Application of Artificial Neural Networks - Digital Signal Processing in Power System Protection and Control - page 260

Fig. 12.15 Block scheme of
the CT saturation detection
and compensation
• Correction is based on information extracted from the secondary current during
the non-saturation interval, which aims at determination of the fundamental and
DC components of the primary current [ 23 ].
• CT saturation detection and/or correction is done with use of the Artificial
Neural Networks, trained with the simulation cases of CT transients [ 3 , 17 ,
39 , 44 ].
Below an approach based on application of artificial neural networks is
described. The performance of ANN-based decision units depends to a large
degree on the cases used to train the networks. With the variety of possible cases it
may be hard ensuring that all possible CT characteristics, residual fluxes, loads and
primary current parameters are covered. The solutions described in [ 17 , 39 , 44 ] are
quite promising, however, they are based on recurrent networks (prone to insta-
bility under some circumstances, [ 39 ]) or a set of ANNs designed for sublevels of
measured current amplitudes [ 3 ], which the authors find as non-optimal.
The basic idea of handling the CT saturation problem adopted here is based on
splitting the task into two subtasks, namely saturation detection and correction of
the distorted secondary current (Fig. 12.15 ). When the CT is not saturated (i.e. the
detection block has not detected it) the correction block is not activated. Starting
from the point of saturation beginning the procedure of secondary current cor-
rection is activated. The procedure is operative until the CT goes out of saturation.
The process is repeated when the detection block affirms saturation beginning
again.
Preparation of the neural networks for both subtasks, choosing proper size of
the networks and their input vectors as well as training and testing of the solutions
are outlined. The neural CT saturation detection and correction units have been
tested with EMTP-ATP-generated current signals. Special attention is also paid to
ANN implementation problems. Practical implementation of the neural schemes in
real-time with use of a signal processor is also described.