[1M3-GS-13] AI application: Social data and prediction

[1M3-GS-13-05] Study on the relationship between Radar AMeDAS analysis rainfall and dam inflow discharge using LSTM

〇Masazumi Amakata¹, Junichiro Fujii¹, Ryuto Yoshida¹, Takato Yasuno¹, Junichi Okubo¹ (1. Research Institute for Infrastructure Paradigm Shift)

Keywords:LSTM, time series data, dam inflow discharge, Radar AMeDAS analysis rainfall, multicolliniality

To reduce heavy rain disasters which happen frequently in recent years, to use existing dams effectively becomes a topic. One of the effective uses of existing dams is to optimize quantitative PDCA cycles for dam management without depending on the dam manager's experiences. Therefore we try to improve the prediction accuracy of dam inflows using radar rainfalls. Radar rainfalls have real-time rainfalls and predicted rainfalls. It is difficult to use predicted rainfalls for us because predicted rainfall accuracy is improved by the Japan Meteorological Agency every year. Therefore we study the possibility of dam inflow accuracy improvement using Radar AMeDAS analysis rainfall which is a representative of realtime radar rainfalls. The way which predicts dam inflows is that we extract the relationship between radar rainfalls and dam inflows by LSTM and predict several hours ahead dam inflows from Radar AMeDAS analysis rainfall using time-lag between rainfalls and dam inflows. In this article, we study the effect which input conditions to LSTM, ranges, data representation and data summation of radar rainfalls give dam inflow prediction accuracy.