Detail Abstract

Developing a Recurrent Neural Networks Model for the Da River Flowrate Forecasting in Vietnam

Abstract

Real-time streamflow prediction, especially flood forecasting, is an essential requirement in the management of river basin. However, this problem is still a challenge for regulators in developing countries, where access to real-time remote sensing data is limited. In order to solve the above problem, an effective approach to predicting incoming flows based on the Recurrent Neural Network (RNN) model has been introduced in this study. We have developed a Long Short-Term Memory (LSTM) Neural Network model, which is a special kind of RNN. Along with the Gated Recurrent Unit (GRU), LSTM model is one of the most widely used models in deep learning for solving sequential data problems. In this study, Python is the programming language of choice and TensorFlow is the framework of the selected model. TensorFlow is an open-source software library for dataflow programming across a variety of different complex tasks. The LSTM model developed in this study has been using to forecast the Da River flowrate one and two days ahead at the Lai Chau hydrological station located in the northwest mountainous area of Vietnam. The model was constructed without requiring rainfall data, land cover as well as topography. The input data is only daily flowrate measured at Lai Chau hydrological gauge station. All data available for 33 years (1984-2016) is divided into three non-overlapping sets for different purposes. The data set of 32 years (1984-2015) was employed for training and validation purposes, with corresponding rates of 80% for training and 20% for validation respectively. Another data set of one year (2016) was used for testing purpose to objectively verify the performance and accuracy of the model. Though only a modest amount of input data is required, the forecasted results are in satisfactory agreement with observed data. For the validation phase, the value of Nash - Sutcliffe efficiency (NSE) reaches about 0.946 when forecasting for one day in advance, while the corresponding figure in the testing phase is well above 0.926. The findings of this study illustrate the high performance of the LSTM neural networks model in predicting daily flowrate at the Lai Chau station. As a result, Recurrent Neural Networks models, as well as LSTM models, have great potential for flood forecasting and disaster prevention.