You are on page 1of 6

Machine Learning of Factors Influencing Damping

and Frequency of Dominant Inter-area Modes in the


WECC Interconnect
Zhangshuan Hou, Jim Follum, Pavel Etingov, and Francis Tuffner Dmitry Kosterev and Gordon Matthews
Pacific Northwest National Laboratory Bonneville Power Administration
Richland, Washington, USA Portland, Oregon, USA
Email: Zhangshuan.Hou@pnnl.gov, James.Follum@pnnl.gov

Abstract—The stability of inter-area electromechanical os- difficult to attain, leading to differences between modeled and
cillations are critical to power system reliability. Due to the observed system behavior. The 1996 breakup of the WECC
complexities of power systems, relationships between system serves as a dramatic example of this type of mismatch [1].
conditions and oscillation characteristics, such as damping and
frequency, tend to be expressed only in generalities. In this study, To address this challenge, significant effort has been in-
a list of influential factors on Western Electricity Coordinating vested to complement model-based approaches with methods
Council interconnect modal characteristics are identified and based on Phasor Measurement Unit (PMU) data [5]–[9]. Such
evaluated with advanced machine learning techniques including methods apply advanced signal processing techniques to ex-
principal component analysis, analysis of variance, classification tract information about the system’s modes from data collected
and regression trees, and support vector machine approaches.
The predictive relationships between the influential factors and immediately after system events or during ambient conditions
modal characteristics could be used in the future to develop when the system’s dynamics are continuously excited by
alert thresholds earlier than measurement-based mode estimation random load changes. The methods applicable during ambient
approaches. conditions are particularly useful for ongoing monitoring of
Index Terms: Electromechanical modes, machine learning, the system’s stability. Still, the methods tend to rely on long
principal component analysis, analysis of variance, classification
and regression trees, support vector machine record lengths to provide accurate estimates. As a result, even
when modes are monitored in the online environment there is
I. I NTRODUCTION a delay before estimates reflect changes in the modes [9].
Studies that seek to systematically understand the mecha-
The stability of electromechanical oscillations in power sys-
nisms and factors that influence the modes help address this
tems is a critical concern and a prerequisite for secure system
challenge [10]–[13]. For example, if poor damping is found
operation [1]. A system’s modes dictate the periodic exchanges
to coincide with a particular set of power flows, the system
of energy between groups of generators in disparate parts of
can be operated in a way that avoids such system conditions.
the system. They are often characterized with by frequency
Because real-time information about power flows is available
and damping ratio terms. If a mode becomes unstable, i.e., its
via PMUs, adjustments could be made faster than relying
damping ratio becomes negative, the resulting oscillations in
exclusively on mode estimates based on long record lengths
power and synchronous frequency will continue to grow until
of ambient data.
protective actions are taken. The result can be catastrophic,
In this study, we adopt a data-centric approach to eval-
as evidenced by the 1996 breakup of the Western Electricity
uate the impacts of system conditions on the damping and
Coordinating Council (WECC) system, where 7.49 million
frequency of the interconnect modes. A system model is
customers were affected by the loss of 30,930 MW of load
used in this initial study to prove the concept, but similar
[1].
approaches could be applied with measured power system
Methods of analyzing electromechanical modes can typ-
data. The system condition is characterized with a set of
ically be categorized as model- or measurement-based ap-
voltage angle pairs, generator outputs, and line flows. The term
proaches. Model-based approaches provide useful insight as
angle pair refers to the difference between two voltage angles
long as the model is accurate, but due to the size and
at different buses, typically separated by several transmission
complexity of power systems, the required accuracy can be
lines. Because all of these parameters can be updated in real-
The Pacific Northwest National Laboratory is operated for the U.S. De- time using PMU data, their evaluation serves as a complement
partment of Energy by Battelle Memorial Institute under Contract DE-AC05- to algorithms that directly estimate the modes from the past
76RL01830. several minutes of data. The parameters are analyzed using
This work was funded by the U.S. DOE through the Grid Modernization
Lab Consortium and through Bonneville Power Administration’s Technology various machine learning (ML) techniques for feature selection
Innovation program. and variable ranking [14].

2018,
c Battelle Memorial Institute PMAPS 2018
ML techniques, such as cross-correlation, principal compo-
nent analysis (PCA), and analysis of variance (ANOVA) are
typical approaches for feature selection [14], [15]. To better
characterize the non-linear relationships among the multiple
variables, people have adopted more advanced techniques such
as decision-tree (DT) and support vector machines (SVM)
techniques [22]. These analyses are usually integrated to
achieve a comprehensive set of critical features; however, such
integrated analyses are absent for power system oscillation
problems.
In this study, we focused on using these advanced learning
techniques to establish a comprehensive set of influential Fig. 1. Schematic of angle pairs (lines), generation sites (G), and transmission
factors on interconnect modal frequency and damping, and lines/paths (arrows) associated with the MT mode.
identifying the most appropriate predictive models to facilitate
modal frequency and damping predictions.
were first computed, resulting in a (21 × 21) cross-correlation
II. M ETHODOLOGY matrix as a measure of similarities among the 19 factors and
Disturbance datasets were generated based on a full-scale the two modal attributes. Elements within the matrix range
WECC model using General Electric’s Positive Sequence Load from -1 to 1, which gives the strength of the relationships. A
Flow (PSLF) tool. The simulations were performed for 125 value of zero indicates that there is no relationship between a
different sets of system conditions, including diverse genera- pair of variables, while a value near unity indicates a nearly
tion patterns and planned transmission outages. Oscillations perfect positive correspondence between the pair.
were induced by inserting a 1400 MW dynamic braking Another unsupervised approach, PCA, was then used to
resistor [17] at its true location within the system, and at target linear correspondence among the variables. PCA is an
a second bus where the mode of interest was known to be orthogonal linear transformation that transforms the data to
excitable. Analysis of post-disturbance ringdown responses a new coordinate system, such that the greatest variance by
were conducted using Prony analysis [5], [18] for estimating some projection of the data lies on the first coordinate (i.e.,
the system modes. WECC has greatly advanced its under- the first principal component), the second greatest variance
standing of the inter-area modes of power oscillations in on the second coordinate, and so on. Mathematically, the
the Western Interconnection [19]. Damping and frequency transformation is defined by a set of p-dimensional vectors
attributes were estimated for the following WECC modes: of weights or loadings w(k) = (w1 , . . . , wp )(k) that map each
Montana-Northwest, North-South A, North-South B, British row vector x(i) of X to a new vector of principal component
Columbia, and East-West. This paper focuses on the Montana- scores t(i) = (t1 , . . . , tm )(i) , given by tk(i) = x(i) · w(k) , for
Northwest (MT) oscillation mode. i = 1, . . . , n, k = 1, . . . , m. The new t vector inherits the
After applying Prony analysis, the relationships between maximum possible variance from x, with each loading vector
mode estimates and system conditions were evaluated. The w constrained to be a unit vector. Therefore, PCA converts
system conditions were characterized by a set of 19 variables the set of observations with possible correlations into a set of
associated with five voltage angle pairs, active power outputs values of linearly uncorrelated principal components such that
of six generation sites, and active power flow along four the most variance lies on the first several coordinates [20].
transmission lines and four transmission paths (several lines ANOVA was then applied to express a dependent variable
summed together). Geographical locations are depicted in Fig. (e.g., modal damping or frequency) as a linear combination
1. Note that G3, G4, G5, and G6 are units within the same of other features or measurements. It partitions the observed
plant. variance of a dependent variable into components attributable
In some cases, it was important to represent a component’s to different sources of variation. ANOVA is a special case of
status (in or out of service) and the amount of power flowing general linear model, formulated as
or being generated. This was accomplished by using two
Y = Xβ + , (2)
variables: one for the status and the other for the amount of
power. Where power appeared in discrete levels across the where coefficients β can be estimated with maximum like-
simulations, categorical variables were used. This approach lihood or Bayesian techniques. Xβ are the fitted values of
prevented the analysis methods from assigning undue weight observations Y. The resulting model is then used to compute
to minor changes within a level. variance contributions from each predictor (a column of X)
In order to understand the relationships between the modal [15].
attributes and the system conditions, various ML approaches In practice, the relationships among the dependent and
were applied. Paired cross-correlations independent variables might be non-linear. For the modal
E [(X − µX ) (Y − µY )] analyses in this study, two approaches were used to study
ρ(X, Y ) = (1) the potential non-linearity in the cross-dependence of data.
σX σY
Classification and regression trees (CART) is an example of
a DT and is adopted in the study [21]. It is formed by a
collection of rules to achieve the best split to differentiate
damping/frequency observations with respect to the dependent
variables, and the splitting is done recursively for each child
node until no further gain can be made. Mathematically, CART
tries to minimize
ΣN M 2
i=1 [yi − Σm=1 cm I(x ∈ Rm )] , (3)
with M the number of subregions Rm of the input variable
space, and cm the estimated values of the output y in each
subregion.
The above feature selection studies were integated to yield
a short list of influential factors, which were used to develop
nonlinear predictive models using SVM [22]. SVM regression
maps the input variables into high-dimensional feature spaces
via non-linear kernel transformations, then a linear model is
constructed in this feature space in the form of
f (x, w) = Σm
j=1 wj gj (x) + , (4)
where gj (x), j = 1, . . . , m denotes a set of nonlinear trans- Fig. 2. Cross-correlation of the data matrix with distance-based clustering.
formations (e.g., radial basis function), and  is the bias Absolute values are shown as we are interested in the correlation strength.
White color represents a correlation of 1.0 and dark green represents 0. Factors
term. SVM regression has an advantage by utilizing the between 0 and 1.0 are on a linear scale, from dark green to orange to yellow
regularization term in the cost function to avoid over-fitting. to white.
The predictive models were multi-fold validated by dividing
the datasets randomly into training and testing subsets. Leave-
one-out validation tests were also performed. The goodness
of fit (e.g., the root mean squared errors) for the training and
testing sets are expected be comparably low in the finalized
models.
III. R ESULTS
To aid in the discussion of results, the following nomen-
clature for the variables in Fig. 1 has been adopted. The
difference in voltage angle between points i and k is denoted
as APi,k . Active power generation at site i is denoted as Gi,
while active power flow along transmission lines and paths
is denoted as T i and P i, respectively. Asterisks are used
to indicate parameters related to a line or generator’s status
(in or out of service). Estimates of the damping ratio and
frequency of the Montana-Northwest mode are denoted as
M T.D and M T.F , respectively. It is a long list of factors;
feature selection with shortened parameter/factor list enables
ML algorithms to train faster, reduces the complexity of the
models, makes them easier to interpret, and reduces overfitting.
A comprehensive set of unsupervised and supervised ML
Fig. 3. PCA biplot of the data matrix. The scores of all system conditions
techniques were adopted and integrated to help quantify the (black dots) and the projections of all variables (red arrows) are shown with
complex relationships among the high-dimensional factors and respect to the first two principal components.
modal attributes. Cross-correlation analysis of the complete
data matrix shows that voltage angle pairs (e.g., AP1,5 , AP1,4 ,
and AP2,5 ) are among the most correlated to both frequency properties and all the factors; Fig. 3 shows that with respect
(MT.F) and damping (MT.D) of the MT mode, followed by a to the first two principal components, several angle pairs (e.g.,
few other power flows and voltage angle pairs, as shown in Fig. AP1,4 ) and active power generation (e.g., G5) align well with
2, where the most influential factors are generally grouped next MT.D and MT.F.
to MT.F and MT.D, respectively. These findings are supported The factor contributions were quantified via ANOVA with
by PCA which helps visualize the similarities among modal generalized linear model fitting (see Equation 2) and a back-
Fig. 4. Scatter plots of damping against a selection of the influential Fig. 5. Scatter plots of frequency against a selection of the influential
factors. The bottom-right panel shows the goodness of fit (i.e., the damping factors. The bottom-right panel shows the goodness of fit (i.e., the frequency
observations vs the fitted values of damping with the influential condition observations vs the fitted values of frequency with the influential condition
parameters using the generalized linear model for damping). parameters using the generalized linear model for frequency).

TABLE I
T HE RELATIVE CONTRIBUTIONS OF THE IMPORTANT FACTORS TO MT
ward removal approach. The model was fitted with the system MODAL DAMPING AND FREQUENCY ACCORDING TO ANOVA BACKWARD
condition factors including the in and out of service indicators, REMOVAL .
and each trivial factor was eliminated one by one using the
backward removal approach based on the Akaike informa- Damping Frequency
tion criteria (AIC, [23]). The finalized model contains only Factor Contribution Factor Contribution
the influential factors with relatively significant contributions. T 1∗ 15.9% AP1,4 19.9%
Among all the system condition parameters, 12 are considered AP1,5 13.8% AP2,5 19.8%
to be significant for damping and 13 for frequency variations. G1 11.1% G4∗ 13.1%
A selection of scatters are plotted in Figs. 4 and 5, which AP1,4 10.9% G4 11.9%
illustrate either positve, negative, or nonlinear relationships G5 10.6% G1 6.9%
between the system conditions and modal properties. The G6 9.5% T 4∗ 6.8%
contributions of all significant parameters are provided nu- G6∗ 7.7% T 1∗ 5.9%
merically in Table I and geographically in Fig. 6. There are P4 6.8% G5∗ 5.0%
negligible linear or nonlinear patterns in the scatters between P3 5.0% G6∗ 4.7%
the trivial factors and modal properties. Figs. 4 and 5 show G2 3.4% T4 3.1%
that a subset of the system condition parameters can linearly T 2∗ 2.9% G3 1.1%
explain up to 38% and 50% (i.e., R2 ) of the variability of G3 2.4% G3∗ 1.1%
modal damping and frequency, respectively. Here R2 (i.e., R- T 2∗ 0.7%
squared) = explained variation / total variation, is a statistical
measure of how close the data are to the fitted regression
line, and is also known as the coefficient of determination. AP1,5 voltage angle pairs is considered to be the dominant
The adjusted R2 was also calculated, which is equal to factor, according to CART. Regression using SVM yields
1 − (1 − R2 ) ∗ (n − 1)/(n − p − 1), where n is the sample size quantitative nonlinear relationships between the influential
and p is the number of explanatory variables. The formula is factors and modal attributes, by integrating nonlinear kernel
to penalize a big model having extra terms. Here, the adjusted functions (e.g., radial basis functions). Validation tests show
R2 are 0.31 and 0.46, respectively. that reasonable surrogate models can be developed using
The performance of the linear models is satisfactory, but SVM, and that the R2 can be improved to 0.76 and 0.75 (see
can be potentially improved given the presence of nonlinear Fig. 8), with the adjusted R2 0.75 and 0.74, respectively, much
correspondence between several system condition parameters higher than achieved using generalized linear model fitting.
and modal properties (e.g., AP2,5 and MT frequency), and the Such an accurate predictive model can be used to develop alert
idea is to integrate nonlinear relationships (e.g., using CART thresholds earlier than mode-estimation-based approaches.
or SVM models) to achieve higher goodness of fit. CART-
based parameter ranking for MT modal damping is shown in IV. C ONCLUSION
Fig. 7. It generally agrees with the ANOVA results, except that In this study, we applied multiple feature selection ap-
it represents more of the nonlinear relationship, and therefore proaches to identify the most influential factors (system condi-
the ranks of the factors are slightly different. For example, the tions and equipment status) on the damping and frequency of
the Montana-Northwest mode in the Western Interconnection.
The approaches were integrated to improve confidence on the
resulting short list of the most important factors. Several volt-
age angle pairs and units at a generation site were identified
as the key factors. These factors can be monitored in real
time. Pairing with mode estimates, which are usually delayed,
will improve situational alerts. A set of alert thresholds can be
developed dependent on system conditions and mode estimates
with the SVM models developed in the study.
In future work, the methods presented here will be incorpo-
rated in the Oscillation Baselining and Analysis Tool (OBAT)
within the open-source suite for advanced synchrophasor anal-
ysis discussed in [24], allowing them to be applied to measured
data. The methods will also be extended to mode estimates
from ambient data for a more comprehensive analysis and
comparison of ambient and event signals.
Bonneville Power Administration developed a Mode Me-
ter application that has been deployed in its control room
since 2013 [25]. The application uses ambient oscillations
to estimate damping and frequency of inter-area oscillations,
including the MontanaNorthwest mode. The application is
currently in a monitoring mode. This work will help BPA to
Fig. 6. Relative importance of angle pairs, generation sites, and transmission
develop alarming thresholds and to operationalize the Mode
lines/paths associated with the damping (top) and frequency (bottom) of the Meter application.
MT mode.
R EFERENCES
[1] D. N. Kosterev, C. W. Taylor, and W. A. Mittelstadt, “Model validation
for the August 10, 1996 WSCC system outage,” IEEE Transactions on
Power Systems, vol. 14, no. 3, pp. 967–979, 1999.
[2] D. J. Trudnowski, “Estimating electromechanical mode shape from
synchrophasor measurements,” IEEE Transactions on Power Systems,
vol. 23, no. 3, pp. 1188–1195, 2008.
[3] J. F. Hauer, D. Trudnowski, and J. G. DeSteese, “A perspective on
WAMS analysis tools for tracking of oscillatory dynamics,” in Power
Engineering Society General Meeting, 2007. IEEE. IEEE, 2007,
Conference Proceedings, pp. 1–10.
[4] C. Grund, J. Paserba, J. Hauer, and S. Nilsson, “Comparison of prony
and eigenanalysis for power system control design,” IEEE Transactions
on Power Systems, vol. 8, no. 3, pp. 964–971, 1993.
[5] J. F. Hauer, C. Demeure, and L. Scharf, “Initial results in prony analysis
of power system response signals,” IEEE Transactions on power systems,
vol. 5, no. 1, pp. 80–89, 1990.
Fig. 7. CART ranking of the most influential parameters/factors. The relative [6] I. Kamwa, R. Grondin, E. J. Dickinson, and S. Fortin, “A minimal
importance scores of all parameters add up to 1, and the dominant parameters realization approach to reduced-order modelling and modal analysis for
are shown in red. power system response signals,” IEEE Transactions on Power Systems,
vol. 8, no. 3, pp. 1020–1029, 1993.
[7] J. J. Sanchez-Gasca and J. H. Chow, “Computation of power system
low-order models from time domain simulations using a Hankel matrix,”
IEEE Transactions on Power Systems, vol. 12, no. 4, pp. 1461–1467,
1997.
[8] L. Guoping, J. Quintero, and V. Venkatasubramanian, “Oscillation mon-
itoring system based on wide area synchrophasors in power systems,” in
Proceedings of the iREP Symposium on Bulk Power System Dynamics
and Control - VII. Revitalizing Operational Reliability, 2007, pp. 1–13.
[9] D. J. Trudnowski, J. W. Pierre, N. Zhou, J. F. Hauer, and M. Parashar,
“Performance of three mode-meter block-processing algorithms for
automated dynamic stability assessment,” IEEE Transactions on Power
Systems, vol. 23, no. 2, pp. 680–690, 2008.
[10] I. Dobson, J. Zhang, S. Greene, H. Engdahl, and P. W. Sauer, “Is
strong modal resonance a precursor to power system oscillations?”
IEEE Transactions on Circuits and Systems I: Fundamental Theory and
Applications, vol. 48, no. 3, pp. 340–349, 2001.
[11] F. Blaabjerg, R. Teodorescu, M. Liserre, and A. V. Timbus, “Overview
of control and grid synchronization for distributed power generation
Fig. 8. SVM goodness of fit for MT modal damping and frequency.
systems,” IEEE Transactions on industrial electronics, vol. 53, no. 5,
pp. 1398–1409, 2006.
[12] S. Mendoza-Armenta and I. Dobson, “A formula for damping interarea [19] “Modes of inter-area power oscillations in Western Intercon-
oscillations with generator redispatch,” in 2013 IREP Symposium Bulk nection,” WECC Joint Synchronized Information Subcommittee,
Power System Dynamics and Control - IX Optimization, Security and https://www.wecc.biz/Reliability/WECC, Tech. Rep., 2013.
Control of the Emerging Power Grid, Aug 2013, pp. 1–18. [20] I. Joliffe and B. Morgan, “Principal component analysis and exploratory
[13] R. Diao, Z. Huang, N. Zhou, Y. Chen, F. Tuffner, J. Fuller, S. Jin, and factor analysis,” Statistical methods in medical research, vol. 1, no. 1,
J. E. Dagle, “Deriving optimal operational rules for mitigating inter- pp. 69–95, 1992.
area oscillations,” in 2011 IEEE/PES Power Systems Conference and [21] L. Breiman, J. Friedman, C. J. Stone, and R. A. Olshen, Classification
Exposition, March 2011, pp. 1–8. and regression trees. CRC press, 1984.
[14] I. Guyon and A. Elisseeff, “An introduction to variable and feature
selection,” Journal of machine learning research, vol. 3, no. Mar, pp. [22] I. Steinwart and A. Christmann, Support vector machines. Springer
1157–1182, 2003. Science & Business Media, 2008.
[15] W. N. Venables and B. D. Ripley, Modern applied statistics with S- [23] H. Akaike, “An information criterion (AIC),” Math Sci, vol. 14, no. 153,
PLUS. Springer Science & Business Media, 2013. pp. 5–9, 1976.
[16] B. Yegnanarayana, Artificial neural networks. PHI Learning Pvt. Ltd., [24] P. Etingov, F. Tuffner, J. Follum, X. Li, H. Wang, R. Diao, Y. Zhang,
2009. Z. Hou, Y. Liu, D. Kosterev, S. Yang, and G. Matthews, “Open-source
[17] J. F. Hauer, W. A. Mittelstadt, K. E. Martin, J. W. Burns, H. Lee, J. W. suite for advanced synchrophasor analysis,” in IEEE PES Transmission
Pierre, and D. J. Trudnowski, “Use of the WECC WAMS in wide-area and Distribution Conference and Exposition, 2018, accepted.
probing tests for validation of system performance and modeling,” IEEE [25] “Identification of electromechanical modes in power
Transactions on Power Systems, vol. 24, no. 1, pp. 250–257, 2009. systems,” IEEE Power System Dynamic Perfor-
[18] D. J. Trudnowski, J. M. Johnson, and J. F. Hauer, “Making Prony mance Committee, http://sites.ieee.org/pes-resource-
analysis more accurate using multiple signals,” IEEE Transactions on center/files/2013/11/TR15 Modal Ident TF Report.pdf, Tech. Rep.,
Power Systems, vol. 14, no. 1, pp. 226–231, 1999. 2012.

You might also like