Loading...

Evaluating The Efficiency of Machine Learning Methods for Variance-Based Sensitivity Analysis in Hydrological Models

Khanjani, Mohammad | 2024

4 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 56979 (09)
  4. University: Sharif University of Technology
  5. Department: Civil Engineering
  6. Advisor(s): Sheikholeslami, Razi
  7. Abstract:
  8. It is inevitable to simulate the phenomena that govern nature to make decisions and policies on water and environmental issues. Therefore, researchers have tried to make a good prediction of water's quantitative and qualitative status by using modeling and computer processing power. One of the primary and essential issues in modeling is using sensitivity analysis techniques. Using this knowledge, much information will be obtained on how to influence the model's output regarding various factors such as parameters, input data, and initial and boundary conditions. Global sensitivity analysis based on variance as one of the most well-known and, at the same time, the most widely used sensitivity methods has attracted the attention of water and environmental researchers in recent years, and one of the most common methods is the Sobol method. With the growth of machine learning algorithms, researchers have introduced the data-driven approach as a new method of sensitivity analysis, which, in addition to the lower computational load, does not require the use of a physical model, as well as the lack of a sampling strategy. Recently, the variable importance index obtained from the random forest algorithm has been widely used as one of the critical indicators for sensitivity analysis with a data-driven approach. Mathematically, it can be shown that the index calculated by this algorithm is theoretically related to the total-order sensitivity index calculated by the Sobol method. By reviewing previous sources, no research evaluated the efficiency of these two indicators about the parameters of hydrological models. Therefore, the main goal of this research is to comprehensively evaluate and compare global sensitivity analysis based on variance (Sobol) and data-driven sensitivity analysis of random forest (variable importance index) in mathematical and hydrological models, using four criteria of convergence, precision, robustness and the computational time of the two sensitivity methods. For this purpose, three mathematical models and two conceptual hydrological models of rainfall-runoff HBV and TANK were used. Numerical results show that for all models, the total-order sensitivity index obtained by random forest performs better than the traditional Sobol method in terms of convergence and robustness of the results. This is while the precision criterion performed slightly better in the random forest method. The computing time of the random forest method was 25 times that of the Sobol method in mathematical models. In contrast, in hydrological models, the computing time of the Sobol method for the HBV model was 46 times, and for the TANK model, it was 548 times that of the random forest method. Based on this, the cost-effectiveness of the random forest method compared to the Sobol method in models with high complexity and parameters was determined
  9. Keywords:
  10. Hydrological Modeling ; Random Forest Algorithm ; Global Sensitivity Analysis ; Variable Importance Index ; Variance-Based Global Sensitivity Analysis ; Machine Learning

 Digital Object List

 Bookmark