Scholarlink

<p>Accurate suspended sediment load (SSL) prediction is crucial for effective basin management. Given the spatial variability of hydrological conditions, multi-site model evaluation is essential for robust SSL modeling. This study aims to determine the most accurate model for SSL prediction across diverse basin locations. It compares traditional machine learning methods, including artificial neural networks (ANN), adaptive neuro-fuzzy inference systems (ANFIS), and support vector machines (SVM), and advanced deep learning techniques, particularly long short-term memory (LSTM) and gated recurrent units (GRU). SSL, daily flow discharge (Q), and rainfall (R) data were collected from four stations in different regions of the Kashkan basin, Iran. Recognizing the significant influence of flow discharge on suspended sediment load, a novel approach was introduced by using one-day-ahead Q predictions as inputs for SSL modeling. The study identified that past flow data, Q(t-1), Q(t-2), Q(t-3), and rainfall data from the previous day, R(t-1), were most effective for predicting current-day flow discharge across all stations. Using the weighted aggregates sum product assessment (WASPAS) method, ANN performed best for Q prediction at all stations. For SSL, ANN outperformed other models at Kaka Reza (<i>R</i><sup>2</sup> = 0.629) and Sarab Seydali (<i>R</i><sup>2</sup> = 0.679) stations using Q(t), Q(t-1), and R(t-1) as inputs, while SVM performed better at Afarineh (<i>R</i><sup>2</sup> = 0.896) and Poldokhtar (<i>R</i><sup>2</sup> = 0.902) with Q(t) and R(t-1). These results indicate that traditional models outperform deep learning models Q and SSL prediction in the Kashkan basin.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Flow discharge and suspended sediment load prediction in multiple sites of a basin: a comparative study of classical AI and deep learning approaches

  • 期刊论文

  • 发表日期：

    2025年7月7日

查看全文
  • Fatemeh Avazpour,
  • Mohammad Reza Hadian,
  • Ali Talebi,
  • Ali Torabi Haghighi

摘要

Accurate suspended sediment load (SSL) prediction is crucial for effective basin management. Given the spatial variability of hydrological conditions, multi-site model evaluation is essential for robust SSL modeling. This study aims to determine the most accurate model for SSL prediction across diverse basin locations. It compares traditional machine learning methods, including artificial neural networks (ANN), adaptive neuro-fuzzy inference systems (ANFIS), and support vector machines (SVM), and advanced deep learning techniques, particularly long short-term memory (LSTM) and gated recurrent units (GRU). SSL, daily flow discharge (Q), and rainfall (R) data were collected from four stations in different regions of the Kashkan basin, Iran. Recognizing the significant influence of flow discharge on suspended sediment load, a novel approach was introduced by using one-day-ahead Q predictions as inputs for SSL modeling. The study identified that past flow data, Q(t-1), Q(t-2), Q(t-3), and rainfall data from the previous day, R(t-1), were most effective for predicting current-day flow discharge across all stations. Using the weighted aggregates sum product assessment (WASPAS) method, ANN performed best for Q prediction at all stations. For SSL, ANN outperformed other models at Kaka Reza (R2 = 0.629) and Sarab Seydali (R2 = 0.679) stations using Q(t), Q(t-1), and R(t-1) as inputs, while SVM performed better at Afarineh (R2 = 0.896) and Poldokhtar (R2 = 0.902) with Q(t) and R(t-1). These results indicate that traditional models outperform deep learning models Q and SSL prediction in the Kashkan basin.