Design And Implementation Of Heterogeneous Parallel Calibration Algorithm For Large-scale Hydrological Simulation

Hydrological model refers to a scientific model obtained by simply conceptualizing the process of complex hydrological phenomena that people cannot directly recognize in nature using modern simulation methods.The hydrological model is an important tool for hydrological simulation and the study of the water cycle process.Its development is affected and restricted by a variety of factors,of which the more important aspect is the parameter calibration of the hydrological model.It plays a vital role in the distributed hydrological model,and the calibration algorithm and the computing platform also play a decisive role in the effect and efficiency of the calibration.Since most of the hydrological models are composed of many complex physical models,there are a large number of parameters that need to be calibrated,which also makes the calibration of hydrological simulation parameters require considerable computing resources.Therefore,transplanting hydrological models to heterogeneous platforms with high-performance computing capabilities has broad development prospects for research on hydrological work.The current hydrological model parameter calibration process has the problem of too slow calculation and convergence speed.In view of the current development status and future development needs of hydrological model parameter calibration,this paper proposes a hydrological simulation parallel calibration method for heterogeneous platforms and a large-scale multi-stage heterogeneous parallel parameter calibration algorithm based on domestic accelerators.First,the traditional L-BFGS algorithm and genetic algorithm are parallelized and transformed to adapt their structure to the parameter calibration requirements of parallel calculation in hydrological simulation;secondly,the HIMS model is used as the basic hydrological simulation parameter calibration model,and the models are different.The structure transplantation makes the entire hydrological model run concurrently on the coprocessor side in the form of multi-threading,combined with the improved L-BFGS algorithm and genetic algorithm to realize the parallel heterogeneous realization of parameter calibration;again,The large-scale multi-level parallel algorithm with parameter calibration is combined with heterogeneous parallel algorithm to realize the cooperative work of main processor and coprocessor;finally,taking the Lhasa River Basin as an example,using a GPU server equipped with Intel+Nvidia and a domestic accelerator equipped with a domestic advanced computing platform Models and test cases are deployed on heterogeneous platforms.The test results show that the calibration method described in the article is suitable for the current mainstream processor + coprocessor architecture and can obtain better parameter calibration results.