Professional Documents
Culture Documents
Power Quality
• Artificial Intelligence
• Applications in Power System
• Issues and Challenges
• Power Quality
• Case Study
Artificial Intelligence
What is Intelligence?
It can be defined as the ability to learn and understand, learn from
experience, apply knowledge and skills to adapt to new situations or
solve problems, make judgements or have opinions that are based on
reason.
Chatbots for customer service Social media monitoring tools for Song or TV show or video
dangerous content or false news recommendations
• Model-based: uses a
previously learned model
to accomplish the task
Image Source: Dayan P, Niv Y, Reinforcement learning: The Good, The Bad and The Ugly, Curr Opin Neurobiol (2008),
doi:10.1016/j.conb.2008.08.003
Applications in Power Systems
Necessity of AI in power system
Power System AI
• Complex in structure and • Data-driven method capable to
mathematical model, non- model any degree of non-linearity
linear behavior without requiring apriori
knowledge of system model
Stationary Signals
Non-stationary Signals
Voltage Interruption
Interharmonics: non-integral
multiples of the fundamental
frequency
Ex: 82 Hz, 142 Hz, 412 Hz etc.
Sub-harmonics: frequency<50 Hz
Ex: 12 Hz, 35 Hz etc.
Power Quality Disturbances
Transients
• Impulsive
• Oscillatory
Impulsive Oscillatory
Power Quality Disturbances
Notch
Noise
Imbalance
Effects of PQ Disturbances
1. Process or equipment shutdown
2. Failure of electronic devices and controllers
3. Light flicker
4. Overheating of motor and transformer windings
5. Premature operation of protective devices
6. Metering inaccuracies
7. Insulation breakdown and damage of semiconductor devices
8. Subsynchronus oscillations
A/D
Voltage/ Signal Processing and Power Quality
Process
Current Signal Intelligent Techniques Analysis
Processing
1. Estimation of PQ Indices
Non-linear loads or
2. Monitoring of amplitude &
Renewable Energy Sources
frequency
3. Detection of PQ disturbances
Ω1 Ω2 ΩM-1
Compute the FFT, X(k)
Frequencies,
𝜔 = {𝜔! , 𝜔" , 𝜔# , … . . 𝜔$ }
Fourier segmentation
Boundaries Ω% = 0, Ω$ = 𝑓& +1000
Scaling function
ì 1 if w £ (1 - g )W1 ü
ï ï
ï æp ö ï
f1 (w ) = ícos ç b (g , w , W1 ) ÷ if (1 - g )W1 £ w £ (1 + g )W1 ý
ï è2 ø ï
ï 0 otherwise ï
î þ
Frequency
Apply FFT
X(ω) estimation
+ - X1(ω)
X2(ω)
Estimate the interharmonics
2nd set of frequencies Λ2
Process the nearest frequencies
Where, Xrms(n) is nth half cycle RMS of the discrete signal x(i)
and S is the number of samples in a half cycle. The windowed
signal of 200 ms duration having 50 Hz as the fundamental
frequency will give a signal RMS vector Xrms having values
calculated for 20 half cycles as
X rms = [ X rms (1),..., X rms (n),..., X rms (20)]
F4: This feature represents the amount of deviation of the disturbance half cycle (h)
with its consecutive half cycle, computed as
F 4 = X rms (h) - X rms (h +1)
F5: The RMS value of the fundamental frequency component extracted from the
adaptive filtering technique plays a key role in the discrimination of flicker, notch, and
spike from the fundamental magnitude related disturbances. Thus, the feature F5 is the
RMS value of the fundamental component calculated for only h half cycle.
1 hS
1
F 5 = X1rms (h) = å x1 (i)2
2 S i =(h-1) S +1
Feature Extraction
F6: The high-frequency harmonic distortion HHD(h) separates all the harmonics
related signals from other class of signals
X Hrms (h)
HHD(h) =
X1rms (h)
Where, X1rms(h) represents the h half cycle RMS value of the fundamental frequency
component. XHrms(h) is the h half cycle RMS value of the high-frequency harmonic
signal xH, which contains only the components with frequency ≥ 300 Hz.
1 1 hS
X Hrms (h) = å xH (i)2
2 S i =(h-1) S +1
F7: This novel feature, mean of the detected half cycle h and its consecutive half
cycle of the input signal, x(i) helps in detection of transient, notch, and spike.
1 i =( h +1) S
F7 = å x(i)
2S i =( h -1) S +1
Feature Extraction
F8: The discrimination of voltage interruption, sag and swell from each other is
possible only with features reflecting the magnitude of the fundamental frequency
component. Hence, the per cycle RMS value for ten cycles of the fundamental
frequency component is calculated as
F8 = [ X rms (1),..., X rms (n),..., X rms (10)]
F9: The last feature, mean of the twenty half cycle HHD values reflect the presence
of high-frequency harmonic distortion over the complete input signal.
F9 = mean(HHD)
Data Flow
Signal
Xrms Vector
Adaptive
Filtering
Find the half cycle “h” Technique
Vertical and
Horizontal
histograms of Xrms(h) Xrms(h+1) F7
half cycle
per cycle per cycle
X1rms Vector HHD Vector
+ -
X1rms(h)
Mean(HHD)
HHD(h)
FFNN 1 FFNN 2
Oscillatory Harmonics/
Flicker Notch Spike Interruption Sag Swell
Transient Interharmonics
1 1 1
Dual FFNN F1
2
3
4
2
.
2
. 1 Flicker
. .
5
. Oscillatory
6 . 2
Transient
7 . .
F2 8 . . 3 Notch
9
. .
10
F3 11 . . 4 Spike
F4 12
. .
F5 13
29 29
F6 14
F7 15 30 30
1 1
1
2 2
2
. . 1 Interruption
3
. .
4
. . 2 Sag
5
F8 . .
6
. . 3 Swell
7
. .
8
. . 4 Harmonics
9 . .
10 29 29
F9 11 30 30
Results
• Eight single and ten two-combination disturbances are considered for
analysis along with a pure sinusoidal signal.
• A complete set of 7,600 patterns with 400 patterns of each class are
generated by varying the parameters and duration of an event.
• The synthetic signals are contaminated with the additive white
Gaussian noise of SNR varying from 25 dB to 55 dB
• Different phase angles are also considered with a fundamental
frequency variation of ± 0.5 Hz
• The 50 Hz fundamental frequency signals of 200 ms are generated at
a sampling rate of 10 kHz, yielding 2000 samples per window
• The dual FFNN classifier is initially trained with 300 patterns of each
disturbance and the remaining for evaluating the dual FFNN.
Classification Accuracy
• For noiseless case, minimum accuracy for sag
with transient and swell with transient.
• The use of ANN having one output neuron dedicated for a single
disturbance helps in recognising the combined disturbances and also
improves the robustness to noise.