Double-layer parallelization for hydrological model calibration on HPC systems

Abstract

s u m m a r yLarge-scale problems that demand high precision have remarkably increased the computational time of numerical simulation models. Therefore, the parallelization of models has been widely implemented in recent years. However, computing time remains a major challenge when a large model is calibrated using optimization techniques. To overcome this difficulty, we proposed a double-layer parallel system for hydrological model calibration using high-performance computing (HPC) systems. The lower-layer parallelism is achieved using a hydrological model, the Digital Yellow River Integrated Model, which was parallelized by decomposing river basins. The upper-layer parallelism is achieved by simultaneous hydrological simulations with different parameter combinations in the same generation of the genetic algorithm and is implemented using the job scheduling functions of an HPC system. The proposed system was applied to the upstream of the Qingjian River basin, a sub-basin of the middle Yellow River, to calibrate the model effectively by making full use of the computing resources in the HPC system and to investigate the model's behavior under various parameter combinations. This approach is applicable to most of the existing hydrology models for many applications.

Analysis of the Influence of Rainfall Spatial Uncertainty on Hydrological Simulations Using the Bootstrap Method

“…Former studies focused on the topic of hydrological simulation and soil erosion on the Loess Plateau using the DYRIM, as well as the parallelization of the DYRIM based on dynamic basin decomposition [53][54][55]. Moreover, the DYRIM has been widely employed to the simulations of streamflow, floods and sediment, which demonstrated its applicability with satisfactory results in the temporal scales ranging from daily to monthly in many river basins including the Yellow River basin [4,[56][57][58][59].…”

“…Several representative adjustable parameters are listed in Table 3. The actual physical meaning of the parameters and the effect on the simulation results are studied in detail in our previous work [55]. The topsoil initial moisture (θ u,0 ) is the initial state of the topsoil layer and θ d,0 represents that of the subsoil layer.…”

“…The calibration of model parameters is parallelized with a double-layer structure on an HPC system [55]. A dynamic sub-basin decomposition method [53] was developed to parallelize the hydrological simulation of the DYRIM, which contributes to the lower-layer parallelism.…”

“…When all of the jobs are completed, the model efficiency will be estimated using observed data to propose the fitness evaluation list. Generations of the GA will be run to explore more parameter combinations until the stop criterion is reached [55]. Because the fluctuation range of the NSE (Nash-Sutcliffe coefficient of efficiency) is needed to be explored in uncertainty estimation, so the maximum number of generations (e.g., 10 generations in this study) is set as the stop criterion.…”

A bootstrap method to estimate the influence of rainfall spatial uncertainty in hydrological simulations

“…Former studies focused on the topic of hydrological simulation and soil erosion on the Loess Plateau using the DYRIM, as well as the parallelization of the 25 DYRIM based on dynamic basin decomposition (Li et al, 2011;Wang et al, 2011;Zhang et al, 2016). Moreover, the DYRIM has been widely employed to the simulations of streamflow, floods and sediment, which demonstrated its applicability with satisfactory results in the temporal scales ranging from daily to monthly in the Yellow River basin (Shi et al, , 2016.…”