Genetic algorithm-based strategy for the steam reformer optimization

The steam reforming reaction is widely used for obtaining hydrogen. The reforming re-action has a strong endothermic character, which means it requires a considerable and continuous heat supply to proceed. Due to the process character, a highly non-uniform temperature field develops inside the reactor. It has a consequence in large temperature gradients, leading to the catalyst degradation and a reduced lifetime of the reforming unit. The aim of the presented research is to unify the temperature field developing in the reactor, for easier control of the process and extension of the reformer's life expectancy. A conventional plug-flow reactor consists of a cylindrical pipe body filled with catalyst. The presented methodology included optimizing the catalyst distribution in the reactor to ac-quire the most uniform temperature field possible. A genetic algorithm is selected as an optimization technique for finding the most advantageous alignment of the catalyst. It is an example of evolutionary algorithms, basing on rules similar to natural selection. The algorithm generates a random, initial population of reactors and the reforming simulation is executed for each of them. The computation results are then evaluated and ranked using predefined fitness functions. The ranked reactors parameters' are further recombined with selection probability based on the fitness values, until a whole new population is created and the algorithm's loop restarts. The higher the fitness value of a specific reactor, the higher are the chances of passing its segments composition to the proceeding generation. The fitness computation leaves a vast space for improvements, as it may be computed based on many different process' parameters. This work focuses on distinguishing differ-ences in the algorithm performance, depending on the formula for fitness calculation. The algorithm's converging speed, overall fitness values of specimens and optimization results were investigated and compared. The results show that the algorithm with an updated fitness calculation procedure performs considerably better. Only about 50% of