Real-Time Water-Level Forecasting Using Dilated Causal Convolutional Neural Networks

Accurate forecasts of hourly water levels during typhoons are crucial to disaster emergency response. To mitigate flood damage, the development of a water-level forecasting model has played an essential role. We propose a model based on a dilated causal convolutional neural network (DCCNN) that can yield water-level forecasts with lead times of 1- to 6-h. A DCCNN model can efficiently exploit a broad-range history. Residual and skip connections are also applied throughout the network to enable training of deeper networks and to accelerate convergence. To demonstrate the superiority of the proposed forecasting technique, we applied it to a dataset of 16 typhoon events that occurred during the years 2012-2017 in the Yilan River basin in Taiwan. In order to examine the efficiency of the improved methodology, we also compared the proposed model with two existing models that were based on the multilayer perceptron (MLP) and the support vector machine (SVM). The results indicate that a DCCNN-based model is superior to both the SVM and MLP models, especially for modeling peak water levels. Much of the performance improvement of the proposed model is due to its ability to provide water-level forecasts with a long lead time. The proposed model is expected to be particularly useful in support of disaster warning systems.