Fast Model Predictive Control Algorithm Based On Neural Networks

Model predictive control(MPC)has been widely used in the petroleum,chemical and other process industries due to its excellent control capabilities in complex multi-variable systems.In essence,the ability to handle constraints explicitly of MPC mainly arises from solving the quadratic programming(QP)problem.Although traditional numerical methods for QP problems have been applied widely,they mostly involve operations such as matrix inversion and factorization.Therefore,traditional numerical methods occupies more resources,have lower real-time performance and are relatively complicated to implement.Currently,most of the MPC algorithms are implemented via high-performance computing equipment such as PCs or workstations,but difficult to implement on embedded platforms with.limited resources.And it is difficult to apply most of them to the fast sampling process.One promsing approach to handle these problems is replacing traditional numerical methods with recurrent neural networks for QP problems.Neural networks have distinctive features of lower resource consumption,natural parallelism and simple implementation,thus providing favorable conditions for implementing complex MPC algorithms on embedded platforms.With the ultimate aim of fast MPC algorithm,this paper studies the topics of the optimization algorithm and embedded MPC.The main contributions are as follows:(1)In view of the problem of traditional numerical methods occupying more resources and complicated to implement,a model predictive control algorithm based on discrete neural network is applied and improved.This algorithm uses discrete simplified dual neural networks to solve QP problems.It has the characteristics of lower resource consumption.Particularly,it is flexible and easy to implement.The various discretization methods are analyzed and compared.And the advantages and disadvantages of the discrete neural networks are pointed out.The algorithm proves well performance in simulation experiments.(2)Aiming at the problem of large computational burden which exists in the traditional predictive control algorithm for zone control,an embedded and layered model predictive control algorithm based on discrete neural network is proposed.The large computational burden is reduced significantly by layered optimization which can conveniently transforms zone control into setpoint control.The neural network for quadratic programming is adopted to effectively solve the problem of limited resources on embedded platform.An analytic method which is more efficient compared with the neural network method is proposed for steady-state optimization in SIMO processes.The system is implemented and verified on the embedded platform.Experimental results show that the embedded,layered and neural-network-based model predictive control system designed in this paper possesses good control performances and practical significance.(3)For the convergence and limited solution speed of discrete neural networks,a fast model predictive control algorithm based on FPAA analog neural network is proposed.To tackle the signal limitation of FPAA analog circuits,the translation and scaling methods are proposed,which provide a theoretical basis for analog circuits to realize continuous neural networks.Therefore,a comprehensive software and hardware implementation method of model predictive control based on FPAA analog neural network is developed.Experimental results verify the effectiveness and rapidity of the algorithm.