Application of Artificial Neural Networks - Digital Signal Processing in Power System Protection and Control - page 264

• Fault instant: 20:2:40 ms (various angles of fault inception),
• R/X ratio of the primary system side: [0.07, 0.08, 0.09, 0.1],
• Time constant T s of the CT secondary side during saturation: [0.3, 0.5, 1, 1.5, 2,
3, 4, 5] ms.
With the above mentioned options a number of 4224 simulation cases have
been generated. Further signal processing and testing of the CT detection and
correction methods was performed in Matlab environment.
Thorough investigations on efficiency of designed ANN-based CT saturation
detector are described in [ 30 ]. Its operation was compared with chosen deter-
ministic CT saturation identification scheme based on calculation of the 3rd
derivative of CT secondary current [ 15 ]. Despite small size of the ANN applied
(14 neurons) all starting and ending points of CT saturation intervals were properly
detected. In addition, no unnecessary excitations were observed at the beginning of
fault, which took place when the non-AI method was applied. The latter feature of
the ''classic'' method was an effect of the fact that the method responds with high
peaks in 3rd derivative to any sudden change in current waveform. The method
could react properly (with visible impulses) at the beginning of saturation, but had
problems with detecting saturation endings, when the CT turns to unsaturated
operation mode with much smoother waveform change.
Below the features of designed ANN-based CT saturation compensators are
presented with higher attention. The investigations have revealed that the per-
formance of neural compensation units was almost perfect when the network
ANN2 (trained with a large number of EMTP cases) was applied. Although only
one-third of all patterns were shown to the ANN during training, the robustness
and ability to knowledge generalization were more than enough to enable correct
operation of the compensator for most cases. The correction errors increase a little
bit only for high amplitudes of the primary current for which the internal ANN
signals may fall very deep into the regions where the neurons' activation functions
are heavily saturated. The network trained on narrow base of patterns (ANN1) was
somewhat worse than the ANN2, however the results of operation were not far
away from the ideal ones. The network ANN3 (without current scaling) performed
badly, which can be seen in Fig. 12.21 . The quality of primary CT current
reconstruction with use of ANN2 is several times higher than with other ANNs.
Definitely the worst performance was observed on average for the ANN3 solution
(without scaling), which confirms the significance of properly selected pre-pro-
cessing of the ANN input signals. It was observed that all the ANNs perform with
higher accuracy when the ratio of CT saturation time constant T s to the sampling
period T S is higher and when the time constant of decaying DC component in CT
primary current is higher.
Additional analyses have been done for the network ANN4 (activation function
satlins) that was not compared in broader sense with other ANNs. The aim of the
studies was to check the influence of the simplified activation function on
the primary current reconstruction accuracy. As seen in Fig. 12.22 , the operation
of compensator ANN4 is basically good (the CT primary current waveshape is