Design of fuzzy neural network predictive controll

  • Detail

Design of fuzzy neural network predictive controller for superheated steam temperature

1 introduction

because the boiler is a typical nonlinear, slow time-varying and complex thermal system that meets the long-term and harsh quality requirements of vanadium battery, the traditional PID control parameters are difficult to set and lack the adaptive ability to dynamic changes. The general fuzzy control is a control system based on fuzzy rules. Due to the influence of nonlinear, time-varying and random interference in the control process, the fuzzy control is not suitable and incomplete, which affects the control effect. Neural network has opened up a broad application prospect for its application in the field of control with its nonlinear processing ability, self-learning, self-organization and self-adaptive ability [1~3]. In order to solve the nonlinear and large time-delay of the system, this paper combines neural network with predictive control, establishes the neural network model according to the input and output data of the controlled object, uses an improved hierarchical genetic algorithm to train the weight and structure of the neural network at the same time, and carries out multi-step recursive prediction, realizing the accurate neural network model prediction of the nonlinear and large time-delay complex system. On this basis, this paper combines neural network and fuzzy logic control to design a fuzzy neural predictive controller, that is, according to the output of the predictive model, the fuzzy rules of the predictive deviation and control quantity are obtained, and the structure of the fuzzy logic controller is realized through the neural network. At the same time, in order to avoid BP algorithm falling into local minima in the training process, genetic algorithm is used to train the fuzzy neural network, so as to achieve the best control effect. Finally, the controller is used as the main controller of the feed-forward feedback cascade control system of boiler superheated steam temperature, and the secondary controller adopts a two degree of freedom PID controller, in order to further enhance the robustness of the system. The simulation shows that the cascade control system with a certain proportion of change has strong robustness, real-time performance and anti-interference ability, and maintains good control performance and operation effect, which is effective in engineering

2 design of feedforward feedback cascade control system

2.1 mathematical model of main and auxiliary objects

Figure 1 is the block diagram of feedforward feedback cascade control system, R1 (s) and R2 (s) are the auxiliary controller and main controller respectively; G1 (s) and G2 (s) are the transfer functions of the leading region and the inert region respectively; WF (s) is the transfer function of load D feedforward compensator

in the formula, K1 and K2 are the amplification coefficients of the leading region and the inert region respectively; T1 and T2 are the inertia time of the leading zone and the inert zone respectively; N1 and N2 are the order of the pre conduction region and the inert region, respectively

transfer function of mathematical model of superheated steam temperature and flow feedforward compensator

where TF1 and TF2 are the differential coefficient and inertia time of feedforward compensator respectively

2.2 design of main and auxiliary controllers

2.2.1 main controller

the main controller is fuzzy neural predictive control based on neural network model, which is composed of two neural networks, the former is neural network predictive model, and the latter is fuzzy neural network controller

(1) the neural network prediction model trained by the improved hierarchical genetic algorithm

assumes that the input and output characteristics of the time-delay nonlinear system are described by the following general form of time discrete equations

where u (k) and Y (k) are the input and output of the control object respectively, m and N are the order of the system input and output respectively, D is the system time delay, and F is an unknown nonlinear mapping of rn+m+1 → R

if the neural network model is

, where NN is the known mapping of rn+m+1 → R, and W is the weight of the neural network model

the neural network prediction model is

the multi-step prediction of the neural network model is actually d-times mapping when rn+m+1 → R

at time k, due to y (k+d-1) in the prediction model (clear pattern 3), Y (k+d-n) is the future output value, which cannot be measured, so the neural network model prediction value YM (k+d-1) was used. After the scandal of tampering with product data broke out by Japanese nonferrous metal giant Mitsubishi Integrated Materials Co., Ltd. (hereinafter referred to as Mitsubishi materials), YM (k+d-n) is used to replace the corresponding output value, so the multi-step prediction of k-time neural network model based on one-step prediction is realized

the hierarchical genetic algorithm used here is a new training method, which is based on the following facts: biological chromosomes are composed of changes in genes, and genes are arranged in a certain hierarchical manner, which are divided into sequence genes and structural genes. Therefore, the recursive algorithm (HGA) can represent the topology and parameters of the solution at the same time, that is, the weights and structures are coded at the same time during training, and the weights and structures are optimized at the same time through global search, which overcomes the defect of the structure pre-determined in the traditional neural network training algorithm, and the neural network model established can more accurately approximate the real system, providing a guarantee for the high-precision predictive control of nonlinear systems. In addition, in many optimization problems, finding the optimal solution is not

Copyright © 2011 JIN SHI