Developing a Cloud-based Toolbox: Sensitivity/Uncertainty Analysis of a Water Quality Model

Abstract

The complexity associated with water quality models (WQMs) have increased owing to the introduction of numerous physical and biological mechanisms in the models. In response, global sensitivity analysis (GSA) and uncertainty analysis (UA) have been conducted to identify influential parameters in these mechanisms and quantify model uncertainty. To implement GSA and UA, a substantial amount of computational power and time is required to obtain numerical solutions during model simulations. By using cloud computing systems, these issues can be solved to some extent in that cloud systems provide massive computational power and service. For these reasons, this study has designed a graphical user interface (GUI) toolbox to allow users to perform GSA and UA of the WQMs more conveniently and effectively on cloud computing. In addition, different monitoring data of waterbodies (i.e., grab samples and hyperspectral imaging (HSI) data) were compared with the WQM results to explore the data dependency of the SA and UA results. This is because grab sampling techniques do not yield spatial distribution of algal blooms, thus limiting the scope of SA and UA, while HSI data identify spatial and spectral information of water quality. The WQM of this study was a modified version of Environmental Fluid Dynamic Code (EFDC), a three-dimensional hydrodynamic and water quality model, referred to as EFDC-NIER. First, the development of our toolbox was described, and the computational efficiency of the toolbox was explored (Chapter 3). The toolbox adopted 1) Latin Hypercube One-factor-At-a-Time (LH-OAT) sampling and multiple model executions; 2) GSA using Elementary Effect Test (EET); and 3) UA using Generalized Likelihood Uncertainty Estimation (GLUE). In addition, the toolbox included multiple executions of the EFDC model and the use of grab sampling and HSI data. Coupling of multiple execution and cloud computing approaches overcomes the limitation of computing power availability. The computing capacity of the cloud system (i.e., 72 Intel Xeon Platinum 8000 processor and 192 GB memory) was compared to a desktop computer (PC) (i.e., two Intel Xeon CPU E5-2680 v3 and 128 GB memory). To simulate the 5,100 EFDC-NIER models, the PC with 50 execution showed a total of 837 hours, while 10 instances with 50 multiple executions required a total of 47 hours. Therefore, this toolbox can reduce the simulation time by up to 20 orders of magnitude than the PC. In addition, the simulation time ratio of the cloud system (10 instances) to PC reached a factor of 20 as the number of multiple executions increased. Second, this case study was conducted in the Yeongsan River, South Korea (Chapter 4). With the consideration of the spatial distribution of algae blooms, the SA and UA were performed with the HSI-based chlorophyll-a (Chl-a) concentration data which was compared with that of EFDC-NIER model results. The results of HSI-based SA were not exactly the same as the grab-based SA. The HSI-based SA results showed that a common sensitive parameter is TMD2, which is related to the optimal temperature for the algal growth. In this regard, TMD2 should be the priority during the calibration of the EFDC-NIER model with HSI-sampling data. The nutrient parameters (i.e., CPprm3 and ANCC) and algal kinetic parameters (i.e., PMD and PMM) should also be considered in the calibration of algal concentrations. In the HSI-based SA results, in addition, WQR shows less sensitivity than the grab-based SA results. This is likely because the HSI events were conducted under moderate wind speed and cloudless weather conditions; hence, there was less variability in the temperature values. This may result in less sensitivity of WQR, because the velocity of the vertical movement was largely affected by the temperature. These findings could be effective in water management by 1) identifying the influential parameters in algal blooms, and 2) reducing the computational cost for implementing SA and UA of the WQM. This could contribute to minimizing the operational and labor costs of future stakeholders.