Cg. Windsor et al., REAL-TIME ELECTRONIC NEURAL NETWORKS FOR ITER-LIKE MULTIPARAMETER EQUILIBRIUM RECONSTRUCTION AND CONTROL IN COMPASS-D, Fusion technology, 32(3), 1997, pp. 416-430
The plasma position and shape on the COMPASS-D tokamak have been contr
olled simultaneously with a 75-kHz bandwidth, hard-wired, real-time ne
ural network. The primary network operates with lip to 48 selected mag
netic inputs and has been used in the vertical position control loop t
o control the position of the upper edge of the plasma at the radios o
f a reciprocating Langmuir probe and to keep this constant during a pr
ogrammed shape sequence. One of the main advantages of neural networks
is their ability to combine signals from different types of diagnosti
cs. Two coupled networks are now in use on COMPASS-D. A dedicated soft
-X-ray network has been created with inputs from 16 vertical and 16 ho
rizontal camera channels. With just four hidden units, it is able to a
ccurately determine three output signals defining the plasma core radi
us, vertical position, and elongation. These signals are fed to the pr
imary network along with selected magnetic inputs and four poloidal fi
eld coil control current inputs. The core data are expected to help ch
aracterize the equilibrium by providing information on the Shafranov s
hift and gradient of elongation, related to the equilibrium parameters
beta(p) and l(i). This network, with 15 hidden units, is able to defi
ne 10 outputs capable of giving a parameterized display of the plasma
boundary. This paper describes results from several networks trained o
n various combinations of inputs with (a) simulated inputs and output
values, where the precision of the network can be tested; (b) experime
ntal inputs and calculated output values, where operational precision
can be tested; and (c) hardware networks, where real-time performance
can be tested. The results confirm that the neural network method is c
apable of giving excellent precision in tokamak boundary reconstructio
n but that the necessary accuracy in the experimental inputs for this
task is not easily achieved.