Author(s): Tomoki Izumi, Noriyuki Kobayashi
Linked Author(s): Tomoki Izumi
Keywords: Runoff analysis, neural network, deep learning
Abstract: A deep neural network (DNN) model for runoff analysis was developed. The model consists of one input layer, three middle layers, and one output layer. In order to avoid the gradient vanishing, the layer-wise pre-training by the auto-encoder was employed. For the problem of overfitting, the number of training run is limited based on the early-stopping technique. The model validity was examined through runoff analysis. The model predicted the daily river discharge from the daily rainfall. The study area was at Shigenobu River watershed, Ehime prefecture, Japan. The daily discharge and rainfall data were obtained at the observatory of Deai and Matsuyama, respectively. The training (calibration) period was from 2001 to 2010, and prediction (verification) period was from 2011 to 2013. To assess effectiveness of the model, calibration and verification results were compared with those of a hierarchical neural network (HNN) model, which consisted of the same number of layers without employing the layer-wise pre-training by the auto-encoder. From the results, it was found that the statistical performance indices of DNN model in verification period is better than that of HNN model
Year: 2017