Research Based On Neural Network In Predictive Control

Predictive control is a calculation control method proposed in the 20th century70s, which is made up of three parts: model prediction, moving horizon strategy and feedback compensation. At present, model predictive control based on linear systems are already quite mature both in theory and practical applications. But It still has a lot of works before finding a unified method directly for solving nonlinear systems, because of the complexity of its own special structure.Combing with neural network approach, this paper attempts to make in-depth study for three key elements of predictive control including predictive model, moving horizon strategy and feedback compensation. The main research study includes the following aspects:1. For the problems of slow convergence, easy to fall into local minimum using Gradient Descent (GD) method and large calculation, low robust using Newton Iteration (GI) method. This paper using PID Neural Network(PIDNN) combined with Levenberg-Marquardt (LM) algorithm and BP algorithm to establish the predictive model, which has a simple structure and dynamic self-adaptive. The method not only greatly reduce the computational capacity but also speed up the convergence and improve the model's accuracy.2. To reduce the calculation of controller for solving multivariable nonlinear systems. We proposed a new neural network predictive control algorithm. The algorithm introduced inverse dynamic control method and decomposition of control strategy, which has a fixed control structure and optimize performance index directly. Thus, it can reduce the storage space of variables and the complexity of calculation when deal with multi-step predictive control.3 To improve the revising performance, a self-adaptive neural network method is presented to reduce the error of the model. The method can get the error's changing trends by training the network's weights over the series accumulated error data, then more effective error correction model is obtained.