Author(s): Go Ohno; Kazunori Ito
Linked Author(s):
Keywords: River construction; Safety management; Flood prediction; Neural network; Rain distribution maps
Abstract: When flooding is expected at a river construction site, workers are able to evacuate to a place of safety within a few hours. If heavy machinery and materials are present, several hours are required to ensure safety. One responsibility of the construction manager is to make a judgement as to whether evacuation is necessary. Recently, a flood forecast system has been developed and applied to river work; based on forecast water levels, the man-ager can make a judgement based on personal experience. However, this system has some problems: 1) tuning of system parameters and the collection and selection of data takes a lot of time; 2) to secure the robustness, water levels are predicted by multiple methods and the manager’s judgement may be affected if there are discrepancies among the predictions. In this paper, a new forecasting technique for judging whether water levels will exceed a ‘flood’ threshold or not is developed using deep learning based on weather forecast rain distribution maps. The input data are the center of gravity of rainfall and the amount of rainfall. The method is applied to the Abukuma River in Fukushima prefecture and is able to judge flooding in excess of the threshold value, giving a correct evaluation rate of 60%. The precision of the judgement can be improved by selecting the learning data. This technique is suitable for application to safety management during river construction work.
Year: 2020