Evaluation of GPR-PSO and KNN-PSO data-mining models for prediction of suspended sediment concentration distribution

Document Type : Research Paper

Authors

• Mohsen Nasrabadi ¹
• Yaser Mehri ²
• Ali Abdolrazaq Sabbar ³
• MohammadJavad Nahvinia ³

¹ Department of Water Science and Engineering, Faculty of Agriculture and Environment, Arak University, Arak, Iran.

² Depratment of Irrigation and Reclamation Engineering, University College of Agriculture and Natural Resources, University of Tehran, Karaj, Iran.

³ Department of Water Science and Engineering, Faculty of Agriculture and Environment, Arak University, Arak, Iran

Abstract

Keywords

• Concentration distribution
• suspended sediments
• data mining models
• Gaussian process regression

Main Subjects

• سازه‌های آبی

Introduction

The vertical distribution of suspended sediment concentration (SSC) is one of the most fundamental parameters in the hydraulics of sediment transport in rivers. This parameter plays an important role in calculating the total sediment discharge in channels and rivers. For this reason, accurate measurement of this parameter has always been one of the goals of researchers. One way to accurately estimate this parameter is to use intelligent models.

The main aim of the present study is to present and evaluate highly accurate models for predicting suspended sediment concentration under laboratory conditions. Calculation of suspended sediment distribution is one of the important parameters in river engineering, and its accurate calculation leads to an acceptable estimate of suspended sediment discharge. Considering the necessity of accurate measurement of suspended sediment distribution and the difficulty of laboratory works, the use of intelligent models can be useful. To solve this problem, four intelligent models of KNN, KNN-PSO, GPR, GPR-PSO have been used in the present study.

Methodology

In this study, four intelligent models KNN, KNN-PSO, GPR, GPR-PSO were used to predict the distribution of suspended sediment concentration. For this purpose, all models were coded in the MATLAB software environment. Laboratory data of Vanoni (1946) and Einstein and Chien (1955) were used for modeling (Table 1). These data include 203-point data of sediment concentration distribution. Einstein and Chien’s experiments were conducted in a flume with a width of 30.7 cm and a smooth bed with different slopes of 0.00185 and 0.0025. Sediment particles with diameters of 1.3, 0.94 and 0.274 mm were used. In addition, Vanoni’s experiments were conducted in a flume with a rough bed with a width of 84.5 cm and a constant slope of 0.0025. The diameters of sediment particles used in this study were 0.1, 0.13, and 0.16 mm.

Results and Discussion

The data were first normalized and adjusted between 0 and 1. Then, 80% of the data were used for training and 20% of them were used for testing. The results showed that the optimization by PSO method has increased the accuracy of GPR and KNN models. It was found that the superior model is the GPR-PSO model, which has an accuracy of RMSE = 0.0297, R² = 0.9878 and KGE = 0.9776 in the training phase and RMSE = 0.0226, R² = 0.9907 and KGE = 0.9715 in the testing phase. After the aforementioned model, the KNN-PSO model was ranked. These results show that if the optimal values ​​are selected for the GPR model, it will have more accuracy than KNN model. This is because the prediction level in KNN is discrete, which makes the search space for PSO limited, while in the GPR model the prediction level is continuous, which makes the PSO optimization search a larger space of parameters, which makes it more accurate in regression and interpolation problems. Also, the modeling method in KNN is linear, which makes the optimization performed less effective; while the GPR model has a greater ability to model nonlinear relationships such as sediment problems, which makes this model more flexible than other models. In addition, the GPR model has the ability to consider uncertainty in modeling, while this is not possible with KNN. This is because the GPR model is a probabilistic model that models the distribution of the output variable according to the input variable, while KNN only considers the K-nearest neighbors for the prediction.

Conclusions

According to the results obtained, it was found that PSO optimization had an important effect on the performance of the GPR and KNN models and increased the accuracy of the models. Among the studied models, the GPR-PSO model was recognized as the best model. The accuracy of this model in the test phase was equal to RMSE = 0.0226, R² = 0.9907 and KGE = 0.9715. Also, by analyzing the results of the studied models, it was determined that the parameters y/D and y/a in most models were recognized as important parameters in determining the highest accuracy.

The results indicate that PSO optimization impacted the performance of the GPR and KNN models, resulting in increased accuracy. The GPR-PSO model was identified as the most effective model, with test phase accuracy metrics of RMSE = 0.0226, R2 = 0.9907, and KGE = 0.9715. Analysis of the results also revealed that the parameters y/D and y/a were consistently identified as crucial factors in achieving optimal accuracy across most models.

Author Contributions

“Conceptualization, Y.M. and M.N.; methodology, A.S.; software, Y.M.; validation, Y.M., M.N. and M.J.N.; formal analysis, Y.M.; investigation, M.N.; resources, A.S.; data curation, M.N.; writing—original draft preparation, Y.M.; writing—review and editing, M.N.; visualization, Y.M.; supervision, M.N.;

All authors have read and agreed to the published version of the manuscript.”

Data Availability Statement

Not applicable

Acknowledgements

The authors would like to thank all participants of the present study.

Ethical considerations

The study was approved by the Ethics Committee of the Arak University. The authors avoided data fabrication, falsification, plagiarism, and misconduct.

Conflict of interest

The author declares no conflict of interest.