Research On Particle Swarm Optimization Algorithm And Dynamic Neural Network Modeling Prediction

With the rapid development of modern science and technology, the complexity of the control system is getting higher and higher, and the scale is also increasing. Thus it has been unable for industrial controlled plants to be described by a precise mathematical model. It is difficult for conventional control approaches to meet human’s requirements on the automatic control systems. It is a key problem how to model unknown nonlinear delay system quickly and accurately, and implement the efficient and accurate predictive control. As an emerging optimal approach in the swarm intelligence field, particle swarm optimization algorithm has shown the excellent performance in model parameter optimization. Therefore, the focus of this paper is the particle swarm optimization algorithm. And then it is combined with fuzzy C-means clustering intelligent methods which are employed to achieve large-scale dynamic neural network offline or online parameter optimization. After that, we construct unknown nonlinear delay system identification approaches and predictive control frameworks. The research topics include the following three aspects:(1) The online model parameter optimization based on Gaussian particle swarm optimization is proposed. Particle swarm optimization algorithm has the advantages of fast calculation and strong scalability, whose structure is simple and independent of continuously differentiable constraints of objective functions. On the basis of Gaussian functions and chaotic mappings, a novel Gaussian particle swarm method is presented, which could adjust particle swarm searching adaptively. It could maintain the global search capability and the late continued local search capability. Hence, the "prematurity" problem is solved. Further, the robust control theory is adopted to analyze the stability conditions of the Gaussian particle swarm optimization. Through various types of standard function test, the proposed Gaussian particle swarm optimization (GPSO) method significantly improves the accuracy. On this basis, the GPSO approach is used for echo state neural network online parameter optimization problem. The network parameters are adjusted online based on the input data, so the drawbacks of offline training can be overcome. Nonlinear delay system and chaotic time series data are employed to verify the effectiveness of the proposed method.(2) The large-scale optimization based on fuzzy C-means cluster and particle swarm optimization is proposed. To solve the problem of "curse of dimensionality" faced by traditional optimal methods, the particle swarm optimization algorithm and fuzzy C-means clustering method are combined to complete multi-group cooperative particle swarm optimization algorithm. First of all, considering that the fuzzy C-means clustering algorithm is sensitive to the initial points, a two-stage fuzzy clustering algorithm with a linear assignment strategy is presented for the initial point selection, which enhances the stability of the whole algorithm. Then with the part of some prior samples, a novel single point approximation and the weighted semi-supervised fuzzy C-means algorithm is proposed. The validity of the proposed approach is verified by using the machine learning, image segmentation standard database and the actual wetland remote sensing data to verify the validity of the method. On this basis, combined with trust region methods, the modified fuzzy C-means algorithm is adopted to construct multi-group structure, and to decompose the large-scale parameter optimization problem into sub-problems. Each group collaborative optimizes and shares the information. The large-scale standard test functions are used to verify the optimization capability of the proposed method.(3) A dynamic feedforward neural network predictive control based on particle swarm optimization is proposed. In order to improve the predictive accuracy, a new dynamic feedforward neural network structure is presented. The added dynamic delay operators not only enhance the ability of the dynamic expression, but also identify the pure delay time in the system. In addition, GPSO is adopted for the above parameter optimization of the mentioned dynamic feedforward neural network. It could speed up the network convergence. The stability of the combined model is also analyzed deeply. Furthermore, when the neural network consists of higher number of neurons, the cooperative particle swarm optimization is employed to the large-scale neural network parameter optimization. The proposed neural network model as an identifier and predictor is used in Smith predictive control with dual controllers and multivariable constrained model predictive control structure, respectively. The simulation examples demonstrate that the proposed control method can perform effectively on the unknown nonlinear delay system identification and predictive control.