Nonlinear model identification and adaptive model predictive control using neural networks (original) (raw)

This paper presents two new adaptive model predictive control algorithms both consisting of an on-line process identification part and a predictive control part. Both parts are executed at each sampling instant. The predictive control part of the first algorithm is the Nonlinear Model Predictive Control strategy and the control part of the second algorithm is the Generalized Predictive Control strategy. In the identification parts of both algorithms the process model is approximated by a series-parallel neural network structure which is trained by a recursive least squares (ARLS) method. The two control algorithms have been applied to: 1) the temperature control of a fluidized bed furnace reactor FBFR) of a pilot plant and 2) the auto-pilot control of a F-16 aircraft. The training and validation data of the neural network are obtained from the open-loop simulation of the FBFR and the nonlinear F-16 aircraft models. The identification and control simulation results show that the first algorithm outperforms the second one at the expense of extra computation time. instead of the first principles model reduces the computational burden when an accurate model is required for real-time control implementation [23], [25]. This is because a nonlinear discrete NN model of high accuracy is available immediately after or at each instant of the network training process. The accuracy of a nonlinear NN model used to identify a physical process instead of a first principles model of a physical process [9]-[12], [21] depends on the model structure of the system to be identified by the NN; the NN architecture; the selection of the inputs to the NN; the NN training algorithm; how the model will be used in an adaptive control algorithm and the structure of the closed-loop system i.e. the arrangement of the system, the NN model and the adaptive controller. Over the years, different architectures of NNs have evolved [19], [26] and it has been shown that a multilayer perceptron (MLP) NN with one hidden and output layers is capable of approximating any continuous and/or nonlinear functions reasonably well with an arbitrary degree of accuracy. The most widely used NN architecture for dynamic system modeling is the dynamic feedforward NN (DFNN) [23], [27]-[29]. The use of recurrent NN (RNN) for modeling nonlinear dynamic systems has also been reported [25], [30]-[32]. RNNs are more powerful than DFNNs because they contain the basic FNN structure with feedback connections from the output to the input layer via a state layer (the so-called Jordan network [33]) or from the output unit to the input unit of the hidden layer via a context layer (the so-called Elman network [34]). However, training these networks presents difficulties due to their feedback structures [35]. The usual methods for training RNNs are: 1) the real-time recurrent learning (RTRL) [36] which is an unrestrictive on-line, exact, stable but computationally expensive method for determining the derivatives of the state functions of a dynamic system with respect to the internal parameters of the system; and 2) the backpropagation through time (BPTT) [37] where the network is unfolded into a multilayer feedforward network that increases by one at each time step with growing memory requirements [33], [34]. Both RTRL and BPTT are variations of the backpropagation (BP) algorithm [38] which is a basic gradient descent algorithm [39], [40].