Dynamic modeling and model predictive control using generalized perceptron networks.

摘要

Based on recurrent neural network architectures and classical system identification approaches, this thesis presents a Recurrent Neural Network (RNN) modeling approach for developing nonlinear dynamic models. First, Generalized Perceptron Networks (GPN) is defined, which include Multilayer Perceptron Networks (or Multilayer Feedforward Networks, FFN) and Recurrent Neural Networks (RNN). Using ordered derivatives, a parallel neural network training algorithm called Universal Delta Rule is derived for the Generalized Perceptron Networks. The Universal Delta Rule unifies the delta rule, the generalized delta rule (or backpropagation), and the back-propagation through time. As an alternative, the proposed modeling approach is found to have significant advantages over the conventional feedfoward network (FFN) modeling approach. For example, the proposed modeling approach can lead to a much better long-range predictive model than the conventional approach, and it is less sensitive to measurement noise. A neural network initialization method is developed because presently there is no appropriate approach for initializing a neural network other than using a set of random weights. The initialization method first uses a linear dynamic model to learn from the data and then uses Partial Least Squares for converting one coefficient matrix of the linear model into two initial weight matrices of the neural network model. Furthermore, a Neural Network Hammerstein modeling technique is developed. In particular, a static neural network is integrated with a linear dynamic model in series. Because the model structure is simplified, the computation time required to obtain a good nonlinear model can be reduced. The proposed neural network Hammerstein (NNH) modeling approach can also account for the difficulty of data collection. Using Neural Network Hammerstein Modeling technique, nonlinear steady state data can be used for developing a dynamic model in the situation where the dynamic data does not contain sufficient nonlinear information. Finally, a general Neural Network Model Prediction Control (NNMPC) algorithm is developed using Non-Linear Programming (NLP) techniques. In particular, a Feasible Sequential Quadratic Programming (FSQP) technique is used for solving the NNMPC control problem.