The Use of Recurrent Neural Networks (S-RNN, LSTM, GRU) For Flood Forecasting Based on Data Extracted from Classical Hydraulic Modeling

Abstract

Floods are natural disasters that have a significant impact on everyday human life, both through material losses and loss of life. In the context of climate change, these events may be more frequent or more dangerous. For real-time flood forecasting, fast methods for determining flood hydrographs along watercourses are needed. Classic hydraulic modeling software provides satisfactory results, but in many cases the calculation time can be high. Another approach, different from classical hydraulic modeling is the use of neural networks for forecasting hydrographs. Thus, the present study aims to analyze three different types of recurrent neural networks, including SRNN, RNN-LSTM, RNN-GRU. For each network type, flow hydrographs and level hydrographs resulting from hydraulic modeling were provided as input and training data. Using the deep learning environment, based on previous calibration and validation of recurrent neural networks, flood hydrographs for 2 historical events were modeled. The obtained hydrographs are extremely close to those recorded, while the running time is tens of times smaller.

Keywords

• deep learning
• neuronal networks
• synthetic hydrographs
• hydraulic modeling
• forecasting hydrographs