The real-time river stage prediction model was developed using the artificial neural network model, with deep learning as the training method. The main component of the model was the four-layer feed-forward network. As a network training method, the stochastic gradient descent method based on the back propagation technique was applied. The denoising autoencoder was applied as a pre-training method. The developed model was applied to one catchment of the Ooyodo River, one of the first-grade rivers in Japan. The hourly change in river stage and hourly rainfall were used as input to the model, while output data was the river stage of Hiwatashi. To clarify the suitable configuration of the model, a case study was done. The prediction result was compared with those of other prediction models. Consequently, the developed model showed the best performance.

正会員 工修 日本工営株式会社 中央研究所(〒300-1259 茨城県つくば市稲荷原2304)2 正会員 博(工) 日本工営株式会社 中央研究所(〒300-1259 茨城県つくば市稲荷原2304)The real-time river stage prediction model is developed, using the artificial neural network model which is trained by the deep learning method. The model is composed of 4 layer feed-forward network. As a network training method, stochastic gradient descent method based on the back propagation method was applied. As a pre-training method, the denoising autoencoder was applied. The developed model is applied to the one catchment of the OOYODO River, one of the first-grade river in Japan. Input of the model is hourly change of water level and hourly rainfall, output data is water level of HIWATASHI. To clarify the suitable configuration of the model, case study was done. The prediction result is compared with the other prediction models, consequently the developed model showed the best performance.

We developed a real-time river stage prediction model using a hybrid deep neural network and physically based distributed rainfall-runoff model. The main component of the hybrid model was a four-layer feed-forward artificial neural network. Using the predicted flow of the rainfall-runoff model as the input data of the neural network, we integrated the two models into the hybrid model. The input data of the hybrid model included upstream water level, hourly change in water level, and estimated hourly change in catchment storage. The output was the change in water level at the prediction point. In the training phase, input data and supervised data were formed using the observed data. In the prediction phase, input data were formed using a combination of the observed data and flowrate calculated using the distributed mode.The result of the hybrid model outperformed those of the ANN and distributed models. Especially in the largest flood event, the performance of the hybrid model was significantly stronger.

No abstract

Although artificial neural networks (ANN) is widely used for real-time flood prediction model, it is pointed out that the weak point of the model is poor applicability for the inexperienced magnitude of flood. In this study, the ANN models were applied to first-grade rivers in Japan, Tokoro River catchment and Abashiri River catchment. The training data of the ANN models were all the rainfall-runoff event which exceeded the Flood Watch Water Level during the period of 1998-2015. Types of observation data were river-stage and rainfall at 1-hour pitch. The validation data was the largest flood since the river-stage observation had started. The main component of the model was the four-layer feed-forward network. As a network training method, the deep learning based on the denoising autoencoder was applied. The output of the neural network was change in river-stage in T hours at the prediction point. The input data was the upstream river-stage, hourly change in river-stage and hourly rainfall. The riverstage prediction up to 6 hours showed very good accuracy, and It was proved that it can be nicely predicted even for the past largest flood.Engineering