CONSTRUCTION OF A DAM INFLOW FORECASTING SYSTEM BY USING DEEP LEARNING AND HAVING APPLICATION FOR UNUSUAL AND INEXPERIENCED FLOOD

Abstract

 At a multipurpose dam, it is necessary to forecast inflow to control increasingly severe and frequent floods. For more effective dam operation, we developed Maruyama Dam inflow forecasting system using deep learning. To improve forecast accuracy and to decide the structure of the forecasting system, we identified input data that are highly correlated with dam inflows and optimized hyperparameters with a large impact on forecast accuracy.

 The deep learning model has low forecast accuracy for unusual and inexperienced floods because of little training data. Therefore, the system has two applications to improve this problem. One is the forecasting system using the storage function model during the severe floods over the limit of the deep learning model forecasting. The other is to train the unusual and inexperienced floods to improve the range of forecast by the deep learning model.