Artificial Intelligence: Summary and Hybrid Schemes - Digital Signal Processing in Power System Protection and Control - page 309

Fig. 15.2 Hybrid schemes
for OS detection: a fuzzy-
neuro, b neuro-fuzzy

(a)

FIS
output

Other
input
signals

OS
det.

ANN

(b)

ANN
output

Fuzzy
Inference
System

Other
input
signals

OS
det.

be found in the literature [ 2 , 8 , 13 ]. The ANN may learn the parameters from the
fuzzy system. This can be either performed offline or online while the fuzzy
system is applied. The neural network can also provide fuzzy sets or even fuzzy
rules for a fuzzy system by training on examples [ 8 ]. Rule weights can also be
determined by a neural network [ 13 ].
The investigation on application of various AI techniques to out-of-step detec-
tion/prediction of synchronous machine has been presented in [ 16 ]. The basic
protection scheme was employing a single neural network. Relatively compact
structures (ANN with 9-9-1 neurones) were achieved which implies that they may
be easily implemented on-line on traditional signal processors. The developed
neural OS detectors proved to be robust against changing power system conditions
(different fault types) and may be used for synchronous machines of various ratings.
Further optimization of the protection scheme has been done by introducing
some concepts of FL, thus making the scheme an AI hybrid. Two cascade-type
neuro-fuzzy configurations have been examined. The fuzzy inference system (FIS)
module was used as a pre-processor of natural power system signals (the FIS
output was further processed by an ANN) or as an interpreter of the ANN outputs
(Fig. 15.2 ). Apart from the output signal of the pre-processing unit (FIS or ANN,
respectively) some additional inputs may also be added. Their character, number
as well as time position are to be examined in order to become the most optimal
operation of the scheme under investigation.
The EMTP simulations confirmed superior performance of the hybrid FIS-ANN
scheme. Additional processing of FIS output by an ANN brought about significant
narrowing of the uncertainty area around tripping threshold. The ANN outputs are
closer to 0 and 1 and thus more reliable decision could be issued. As a result the
hybrid system became more robust than the other single-technique schemes
developed. The developed hybrid OS detectors were able to recognize coming OS
conditions 100-1,000 ms before they actually took place (a prediction action) thus
providing a chance to undertake proper preventive operations to maintain system
stability and to avoid presumable stresses to the protected machine.