A Distributed Parallel Genetic Algorithm With Application To Optimization Of Strongly Coupling Separation Processes

The separation process is one of the most important processes in chemical industry, and also one that consumes a considering amount of energy and resources. Heat and water integration are widely used to reduce energy and resource consumption, which may result in strongly coupling between operating parameters of some units in separation process. The application of mathematical programming tools to improve the efficiency of optimizing such kind of processes has been vastly studied in the past years. Most researchers pay much attention to the generation of superstructures of separation sequences and the solution of the mathematical models which is usually named as MENLP by simplifying the topology structures of the separation processes or supplying a near-optimal initial value. However the simplification of the processes would lead to the loss of the accuracy of solutions, while the supplement of a reasonable initial value for the convenience of the convergence sounds unrealistic for many particular processes. Moreover, the terrible complexity and time-consumption for MINLP also play a key role as the barrier of putting it into a wide practice.A Distributed-Parallel-Genetic Algorithm (DPGA) framework integrating with commercial software of ASPEN PLUS~@ is presented in this paper which is based on the PC cluster. The simpleness and goodness of fit for an engineer as a powerful tool could be illustrated by the following factors: First of all, the adoption of the process simulation software enables us to achieve an accurate and convergent result without any process simplifications and construct, modify and maintain processes as an engineer wants; Secondly, The rigid mathematical models for the calculation and the objective function adopted by DPGA enhance the credibility and reliability of the results; Thirdly, the parallel computing strategy reduces the time consumption dramatically and guarantees the computation efficiency of the optimization approach highly; Last but not least, the apply of the distribute environment based on theEthernet intranet promises users a handy way to compose and maintain it.This paper is composed of three parts. For the first part, the genetic algorithm (GA) is introduced by the detail description of the suitable design of chromosomes, the specialized genetic operators, the mixed code and random parent-number fitness-weighted crossover. To improve the performance of the standard GA, several strategies such as the multi-step evolution strategy, the adaptive crossover and mutate probability are implemented. To avoid the trouble of shrinking the search space greatly in the early generation and being easily trapped in a local optimal solution for the traditional GA, a multi-step evolution method is introduced which additionally compensates for the low evolution speed based on the feedback control principle. By analyzing the consumption time of a single chromosome, we found it varies widely but with most concentrating in a small interval which could leads to a safe conclusion of reducing the waiting time is the key factor of improving the computation efficiency of DPGA, so a Load-Balance strategy of abandoning individuals whose computation times are larger than a fixed value is integrated into the DPGA.The second part presents a detailed description on how the DPGA system works. The standard client-server model is used, in which the server is responsible for the implementation of GA while the individual clients take charge of the much computation of the simulation mainly by APEN PLUS and communicate with the server for the required parameters derived from GA.To illustrate the performance and efficiency of the DPGA, a case is studied with a process of three-column distillation for the purification of synthetic methanol which strongly couples and is difficult to be optimized using MINLP due to the water and heat integration. The Optimization Goal of the process is to minimize the annual cost. The result illustrates that DPGA integrated with the Load-Balance strategy, the adaptive crossover and mutate probability strategy could promote the computation performance and efficiency dramatically compared with the traditional GA as the computation consumption can be reduced by 48.3%. Compared with the original cost, the adoption of DPGA could help save 23.1% of the total cost annually.To achieve and validate the reliability of DPGA, the varieties of some factors including the population size of chromosomes, the number of PCs and the number of the evolution of generation have been testified in a wide range which finally gives us a sound result with a promise of being able to find a optimal solution for a strong-coupling separation system. For the above case, we found the optimizationefficiency of DPGA would be highest and the randomicity could be controlled efficiently in a low level when the population size is between 140 and 160; while the computation efficiency is best when the proportion of the number of computers to the population size is near 0.08.