Urban flood depth prediction using an improved LSTM model incorporating precipitation forecasting

Floods are among the most devastating natural disasters, causing extensive loss of life and property. Accurate flood inundation prediction is essential for reducing the impacts of flood disasters. Although many studies have applied various statistical and machine learning methods to predict future flood depths, the precision and timeline of these predictions are still insufficient for effective disaster prevention and emergency response. This paper introduces an improved LSTM model that incorporates precipitation forecasts to increase the accuracy of flood depth prediction and extend the prediction timeline. To capture time series dependencies and generate future precipitation data, a precipitation forecast model is developed and integrated into the LSTM-based flood depth prediction framework. The single-step recursive method is used to predict future flood depths. The model is validated using data from Shenzhen’s precipitation observations and flood monitoring stations. The results demonstrate that, while ensuring a prediction accuracy with an R² greater than 0.75, the improved LSTM model successfully extends the prediction timeline to 8 time steps (40 min), with an R² increase of 6.5% and a reduction in the RMSE of 13.8% in such an interval, thereby allowing for a longer prediction span without compromising accuracy. The study also revealed that improving the accuracy of precipitation forecasts, particularly through the use of ANN models, significantly enhances the performance of the flood depth prediction model. Specifically, a 20% increase in the precipitation forecast accuracy results in a 3.1% improvement in the flood depth prediction accuracy. These findings demonstrate that more accurate precipitation forecasts play a crucial role in enhancing the model’s ability to predict flood depths.