Urban inundation prediction using seasonal autoregressive integrated moving average model under extreme rainfall events

Abstract

Urban inundation prediction enables the timely implementation of emergency measures and mitigates casualties and property losses caused by disasters. Physical models, hindered by their complex computational requirements, can no longer satisfy the demands for high-efficiency applications. In recent years, machine learning algorithms—particularly convolutional neural networks and recurrent neural networks have demonstrated substantial advancements in time series prediction tasks. This paper employs the Seasonal Autoregressive Integrated Moving Average (SARIMA) model to forecast inundation scenarios under a 100 year extreme rainstorm event in Jinan City, Shandong Province, China. Four water inundation points within the jurisdiction of the Yuxiu River Basin were selected as the prediction targets, and the inundation depth, unit flow rate, and flow velocity were predicted. The prediction accuracy for both long and short durations was analyzed. Additionally, it was discussed that variations in data volumes, rainfall durations, and rainfall peak coefficients had minimal impact on the prediction results. The results show that prediction accuracy improves with more available data and less data fluctuation within the time period. Under the 100 year return period rainstorm scenario, the global average RMSE of predicted inundation depth, flow and flow velocity at four waterlogging points for long-duration and short-duration rainfall are 0.011 m, 0.037 m3/s and 0.056 m/s, respectively. These efforts are expected to contribute to the advancement of inundation prediction, offering a new perspective on rainfall events preprocessing to achieve more accurate predictions.