Prediction of Flow Based on a CNN-LSTM Combined Deep Learning Approach

Abstract

Although machine learning (ML) techniques are increasingly used in rainfall-runoff models, most of them are based on one-dimensional datasets. In this study, a rainfall-runoff model with deep learning algorithms (CNN-LSTM) was proposed to compute runoff in the watershed based on two-dimensional rainfall radar maps directly. The model explored a convolutional neural network (CNN) to process two-dimensional rainfall maps and long short-term memory (LSTM) to process one-dimensional output data from the CNN and the upstream runoff in order to calculate the flow of the downstream runoff. In addition, the Elbe River basin in Sachsen, Germany, was selected as the study area, and the high-water periods of 2006, 2011, and 2013, and the low-water periods of 2015 and 2018 were used as the study periods. Via the fivefold validation, we found that the Nash–Sutcliffe efficiency (NSE) and Kling–Gupta efficiency (KGE) fluctuated from 0.46 to 0.97 and from 0.47 to 0.92 for the high-water period, where the optimal fold achieved 0.97 and 0.92, respectively. For the low-water period, the NSE and KGE ranged from 0.63 to 0.86 and from 0.68 to 0.93, where the optimal fold achieved 0.86 and 0.93, respectively. Our results demonstrate that CNN-LSTM would be useful for estimating water availability and flood alerts for river basin management.