Seminar Course 392N ● Spring2012
Lecture 2 Intelligent Energy Systems: Monitoring Basics Dimitry Gorinevsky
ee392N - Spring 2012 Stanford University
Intelligent Energy Systems © Dimitry Gorinevsky
1
Traditional Grid • Worlds Largest Machine! – 3300 utilities – 15,000 generators, 14,000 TX substations – 211,000 mi of HV lines (>230kV)
• A variety of interacting information decision and control systems
ee392N - Spring 2012 Stanford University
Intelligent Energy Systems © Dimitry Gorinevsky
2
Smart Energy Grid Intelligent Energy Network Source IPS
energy subnet
Load IPS Intelligent Power Switch
Generation Transmission Distribution Load Conventional Electric Grid Conventional Internet ee392N - Spring 2012 Stanford University
Intelligent Energy Systems © Dimitry Gorinevsky
3
Outline 1. 2. 3. 4.
Monitoring Applications Statistical Process Control - SPC Multivariate SPC – MSPC Principal Component Analysis - PCA
ee392N - Spring 2012 Stanford University
Intelligent Energy Systems © Dimitry Gorinevsky
4
Internet Applications Tablet
Computer
Smart phone
Internet
Presentation Layer Backend
Business Logic
CRM and ad analytics Portfolio optimization Decision support Fraud detection
Database ee392N - Spring 2012 Stanford University
Intelligent Energy Systems © Dimitry Gorinevsky
5
Intelligent Energy Applications Tablet
Computer
Smart phone
Communications Internet Energy Application
Presentation Layer Application Logic Business Logic (Intelligent Functions) Database ee392N - Spring 2012 Stanford University
Intelligent Energy Systems © Dimitry Gorinevsky
6
Control Functions • Control function in a systems perspective – Closed loop
ee392N - Spring 2012 Stanford University
Intelligent Energy Systems © Dimitry Gorinevsky
7
Monitoring & Decision Support • Monitoring functions are open-loop - Data presentation to a user
ee392N - Spring 2012 Stanford University
Intelligent Energy Systems © Dimitry Gorinevsky
8
Power Generation Time Scales • Power generation and distribution • Energy supply side
1/10 ee392N - Spring 2012 Stanford University
1
10
100
Anomalies & Sustainment Power Supply Scheduling
1000 Time (s)
http://www.eeh.ee.ethz.ch/en/eeh/education/courses/viewcourse/227-0528-00l.html Intelligent Energy Systems 9 © Dimitry Gorinevsky
Power Demand Time Scales • Power consumption – DR, Homes, Buildings, Plants
• Demand side
Anomalies & Sustainment
Enterprise Demand Scheduling Building HVAC Home Thermostat Demand Response
100 ee392N - Spring 2012 Stanford University
1,000 Intelligent Energy Systems © Dimitry Gorinevsky
10,000
Time (s) 10
Monitoring Goals • Situational awareness – Anomaly detection – State estimation
• System health management – Fault isolation – Condition based maintenances
ee392N - Spring 2012 Stanford University
Intelligent Energy Systems © Dimitry Gorinevsky
11
Condition Based Maintenance • DOD CBM+ Initiative
ee392N - Spring 2012 Stanford University
Intelligent Energy Systems © Dimitry Gorinevsky
12
Outline 1. 2. 3. 4.
Monitoring Applications Statistical Process Control - SPC Multivariate SPC – MSPC Principal Component Analysis - PCA
ee392N - Spring 2012 Stanford University
Intelligent Energy Systems © Dimitry Gorinevsky
13
Anomaly Detection - SPC • SPC - Statistical Process Control – Introduced for monitoring of manufacturing processes – Warning for off-target quality
• SPC vs. EPC – Engineering Process Control = feedback control
• Main SPC method – Shewhart Chart (Control Chart)
• Other SPC methods – EWMA, CuSum, Western Electric Rules ee392N - Spring 2012 Stanford University
Intelligent Energy Systems © Dimitry Gorinevsky
14
Exceedance Monitoring • Currently used in most monitoring systems • Example: grid frequency deviation from 60Hz – Empirical exceedance threshold
Plant/System Data
ee392N - Spring 2012 Stanford University
Monitoring Function: Exceedance
Intelligent Energy Systems © Dimitry Gorinevsky
Detected Anomalies
15
SPC: Shewhart Control Chart W.Shewhart, Bell Labs, 1924 Statistical Process Control (SPC) UCL = µ + 3·σ LCL = µ - 3·σ Upper Control Limit
quality variable
• • • •
mean µ Lower Control Limit 3
ee392N - Spring 2012 Stanford University
6
9
sample
12
Walter Shewhart (1891-1967)
Exceedance / Out of control
15
Intelligent Energy Systems © Dimitry Gorinevsky
16
Shewhart Chart, cont’d • Quality variable assumed randomly changing around a steady state • Detection: y(t) > UCL = µ + 3·σ • For a normal distribution, false alarm probability is 0.27% LCL
z (t ) =
y (t ) − µ
σ
UCL
P(z > 3) = 1-Φ(3) = 0.1350·10-2 P(z < 3) = Φ(-3) = 0.1350·10-2 ee392N - Spring 2012 Stanford University
Intelligent Energy Systems © Dimitry Gorinevsky
17
SPC: Use Examples • SPC in manufacturing • Fault monitoring for PHM/CBM • Sensor integrity monitoring – Fault tolerance and redundancy management
Sensor Reference ee392N - Spring 2012 Stanford University
+ -
|v| < 3σ
Intelligent Energy Systems © Dimitry Gorinevsky
no
Fault
yes
Normal
18
Outline 1. 2. 3. 4.
Monitoring Applications Statistical Process Control - SPC Multivariate SPC – MSPC Principal Component Analysis – PCA
ee392N - Spring 2012 Stanford University
Intelligent Energy Systems © Dimitry Gorinevsky
19
Multivariate SPC • Univariate process: y(t) z = 2
y − µ0 2 ~ χ σ 2
Chi-squared CDF
P (z 2 > c 2 ) = 1 − F ( c 2 ,1) = Φ ( −c) + 1 − Φ (c)
• Two univariate processes
ee392N - Spring 2012 Stanford University
Intelligent Energy Systems © Dimitry Gorinevsky
20
MSPC Explanation • MSPC=Multivariate Statistical Process Control • Scatter plot for correlated channels Time series data
Keep the data values, ignore the time stamp
y1(t) y2 y2(t) y1 ee392N - Spring 2012 Stanford University
Intelligent Energy Systems © Dimitry Gorinevsky
21
Multivariate SPC • Two correlated univariate processes y1(t), y2(t) y1 µ1 y= µ= y2 µ2 cov(y) = P multivariate outlier: out of control
ee392N - Spring 2012 Stanford University
Intelligent Energy Systems © Dimitry Gorinevsky
22
Whitened Variables • Uncorrelated linear combinations
z = ( y − µ ) P −1 ( y − µ ) ~ χ 22
z(t) = L·[y(t)-µ] LTL= P-1
2
T
cov(z) = I
• Declare fault (anomaly) if
( y − µ )T P −1 ( y − µ ) > c 2 P (z 2 > c 2 ) = 1 − F ( c 2 ;2) CDF for Chi-squared with 2 DOF ee392N - Spring 2012 Stanford University
Intelligent Energy Systems © Dimitry Gorinevsky
23
Multivariate Monitoring Plant / System y (t ) Data
Monitoring data processing: x = y − µˆ T 2 = x T Pˆ −1 x
Historical Data Set
Models: Performance, Noise µˆ , Pˆ
ee392N - Spring 2012 Stanford University
Intelligent Energy Systems © Dimitry Gorinevsky
Advisory Info: • Anomaly T 2 > c2
25
Hotelling's T2 • Empirical parameter estimates 1 µˆ = N 1 ˆ P= N
N
∑ y (t ) ≈ E ( y ) t =1 N
T ˆ ˆ ( ( ) )( ( ) y t − µ y t − µ ≈ cov ( y − µ ) ) ∑ t =1
• Hotelling's T2 =
•
T2
N N +1
T2
two-sample statistics is
Harold Hotelling (1895-1973)
T ⋅ ( y ( N + 1) − µˆ ) Pˆ −1 ( y ( N + 1) − µˆ )
2 χ distribution differs from since Pˆ , µˆ are
considered as random variables, y(t) ~ N(µ,P)
ee392N - Spring 2012 Stanford University
Intelligent Energy Systems © Dimitry Gorinevsky
26
Multivariate SPC with
2 T
• The anomaly detection decision is T 2 > c2
• Threshold c is defined by the false positive/false negative tradeoff based on the distribution ( N − 1) p Fp , N − p T ~ ( N − p) 2
where F is the Fisher-Snedecor’s F-distribution p is the dimension of the data vector y N is the size of the training data set ee392N - Spring 2012 Stanford University
Intelligent Energy Systems © Dimitry Gorinevsky
27
Outline 1. 2. 3. 4.
Monitoring Applications Statistical Process Control - SPC Multivariate SPC – MSPC Principal Component Analysis – PCA
ee392N - Spring 2012 Stanford University
Intelligent Energy Systems © Dimitry Gorinevsky
28
PCA • PCA = Principal Component Analysis X = [ y (1) − µˆ
y ( 2) − µˆ y ( N ) − µˆ ]
• What if empirical covariance P=XXT/N is not invertible? Cannot compute xTP-1x – This happens in most real cases
• SVD of the data and covariance matrix X = U⋅ S⋅ VT = ∑k uk skvkT XXT= U⋅ S2⋅ UT = ∑k uk sk2ukT VTV = I ee392N - Spring 2012 Stanford University
Intelligent Energy Systems © Dimitry Gorinevsky
Scores Loadings 29
PCA Structure • Singular vectors (principal components) U = [UR U0] UR - Range Space; nonzero singular values, sk > 0 U0 - Null Space; zero singular values, sk=0 • Singular values
nonzero
sk ‘zero’
k >>[U,S,W]=svd(X*X’) % 2ms for 100x100 ee392N - Spring 2012 Stanford University
Intelligent Energy Systems © Dimitry Gorinevsky
30
PCA,
2 T,
and Q statistics
• T2 statistics is used in Range Space of P xR = U RTU R x is Range Space projection of x • Range Space: covariance is invertible
T 2 = xRT PR−1 xR
• Must also monitor Null Space projection • Q statistics: Q = xTU0U0 x – a.k.a. SPE (squared prediction error)
Q = x − xR ee392N - Spring 2012 Stanford University
2
Intelligent Energy Systems © Dimitry Gorinevsky
31
PCA, T2, and Q Summary Q
T2
principal component #1
principal component #2
ee392N - Spring 2012 Stanford University
Intelligent Energy Systems © Dimitry Gorinevsky
32
PCA Prediction Model • Null space defines linear dependency between monitored variables U0x = v ≈ 0 m linear equations • Can be interpreted as a dependence between two subsets of variables y m
x= z k
y = bz + v
• SPE yields model prediction error:
y − bz ee392N - Spring 2012 Stanford University
2
Intelligent Energy Systems © Dimitry Gorinevsky
33
End of Lecture 2
ee392N - Spring 2012 Stanford University
Intelligent Energy Systems © Dimitry Gorinevsky
34