Research Of Improved Particle Swarm Optimization And Its Applications

Chemical industry is very significant in the national economy, which has respect to our daily life as well as other industries even national security. Nowadays, the complication of the chemical processes results in the inefficiency of the traditional optimization methods. Due to this, intelligent modeling and optimization methods begin to attract more and more attention. In this dissertation, for the complex coal chemical processes, such as ammonia synthesis, Texaco gasification, and methanol synthesis, the neural network (NN) based intelligent modeling and particle swarm optimization (PSO) methods are investigated, and the proposed novel methods are applied to soft sensor modeling in these processes. Furthermore, an intelligent technique is proposed to pallet grouping problem. The main results in this dissertation can be summarized as follows:(1)The features of popular intelligent optimization algorithms are discussed as well as the principle of the particle swarm optimization and soft senor modeling. The development and improvement of PSO are underlined. Moreover, the processes of ammonia synthesis and methonal industy are reviewed, and then the material feeding especially the pallet grouping problem is outlined. Finally, the application of intelligent optimization algorithm in these areas is introduced.(2)In order to overcome the premature convergence of PSO algorithm, a self-government particles swarm optimization (SGPSO) is proposed. In the SGPSO, the update of particle position not only depends on the particle personal best position currently and the swarm global best position found so far, but also is related to the local best information found in the previous experiments, thus greatly improving the optimization capability. The simulation results in the benchmark function tests indicate that the proposed SGPSO and RSGPSO algorithms are superior to the standard algorithm for convergence speed and acurracy. The SGPSO and RSGPSO are integrated with BP neural network, and SGPSO-NN based soft sensor model for gasifier temperature is established. The results of the model show that the soft sensor based on SGPSO-NN model provides small testing error and has good generalization capability, which is very suitable for measurement of gasifier temperature in the real-world application.(3)The learning factors directly influence the optimizatioin capability of PSO, thus two random tuning rules of learning factors, i.e. radical tuning and conservative tuning, are presented. Based on the rules, the random learning factor particle swarm optimization (RLFPSO) is proposed. In order to further improve the performance of RLFPSO and reduce the risk of being trapped to local optima, two random learning facto particle swarm optimization algorithms with chaos (RLFPSOC1and RLFPSOC2) are proposed. In the two improved algorithms, the ergodicity of chaos is introduced at the prophase and anaphase of evolution, respectively. The performance of RLFPSO, RLFPSOC1and RLFPSOC2is tested by benchmark functions, and then compared with other methods. The test results reveal that RLFPSO outperforms PSO, and two RLFPSOC algorithms have further improvements on the basis of RLFPSO. Finally, RLFPSO and RLFPSOC based neural networks are used for soft-sensing the methanol conversion rate, and compared with PSO-NN and CenPSO-NN. The comparison results indicate that the prediction models of methanol conversion rate based on RLFPSOC1-NN and RLFPSOC2-NN have better prediction capability than other methods, and can provide accurate estimation of the methanol conversion rate.(4) The traditional PSO algorithm has the drawback of falling into local minima easily. In order to improve the performance, a novel PSO algorithm——Velocity Share Historical Best Particle Swarm Optimization (VSHBPSO) is proposed. The basic idea of VSHBPSO is that the particle is attracted by the global historical best position searched in the former experiment as well as the current global best particle. The simulation studies on VSHBPSO and two variants VRSHBPSO, AVRSHBPSO are executed, and are compared with PSO. The simulation states that VSHBPSO and two variants are applicable and effective for both the low and high dimensional optimization problems. Meawhile, AVRSHBPSO performs best among the algorithms. The AVRSHBPSO-NN soft sensing model is applied to estimation of the ammonia concentration at the converter outlet. The experiment results verify the reliability and effectiveness of the proposed model, which is capable of providing the instruction in the real-world production.(5) A new modeling method is presented to describe the pallet grouping problem, whose complexity is analyzed. a new learning algorithm, which is defined as Learning Algorithm Only from Excellent-Pbest, Gbest and Gpbest, for short LAOE～PGpG is developed to solve the pallet loading problem. A new encoding scheme namely0-1encoding is adopted in LAOE～PGpG, and several Inter-learning schemes are developed to update the individual. Two application cases with different dimensions are employed to verify the performance of the proposed algorithm, and the optimal pallet grouping is given. The results indicate that the proposed algorithm outperforms the compared methods, and can be integrated into the logistics management system for realizing the automatic collection and distribution.