Data Driven Water Surface Elevation Forecasting Model with Hybrid Activation Function—A Case Study for Hangang River, South Korea

Abstract

To date, physical, numerical or data-driven models have been used to forecast water surface elevation in rivers for specific times or locations in the literature. Recently, the trend of forecasting water surface elevation changed from physical and numerical models to data-driven models with the help of the development of big data processing technology and fast simulating time of data-driven models. In this study, a data-driven model with Long Short-Term Memory (LSTM) was developed using TensorFlow, one of the famous deep learning frameworks and forecasting of water surface elevation affected tidal river was performed in Hangang River, Korea. From many types of field measurements, the hourly hydrological data, precipitation, outlet discharge of dam upstream and tidal levels were selected as the input dataset through a t-test and a p-value. In particular, the hybrid activation function was proposed to alleviate the vanishing gradient and dying neuron problems generally issued in the application of the activation function. The model showed that the root mean square error (RMSE) and peak error (PE) decreased by 0.22–0.25 m and 0.11–0.21 m, respectively, and the Nash-Sutcliffe efficiency (NSE) increased up to 79.3%–97.0% compared with the single activation functions. For w 1 = 0.6 and w 2 = 0.4 in the hybrid activation function, the improvement of accuracy and the enhancement of the application range of the leading time interval were obtained through a sensitivity analysis. Moreover, the hybrid activation function showed a good performance. The forecasting results provided by this model can be used as reference data for the establishment of the emergency action plan (EAP).