IMPROVEMENT OF WATER LEVEL PREDICTION USING NEURAL NETWORK BY SPATIAL SEGMENTATION OF RAINFALL

Abstract

 In river construction site, it is required to assess whether there will be flooding 24 hours ahead in order to safely evacuate construction workers, heavy machinery, and construction materials.We have already developed the method for predicting water levels 24 hours ahead using Neural Network based on the spatial distribution of rainfall as input.In this method, the water level is predicted from the mean and variance of the rainfall, the center of gravity of rainfall, and the distance from the prediction point to the center of gravity, which are the parametes representing the spatial distribution of the rainfall. However, this method cannot take rainfall distribution into account in detail, resulting the predicted water level being lower than the actual water level, which is unsafe. In this study, a method was considered in which the rainfall area was divided into several parts to take into account the rainfall distribution, and was applied to water level predictions for four watersheds. As a result, it is found that underpredictions were improved and that the improvement was more effective when the rainfall distribution was uneven, i.e., when the variance values of the divided rainfall were large.