River Flood Prediction Based on Physics-Informed Long Short-Term Memory Model

Flooding is one of the major natural disasters. Predicting river water levels and flooding is an effective way of enabling proactive flooding response measures. Although machine learning-based prediction models in prior studies have obtained a low error rate, they do not perform well during the rapid and significant water level rising (i.e., flooding). To provide a better flooding prediction tool, the present study first evaluates the commonly used long short-term memory model and points out the limitation of prior studies. Then, a novel model named physics-informed (PI) LSTM is proposed. The PI-LSTM integrates hydrological knowledge into the neural network as well as extends the current physics-informed river water level prediction neural networks to a recurrent one. Compared with LSTM, PI-LSTM has a better performance in predicting rapid and significant water level rising. The study is expected to increase the accuracy of flooding prediction and provide better decision-making support to agencies responsible for flood forecasting and warning.