Optimization Method Of Complex Chemical Process Based On Surrogate

With the advent of the digital twin era,accurate numerical simulations for complex reaction mechanisms in chemical processes have emerged.Because of a large of fitness evaluations in the evolutionary algorithms,there are exist problems such as expensive computation and difficult convergence in the optimization processes based on the high precision models.For the above problems,the entrained flow coal gasification process was taken as the research object in the thesis,the complex simulation model was replaced by the surrogate.The relevant researches on the offline optimization and online optimization were carried out by the surrogate model construction,which provides new technical guidance and reference for the chemical engineering process optimization with high computational costs.The specific research contents are as follows:1.Based on the simulation model of the coal gasification process,a large number of modeling samples were generated by latin hypercube sampling,and the surrogate models including Kriging,radial basis function,polynomial response surface and their ensemble were respectively constructed offline.By the model performance evaluation,the surrogate that meets the approximate accuracy requirements was used in the optimization.It is found that the ensemble which combines the characteristics of multiple surrogates has stronger adaptability and higher fitting accuracy than that of other single surrogates,which can completely replace the simulation model in the optimization.Through genetic algorithm optimization,under the same times of calling the simulation model,the optimization results of the ensemble model are better than those of the simulation model,and are significantly better than the results of the literature.2.In view of the fact that it is only small size samples,the adaptive selection model ensemble algorithm was proposed.First,the existing data can be fully utilized by Bootstrap sampling,and a large number of submodels were established as base learners.From the perspective of the diversity of the model,the submodels were sorted according to their predicted values on the current optimal solution and evenly divide them into several groups,then the submodel with the worst prediction accuracy were selected from some groups according to a certain probability as the representative model for increasing the model diversity,and submodels from the remaining groups were randomly selected.The selected submodels were integrated into the global surrogate by ensemble learning.With the adaptive selection of the submodels and the updation of the global model,the optimal solution will gradually approach the optimal solution.The optimization results of the classic benchmark functions and the coal gasification process show that compared with other algorithms,the proposed algorithm proposed not only considers local approximation accuracy and improves the diversity of the model,but also reduces computational complexity,effectively improving the prediction quality of the global surrogate.The optimal solutions found by the proposed algorithm are the best on most of benchmark functions and the optimization conditions with higher effective syngas yield are obtained.3.It is easy to fall into the local optimum in the optimization process.For jumping out of the local optimum,the prediction accuracy of the surrogate can be improved by exploring the areas with less known samples in the design space and selecting samples with high uncertainty.The sampling strategy for the most uncertain samples was proposed based on leave one cross validation and Voronoi diagram.By leave one cross validation,the point with the highest variance was found,and the region including which has a high uncertainty.The area with high uncertainty was determined using Voronoi diagram and then a large number of random points were generated by the Monte Carlo to approximate its boundary range.For improving the diversity of sampling points and the prediction accuracy of the surrogate model,the most uncertain sample was selected by maximizing the objective function which composites the indexes of the distances between random points and known points and the prediction deviations of the surrogate.In addition to the global optimum,the local optimum was selected by constructing the local surrogate in the particle swarm optimization,which further improves the prediction accuracy of the surrogate near the optimal solution.As new samples were sequentially added into the training dataset,and the surrogate was updated gradually,and further the global search and local search were switched dynamically according to the improvement of the optimization performance.Compared with other online optimization algorithms,the results show that the proposed algorithm considers both the global exploration sampling and the local exploiting sampling,and improves the construction quality of the surrogate,and finds the best quality optimization solutions on most of test functions and coal gasification process optimization.Under the same times of calling the simulation model,the effective syngas yield is higher than the optimization results of the offline surrogate,which improves the efficiency of the engineering optimization design.4.With the acquisition of new samples in the optimization,the known feature of the search area will change,and the surrogate that can accurately describe the features of the region should be also reselected.For solving the problem,the dynamic model management strategy based on adaptive model selection was proposed.K-fold cross validation and minimum root mean square error were used to measure the model performance.Based on the current search area and the known samples in which,the appropriate surrogate with the highest prediction accuracy was selected from the diversity model library composed of typical single models and their ensemble.As the search area changes,the global modeling and local modeling can be dynamically switched by the particle swarm optimization.By verifying on the benchmark functions and the coal gasification process,the proposed algorithm obtains better optimization solutions than other algorithms on most of test function and higher effective syngas yield in the limited function evaluations.It shows that the algorithm can select the surrogate that better describes the characteristics of the current search area,and has good advantages in the optimization quality and efficiency.