The application of neural networks to excitation control of a synchron
ous generator is considered here. A radial basis function (RBF) networ
k was constructed using a hybrid training algorithm, combining linear
optimization of the output layer weights with singular-value decomposi
tion, and non-linear optimization of the centres and widths using seco
nd-order gradient descent BFGS. The Jacobian of the RBF network was ca
lculated to provide instantaneous linear models of the plant, which we
re then used to form linear controllers. Generalized minimum variance,
Kalman, and internal model control schemes were implemented on an ind
ustry-standard VME platform linked to a network of Inmos transputers,
and the performance of the neural models and neural control schemes we
re investigated on a 3 kVA laboratory micromachine system. Comparison
was made with a self-tuning regulator, employing a generalized minimum
variance strategy. The results presented illustrate that not only is
it possible to successfully implement neural controllers on a generato
r system, but also their performance is comparable with a benchmark se
lf-tuning controller, while avoiding the significant supervisory code
needed to ensure robust operation of the self-tuning controller. (C) 1
997 Elsevier Science Ltd.