박찬은∗
(Chan-eun Park)
1
박인석∗
(In Seok Park)
1
박부견
(PooGyeon Park)
†
-
(Department of Electrical Engineering, POSTECH, Korea.)
Copyright © The Korean Institute of Electrical Engineers(KIEE)
Key words
Distribution system state estimation, Dynamic state estimation, Power system modeling, Extended Kalman filter
1. Introduction
Power systems are controlled to maintain the stage which is planned by power system
operators. To operating the power sys- tem efficiently, state estimation on power
system such as distribu- tion system and transmission system is crucial. Thus, many
papers on state estimation for power system have been published such as least squre
algorithm (1), weighted least square estimation [2–4] and Gauss-Newton method (5,6). However, these approaches are static state estimation, which are hard to capture
the properties of system dynamics.
Nowadays, distribution system becomes larger and complex. It may have unexpected changes
such as abrupt disconnection of generator and electric vehicles. Increasing number
of photovoltaic panel and wind turbine also causes the dynamic behavior of power system.
Therefore, instead of static estimation, the researches on dynamic state estimation
algorithm has been published. Especially, the authors’ of (7) studied the extended Kalman filter for distri-algorithm requires state transient
model which describes the state of next time step by using current information. Because
the exact state transient model of distribution power system is unknown, the exist-
ing dynamic state estimation algorithm have used the state transient model by smoothing
the previous states. However, the smoothing model may not reflect the current power
system, which is the moti- vation behind this study.
This paper suggests mathematical distribution power system state transient model by
analyzing a nonlinear function of powerw injection model. Measurement model is also
described as nonlin- ear function of state. Then, extended Kalman filter is adopted
for dynamic state estimation. In this step, jacobian matrices of nonlin- ear functions
are derived as a function of system state. To verify proposed algorithm, IEEE 15 test
bus with scheduled generator is used.
The notations in this paper is fairly standard. For vector $x, x(i)$ denotes $i-\mathrm{th}$
component of $x$. For matrix $A$, $A_{i j}$ and $[A]_{i j}$ denote $(i, j)-$element
of matrix $A$. $\mathbf{0}_{a \times b} \in \mathcal{R}^{a \times b}$ is the zero
matrix, $\mathbf{1}_{a \times b} \in$ $\mathcal{R}^{a \times b}$ is the matrix whose
all components are one.
2. Problem Statements
To use dynamic state estimation algorithm, the state space representation for the
distribution system is required. Namely, setting the following system equations is
essential.
where $x_{k} \in \mathcal{R}^{n}$ is the state, $z_{k}$ is the measurement, $w_{k}$
and $v_{k}$ are modeling error of system model and measurement model, respec- tively.
$f(\cdot)$ and $h(\cdot)$ are model functions. $n$ is the number of nodes in power
system. We set the estimation target as complex voltage in the distribution node and
use power flows/injections in the distri- bution system as measurement values through
phasor measurement units (PMUs). Therefore, $x_{k}$ and $z_{k}$ are defined as follows:
where the notations $\left|V_{i, k}\right|$ and $\theta_{i, k}$ denote voltage magnitude,
voltage phase of bus $i$ at time $k$. Also, $P_{I, k}\left(Q_{I, k}\right)$ and $P_{F,
k}\left(Q_{F, k}\right)$ mean real part (imaginary part) of power injections and that
of power flows, respectively. While the measurement model $h(\cdot)$ can be derived
by using the topology of power system, the exact model $f(\cdot)$ in (1) is unknown.
Therefore, as an alternative model, the recent papers (7,8) assumed the quisi steady-state behavior of distribution system and used the following
linear time-varying model.
2.1 Previous Works: Linear Time-varying State Transition Model
where $F_{k}$ and $g_{k}$ are the state transition process, and they are up- dated
online as the Holt-Winters double exponential smoothing method (10):
where $I$ is an identity matrix with appropriate dimension, $a_{k}$, $β_{k}$ $∈ (0,
1)$ are smoothing coefficient to be adjusted, $x_{k | k^{\prime}}$ denotes the esti-
mation of $x_{k}$ given the measurements up to time $k^{\prime}$. The terms $m_{k}$,
$l_{k}$ are recursively updated as
Remark 1: There are some reason why this model is not always robust. First, because
the terms (8), (9) are composed of the estimation state $x_{k | k-1}$, the model accuracy may drop if
the estimation state $x_{k | k-1}$ is not correct. Second, it is hard to adjust the
adequate smoothing coefficient $\alpha_{k}, \beta_{k}$ for each time. Finally, quasi
steady-state behavior of distribution system cannot be always assumed since power
systems in nowadays can have some abrupt load changes. Therefore, we propose the following
power system model which comes from mathematical analysis.
3. Distribution system modeling
In this section, the nonlinear models of measurement system and state transient
system are provided via mathematical approach.
3.1 Measurement system model
We use both power injection and flow as measurement values. In distribution power
system, they are represented as complex values and can be modeled by using power flow
equation. First, power injection model is represented as follows:
where $P_{I, k}(i)\left(Q_{I, k}(i)\right)$ is real(imaginary) part of power injection
in $i-\mathrm{th}$ node at time $k$, and $G_{i, j}\left(B_{i, j}\right)$ is real(imaginary)
part of $(i, j)-$component of $\boldsymbol{Y}_{b u s}$, which is an admittance matrix
determined by distribution power system topology.
Next, power flow model is obtained such that:
where $P_{i, j, k}\left(Q_{i, j, k}\right)$ is real(imaginary) part of power flow
from node $i$ to node $j$ at time $k$, and $g_{i j}\left(b_{i j}\right)$ is real(imaginary)
part of the ad-mittance between node $i$ and node $j$. Therefore, the following mea-surement
model is used in this paper.
where $i \in[1,2, \cdots, n], t \in m_{i}$ and $m_{l}$ denotes index that is linked
to node $l$ and larger than $l$, such that $m_{l} \in\left\{i | l<i \leq n, Y_{b u
s}(l, i) \neq 0\right\}$. The modeling error wk is obtained as offline, and it is
assumed that the white Gaussian noise vector with known covariance $R$. Additionally,
the dimension of $z_{k}$ is determined by distribution system topology.
3.2 State transient system
As mentioned in previous section, power injection model is rep- resented as a function
of system state, and it has same dimension as system state.
where nonlinear function $g(\cdot)$ is constructed by using the equations (10)-(11). Then taking partial derivative of $x_{k}$ to both sides earns following relations:
By setting $\partial y_{k} \approx y_{k+1}-y_{k}$ and $\partial x_{k} \approx x_{k+1}-x_{k},
x_{k+1}$ is calculated as follows:
(18), additional assumption is used such that
which is valid under the $PMU$. The matrix $J_{k}\left(x_{k}\right)$, which is a jaco-bian
matrix of $g\left(x_{k}\right)$, is defined as
It is known that active power injection $P_{I, k}$ and reactive power injection $Q_{I,
k}$ sensitivity of voltage magnitude $\left|V_{k}\right|$ and voltage phase $\theta_{k}$
is very small (11). Therefore, by considering the computation efficiency, we suppose that $\frac{\partial
P_{I, k}}{\partial\left|V_{k}\right|}=0$ and $\frac{\partial Q_{I, k}}{\partial \theta_{k}}=0$.
This supposition reduces the computational complexity of $J_{k}^{-1}\left(x_{k}\right)$
from $O\left(8 n^{3}\right)$ to 2$O\left(n^{3}\right)$. Then state transient model
of distribution system is determined as follows:
where $I_{a} \in \mathcal{R}^{a \times a}$ is the identity matrix, and $w_{k}$ is
state transition modeling error which is assumed to be white Gaussian noise vector
with known covariance matrix Q. The components of matrices $J_{1}\left(x_{k}\right),
J_{2}\left(x_{k}\right)$ are described in (23), (24). The detailed proof is omitted due to lack of page limit.
Fig. 1. Modeling result of state transient system
Fig. 2. Modeling result of measurement system
where $\mathbf{m}=\left\{(x, y) | x \in[1,2, \cdots, n], y \in m_{i}\right\}$ and
$\mathbf{m}_{i} \triangleq\left(i_{x}, i_{y}\right)$.
3.3 Modeling result
By using a simulation data from Matlab Simulink of IEEE 15 radial test bus, we verify
the effect of the proposed models. Fig. 1 shows the actual states from Matlab simulink (red-solid) and cal- culated states
from the proposed model in (18) (blue-dashed). The used data in (18) is sampled as 30Hz. Fig. 2 illustrates the actual measurement from Matlab simulink (red-solid) and calculated
val- ues by using states in Fig. 1 (blue-dashed). It looks like that the measurement error which is the difference between
actual value of measurement and calculated measurement value follows white Gaussian
distribution. Thus, this model can be used to extended kalman filter.
4. Dynamic state estimation algorithm for power system
represented as nonlinear function and model errors are assumed as white Gaussian
noise vector, extended Kalman filter (EKF) can be adopted (12).
4.1 Jacobian Calculation
Jacobians of nonlinear functions $f(\cdot), h(\cdot)$ are defined such that
First, $H_{k}$ is derived as follows:
where each component of $J_{1}\left(x_{k}\right), J_{2}\left(x_{k}\right), H_{1}\left(x_{k}\right),
H_{2}\left(x_{k}\right)$ is given in (23),(24),(25),(26). Here, the assumption that
$\frac{\partial P_{F, k}}{\partial\left|V_{k}\right|}=0$ and $\frac{\partial Q_{F,
k}}{\partial \theta_{k}}=0$ is still valid. Next, it is hard to calculate $F_{k}$
because $f\left(x_{k}, z_{k}, z_{k}-1\right)$ consists of inverse of $J_{1}\left(x_{k}\right),
J_{2}\left(x_{k}\right)$. Therefore, instead of directly calculating partial derivative
of $f\left(x_{k}, z_{k}, z_{k}-1\right)$, the complex step differentiation algorithm
is used.
Lemma 1 (13): Let $f(z)$ be a function with complex variable $z=r_{0}+i h$, where $r_{0}$ is a
point on the real axis and $h$ is a real parameter. By taking Taylor series expansion
on $f(z)$ off the real axis, then
where $f^{(n)}(\cdot)$ is $n-\mathrm{th}$ order partial differential function. After
that, taking the imaginary part of both sides, dividing by $h$, and setting $h$ as
very small scalar holds the following relation:
where $\mathfrak{I}\{Z\}$ denotes the imaginary part of $Z$.
4.2 State Prediction
This stage is to propagate the state into the next time step by using state transition
model:
where $P_{a | b}$ is the covariance matrix of predicted state between time step $a$
and $b$, and $Q$ is covariance matrix of transition model noise. $x_{k | k-1}$ is
predicted state estimation at time step $k − 1$ and also called as state forecast
value.
Fig. 3. Topology of IEEE 15 radial test bus
4.3 State Update
After acquiring measurement $z_{k}$, state estimation can be updated:
where $K_{k}$ is Kalman filter gain and $R$ is covariance matrix of mea- surement
noise, which is estimated as off-line.
5. Simulation Results
To verify the validity of the proposed dynamic model (21) and the effectiveness of dynamic state estimation of distribution power system, IEEE
15-bus radial test system is considered. In Fig. 3, bus 1 is substation, where its voltage phase can act a role of reference, and there
are 3 generators (G1,G2,G3) which can be scheduled by system operator. To show the
robustness of the model under the suddenly changes in power system, power generation
from G1-G3 is scheduled as (Fig. 4). The measurements are obtained as MATLAB simulink model, and the sampling frequency
of measuring is 30Hz, which is same as PMU data.
By conducting several simulations under difference generator scheduling data, and
taking its average values, we also verify that the modeling error has a pattern of
normal distribution, such that $w_{k} \mathcal{N}(\overline{\mathbf{m}}, \sigma)$,
where $\overline{\mathbf{m}}=\left[\begin{array}{ll}{\mathbf{0}_{1 \times 15}} & {3
\cdot 10^{-3} \cdot \mathbf{1}_{1 \times 15}}\end{array}\right]^{T} \approx\left[\mathbf{0}_{30
\times 1}\right]$, $\sigma=$ $\left[4 \cdot 10^{-4} \cdot 1_{1 \times 15} \quad 4.5
\cdot 10^{-3} \cdot 1_{1 \times 15}\right]^{T}$, and the unit is p.u for voltage magnitude
and degree for voltage phase.
Next, state estimation result by using proposed model and EKF algorithm is shown in
Fig. 5. They show the voltage magnitude and phase of node 13. The actual signal is acquired
from Matlab simulink.
Fig. 4. Active/Reactive power scheduling
Fig. 5. Estimation (blue) and actual value (red) of voltage magnitude and phase at
node 13
6. Conclusion
This paper suggested the distribution system state transient model by using mathematical
analysis of nonlinear functions, whereas the existing papers have been used only smoothing
model. Also, setting the measurement model as a set of nonlinear function of system
state was done for dynamic state estimation for power system. Jacobian matrices for
nonlinear state transient model and measurement model were calculated to apply extended
Kalman fil- ter algorithm. The validity of proposed modeling and dynamic es- timation
algorithm were verified by using simulation data of IEEE 15 bus radial test system.
Our future work is extending our result under the bad data problem and cyber attack.
Acknowledgements
본 연구는 2019년도 삼성전자의 재원으로 지원받아 수행되었음.
References
Giannakis G. B., Kekatos V., Gatsis N., Kim S.-J., Zhu H., Wollenberg B. F., 2013,
Monitoring and optimization for power grids: A signal processing perspective, IEEE
Signal Process- ing Magazine, Vol. 30, No. 5, pp. 107-128
Vishnu T., Viswan V., Vipin A., 2015, Power system state es- timation and bad data
analysis using weighted least square method, in 2015 International Conference on Power,
Instru- mentation, Control and Computing (PICC), IEEE, pp. 1-5
Primadianto A., Lu C.-N., 2017, A review on distribution sys- tem state estimation,
IEEE Transactions on Power Systems, Vol. 32, No. 5, pp. 3875-3883
Muscas C., Pau M., Pegoraro P. A., Sulis S., 2014, Effects of measurements and pseudomeasurements
correlation in distri- bution system state estimation, IEEE Transactions on Instru-
mentation and Measurement, Vol. 63, No. 12, pp. 2813-2823
Kekatos V., Giannakis G. B., 2013, Distributed robust power sys- tem state estimation,
IEEE Transactions on Power Systems, Vol. 28, No. 2, pp. 1617-1626
Teng J.-H., 2002, A modified gauss-seidel algorithm of three-phase power flow analysis
in distribution networks, International journal of electrical power & energy systems,
Vol. 24, No. 2, pp. 97-102
Do Coutto Filho M. B., de Souza J. C. S., Freund R. S., 2009, Forecasting-aided state
estimation part ii: Implementation, IEEE Transactions on Power Systems, Vol. 24, No.
4, pp. 1678-1685
Guo Z., Li S., Wang X., Heng W., 2016, Distributed point-based gaussian approximation
filtering for forecasting-aided state es- timation in power systems, IEEE Transactions
on Power Sys- tems, Vol. 31, No. 4, pp. 2597-2608
Valverde G., Terzija V., 2011, Unscented kalman filter for power system dynamic state
estimation, IET generation, transmis- sion & distribution, Vol. 5, No. 1, pp. 29-37
Da Silva A. L., Do Coutto Filho M., De Queiroz J., pp 237-244, State forecasting in
electric power systems, in IEE Proceedings C (Generation, Transmission and Distribution),
Vol. 130, No. 5. iet, pp. 1983
Grainger J. J., Stevenson W. D., Stevenson W. D., et al. , 2003, Power system analysis
Welch G., Bishop G., et al. , 1995, An introduction to the kalman fil- ter
Martins J. R., Sturdza P., Alonso J. J., 2003, The complex-step derivative approximation,
ACM Transactions on Mathemati- cal Software (TOMS), Vol. 29, No. 3, pp. 245-262
저자소개
She received her B.S. degree from School of Electrical En- gineering and Computer
Science at Ulsan National Insti- tute of Science and Technology (UNIST), Korea, in
2014, and the M.S. degree in the Division of Electrical and Com- puter Engineering
at Pohang University of Science and Technology (POSTECH), Korea, in 2016.
She is currently a Ph. D candidate in Electric and Electrical Engineering at POSTECH.
Her research interest includes stochastic system, singular system and robust control
theory.
E-mail: chane2727@postech.ac.kr
He received his B.S. degree from School of Electrical En- gineering and Computer Science
at Kyungpook National University (KNU), Korea, in 2014, and the M.S. degree in the
Division of Electrical and Computer Engineering at Po- hang University of Science
and Technology (POSTECH), Korea, in 2016.
He is currently a Ph. D candidate in Elec- tric and Electrical Engineering at POSTECH.
His research interest includes fuzzy system, stochastic system and ro- bust control
theory.
E-mail: ispark90@postech.ac.kr
He received his B.S. degree and M.S. degree in Control and Instrumentation Engineering
from Seoul National Uni- versity, Korea, in 1988 and 1990, respectively, and the Ph.D
degree in Electrical Engineering from Stanford Uni- versity, U.S.A., in 1995.
Since 1996, he has been affili- ated with the Division of Electrical and Computer
Engi- neering at Pohang University of Science and Technology (POSTECH), where he is
currently a professor.
His cur- rent research interests include robust, LPV, and network- related control
theories, delayed systems, fuzzy systems, and signal processing.
E-mail: ppg@postech.ac.kr