APPLICATION OF DEEP LEARNING TO DAM INFLOW FORECASTS WITH DIFFERENT CHARACTERISTICS IN THE CHUBU AREA AND THE EFFECT OF FORECAST ACCURACY CAUSED BY MIX OF INPUT RAINFALL TYPES

Abstract

 We classified dams in the Chubu region into three patterns with similar characteristics in the upstream areas, based on the influence of upstream dam discharge operations. By constructing deep learning models for representative dam in each pattern and evaluating their prediction accuracy, we assessed the validity of the model construction method based on the input conditions and its applicability to predicting dam inflow volume. To address the challenge of limited data for low-frequency and unprecedented floods, we supplemented the training data with large-scale flood events during periods with-out radar rainfall data, utilizing ground-based rainfall data. Additionally, we analyzed the impact of using different types of rainfall data in training, verifying the effectiveness of our approach.