|
River Flow Prediction Using Machine Learning: A Case Study of the Middle Qareh-Sou Basin
|
Sayyad Asghari Saraskanroud *    , Amin Abdolalipouradl |
| Professor ، Department of Physical Geography ، Faculty of Social Sciences ، University of Mohaghegh Ardabili ، Ardabil ، Iran |
|
|
Abstract: (113 Views) |
| The Qareh-Sou River is one of the most important rivers in agricultural water supply in Ardabil Province. The present study was conducted to evaluate five machine learning models, namely Random Forest (RF), K-Nearest Neighbors (KNN), Linear Regression (LR), Gradient Boosting (GBoosting), and Artificial Neural Network (ANN), for daily streamflow prediction at two stations, Samian (1999–2018) and Arbāb-Kandi (1976–2018), located in the lower basin of Qareh-Sou River. In this research, 17 hydro-climatic features were used as model inputs, and daily discharge was considered as the output.To assess model performance, statistical evaluation parameters including the coefficient of determination (R²), mean absolute error (MAE), and root mean square error (RMSE) were employed. Furthermore, uncertainty analysis was conducted using the prediction interval coverage probability (PICP), normalized mean prediction interval width (PINAW), and mean prediction interval width (MPIW). results for the Samian station indicated that the GBoosting model, with R² = 0.8530 and RMSE = 0.651, provided the best performance among all models, while KNN exhibited the weakest performance, with R² = 0.4380 and RMSE = 1.273. Uncertainty indices also revealed that GBoosting, with MPIW = 1.7577 and PINAW = 0.13020, produced a narrower prediction interval and achieved the 95% coverage target. In contrast, KNN, despite achieving 95% coverage, yielded MPIW = 3.8413 and PINAW = 0.28454, indicating a wider interval and poorer performance.At the Arbāb-Kandi station, Random Forest outperformed the other models with R² = 0.7930 and RMSE = 1.0906, whereas KNN again showed the weakest performance, with R² = 0.6663 and RMSE = 1.386. Uncertainty indices confirmed that RF placed approximately 94% of the data within the prediction interval, with PINAW = 0.06533 and MPIW = 1.575. In comparison, KNN covered 93% of the data, with PINAW = 0.10325 and MPIW = 2.0488, further confirming its inferior performance at this station.Overall, tree-based models (GBoosting and RF) demonstrated superior accuracy compared to KNN at both stations. Finally, it is recommended that hourly hydro-climatic data, including discharge, temperature, precipitation, and humidity, be utilized to improve prediction accuracy. |
Article number: 4 |
| Keywords: Ardabil, Machine learning, discharge, prediction |
|
|
|
Type of Study: Research |
Subject:
Special
Received: 2026/07/8 | Revised: 2026/09/23 | Accepted: 2026/09/18 | ePublished ahead of print: 2026/09/18
|
|
|
|
|
|
|
| Send email to the article author |
|
|
|
| Add your comments about this article |
|
|
|
