Fuzzy Neural Network Model Predictive Control Based on Dynamic Partial Least Squares Framework

An improved fuzzy neural network model predictive control (MPC) method based on dynamic partial least squares (DPLS) framework is proposed for the control of highly coupled nonlinear systems. Firstly, the PLS framework of dynamic neural network is established by adding neural network into the traditional DPLS framework. Secondly, the internal model of latent variable space is established by using the Dung Beetle optimized fuzzy neural network algorithm, and the parameters of the model are optimized by the improved adaptive LM algorithm to obtain the accurate prediction model. Furthermore, the adaptive learning rate gradient descent algorithm is used to optimize the control quantity of the latent variable loop. Finally, the PH neutralization titration process was used for the experiment. The results show that the fuzzy neural network model predictive control based on dynamic PLS framework has good control performance.

I. Introduction

As industrial production becomes more and more automated, production processes become more and more complex, especially in chemical, electric power, steel and other fields. These industrial processes tend to be nonlinear, strongly coupled, high dimensional and possibly non-square. As these production processes are often accompanied by the generation of a large amount of data during operation, and with the popularization of the Internet, digital technology has made great progress in the industrial field, data-driven methods have received extensive attention from the academic community^[1]. Model predictive control (MPC) is the most widely used advanced control. It can find the optimal control quantity by modeling, feedback correction and rolling optimization. Researchers have demonstrated the superiority of MPC over traditional control methods, which have been widely used in industrial and aerospace fields^{[2], [3].} However, when MPC is solving complex problems, the estimation accuracy of the model greatly affects the control effect, and a large amount of time cost will be generated in rolling optimization due to the high system dimension.