Development of the Real-Time River Stage Prediction Method Using Deep Learning

Abstract

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.

Convolutional Neural Network Coupled with a Transfer-Learning Approach for Time-Series Flood Predictions

“…The time-series water-level predictions of A1-A4 using the CNN in Domain A showed acceptable agreement with the observed data, although the higher peaks of each of A1-A3 were poorly captured. Comparisons with other neural networks, such as the fully connected deep neural network (FCDNN) [7] and recurrent neural network (RNN) [31] indicated that our CNN prediction was poorer than the performance of these models. For example, Hitokoto and Sakuraba [7] reported that their FCDNN provided better predictions of water levels for the top-four largest flood events from 1990 to 2014 in the same watershed.…”

“…Comparisons with other neural networks, such as the fully connected deep neural network (FCDNN) [7] and recurrent neural network (RNN) [31] indicated that our CNN prediction was poorer than the performance of these models. For example, Hitokoto and Sakuraba [7] reported that their FCDNN provided better predictions of water levels for the top-four largest flood events from 1990 to 2014 in the same watershed. The RMSE in an hour prediction of water level was 0.12 m, which is approximately 1/3 of the mean RMSE of our CNN prediction, due to the use of two different verification methods.…”

“…Numerous studies of ANN models have been performed for flood predictions [6]. ANN models that adopt a deep learning approach have been recently employed in practical flood predictions [7]. As another type of ANN model, the long, short-term memory (LSTM) architecture [8,9] was utilized for water-level predictions during flood events [10][11][12].…”

Practical Data‐Driven Flood Forecasting Based on Dynamical Systems Theory

“…We compared the proposed forecast approach with three existing methods: The multilayer neural network, the distributed runoff model with particle filter, and the conventional local linear method. Note that the results of the multilayer neural network and the distributed runoff model are scanned values in Hitokoto et al (2017).…”

“…The first relates to performance when considering previously unexperienced flood magnitudes. As neural network models assume black box functions and perform interpolations using past data with these functions, they may show a poor performance when using data that exceed the range of the training data (Hitokoto et al., 2017), and it is difficult to analyze the real dynamics of these data. The other issue concerns the amount of data required.…”