Streamflow Estimation in a Mediterranean Watershed Using Neural Network Models: A Detailed Description of the Implementation and Optimization

Abstract

This study compares the performance of three different neural network models to estimate daily streamflow in a watershed under a natural flow regime. Based on existing and public tools, different types of NN models were developed, namely, multi-layer perceptron, long short-term memory, and convolutional neural network. Precipitation was either considered an input variable on its own or combined with air temperature as another input variable. Different periods of accumulation, average, and/or delay were considered. The models’ structures were optimized and automatically showed that CNN performed best, reaching, for example, a Nash–Sutcliffe efficiency of 0.86 and a root mean square error of 4.2 m3 s−1. This solution considers a 1D convolutional layer and a dense layer as the input and output layers, respectively. Between those layers, two 1D convolutional layers are considered. As input variables, the best performance was reached when the accumulated precipitation values were 1 to 5, and 10 days and delayed by 1 to 7 days.

Keywords

• neural networks
• MLP
• LSTM
• CNN
• streamflow estimation