Objectives With the increasing frequency of extreme rainfall events and rapid urbanization, urban waterlogging has become a growing concern. Accurate prediction of water accumulation depth is crucial for disaster prevention and mitigation. Traditional statistical models struggle to capture complex spatiotemporal features, while conventional machine learning methods lack effective modeling capabilities for spatiotemporal correlations.

Methods To address these limitations, this paper proposes a ChebNet-LSTM (long short-term memory network) model that effectively integrates spatial and temporal features for improved prediction accuracy. ChebNet leverages the relational matrix between monitoring stations to extract spatial correlations, while LSTM captures temporal patterns in the data. By combining these two approaches, the proposed model fully exploits the spatiotemporal characteristics of water accumulation depth data, overcoming the shortcomings of traditional models that fail to account for both spatial and temporal information.The study utilizes hourly water depth observations from 15 monitoring stations, hourly precipitation data from six meteorological stations, and distance data between the monitoring stations in Kaifeng City, Henan Province, from 2020 to 2021.The ChebNet-LSTM model is constructed using these datasets and subsequently applied to predict water accumulation depths at monitoring stations in the test set.

Results The experimental results demonstrate that the ChebNet-LSTM model achieves outstanding predictive performance， with the water depth predictions showing mean absolute error of 0.62， mean absolute percentage error of 6.23%， and root mean square error of 1.06. These evaluation metrics conclusively prove that the proposed model maintains high accuracy and reliability in practical water accumulation depth prediction tasks.

Conclusions Comparative experiments with models such as graph attention network， LSTM， and spatiotemporal graph convolutional network show that ChebNet-LSTM model outperforms these alternatives， effectively predicting waterlogging trends. The findings provide valuable support for urban drainage planning and flood prevention decision-making， and offer a robust tool for managing urban waterlogging risks.