DEVELOPMENT OF FLOOD FORECASTING TECHNIQUE BY COMBINING DNN AND DATA ASSIMILATION TO ENABLE SEQUENTIAL SELECTION OF ASSIMILATION VARIABLES

Abstract

 A flood forecasting system using physical models (such as rainfall runoff model, indefinite flow model) and data assimilation has been introduced in rivers in Japan. On the other hand, the application of DNN which empirically predicts unknown quantity from known quantity without depending on physical model has been observed. By using data assimilation based on physical interpretation and combined use of DNN to ensure accuracy from correlation between various quantities, it can be expected to simultaneously improve explainability and accuracy.

 In this paper, first, the results of prediction calculation by assimilation variables were compared and analyzed, and it was grasped that assimilation variables which affect the prediction accuracy most change even during flood. Based on this, a model to select an optimum assimilation variable by DNN was constructed, and a combination of assimilation variables selected successively was applied to carry out a water level and flow rate prediction simulation by data assimilation. As a result, the prediction accuracy improvement by the combined use of DNN and data assimilation was confirmed.