9:45 AM - 10:00 AM
[AHW20-04] Short-term water level forecasting model for low-lying agricultural areas using DNN trained on observed and artificially generated data
Keywords:Deep learning, Drainage pump operation, Preventing inundation damage, Preliminary drainage
There has been a serious issue under the progressing climate change that the operators of agricultural drainage facilities have faced urgent requirements of appropriate operation under extreme heavy rain events, which they have never experienced. In such a situation, preliminary drainage has been recommended. This process involves draining water from an area in anticipation of heavy rainfall or other events that could lead to flooding to ensure that the drainage system has enough capacity to handle the incoming water. Therefore, it is necessary to develop a simulation model that can immediately predict the time series of water levels under various scenarios of possible drainage operations.
In this study, considering the Lake Toyanogata basin in Niigata Prefecture, Japan, which is a low-lying agricultural area, we developed a DNN (Deep Neural Network) model that generates the time series of inflows to the lake based on rainfall data and drainage data pumped out from the lake to the external river. Then, the applicability to a short-term water level prediction model was examined. For input learning data of the machine learning model, we used the artificially generated mock data acquired using the process-based simulation for drainage analysis and the observed data of the past rainfall event. As a result, it became possible to perform rainfall-runoff simulations in significantly less computational time compared with the process-based model. The developed DNN model can generate data that would well agree with the test data of big rainfall events that were not included in the measured data, suggesting the effectiveness of the method of complementing the learning data with the artificially generated data. In addition, the proposed short-term water-level prediction model can be applied to a tool that supports the operation of drainage pump stations. Moreover, in this study, we will discuss the advantages and disadvantages of process-based models and machine learning approaches considering the characteristics of each model.
In this study, considering the Lake Toyanogata basin in Niigata Prefecture, Japan, which is a low-lying agricultural area, we developed a DNN (Deep Neural Network) model that generates the time series of inflows to the lake based on rainfall data and drainage data pumped out from the lake to the external river. Then, the applicability to a short-term water level prediction model was examined. For input learning data of the machine learning model, we used the artificially generated mock data acquired using the process-based simulation for drainage analysis and the observed data of the past rainfall event. As a result, it became possible to perform rainfall-runoff simulations in significantly less computational time compared with the process-based model. The developed DNN model can generate data that would well agree with the test data of big rainfall events that were not included in the measured data, suggesting the effectiveness of the method of complementing the learning data with the artificially generated data. In addition, the proposed short-term water-level prediction model can be applied to a tool that supports the operation of drainage pump stations. Moreover, in this study, we will discuss the advantages and disadvantages of process-based models and machine learning approaches considering the characteristics of each model.