Optimization Of Complex Distillation Processes Using Stochastic Algorithms And Machine Learning

The significance of the distillation process optimization by stochastic algorithms is not limited to finding promising solutions with improved energy efficiency,but a distinguished method for computers themselves to explore new ideas.However,as the complexity and dimensionality of real-world problems keep increasing,the inefficiency of stochastic algorithms becomes more and more eye-catching.The development of machine learning technology provides an opportunity to improve conventional optimization methods.However,it still suffers from the long sampling time and low prediction accuracy,hindering the optimization of complex distillation processes.To overcome the drawbacks,this thesis herein makes progress in three aspects: parallel computing,superstructure modeling,and surrogate model accuracy.Firstly,to utilize hardware more efficiently,a population-distributed parallel computing framework that eases the restriction of interface invoking is proposed and improves the optimization efficiency by 2.2-9.8 times.Moreover,the influence of subpopulation synchronization on optimization is evaluated,demonstrating the synchronously distributed differential evolution to be a good trade-off between efficiency and stability.Then,superstructure models are improved with an adaptive strategy.The logical disjunction is introduced into the sequential modular approach,to adaptively reduce the redundancy caused by logical variables to predefine the configuration,a priori in the conventional method.Benefit from the reduced complexity of superstructure models,the optimization by stochastic algorithms could find better solutions with 8.3%-24.0% lower computing costs.Finally,a predicted solution sets guided machine learning method is proposed for process optimization.This method repeats prediction and obtains a solution set that consists of dispersed solutions and defines a subregion that includes the maximum-likelihood positions of the optima.With the solution set,the machine learning method can define smaller subregions for more effective model refining and approximate the high-quality solution more robustly.Case studies demonstrated that the improved machine learning method requires 44.7%-68.1%fewer individual evaluations by rigorous simulation compared to the conventional optimization,and the relative deviations of solutions are lower than 0.9%.Additionally,the parallel computing efficiency of sampling is 52.3% higher than that of population evaluation due to the significantly larger sampling scale than the population size.In summary,the machine learning framework using 20 parallel threads could reduce the time consumption by 82.4%-95.7%,significantly improving the performance of stochastic algorithms for process optimization.In a conclusion,this thesis overcame the shortage of inefficiency of stochastic algorithms dealing with process optimization,and made a significant contribution to the field of process systems engineering,by enriching the approach of applying artificial intelligence in this field.