River water level forecast based on deep learning and DIEX-Flood

Abstract:

In Japan, river flood occurs frequently and causes severe damage. The intensity and frequency of heavy rainfalls that cause river flood are expected to increase in the future due to climate change, and at the same time, flood hazards are expected to be more severe. In view of the present situation, it is urgent to take countermeasures against flood hazard. In order to forecast when and where rivers will flood, we are working on river flood prediction using DIEX-Flood, which is a numerical model based on 1-D unsteady flow calculation. This method is capable of calculating the longitudinal distribution of water level with assimilating water level data at multiple points. One of the issues with this method was the setting of boundary conditions for future prediction, so we had to introduce a new technique that does not involve the uncertainty of the boundary conditions. For solving the above issues, we developed a new flood-forecast model based on the DIEX-Flood and a deep learning that can predict future water level at multiple points by using only water level data of the same points. This deep learning future prediction model can represent a complex spatial and temporal structure by connecting multiple LSTM (long short-term memory) layers. The parameters in the neural network were determined by learning past water level data. By inputting real-time hourly water level change of multiple point into this model, we can predict them up to 6 hours in the future. Furthermore, we forecast the longitudinal profile of river water level by combining this deep learning model and DIEX-Flood. We applied this combination forecast model to the Kinu River in Japan. As a result of water level prediction at several observed points for the largest floods in September 2015, it was shown that the prediction accuracy of the combined method is better than that of using only DIEX-Flood. In addition, the prediction of water level longitudinal distribution for 6 hours ahead agrees with the trace water level observed during the flood. By using this method, we were able to predict water level spatially and temporally and detect inundation points in the longitudinal direction of a river. This method uses only past and real time water level observed data, so it is assumed that the same study can be conducted for other rivers.