An Effective Rainfall–Ponding Multi-Step Prediction Model Based on LSTM for Urban Waterlogging Points

With the change in global climate and environment, prevalence of extreme rainstorms flood disasters has increased, causing serious economic property losses. Therefore, accurate rapid prediction waterlogging become an urgent problem to be solved. In this study, Jianye District Nanjing City China is taken as study area. The time series data recorded by rainfall stations ponding monitoring from January 2015 August 2018 are used build a model based on long short-term memory (LSTM) neural network. MSE (mean square error), MAE absolute error) MSLE squared logarithmic were loss functions conduct train LSTM model, then three models built, namely (mse), (mae) (msle), multi-step was predict depth next 1 h. Using measured evaluate results, we selected rmse (root mean mae, mape percentage NSE (Nash–Sutcliffe efficiency coefficient) evaluation indicators. results showed that (msle) best among models, with indicators follows: 5.34, mae 3.45, 53.93% 0.35. At same time, found better effect than (mse) when exceeds 30 mm.