978-1-4799-3561-1/14/$31.00 ©2014 IEEE PMAPS 2014
Probabilistic Assessment of the Process-Noise
Covariance Matrix of Discrete Kalman Filter State
Estimation of Active Distribution Networks
L. Zanni, S. Sarri, M. Pignati, R. Cherkaoui and M. Paolone
École Polytechnique Fédérale de Lausanne (EPFL)
Lausanne, Switzerland
lorenzo.zanni@epfl.ch
Abstract—The accuracy of state estimators using the Kalman
Filter (KF) is largely influenced by the measurement and the
process noise covariance matrices. The former can be directly
inferred from the available measurement devices whilst the latter
needs to be assessed, as a function of the process model, in order
to maximize the KF performances. In this paper we present
different approaches that allow assessing the optimal values of
the elements composing the process noise covariance matrix
within the context of the State Estimation (SE) of Active
Distribution Networks (ADNs). In particular, the paper considers
a linear SE process based on the availability of synchrophasors
measurements. The assessment of the process noise covariance
matrix, related to a process model represented by the ARIMA
[0,1,0] one, is based either on the knowledge of the probabilistic
behavior of nodal network injections/absorptions or on the a-
posteriori knowledge of the estimated states and their accuracies.
Numerical simulations demonstrating the improvements of the
KF-SE accuracy achieved by using the calculated matrix Q are
included in the paper. A comparison with the Weighted Least
Squares (WLS) method is also given for validation purposes.
Keywords—Active distribution networks; Kalman filter;
probabilistic assessment; process noise covariance matrix; real-time
state estimation; phasor measurement unit
I. INTRODUCTION
One of the main challenging tasks related to the operation
of Active Distribution Networks (ADNs) is the development of
accurate and fast (i.e. sub-second) State Estimation (SE)
processes (e.g., [1], [2]). In what follows we make reference to
a SE process that uses time-tagged measurements available
with high refresh rate (typically those streamed by Phasor
Measurement Units – PMUs – with rates of tens of frames per
second [3], [4]). Indeed, all the functionalities associated to the
Real-Time (RT) operation of ADNs like: (i) optimal voltage
control, (ii) feeders congestion management, (iii) losses
minimization, (iv) fault detection and location, (v) post-fault
management, and (vi) increase of the system reliability [5]-[8],
might be largely improved if the knowledge of the network
state is available with high accuracy and refresh rate (e.g., [9],
[10]).
A well-known approach used for the estimation of power
networks state relies on the use of the Kalman Filter (KF) (e.g.
[11]-[17]). As known, compared to traditional Weighted Least
Square (WLS) method (e.g., [18]), it accounts the available
measurements and the time-varying nature of the process to be
identified by means of a suitable-defined process model that
predicts the system state in advance. As a consequence, the KF
is a two-stage algorithm. The first ‘prediction part’ projects the
previous time step state forward in time by means of a pre-
defined process model. The second ‘estimation part’ corrects
the predicted state by accounting the available measurements
and the accuracies of both process model and measurements.
In this respect, one of the key factors that largely influence
the KF accuracy is the knowledge of two error covariance
matrices: (i) the so-called process noise covariance matrix, and
(ii) the measurement noise covariance matrix.
As discussed in [19], if both noise covariance matrices are
not properly defined, the robustness of the KF algorithm is not
necessarily satisfied. Additionally, in [17] the relative influence
of these two uncertainties has been discussed with reference to
the SE of ADNs performed by using the Iterative KF (IKF)
process.
Concerning the measurement noise covariances, they
represent the accuracies of the measurement devices and can be
easily inferred. On the other hand, the process noise
covariances represent the uncertainties introduced by the
process model to predict the next system state. It is worth
observing that, in general, in the literature dealing with power
systems SE using the KF, the values of the process noise
covariance matrix are arbitrarily selected although, in principle,
they need to be computed if the process is known (e.g., [20],
[21]). Therefore, an appropriate assessment of this matrix is of
fundamental importance for the maximization of the KF-SE
accuracy and represents the objective of this paper.
Within this framework, the Authors of [22] have
summarized and discussed some of the novel methods
proposed in the last decade for the estimation of the noise
covariance matrix for non-linear state estimators. In [23], the
Authors have analyzed the tuning of the process and noise
covariance matrices in order to optimize the performance of a
fault detection process based on the Extended KF (EKF). More
recently, in [13] it has been presented a two-stage KF for
power systems SE: in the first stage an adaptive KF algorithm
identifies and corrects the impact of incorrect system modeling
and/or bad PMU measurements; in the second stage the
estimated bus voltages are fed into an EKF to obtain the
dynamic states.
The research leading to the results presented in this paper has received
funding from the NanoTera Swiss National Science Foundation project S
3
-
Grids and from the European Community's FP7-ICT-2011-8 under the grant
agreement n° 318708 (C-DAX).