Net Based Estimation of Power Components

Dynamics
Saumya Raj Satyam Kumar Subham Kumar Singh
Department of Electrical Engineering Department of Electrical Engineering Department of Electrical Engineering
Indian Institute of Technology Roorkee Indian Institute of Technology Roorkee Indian Institute of Technology Roorkee
s [email protected] s [email protected] s [email protected]

Sundram Kumar Prof. Abdul Saleem Mir

Department of Electrical Engineering Department of Electrical Engineering
Indian Institute of Technology Roorkee Indian Institute of Technology Roorkee
s [email protected] [email protected]

Abstract—We propose an Adaptive Gaussian Sum Unscented of smaller-variance Gaussians, which improves the accuracy
Kalman Filter (AGS-UKF) for estimating dynamic states and of the Unscented Transform while maintaining computational
quasi-static parameters in nonlinear power system models. These efficiency.
parameters are typically unobservable under normal conditions
but become identifiable during specific excitation events, resulting Once the refined Gaussians converge, they are merged back
in both persistent modeling drift and occasional abrupt shifts. into a single Gaussian, thereby maintaining computational effi-
To capture this behavior, the AGS-UKF models the process noise ciency. This approach enables the filter to remain robust across
using a two-component Gaussian mixture: one low-variance com- a wide range of operating conditions, accurately tracking both
ponent for gradual changes and one high-variance component for
slow dynamics and sudden changes in system parameters.
sudden jumps. A Bayesian scheme dynamically adjusts the con-
tribution of each component based on measurement likelihoods. The remainder of this report is organized as follows. Section
Additionally, the filter refines high-variance distributions using a 2 formulates the state-space model and outlines the assump-
set of smaller-sigma Gaussians to improve Unscented Transform tions underlying the proposed approach. Section 3 provides
accuracy without sacrificing efficiency too much. Simulations an overview of the AGS-UKF algorithm, highlighting its key
on a three-machine nine-bus system and a DFIG-based wind
turbine demonstrate the algorithm’s improved performance and steps. Section 4 elaborates on the algorithmic framework and
robustness compared to the standard UKF implementation details. Section 5 presents simulation results
on a three-machine nine-bus system and a DFIG-based wind
I. I NTRODUCTION turbine system. Finally, Section 6 concludes the report.
Robust, real-time estimation of power system dynamic
states—such as rotor angles and speeds—as well as quasi- II. P ROBLEM F ORMULATION
static parameters—such as inertia constants and governor
gains—is essential for ensuring reliable and secure system Consider a continuous-time nonlinear state-space model
operation. However, this task is complicated by the presence augmented with quasi-static parameters w whose values be-
of both small modeling inaccuracies and rare but abrupt come identifiable only during particular excitation events. The
changes due to events such as faults or control switching. evolution of w thus exhibits both small, persistent modeling
Traditional nonlinear filters, such as the Extended Kalman errors and occasional large shifts when these events occur. We
Filter (EKF) and Unscented Kalman Filter (UKF), typically model this heavy-tailed behavior by approximating the process
assume Gaussian process noise, making them ill-suited for sce- noise on x and w with a two-component Gaussian mixture:
narios involving heavy-tailed disturbances or discontinuities in one component with low variance to capture gradual drift, and
system behavior. another with high variance to capture abrupt parameter jumps.
To address this limitation, we propose the Adaptive Gaus-      
sian Sum Unscented Kalman Filter (AGS-UKF), designed to xk f (xk−1 , wk−1 , u) r
= + x (1)
operate effectively under both gradual and abrupt parameter wk wk−1 rw
variations. The core innovation lies in modeling the pro- yk = h(xk , wk , u) + q (2)
cess noise of quasi-static parameters using a two-component
Gaussian mixture: one low-variance component to capture where,
continuous modeling drift, and one high-variance component  
to represent sudden jumps that occur during periods of pa- rx
∼ πdrif t Ndrif t (0, Rdrif t ) + πjump Njump (0, Rjump )
rameter observability. The AGS-UKF further adapts to high- rw
variance scenarios by refining the state distribution using a set q ∼ N (0, Q)
 III. A LGORITHM OVERVIEW 2) Splitting into sigma-distributions: The key innovation in
this approach is the decomposition of the prior state distribu-
This section provides an overview of the key components
tion into multiple Gaussian components. This decomposition
underlying the proposed AGS-UKF algorithm. We begin with
allows the filter to capture complex, multimodal behaviors
a brief review of the Unscented Transform (UT), which forms
that are often observed in nonlinear systems with uncertain
the basis for nonlinear propagation in the filtering framework.
parameters. The half-decomposition can be done as follows:
A. Unscented Transform √
Xi = (U s)i , i = 0, ..., r − 1 (12)
Sigma point generation: a set of (2n+1) vectors are cal- 1
culated using the mean value of the state vector, x and its α0 = (13)
2r
covariance matrix, P as follows: 1 − α0
αi = (14)
X0 = xk−1 (3) 2r
p m0 = x̂ (15)
Xi = xk−1 + ( (n + µ)Pk−1 )i , i = 1, ..., n (4) r
ra
p
Xi = xk−1 − ( (n + µ)Pk−1 )i , i = n + 1, ..., 2n (5) mi = x̂ − Xi−1 · , i = 1, ..., r (16)
1 − α0
r
where µ = α2 (n + ν) − n, n is the number of states, α decides ra
mi = x̂ + Xi−r · , i = r, ..., 2r (17)
the spread of sigma points, β represents the prior knowledge 1 − α0
of the distribution of states around x and ν is a scaling factor. Σi = Σx − Σm , i = 0, ..., 2r (18)
The distribution after transformation, f (x), is now approx-
imated by the gaussian: where,
2n
X Σx = U sU T (19)
x̂ = Wim xi (6) 2r−1
i=0 1 X
2n
Σm = (mi − x̂)(mi − x̂)T (20)
X 2r i=0
Σx = Wic (xi − x̂)(xi − x̂)T (7)
i=0 with orthonormal U ; singular values diagonal matrix s; a is the
where, ratio of variance of means to total variance; αi are component
µ weights; r is the number of dimensions in which we wish to
W0m = (8) decompose the distribution.
n+µ
1 The UKF (with Long-tailed process noise approximation)
Wim = i = 1, ..., 2n (9) is then run in parallel on each of the decomposed Gaussian
2(n + µ)
µ distributions. As the estimation process progresses, these dis-
W0c = + 1 − α2 + β (10) tributions gradually converge. Once they become sufficiently
n+µ
close, they are merged back into a single Gaussian distribution.
1
Wic = i = 1, ..., 2n (11)
2(n + µ)
IV. A LGORITHMIC F RAMEWORK
B. Adaptive Gaussian Sum Unscented Kalman Filter
1) Long-tailed process noise approximation: The process This section presents a detailed formulation of the Adaptive
noise is modeled as a Gaussian mixture, with component Gaussian Sum Unscented Kalman Filter (AGS-UKF). The
weights updated after each measurement using a Bayesian algorithm operates in a recursive predict-update framework,
inference scheme. with modifications to account for heavy-tailed process noise
and dynamically adaptive distribution refinement. Each step
y y
of the algorithm is described in sequence, including mixture
y P(y|Ndrif t )P(Ndrif t ) weight estimation, Gaussian merging, adaptive sigma splitting,
P(Ndrif t |y) = and measurement updates. The formulation emphasizes both
P(y)
y y accuracy in nonlinear regimes and computational efficiency.
y P(y|N jump )P(Njump )
P(Njump |y) =
P(y) 1) Each component of the prior state distribution GMM is
y y y y propagated through the system dynamics f , separately
P(y) = P(y|Ndrif t )P(Ndrif t ) + P(y|N jump )P(Njump )
for each component of the noise GMM. The resulting
where, Niy is the measurement distribution when the process intermediate distributions are then passed through the
noise distribution is Ni . The propagated state distribution is measurement function h to obtain the corresponding
obtained by merging the individual state distributions from output distributions of y. The propagated state distribu-
each mixture component, weighted according to their respec- tion for each prior GMM component is constructed by
tive probabilities. aggregating these noise-conditioned sub-components.
 f +rdrif t fx
Nix −−−−−−→ FX i|drif t → Ni|drif t
UT
fx h+q y
Ni|drif t −−→ Yi|drif t → Ni|drif t
UT
x f +rjump fx
Ni −−−−−−→ FX i|jump → Ni|jump
UT
fx h+q y
Ni|jump −−→ Yi|jump → Ni|jump
UT

2) The resulting propagated components are weighted ac-

cording to a Bayesian scheme based on their likelihoods,
merged into one gaussian and then filtered using the
measurement.

y y
P(y|Ni|drif t )P(Ni|drif t )
αdrif t =
P(y)
y y
P(y|Ni|jump )P(Ni|jump )
αjump =
P(y)
[f x,y] [f x,y] [f x,y]
Ni ∼ N (αdrif t µdrif t + αjump µjump ,
[f x,y] [f x,y]
αdrif t Σdrif t + αjump Σjump )
Ki = Σfi x,y Σy−1
i

3) If any component of the posterior distribution exhibits

variance along any axis exceeding a specified threshold,
that component is split into multiple smaller sigma
distributions along those axes.
4) Else if the refined components converge within a pre- Fig. 1. (a) Grid-connected wind turbine-generator system. (b) Equivalent
defined proximity, they are merged back into a single system with shaft stiffness, damping, and a gear train with backlash, along
with a visual representation of typical backlash non-linearity in gears. From
Gaussian distribution. [2]

V. E XPERIMENTS

A. Wind Turbine driven Doubly Fed Induction Generator

1) Backlash modelling: The system’s power balance equa-

tions can be expressed in terms of torque as follows:
Pt
Tt =
1 wt
ẇt = (Tt − Ts ) 1
Ht Te =(E ′ Iqs + Ed′ Ids )
1 ws q
w˙r = (Ts − Te ) 
˙ cd ˙
Hg ks (θd − δ) + cd (θd ), θd + ks (θd ) > δ

θ̇t = wB (wt − ws ) Ts = ks (θd + δ) + cd (θ˙d ), θd + kcds (θ˙d ) < −δ

0, otherwise

θ̇t = wB (wr − ws )

where wt , wr and wB are the turbine, generator and base

electrical speed. Ht and Hg are turbine and generator inertia
constant. θt and θr are angular position of the turbine and
generator shaft. Tt , Ts and Te are turbine, shaft and electrical where, θd = θt − θr . ks and cd are mechanical stiffness and
torque respectively. dampening coefficient of the turbine shaft.
 2) DFIG modelling: The model equations of the DFIG
generator in the d–q frame is given as:
E′
k1 I˙qs = −R1 Iqs + L′s Ids + wr Eq′ − d − Vqs + km Vqr
Tr
′
E q
k1 I˙ds = −R1 Ids − L′s Iqs + wr Ed′ − − Vds + km Vdr
Tr
′
1 ′ Eq
Ėq = R2 Ids − + (1 − wr )Ed′ − km Vdr
wB Tr
1 ′ E′
Ėd = −R2 Iqs − d − (1 − wr )Eq′ + km Vdr
wB Tr
L′
where, k1 = wBs , km = LLmr , Tr = R Lr
r
, R1 = km2
Rr + Rs ,
2
R2 = km Rr . Rs and Rr are stator and rotor resistances. Ls ,
Lr and Lm are stator, rotor and mutual inductances.
3) Infinite grid connection: The DFIG is connected to the
infinite grid through a transmission line and transformer with
a combined reactance of XT , resulting in a current flow Ia Fig. 2. backlash angle, δ, estimates
corresponding to the power transfer. The line current, and
consequently the stator voltage, can be expressed in terms of
the DFIG’s modeling states as follows:
Ia = Iqs + jIds
Vq s = re(Vb + jIa XT )
Vd s = imag(Vb + jIa XT )
Additionally, the total active and reactive power supplied to
the grid can be determined using:

Pe = Vds Ids + Vqs Iqs + Vdr Idr + Vqr Iqr

Qe = Vds Iqs + Vqs Ids + Vdr Iqr + Vqr Idr
where Idr and Iqr are d-axis and q-axis rotor currents,
respectively, given as follows:

′
Eds
Iqr = − − km Iqs Fig. 3. Inertia constants, Ht and Hg , estimates
ws Lm
′
Eqs
Iqr = − − km Ids
ws Lm remained at 0 p.u. for 2 seconds and then returned to 1 p.u.
4) Estimation: The extended state vector and measurement at t=3 s, restoring normal operation. This abrupt fluctuation
is taken as: posed a significant challenge to state estimation, as the system
T dynamics changed drastically within a short period.
x = Iqs , Ids , Eq′ , Ed′ , wr , wt , θd , δ, Hg , Ht

 T
y = Pe Qe B. Three Machine Nine Bus Power System

5) Simulation studies and results: To evaluate the perfor- 1) State space model: A simple state space model for
mance of the proposed Adapted Gaussian Sum Unscented general constant impedance load power system is developed
Kalman Filter (AGS-UKF), simulation studies were conducted as:
on a doubly fed induction generator (DFIG) connected to an Hg d2 δg dδg
infinite bus through a transmission line and transformer. The = Pgm − Pge − Df
2ws dt2 dt
system’s dynamic response was analyzed under varying wind out
power conditions to assess the accuracy and robustness of the Pge = re(Sbus,g )
state estimation process.
out
Sbus = Vbus ◦ (Ybus Vbus )∗
A critical test scenario was introduced where the wind Vbus,l = −(Ybus,l,l + Yl )−1 (Ybus,l,s Vbus,s + Ybus,l,g Vbus,g )
power, Pt , experienced a sudden drop to 0 per unit (p.u.) at
t=1 s, simulating a complete loss of wind input. The power where, s is slack, g is generator and l is load bus.
  T
x = δg δ̇g Yl Df Hg
 T
y = S1 S6 S8
where, Si is complex power output by bus i; Yl is the load
admittances.

Fig. 6. Generator dampening constants, Df , estimates

Fig. 4. WECC 3-machine, 9-bus system; the value of Y is half the line
charging (Copyright 1977. Electric Power Research Institute. EPRI EL-0484.
Power System Dynamic Analysis, Phase I. Reprinted with Permission.) [5]

Fig. 7. Generator rotor angles, δg , estimates

scenario for state estimation. Furthermore, at t=2 s, bus 5 load

admittance is suddenly changed to 5 - 0.1j, causing a sudden
shift in the system operating conditions.

VI. C ONCLUSION

This report introduced the Adapted Gaussian Sum Un-

scented Kalman Filter (AGS-UKF) for improved state es-
Fig. 5. Generator inertia constants, Hg , estimates timation in nonlinear power systems. By decomposing and
merging Gaussian components, AGS-UKF effectively handled
2) Simulation studies and results: Simulations were con- nonlinearity and parameter uncertainties.
ducted on a three-bus power system where key parameters, Simulation results on a doubly fed induction generator
such as generator inertia constants, dampening factors and load (DFIG) and a three-bus power system showed that AGS-UKF
admittances were unknown. The objective was to estimate the significantly outperformed the traditional UKF, providing more
system states accurately under these uncertain conditions. accurate and stable state estimates. These findings demonstrate
In the simulation, the load admittance at bus 5 was subject its potential for real-time power system monitoring and con-
to an error of 35% in its rate of change, creating a challenging trol, with future applications in larger-scale grids.
 R EFERENCES
[1] E. A. Wan and R. Van Der Merwe, ”The unscented Kalman filter for
nonlinear estimation,” Proceedings of the IEEE 2000 Adaptive Systems
for Signal Processing, Communications, and Control Symposium (Cat.
No.00EX373), Lake Louise, AB, Canada, 2000, pp. 153-158, doi:
10.1109/ASSPCC.2000.882463.
[2] Prajapat, G.P., Bhui, P., Senroy, N. and Kar, I.N. (2018), Mod-
elling and estimation of gear train backlash present in wind turbine
driven DFIG system. IET Gener. Transm. Distrib., 12: 3527-3535.
https://doi.org/10.1049/iet-gtd.2017.1377
[3] Ganesh P. Prajapat, N. Senroy, I.N. Kar, Estimation based enhanced
maximum energy extraction scheme for DFIG-wind turbine systems,
Sustainable Energy, Grids and Networks, Volume 26, 2021, 100419,
ISSN 2352-4677, https://doi.org/10.1016/j.segan.2020.100419.
[4] H. Tebianian and B. Jeyasurya, ”Dynamic state estimation in
power systems using Kalman filters,” 2013 IEEE Electrical Power
& Energy Conference, Halifax, NS, Canada, 2013, pp. 1-5, doi:
10.1109/EPEC.2013.6802979.
[5] Sauer, P. W., & Pai, M. A. (1998). Power System Dynamics and Stability.
(1 ed.) Prentice Hall