Implementation of a Machine-Learning-Based Approach for Forecasting Watershed Stream Flow (Case Study: Chehel Chai Watershed, Iran)

Abstract

The importance of the optimal and efficient use of all available water resources becomes noticeable when today due to successive droughts and a decrease in rainfall, the surface water resources are running out. Runoff and surface water resources are some of the primary and vital available water resources, and hence, modeling and predicting their behavior are especially critical. In the current research, the aim was to model the stream flow of the Chehel Chai watershed in Golestan province, Iran, using the data of the stream flow and precipitation for a period of 45 years. For this reason, 4 machine learning algorithms namely, Extreme Learning Machine (ELM), Random Forest (RF), Gaussian Process Regression (GPR), and Gene Expression Programming (GEP) were used. The data were entered into the modeling in the form of different scenarios consisting of the stream flow and precipitation with varying lags of time. The results showed that scenario M2 (using only stream flow data with two time lags) in the ELM (extreme learning machine) model with the values of RMSE (root mean square error) =0.984 (m3/s) and R2=0.613 had the most accurate performance and predictions among all the models and scenarios.