UsingGeneticAlgorithmsasaControllerforHotMetal分析.docVIP

Using Genetic Algorithms as a Controller for Hot Metal Temperature in Blast Furnace Processes By Danny Lai May 22, 2000 Advanced Undergraduate Project (6.199) Supervised by Dr. Amar Gupta Abstract The hot metal temperature (HMT) of a blast furnace is an important indicator of the furnace’s internal state. Because the blast furnace system is non-linear, neural networks have been employed to approximate these systems. Artificial neural nets have been developed to approximate these dynamics and predict future HMT hours into the future, with limited accuracy. This paper focuses on inversion, focusing on the development of an HMT controller. The controller will, given a desired HMT and desired future time, determine the appropriate input values now in order to achieve the goals. Genetic algorithms will be used to derive the controller. Implementation of the genetic algorithm required creating a genome representation, appropriate genome operators, a fitness function, and termination criteria. In each genome, functions were approximated using multiple data points. Roughly 13 data points were used, reflecting a HMT spread of 30 degrees above and below a set operating point. Input variables were limited to the 10-14 that showed highest consistent correlation with HMT for the given operating point. A basic genetic algorithm was employed, using constant mutation and crossover probabilities and a terminate-upon-generation criterion. Preliminary results showed that after a fixed number of generations, the algorithm did not reliably terminate with acceptable solutions. This phenomenon was likely due to the large genome strings (over 500 bits per genome) and loose evolution protocol. It is suggested that a more powerful platform be used in the future. Table of Contents Introduction 3 Background 4 Genetic Algorithms Overview 4 Genetic Algorithms Description 5 Genome Representation 6 Initialization 6 Crossover 7 Mutation 13 Objective (fitness) function and generation tr