Research On Predictive Control Method Based On GAPSO Algorithm And RBF Neural Network

Industrial processes usually has characteristics with strong nonlinearity,time-varying, time-delay and uncertainty. It is difficult to achieve expected targetby using traditional PID control methods and modern control theory. Predictivecontrol is a kind of optimization control method developed from the industrialpractice. It has advantages of strong tracking performance, good control effectand strong robustness, so it has been widely used. In predictive control theestablishment of the prediction model is the basis, and rolling optimization is themost important characteristic of predictive control. But for the complexnonlinear objects, it is difficult to identify structure and parameters of the model.At the same time there has great difficulty in building forecast model. The problemof solving nonlinear control variable is also very complex in the process ofrolling optimization.Aiming at these problems, the modeling method of nonlinear system andoptimization problems of control variable in predictive control are researched. Anonlinear predictive control method based on improved particle swarmoptimization and radial basis function neural network is put forward. Thismethod is applied in the control system of continuous stirred tank reactor and achieves good control effect. Concrete research works are as follows:Based on the problems that particle swarm optimization algorithm (PSO) iseasy to fall into local extremum and search performance excessively relies on theparameter settings. The main parameters of particle swarm optimization(PSO)algorithm are improved firstly. Linear decreasing inertia weight method andasymmetric linear variation method for learning rate are adopted to balance theglobal and local search ability of the algorithm. Genetic algorithm and chaoticoptimization algorithm are introduced into the PSO algorithm. Genetic algorithmand particle swarm optimization(GAPSO), and chaotic particle swarmoptimization(CPSO) algorithm are proposed. In GAPSO algorithm crossoveroperator and mutation operator of genetic algorithm are embedded in the processof PSO. The strategies of adaptive crossover and adaptive mutation are used toenhance the ability of escaping from local optima. In CPSO algorithm, chaoticsequences are used to initialize the particle position, so PSO algorithm cansearch optimization from a good initial value. At the same time a part of particlesare optimized by chaotic algorithm in iterative process to overcome thepremature convergence and slow convergence speed of PSO algorithm. Thesimulation results for standard test functions demonstrate the good optimizationperformance of CPSO algorithm and GAPSO algorithm.In view of the problem that it is difficult to set model parameters of radialbasis function neural network. A modeling method based on RBF neural networkis put forward, which optimized by chaotic particle swarm optimization. In this method centers and width of basis functions, weights and thresholds of RBFneural network are optimized by CPSO algorithm. The proposed method isused to establish model of the nonlinear systems, simulation results arecompared with RBF neural network, BP neural network and support vectormachine, which shows that the modeling method based on CPSO-RBF neuralnetwork has higher precision.A predictive control method for the nonlinear systems is proposed, which isbased on GAPSO algorithm and RBF neural network(GAPSO-RBF). Thismethod utilizes CPSO-RBF neural network to built nonlinear predictive modeland forecast the output values of systems. By combining offline training andonline training of RBF neural network, it realizes the on-line correction of modelparameters. The optimal control values of the nonlinear systems are obtained bythe rolling GAPSO optimization algorithm. The simulation results demonstratethat the predictive control method based on GAPSO-RBF neural network hasexcellent control performance.Finally the predictive control method based on the GAPSO-RBF neuralnetwork is applied to control reactant concentration of continuous stirred tankreactor(CSTR). RBF neural network is used to establish prediction model andpredict the concentration of reactant. Genetic particle swarm optimizationalgorithm is adopt to solve performance indicators. A experiment for CSTRcontrol system is taken. The simulation results show that the predictive control method based on GAPSO-RBF neural network can track reference input ofreactant concentration quickly and effectively.