[3Yin2] Interactive session 1

[3Yin2-16] Convolutional Dam Inflow Forecasting

〇Riku Ogata¹, Masahiro Okano¹, Takashi Izumiya¹, Shigetomo Yamamoto¹, Takato Yasuno¹ (1.Yachiyo Engineering Co., Ltd.)

Keywords:Flood mitigation, Dam inflow forecast, 1D-Convolution

Since 2006, rainfall forecasting such as short-range forecasts of precipitation has been provided to prepare for the occurrence of torrential rains and extreme floods. In order to prepare for uncertain floods, it is important for dam managers to improve the accuracy of dam inflow forecasts. In this study, a one-dimensional convolutional network prediction model is proposed for predicting dam inflows up to six hours in advance, using various dam quantities and short-range forecasts of precipitation. We also apply several methods to improve the accuracy focusing on the loss and activation functions. As a result of comparing the accuracy with the baseline MLP, RNN and LSTM models, it is confirmed that the prediction accuracy of the 1D convolutional network is as good as or better than the baseline. Finally, we mention the usefulness of our method in terms of accuracy improvement and issues for future generalization.