DETECTION OF HIGH IMPEDANCE FAULTS USING ARTIFICIAL NEURAL

NETWORKS

DETECTION OF HIGH IMPEDANCE FAULTS USING ARTIFICIAL NEURAL

NETWORKS

ByIbrahim El-Amin Mohammad

H. Al-Mubarak

ByIbrahim El-Amin Mohammad

H. Al-Mubarak

May 2003

Problem Definition High Impedance Fault (HIF) is

a fault on primary distribution (PD) system that cannot be detected by conventional O/C protection

95% of faults on PD feeders occur on O/H lines

Undetected HIFs may result in public hazard and property damage

Primary motivator to design HIF detectors is public safety

Problem Statement and Paper Objective

O/C relays are set to operate for currents between 125-200% of the normal load current

HIFs draw currents in the range of 0-100 A

Objective of paper is to develop an ANN based HIF detector that can also: Locate the HIF Distinguish HIFs from normal

switching events Identify the faulty phase

Solution Approach

Feeder Simulation with EMTPFeeder Simulation with EMTP

Data Scaling and Preprocessing

Data Scaling and Preprocessing

ANN Trainging and Testing

ANN Trainging and Testing

Fault Diagnosis

Fault Diagnosis

Feeder Simulation

EMTP is used because It can handle switch closing and opening (ideal for

transient analysis or fault simulation) It can simulate unbalanced systems (e.g. single line

to ground faults)Simulation is for 5 cycles (83.33 ms) at a

sampling rate of 0.2778 ms300 samples/phase for each case, i.e. 1800

samples to represent the 3 phases of the current and voltage waveforms

Feeder Simulation (Cont’d)

Feeder Simulation (Cont’d)

Phase-C Current Waveform for Load-4 and C1 Switching

Designing and Testing the ANN

Target Possible Values (Decimal Output)

Target # 1 (Fault Detection) 0, 1

Target # 2 (Fault Location) 0, 1, 2, 3, 4

Target # 3 (Event Type) 0, 1, 2, 3, 4

Target # 4 (Faulty Phase) 0, 1, 2, 3

ANN Targets

Comparison Between the Three ANN Designs

Design No.No. of EMTP Cases for ANN

No. of EpochsTraining Testing

Generalization Check

1D1F

39 13 48414

D1R 335

2

D2F

66 22 24

2872

D2R 844

D2B 404

3

D3F

57 20 36

1812

D3R 1123

D3B 457

Test Results of the ANN Designs w.r.t. Targets

Design No.Overall

Accuracy (%)

Accuracy (%) w.r.t. Target #

1 2 3 4

Te

st Ca

ses

D1F 98.1 100 92.3 100 100

D1R 94.2 100 92.3 92.3 92.3

D2F 97.7 100 90.9 100 100

D2R 96.6 100 86.4 100 100

D2B 86.4 NA NA NA NA

D3F 96.3 100 85 100 100

D3R 100 100 100 100 100

D3B 80 NA NA NA NA

Test Results of the ANN Designs w.r.t. Targets

Design No.Overall

Accuracy (%)

Accuracy (%) w.r.t. Target #

1 2 3 4

Generalization C

heck Cases

D1F 99.3 100 95.2 100 100

D1R 70.3 87.5 52.1 58.3 85.4

D2F 63.5 91.7 45.8 58.3 58.3

D2R 77.1 100 83.3 58.3 75

D2B 86.1 NA NA NA NA

D3F 79.2 100 75 75 83.3

D3R 81.9 100 83.3 66.7 77.8

D3B 81 NA NA NA NA

1

3

2

1 23 4

Conclusions Out of the eight design versions, four are

promising for HIF diagnosis applications. These are D1F, D2B, D3R & D3B

Design D3R is the best because 100% accuracy for all test cases and all targets 100% accuracy for mid-span HIF cases, for extended

feeder cases, for varying fault impedance cases and for varying transmission line impedance cases

All errors are for lightly loaded feeder cases, but none of them is in detecting the HIF occurrence

The design fails to distinguish between two fault events but succeeds in not false-indicating a HIF for normal system operation or vice versa