Prediction of Ferizi River-Flow Using Data-Driven Models

Rivers serve as crucial sources of freshwater on Earth, and precise prediction of river flows plays a vital role in effective water resource management. To address this need, researchers have proposed various models. These models can be categorized into two types: (1) hydrological/hydraulic physically based models and (2) data-driven models. The cited papers provide evidence of the significance of employing data-driven models for hydrological issues, particularly in predicting river discharge.

In this study, the researchers utilized a gene expression programming model (GEP) and support vector machine (SVM) to predict the discharge of the Ferizi River. The dataset used in the study covered a period of ten years (2011-2020) and consisted of daily Ferizi river discharge readings obtained from the Moushang hydrometry station. The study area had an average altitude of 2171 meters above sea level, with a longitude range of 58° 49' 30" to 59° 4' 30" E and a latitude range of 36° 20' 1" to 36° 32' 1". For the prediction models, daily river discharge data from 1 to 5 days ahead were utilized as input variables for both the GEP and SVM models. To ensure data homogeneity and address any deficiencies, the researchers employed the run test and calculated the correlation coefficient between neighboring stations. Subsequently, the dataset was randomly divided into two groups: 80% for model training and 20% for model testing, as well as evaluating the modeling error. The performance of the models was assessed during the training and validation stages using various evaluation metrics, including root mean square error (RMSE), coefficient of correlation (R), bias, Kling Gupta (KGE), and Nash–Sutcliffe model efficiency (NSE). Additionally, in order to estimate the inflow to the Ferizi River using the SVM model, three common types of kernel functions in hydrology were examined, namely linear kernel, polynomial, and radial basis functions.

Among the different functions considered, the radial basis function was selected due to its lower error compared to other functions when applied to the variables. The best input model, Model No. 5, incorporated previous flow variables with a delay of five time steps. In the training phase, both the GEP and SVM models exhibited the highest performance in forecasting the daily flow of Moshang station. The performance of the applied models suggests that SVM (RSME = 1.15, R = 0.985, NSE = 0.85, and KGE = 0.79) demonstrates higher precision in the validation stage for river discharge prediction compared to GEP (RSME = 1.65, R = 0.964, NSE = 0.78, and KGE = 0.69).

Conclusion and Suggestions :

This study has demonstrated that the utilization of soft computing techniques, such as SVM and GEP, is a powerful tool in predicting river flow. These techniques are capable of establishing a favorable relationship between input and output parameters, enabling accurate simulation of average and maximum daily flow.