Evaluating the efficiency of SVM, LS-SVM and SVM-GOA models in simulating the Flood peak discharge at the Poldokhtar station

Document Type : Research Paper

Authors

• Fatemeh Tavakoli ¹
• Hamed Nozari ²
• safar marofi ¹

¹ Dept. of Water Engineering Science, Faculty of Agricultural, University of Bu Ali-Sina , Hamedan, Iran

² Department of Water Sciences and Engineering, Faculty of Agriculture, Bu-Ali Sina University, Hamedan, Iran

Abstract

Keywords

• flood modeling
• support vector machine
• Grasshopper algorithm
• Karkheh basin
• Poldokhtar station

Main Subjects

• هیدرولوژی

Evaluating the efficiency of SVM, LS-SVM and SVM-GOA models in simulating the Flood peak discharge at the Poldokhtar station

EXTENDED ABSTRACT

Introduction

Floods are among the most devastating natural disasters, with profound impacts on ecosystems, human life, and causing environmental damage and economic losses to residential areas, agriculture, and water resources. (Choubin et al., 2019). Among natural disasters, floods cause the most damage in agriculture, fisheries, housing, and infrastructure sectors. Iran is also vulnerable to climate change due to its semi-arid climate and has experienced several devastating floods in the region over the past two decades. Various methods for flood forecasting exist worldwide, with machine learning techniques being particularly popular for their effectiveness in handling large datasets (Ranasinghe and Ilmini., 2020). In this context, based on the conducted research, it can be stated that among the models mentioned, the Support Vector Machine (SVM) model has been identified as a suitable method for modeling non-linear, complex, and dynamic processes like floods, demonstrating acceptable accuracy in flood forecasting. Therefore, considering the satisfactory accuracy of this method and the successful performance of meta-exploratory algorithms in enhancing its accuracy, two hybrid models, LS-SVM and SVM-GOA, were employed to simulate the peak discharge of the Poldokhtar station flood. The accuracy of these models, along with the classical Support Vector Machine (SVM), was assessed.

Methodology

In the research conducted to compare and evaluate the SVM, LS-SVM, and SVM-GOA models in flood modeling at the Poldokhtar hydrometry station, the following steps were undertaken. Initially, statistics and information regarding floods and precipitation in the upstream basin of the Poldokhtar hydrometric station were obtained daily from relevant organizations. Subsequently, a common statistical period was selected between the studied rain gauge and hydrometric stations (years 2009 to 2016). In the next phase, 74 flood events with instantaneous maximum discharge exceeding 60 cubic meters per second were identified within this timeframe, and the precipitation amount on the day of each flood event was determined at each upstream rain gauge station. Considering the impact of rainfall in the days leading up to a flood event on its occurrence, influenced by the time of concentration of the basin, a more detailed analysis included not only the rainfall on the day of the flood but also the rainfall from the five days prior to it. Finally, after selecting and determining the input and output values, the desired models were implemented.

Results and Discussion

In this research, which was conducted in order to compare and evaluate SVM, LS-SVM and SVM-GOA models in the flood modeling of Poldokhtar hydrometry station, the following results were obtained:

Necessarily, the use of rainfall time series with a 5-day scale and the use of statistics from all upstream stations will not increase the accuracy of modeling, and the use of additional series will reduce the accuracy by disrupting the training process. It can be inferred that increased complexity does not necessarily result in improved performance of hydrological models, and the performance may vary based on the hydrological variable or conditions considered

Overall, the results of calculating statistical indices demonstrate the relative superiority of the LS-SVM model compared to the SVM and SVM-GOA models.

The utilization of rainfall statistics from the four stations of Chameshk, Veysian, Khorramabad, and Aleshtar not only enhanced the performance of the LS-SVM model in flood modeling but also facilitated the identification of three key areas with the most significant impact on simulation accuracy.

In this study, the average execution time of the LS-SVM model was approximately a few seconds, whereas the SVM-GOA model took several hours to run. Additionally, manually setting parameters for the classic SVM model is a time-consuming process. Therefore, the LS-SVM model was easier to implement due to its fewer adjustable parameters compared to the SVM-GOA and SVM models.

Although the results show the relative equality of the SVM-GOA and SVM models, but the high execution time of the SVM-GOA model, especially in problems similar to this research, causes it to take months to achieve goals such as sensitivity analysis. From this point of view, the SVM model can be prioritized over the SVM-GOA model.

The output of the LS-SVM model for the same inputs is constant and not random. Meanwhile, the output of SVM-GOA model has a wider confidence band and it shows the randomness of its output in flood simulation.