independent predictions from MLP, CNN, and two other machine learning models (i.e., random forest and support vector machine). Overall, in scenarios where samples are limited, the utilization of hybrid models presents an opportunity for optimizing predictions for extreme precipitation.