Modeling and Simulating for Artificial
Neural Network-Based Direct Torque
Control for Induction Motor Drive
Wang Qunjing Chen Quan Jiang Weidong Hu Cungang
(Hefei University of Technology Hefei 230009 China)
Abstract As a prospective control scheme—direct torque control (DTC) has a control error
caused by the time delays required for the lengthy computations. However, the neural network,
with its parallel computation and robust capabilities, offers a promising means to minimize the
error. This paper presents an artificial neural network-based (ANN) DTC scheme for an induction
motor drive. The neural networks used in this paper are fixed-weight networks and supervised
networks. According to the features of DTC, the local training strategy is adopted in this paper.
Finally, computer simulations of the designed DTC system are presented and discussed. The
experimental results indicate that ANN-based DTC may be a feasible alternative to realization of a
high performance DTC system.
Keywords: Direct torque control, artificial neural network, time delay, stator-flux observer,
voltage vector selector
1 Introduction
DTC offers attractive performance compare to vector control in terms of fast torque response, no
need of coordinate transformation, robust against parameter variation, no need of a PWM
controller and current regulators, but also exists some drawbacks, such as torque and flux ripple,
the harmonic content of the stator voltage and current
[1-3]
. Although involving complicated
calculations can overcome these shortages, it is difficult to implement on some digital signal
processor (DSP) real-timely. Moreover, the lengthy computations affect hardware performance
largely
[4]
. As a result, designers also don’t exploit the capability of modern power switching
devices — high switching frequency
[5-6]
.
The neural network (NN), with its special architecture and inherent computation method, offers a
promising alternative to realization of a highperformance power drive
[7]
. Actually, a neural model
is mathematically represented by a basis function and an activation function. The difference
between neural networks lies in weights and biases. In terms of how The weights and biases are
obtained, the neural network can be classified into three categories: ①
fixed-weight. ②
supervised network.
③ unsupervised network. The characteristic of supervised network is that
weights and biases are trained by a learning mechanism, and the back-propagation (BP) generally
learning algorithm is used to design thesupervised network. In the fixed weight network, the
weights and biases are pre-computed from training data. This kind of network usually is used in
the plant with definite operation law.In this paper, an artificial neural network-based DTC
technique is described. Since DTC of induction motor is a dynamic (there are integrators),
recurrent(there are hysteresis comparators) nonlinear and eight input variables (V
a
, V
b
, V
c
, i
a
, i
b
, i
c
,
T
*
, ψ
*
) system, it is difficult to train the NN-based system with global. With global training, the
training of an individual subnet will be influenced by input/output patterns of other subnets. For
simple and fast design, the individual training method is adopted in this paper. In the following,
section
Ⅱ provides the proposed neural network-based DTC. Programmed computer NN-based