Urban waterlogging is an important urban disaster, and culvert ponding is one of the main manifestations of urban waterlogging, which can easily cause serious casualties and property losses. In order to improve the efficiency and accuracy of predicting the culvert ponding depth, a predicting model naming PSO-SVR is proposed, which the current meteorological data and waterlogging depth data are selected as input and output vectors respectively. Firstly, a support vector regression prediction model (SVR) for the culvert ponding depth is established, and then the particle swarm optimization algorithm (PSO) is used to optimize key parameters, which has the advantages of objectivity and simplicity. The experimental results indicate that the PSO-SVR model is effective in predicting the depth of culvert ponding. Compared with traditional prediction models of SVR and BP neural network, the PSOSVR model has higher fitting accuracy for waterlogging depth. This model can provide technical support for urban waterlogging warning and emergency rescue decision-making.