Nonlinear system identification for predictive control using continuous timerecurrent neural networks and automatic differentiation.

摘要

In this paper, a continuous time recurrent neural network (CTRNN) is developedto be used in nonlinear model predictive control (NMPC) context. The neuralnetwork represented in a general nonlinear state-space form is used to predictthe future dynamic behavior of the nonlinear process in real time. An efficienttraining algorithm for the proposed network is developed using automaticdifferentiation (AD) techniques. By automatically generating Taylorcoefficients, the algorithm not only solves the differentiation equations of thenetwork but also produces the sensitivity for the training problem. The sameapproach is also used to solve the online optimization problem in the predictivecontroller. The proposed neural network and the nonlinear predictive controllerwere tested on an evaporation case study. A good model fitting for the nonlinearplant is obtained using the new method. A comparison with other approaches showsthat the new algorithm can considerably reduce network training time and improvesolution accuracy. The CTRNN trained is used as an internal model in apredictive controller and results in good performance under different operatingconditions.