State Estimation
Lectures prepared by
Prof. S. Shahnawaz Ahmed
For subsequent materials assistance taken mainly from :
[1]. John J. Grainger, William D. Stevenson, Jr. "Power System Analysis",
McGraw-Hill, 1994.
Why state estimation is necessary?
•Power system control needs operational data which are
measured and sent to the server on-line through SCADA.
•These are random noise and/or gross error corrupted
•Cleaning up these is essential
•State estimator does it applying a statistical criterion to a
redundant set of measurements (i.e. M > N; N is the number
of state variables basically V,θ).
•It first estimates the best set of system states and using
those recalculates a set of consistent measurements close to
the exact values with a certain level of confidence.
Load flow also computes unknown V, θ. Then what is the difference between LF and SE?
n number of
[Δθ]i+1 = [JθP]-1 [P-f1(V,θ)]i+1
Pinj and Qinj
measurements n No. of [θ, V]
or database [ΔV]i+1 = [JθP]-1 [Q-f2(V,θ)]i+1 which were
values previously
unknown
Load flow
(no filtering action on measurement noise)
M number (M>N)of
measurements N number of
Z ={injections, line flows, bus [ΔX]i+1 = [G]-1 [H]T [R]-1[Z-h(x)]i+1
states X ={θ, V
voltage magnitudes, at all buses}
phase angles from PMU if any} State estimator using WLS criterion
Statistical properties of random noise in the measurements.
Measured value =True value + error or noise
[Zm]=[hm(x)]+[em] m=1,2….M
The errors or noise em (also denoted by ζ) Y=probability density function
associated with several measurements of X= random variable i.e. em
the same quantity in a power system
usually follow the standard Gaussian
probability distribution function i.e. they
are Gaussian white or Normal noise
characterized by the following features.
•E(em) = 0 i.e. mean or expectation is zero
•E(ememT)=σm2 i.e. variance is equal to the Hatched area is the probability of em being
square of the st.dev. between +3 σm and - 3σm and is the
•Probability of em being ≤±3σm is 99.7%. integral of y between +3 σm and - 3σm
which will be equal to 99.7%
•An example of calculating St. Dev. for a complex (MW and MVAR combined)
line-flow measurement by the meter manufacturers
σm = 1/3 (a1 |S | + a2 D +a3 D)
in units of MVA or pu if divided by Base MVA
where,
S : complex true value (MVA) of line flow
D:Full scale (MVA) deflection corresponding to rated transducer output.
a1 = 2% to 12% due to CT, PT errors
a2 =0.25% to 0.35% due to transducer (meter) errors
a3 = 0.1% due to A/D converters including noise from communication
State estimator works in association with other components of EMS
Prefiltering
Very large
Measurements errors
e.g. ∑Pi ≈ 0 Breaker status
Pij – Pji ≤ d
to be satisfied
Network
Data
configurator
validator Pij ≠ 0 if breaker
of line ij is closed
Data base
Flow of information
State estimator Execution sequence
Display
Weighted Least Square (WLS) Criterion:
Minimize J(x) = ∑wm (zm – hm(x))2 ; m=1,2………….M and wm is the weighting factor
= 1/σm2 for the m-th measurement
Or, J = ([Z] –[h(x)])T [R]-1 ([Z] –[h(x)])
When [h(x)] =[h(x0)] + [H]([X]-[X0])
So that
[Z] –[h(x)] = [Z]-[h(x0)]-[H]([X]-[X0])
=[ΔZ] – [H][ΔX] [ΔZ] is also termed measurement residual or error vector
∂J/ ∂x = 0 eventually leads to the following iterative sequence
[X]i+1 = [X ]i +[G]-1[H]T [R]-1 [ΔZ] i+1 ; [X] means estimated state vector
[R] is the MxM measurement error covariance matrix and is diagonal with elements equal
to the respective measurement variance i.e. st .dev square (σm2 )
hm(x) is mathematical expression for the m-th measurement in terms of x (i.e. v,θ)
[H] = [… ∂h/ ∂x…] is the MxN measurement Jacobian matrix
[G] = [H]T[R]-1[H] is NxN termed gain matrix and is constructed as a constant matrix in a
way similar to the fast decoupled NR load flow Jacobian matrix.
Observability:
Observability of a system is the possibility of determining its states using the
available measurements.
Tests:
1) Gain matrix [G] = [H]T[R]-1[H] should be invertible
2) The measurements and network configuration should provide a connected
graph i.e. unbroken tree that will ensure the existence of [G]-1
I
I : Injection measurement
{i, j,k} Є I F: Line flow measurement
i
F
j
r (reference node)
k
In this way every node should be directly or indirectly connected to the reference node
through measurements.
Bad data or gross errors: these arise mainly due to malfunctioning or failure of
transducers, RTUs, communication link etc.
[Zm]=[hm(x)]+[em] m=1,2….M
If all the errors em are Gaussian random noise then J(x̂) = ∑wm (zm – hm(x̂))2 i.e.
the J (objective or cost function) calculated using estimated state variables will
be a chi-square distribution with k= M-N degrees of freedom.
Bad data detection tests:
Hypothesis H0 : No bad data exists
H1: H0 is not true
(1) Cost function based: Accept H0 if J(x̂) ≤ χp2
Reject H0 if J(x̂) > χp2 p=1-α is the level of confidence
χp2 is available from Tables
(2) Weighted residual based: Accept H0 if |(zm – hm(x̂))/σm | ≤ 3.0; m=1,2….M
Reject H0 if |(zm – hm(x̂))/σm | > 3.0; m=1,2….M
(3) Normalized or standardized
residual based: Accept H0 if |(zm – hm(x̂))/√R̕mm | ≤ 3.0; m=1,2….M
Reject H0 if |(zm – hm(x̂))/√R̕mm | > 3.0; m=1,2….M
When R̕mm = m-th diagonal element of the matrix ([R]-[H][G]-1[H]T)
Identification of bad data:
•i)Remove the measurement with the largest weighted residual |(zm – hm(x̂))/σm |
•ii)Re-estimate [X ]
•iii)Detect bad data by applying the test
•iv)If H0 is rejected remove the measurement with the next largest residual and
go to step (ii)
•v) In this way repeat steps (ii) to (iv) and if necessary replace the removed measurements
by past values in similar conditions(termed pseudo measurements) so that the redundancy
does not fall to an extent making estimation infeasible.
Conceptual example on WLS state estimation using a DC circuit
Solution:
Let x1 = V1, x2 = V2
By applying basic circuit theorem e.g.
superposition principle the measurements
Z can be related to state variables and network
parameters (resistances) as shown.
[Z] = [H] [X]
[R]-1 = [W]
# Bad data detection and identification
if J(x̂) > χp2
The estimated measurements are then
Now degrees of freedom k= M-N = 4-
2=2 and from chi square Table for k
=2 and level of significance α = 0.01,
χ k, α2 = 9.21
Since J(x̂) computed using the
weighted squares of estimated errors
J(x̂) is same as f
is less than χ k, α2 = 9.21 so no bad
data exists with a confidence of (1- α )
i.e. 99% .
Identification based on standardized residuals
Solution:
where R’mm are the diagonal elements
of matrix ([I]- [H][G]-1[H]T [R]-1)[R]
This matrix diagonal elements will always add up to be equal
to the number of state variables.
[R]
R11’= R22’= R33’= R44’=
σ1 2=1/100 σ2 2=1/100 σ3 2=1/50 σ4 2=1/50
It is evident that the largest residual corresponds to z4 so eliminating this state estimation is
repeated using three remaining measurements. Then H, G matrices are recomputed.
Close to estimate obtained without any
bad measurement
The estimated errors are then [e] = [Z] -[Ẑ]
which is less than χ2 threshold 6.64 for a degrees of freedom now being 3-2 = 1 and
Hence no more bad data exists.
AC Power System or Real life system state estimation Formulation
Solution:
Here δ1 = 00 as the reference angle
Then 3 state variables are x1 =δ2
x2 = ІV2І, x3=ІV1І
Now to find h(x) for the measurements we have to use line flow and injections equations.
So applying the above equations to z3 (Q1), z4(P12), z5 (Q21) and substituting x1 =δ2
x2 = ІV2І, x3=ІV1І we can have following expressions for the measured quantities
h1 = ІV2І =x2
h2 = ІV1І =x3
In this 2-bus special case
h3 =
system
h4 = P12 = -P2
Q21 = Q2
h5 =
Iterate until
І [X](k+1) –[X](k) І ≤ Є
Numerical solution of the illustrated example on AC state estimation
Hx (k)
Hx(0)TR-1
Hx(0)TR-1
f>χ2(k=2, α=0.01) [=9.21] So bad data exists
Identification of bad data and re-estimation
Diagonal elements add up to be equal to N =3
largest
Re-estimation eliminating z5
Using z1, z2, z3, z4 in the original measurement vector
and the flat start values
0.598
Z3 and z4 residuals were large due to smearing
effect of WLS estimator
State estimation can also be applied to detect
if a bad data is due to improper meter
connection
Z5 was a reactive
power flow (Q21)
