Intelligent Modeling Methods Based On Evolutionary Algorithms For Complex Chemical Processes And The Applications

Chemical industry is the foundation of our national economy, having respect to daily production activity and life. The modeling and optimization methods in chemical processes have drawn increasing concern because of the complexity. A model directly influences the process control performance, and the optimization plays a significant role in exploring the production potential, improving the efficiency and reducing the energy consumption. However, traditional methods are not always satisfactory in complex chemical processes. In this dissertation, for the complex chemical processes like ammonia synthesis, intelligent modeling approaches including neural network (NN) and support vector machines (SVM), and evolutionary algorithms such as particle swarm optimization (PSO) are studied. Based on these methods, some mathematical models for the process are established and several improved intelligent optimization algorithms are proposed to apply to the modeling. The main results in this dissertation can be summarized as follows:(1) A hybrid particle swarm optimization with prior crossover differential evolution (PSOPDE) is proposed. To overcome premature convergence, PSO and DE are combined, and the trial individuals which are abandoned in the selection operation of DE are used. PSOPDE provides two completely different searching ways for individuals so as to increase the probability of finding good solutions. Moreover, due to the high diversity of the abandoned trial individuals, a prior crossover operation between them and the updated ones based on PSO is implemented before DE computation. The setting of several key parameters in PSOPDE is analyzed through a large number of simulation experiments. The simulation results of typical test functions show that PSOPDE outperforms PSO, DE and several impoved algorithms. A soft sensor for acrylonitrile (AN) ratio based on PSOPDE and BP neural network is established. The empirical results show that this soft sensor has good generalization capability, and demonstrate the global convergence ability of PSOPDE algorithm.(2) Expansion and constriction operations are introduced into basic PSO algorithm and a novel particle swarm optimization (PSOEC) is proposed. The expansion operation encourages the search around the personal best position until then to find better position, which is a good measure to escape from local minimum; if the new position is not good enough, the constriction operation is performed to further search between the new position and current global best position so that better solutions can be found. The appropriate ranges of the parameters in PSOEC are given through many simulations of benchmark functions. In addition, PSOEC is compared with other popular algorithms, and the comparison results are satisfactory. PSOEC algorithm is applied to optimizing the weights and thresholds of BP neural network (BPNN), and a PSOEC-NN based soft sensor for estimating the outlet ammonia concentration of ammonia converter is established. The model results indicate that the soft sensor based on PSOEC-NN provides small errors on testing data and has strong generalization.(3) An improved parallel algorithm (HPSODE) is developed, in which PSO and DE are hybridized. In HPSODE, the population is divided into two subpopulations with irregular individuals in different sizes. The subpopulations search the problem space in parallel according to PSO and DE respectively, and then they are combined to recompose a new population for the next generation. A new PSO method which ignores current velocities in velocity update is introduced for the problem that individuals from DE subpopulation cannot update the velocity following basic PSO strategy. HPSODE enables the particles (individuals) be guided by two different ways in the parallel evolution; meanwhile, population decomposition and recomposition enhance the information exchange between the two subpopulations. The parameter settings of HPSODE algorithm are analyzed and discussed. HPSODE is used to optimize the hyper-parameters of least squares support vector regression (LSSVR). An optimizing model based on HPSODE-LSSVR is constructed, which describes the complex nonlinear relationship between the key operational parameters in the ammonia synthesis process and the ammonia net value. The experiment results demonstrate the effectiveness of HPSODE-LSSVR.(4) For the maximization problem of the ammonia net value in the ammonia synthesis process, a multi-population cultural differential evolution (MCDE) algorithm is proposed. In MCDE, the idea of multiple populations is introduced into the framework of cultural algorithm (CA). Each population separately searches the problem space based on the double-level evolutionary model (population space and belief space) of CA. The regular knowledge exchange between populations promotes the information sharing globally. To reduce the risk of falling into local minima, the concept of culture fusion is used and an adaptive mechanism of population diversity preservation is put forward to prevent the populations from converging prematurely. Eleven typical constrained optimization problems are adopted to verify the performance of MCDE algorithm. Based on the constructed HPSODE-LSSVR optimizing model, MCDE algorithm is employed to solve the problem of maximizing the ammonia net value in the ammonia synthesis process and the results validate the feasibility of MCDE algorithm.(5) The idea of co-evolution is added to CA, and a competitive co-evolutionary cultural differential evolution (CCCDE) algorithm is proposed. With the help of competition between multiple populations, a three-stage judgment criterion of population diversity is presented to find the populations which converge prematurely and the individuals who are inactive. This criterion helps the premature populations escape from local minima and activates the inactive individuals. The simulation experiments using the typical constrained optimization problems show that CCCDE algorithm is applicable to constrained optimization problems. This dissertation studies the application of CCCDE to the optimization problem with the objective to maximize the profit in the butane alkylation process, and the calculated results show the performance of this algorithm.(6) For a real-world ammonia synthesis process, the software of operational optimization system of ammonia converter is designed and developed. The software provides the following services:connecting with DCS system, online calculating outlet ammonia concentration, the suggestive flow rate of purge gas, the optimal operational parameters and their corresponding valve opening values, online displaying these results, and the alarm as well as generating and querying the electronic production reports. Good effects have been achieved since the software is put into operation.