Electrical Power and Energy Systems 106 (2019) 392–400

Contents lists available at ScienceDirect

Electrical Power and Energy Systems

journal homepage: www.elsevier.com/locate/ijepes

Probabilistic load flow methodology for distribution networks including T

loads uncertainty
Giambattista Gruossoa, , Paolo Maffezzonia, Zheng Zhangb, Luca Danielc
⁎

a
Politecnico di Milano, Dipartimento di Elettronica Informazione e Bioingegneria, Piazza Leonardo da vinci, 32 I-20133 Milano, Italy
b
Department of Electrical and Computer Engineering, University of California, Santa Barbara, CA 93106, USA
c
Massachusetts Institute of Technology (MIT), Cambridge, CMA 02139, USA

ARTICLE INFO ABSTRACT

Keywords: Distribution grids probabilistic analysis is an essential step in order to assess the daily network operability under
Distribution network uncertain and stress conditions. It is also functional to the development of new services that require load growth
Load uncertainty capacity or to the exploitation of new energy resources affected by uncertainty. Efficient numerical tools able to
Polynomial chaos forecast the possible scenarios while accounting for loads and sources uncertainty are thus of paramount im-
Probabilistic load flow
portance. The majority of available uncertainty-aware predictive tools are based on Monte Carlo analysis which
Variability analysis
allows probabilistic evaluations of the network state at the price of time consuming simulations. In this paper, a
much more efficient simulation framework is presented. The proposed approach relies on the generalized
Polynomial Chaos algorithm and deterministic Power Flow analysis and allows achieving an at least 100× ac-
celeration compared to standard Monte Carlo analysis for the same accuracy.

1. Introduction loads (i.e. variations in the active power demands at the different
phases of the network) a great number of MC runs is needed to achieve
Distribution networks probabilistic analysis is key to the develop- a satisfactory statistical description. In fact, even though loads un-
ment of new services and ways to exploit the electrical infrastructures. certainty can commonly be modeled as Gaussian distributed parameters
Most of such new services are associated to the low-voltage (LV) dis- [8] the nonlinear nature of the load flow problem leads to state variable
tribution network. Promising services are those related to the charging variations, e.g. maximum voltage at nodes or lines current, that are
of vehicles, the decentralization of resources and the diffusion of new nonGaussian-distributed. In this case, the statistical information about
consumption patterns [1]. The widespread diffusion of such facilities is mean value and variance of an electric variable is not enough to de-
expected to introduce a significant variability/uncertainty of power scribe it properly and the detailed Probability Density Function (PDF)
load profiles compared to those of conventional users. For instance, the shape is required for further meaningful inferences. The accurate de-
increase of electric vehicles, especially considering their usage as sto- termination of PDF with MC method can require tens of thousands load
rage systems, will presumably stress the physical limits (e.g., the flow analyses thus becoming very time consuming. Other approximate
maximum current capability) of the lines. Due to the limited number of techniques for PLF analysis exhist, such as the Point Estimate method
monitoring devices that are commonly deployed along the LV feeders, a and the Cumulant Tensor (CT) [9–15]. However, commonly such
comprehensive view of the overall network state requires the support of techniques do not provide the detailed PDF shape which is instead re-
effective computational tools able to predict bus voltages and line quired in our analysis.
currents under variable loading conditions [2–4]. Computational tools In order to address the above issues, in this paper we describe an
should be able to deal with the uncertainty of power loads and the innovative approach to PLF analysis which is based on generalized
trends of variation [5–7]. These techniques, commonly referred to as Polynomial Chaos (gPC) algorithm and Stochastic Testing (ST) method
Probabilistic Load Flow (PLF) analysis, consist in using appropriate [16–19]. The relevant features of the proposed approach are: (a) the
probabilistic models for the power load profiles as well as in replacing implementation of the gPC + ST method does not need to modify the
deterministic load flow simulation with statistical Monte Carlo (MC) code of the deterministic load flow solver employed; (b) gPC + ST
analysis. method allows handling strongly nonlinear problems, as it is the case
In order to account for the interplay of many independent uncertain for PLF formulated in terms of node voltages and powers, and with

⁎
Corresponding author.
E-mail address: giambattista.gruosso@polimi.it (G. Gruosso).

https://doi.org/10.1016/j.ijepes.2018.10.023
Received 13 March 2018; Received in revised form 10 October 2018; Accepted 21 October 2018
0142-0615/ © 2018 Elsevier Ltd. All rights reserved.
G. Gruosso et al. Electrical Power and Energy Systems 106 (2019) 392–400

many independent statistical parameters. a given Probability Density Function (PDF) r ( r ) [22]. Due to the un-
In this paper, we provide the following original contributions: certainty of the power loads applied to the candidate nodes, each ob-
servable variable describing the state of the network at time t, e.g. the
1. we present in an intuitive way the application of the gPC + ST magnitude of the nth node voltage Vn (t ) = Vn (t ) , becomes a stochastic
method to the PLF problem by considering data-based uncertainty in variable that depends on the uncertainties vector, i.e.
the load profiles;
Vn t , = Vn t , . In conventional MC implementations, the sta-
2. we describe a simulation framework where the gPC + ST algorithm,
implemented in Matlab, is interfaced with the deterministic load
flow solver OpenDSS [20,21], to prove how the metodology is not- tistical description of Vn t, is achieved by generating a very large
invasive and doesn’t require a direct access to the simulation kernel; 1 2 Nmc
3. we show how the proposed method can be exploited to predict the number Nmc of uncertainty vectors , , …, according to the
detailed probability distribution of monitor variables (e.g. Voltages, joint probability distribution of variables in . At time instant tm , For
Currents, Voltage Unbalance or whatever else is needed to be ob- k
served) in the IEEE European low voltage test feeder while ac- each vector , the physical quantity Vn t, , sampled in time over
counting for the interplay of variability in the three phase loads. the tm , can be evaluated by running one deterministic LF analysis. As
the number Nmc of evaluations grows, at limit tending to infinity, the
The above contributions are organized as follows. In Section 2, we distribution of values provided by LF analyses tends to the statistical
review some background material about deterministic load flow ana- distribution of Vn (t ) . However, due to the slow 1/ Nmc convergence rate
lysis and its probabilistic extension with Monte Carlo method. In Sec- of MC method, the number of repeated LF simulations actually needed
tion 3, we describe load uncertainty modeling while in Sections 4 and 5, to obtain a satisfactory statistical description of Vn (t ) (i.e., the detailed
we illustrate the gPC method and its computation details. The im- shape of its PDF) can be very large. The PLF problem is made parti-
plementation and simulation frame are discussed in Section 6. Finally, cularly critical by the nonlinear nature of Eqs. (1). In this case in fact,
Sections 7 and 8 report details about the considered test network and Gaussian-distributed parameters can result in network state variables
the related numerical results.
Vn being not Gaussian distributed. A qualitative example of non
2. Background material: Probabilistic load flow with Monte Carlo Gaussian-distributed variable is shown in Fig. 1. The statistical in-
method ference about the variability interval (with a certain confidence level)

2.1. Deterministic load flow of the network observable Vn requires the complete information
about PDF distribution and it cannot only rely on mean value and
We consider the model of a distribution network made of N buses, variance. As a result, even in the case of a few uncertainty parameters
represented by nodes, and connected by Nl lines described by their (e.g. 2 or 3), several tens of thousands deterministic LF analyses are
impedances. At each candidate node (a candidate is a node of the required, making MC approach time consuming.
network in wich we assign the power load profile), equipment are
connected that may supply or remove power from the electric network.
3. Modeling load uncertainty
This is described by power profiles assigned to the candidate nodes (e.g
it is possible to choose the number of the nodes subjected to the var-
Current practice in probabilistic load modeling relies on an ob-
iation). Deterministic load flow analysis consists in calculating Voltages
servational approach where power profile datasets for different type of
and Currents by solving a set of nonlinear equations of the type:
utilities, area, and time periods are collected and analyzed in order to
N
extract the relevant information that should be reproduced in simula-
Fn (V ) = Sn Vn Yni V i = 0
(1) tions. The theme of loads or energy resources forecasting is a great issue
i=1
that goes beyond the scope of this article. There are several approaches
for n = 1, …, N . In (1), Sn = Pn + jQn denotes the complex power at to forecast load variation with most of them being based on users be-
node n where Pn and Qn are the active and reactive powers respectively, havior analysis [23,24] or on historical data [25], that could be used to
Vn is the node voltage phasor, while Yni are the entries of the bus ad- improve the probabilistic analysis. Here we will suppose to start from
mittance matrix. Node voltage phasors are collected into vector V . the loads presented in the IEEE European low voltage test feeder [26]
Network terminations are specified at the buses by imposing the introducing for each of them, or for groups of loads, a variation de-
known active and reactive powers Pn, Qn absorved or delivered by scribed by a statistical parameter. In order to account for the potential
loads. Load conditions vary in time and thus the associated powers growth or reduction in the power demand, we adopt the following
become function of time, Pn (t ), Qn (t ). Let us consider a given time expression for the active power at nth node in the network:
window (e.g. a day or a week), that is discretized into a sequence of
Ntimes equally-spaced time instants tm = m · t , over which the load Pn (t ) = pn0 (t )[1 + n
p p]
(2)
profiles are given. Node voltage waveforms Vn (t ) are calculated by re-
peatedly solving the nonlinear problem (1) over the sequence of time
instants tm . In doing that, the network state computed at time tm is used
as the solver initialization at next time tm + 1.

2.2. Monte Carlo analysis

Statistical fluctuations and uncertainty of loads can be accounted for

by including into the load flow analysis a set of l stochastic parameters

r that can be collected in the vector = 1, 2, …, l . More details Fig. 1. Qualitative example of a nonGaussian-distributed state variable in the
about probabilistic load modeling will be provided in the next section. network: the area beneath the curve, over a given interval, provides the
Mathematically, each r is a zero-mean stochastic variable described by probability for that variable of falling within the interval.

393
G. Gruosso et al. Electrical Power and Energy Systems 106 (2019) 392–400

where pn0 (t ) is the nominal power profile. In (2), p denotes a zero-mean

Gaussian-distributed statistical parameter having unitary variance, i.e.
p = 0 and ( p )2 = 1 where · is the expectation operator. The

parameter np is a scaling constant that determines the coefficient of

variation.
As a consequence, the active power is a stochastic process whose
mean value and standard deviation are given by [22]:

Pn (t ) = pn0 (t )
(Pn (t ) pn0 (t ))2 = p 0
n pn (t ). (3)

In the analysis that follows, we also consider uncertainty in the power

factor cos( n) by adopting the following statistical model:

cos( n) = cos( n0)[1 + n ], (4)

where cos( is the nominal power factor at nth node while is a zero-
0
n) Fig. 2. Univariate Hermite polynomials and the set of multivariate gPC basis
mean unitary-variance Gaussian-distributed statistical parameter and
functions for the case of two Gaussian-distributed parameters and
n is the scaling constant that determines the power factor degree of
= 1, 2

variability. truncation order = 3.

4. Uncertainty quantification with generalized Polynomial Chaos

order , the degrees ir of univariate polynomials in (6) forming Hi ,
We consider a probabilistic problem where the uncertainty in the
for r = 1, …, l , satisfy the following relation
load power profiles is described by means of l stochastic parameters r
modeling active power and power factor variability as in (2) and (4), l
ir .
respectively. The gPG method consists in adopting generalized poly- (7)
r=1
nomial chaos expansions for the node voltages. Depending on the nu-
merical technique used to solve the gPC problem, the variables that As an example, in Fig. 2, we report the case of two independent
have to be expanded can be all of the node voltages (i.e. the complex Gaussian-distributed parameters 1 and 2 (i.e. l = 2 ) and for expansion
phasors including magnitude and phase information) in the network or order = 3 . In this simple case, the basis functions are the product of
a subset of them. In some cases, the variables that we need to expand couples of Hermite polynomials
are limited to the quantities that we want to monitor: they may be the
magnitude of some node voltages or line currents at a given time or the Hi = i1 ( 1) i2 ( 2 )
peak or minimum value assumed over the time window. In what fol- (8)

lows, we will generically denote as V t, one of such variable. Under whose degrees are such that i1 + i2 3 . In this example, the number Nb

of basis functions Hi is ten. For generic truncation order and

the mild hypothesis that V t, has finite variance (i.e. it is a second-
number of parameters l, the number of gPC basis functions is given by
order stochastic process), it can be approximated by an order- trun- [17]
cated series [16] ( + l)!
Nb = .
Nb ! l! (9)
V t, ci (t ) Hi ,
i=1 (5) Once the coefficients cj (t ) are computed, (using one of the methods
described in the next section) the mean value and standard deviation of
formed by Nb multi-variate basis functions Hi weighted by un- V t, can easily be determined [17]. Furthermore, and even more
known polynomial chaos coefficients ci (t ) . The main feature in ex- importantly, the gPC expansion (5) provides a compact model for the
pression (5) is that the dependence of V (·) on the deterministic variable
time, which is incorporated into coefficients ci (t ) , is separated by its V t, multi-dimensional dependence. For each realization of the

dependence on statistical parameters represented by basis functions

uncertainty vector = 1, 2, …, l , generated accordingly to the joint
Hi .
probability distribution of variables in , the evaluation of polynomials
Each multi-variate basis function is given by the product
l in Fig. 2 and gPC expansion (5) provides a realization of V t, . This
Hi = ir ( r )
(6)
r=1
enables repeated evaluations of V t, for large numbers of un-
where ir ( r ) is a univariate orthogonal polynomial of degree ir whose k
form depends on the density function of the rth parameter r . For in- certainty vector realizations in very short times (one million of
stance, ir ( r ) are Hermite polynomials if r is a Gaussian-distributed evaluations take a few seconds on a quad-core computer) and the de-
variable, while ir ( r ) are Legendre polynomials if r is a uniformly termination of the detailed PDF.
distributed variable. A complete list of correspondence between several
typical stochastic distributions and associated orthogonal polynomials 5. Computing the gPC coefficients
can be found in [16].
For a given number of parameters l and series expansion truncation There are two different mainstream approaches for computing the

394
G. Gruosso et al. Electrical Power and Energy Systems 106 (2019) 392–400

gPC expansion coefficients in (5): Galerkin projection and collocation 1 1

method [27]. H1 … HNb

M= .
5.1. Galerkin Projection (GP)
Ns Ns
H1 … HNb
(13)
Galerkin projection is an intrusive numerical technique that requires
modifying the LF code (1). According to this method, a gPC expansion It is worth noting that the M only depends on the selected basis func-
of the type (5) is adopted for each unknown nodal voltage Vn (t ), i.e. tions and testing points, so that it can be precalculated, inverted and
used for any t = tm as follows:
Nb
Vn t , cin (t ) Hi ,
i=1 (10) c (tm) = M 1 V (tm). (14)

leading to Nb × N unknown cni coefficients that are complex variables. The ST method enables handling PLF problems with larger size and
Such coefficients are determined by plugging the expansions (10) into larger number of parameters. As an example, for expansion order = 3
(1) and then projecting the resulting nodal equations along the Nb basis and number of stochastic parameters l = 6, the ST method needs only
functions. This results in a very large nonlinear system of size Nb × N , repeating 84 deterministic LF analysis.
i.e.

5.3. Testing points selection

Fn V , Hi = 0,
(11) k
The selection of the testing points in the stochastic space is done
for n = 1, …, N and i = 1, …, Nb , where · denotes the inner product in so as to ensure the highest numerical accuracy of the gPC-based in-
the stochastic space [17]. The solution of (11) requires a significant terpolation scheme and of the associated statistical description. This is
computational effort both in terms of time and allocated memory. As an achieved by considering the highest order univariate polynomial ( r )
example, consider the case of a distribution network with N = 100 describing the rth parameter r with PDF r ( r ) . For the univariate case,
nodes, and suppose to perform stochastic Galerkin with l = 3 statistical the testing points are selected in correspondence of the (P + 1) Gauss
parameters and expansion order = 3 . In this case, Nb = 20 , so that the quadrature nodes used in numerical integration rk [17]. When the
nonlinear system to be solved has size Nb × N = 2000 . Due to the pro- multivariate case with l parameters is considered, the testing points
k
blem nonlinearity, equations (11) require a computational time for vectors = k k k
are determined by considering the multi-
1, 2, …, l
solving the system that tends to grow as a power of two of the system
size, i.e. about (Nb × N )2 . dimensional grid of all of the possible combinations (i.e. the tensor
In our example, the computational time for solving Galerkin pro- product) of the univariate quadrature nodes.
blem is about 400× greater than that needed for a single LF analysis. The number ( + 1)l of nodes in the multi-dimensional grid is
The GP computational time grows very rapidly with the number of greater than the number Nb of basis functions defined in (9). For the
statistical parameters thus limiting the applicability of the method to considered example with l = 2 and = 3 , the number of Gauss nodes is
networks of small size and with a few statistical parameters (i.e. 2 3). 4 × 4 = 16 while the number of basis is Nb = 10 . To make problem (13)
well posed, a subset formed by Ns = Nb quadrature nodes has to be
selected as testing points. A possible method for selecting the subset of
5.2. Stochastic Collocation (SC) testing points among the quadrature nodes is presented in [17]. It relies
on the criteria of preferring those quadrature nodes with largest asso-
SC is a nonintrusive method that can be combined with any LF for- ciated Gauss weights and that lead to the best (smallest) condition
mulation (1) without modifying the implementation codes. A second number for the matrix M. Fig. 3 shows the subset of testing points se-
advantage of SC method is that the gPC expansion (5) is adopted lim- lected for the example with 2 stochastic variables.
itedly to the set of network variables that we want to evaluate, i.e. the
peak value of some monitoring node voltages. In what follows, we will
focus on a recently proposed efficient implementation of SC method
referred to as Stochastic Testing (ST) method. According to collocation-
based Stochastic Testing (ST) [17], the Nb unknown coefficients cj (t ) in
the series (5) are calculated by properly selecting Ns = Nb testing points
k k
, for k = 1, …, Ns in the stochastic space where Vk (t ) = V t , is
calculated with a deterministic LF analysis.
At each testing point, the series expansion (5) is enforced to fit ex-
actly (i.e., the polynomials interpolate the samples) the values Vk (t ) .
Mathematically, this results in the following linear system

M c (t ) = V (t ), (12)

where c (t ) = [c1 (t ), …, c Nb (t )]T and V (t ) = [V1 (t ), …, VNs (t )]T are the

column vectors collecting the unknown coefficients and node voltage
values respectively. Fig. 3. (Black Cross Marker) The 2-dimensional grid of quadrature nodes. (Red
k Circle Marker) Subset of nodes used as testing points in ST method. (For in-
The Nb × Nb square matrix M = {ak, i} = Hi collects the gPC terpretation of the references to colour in this figure legend, the reader is re-
basis functions evaluated at the testing points, i.e. ferred to the web version of this article.)

395
G. Gruosso et al. Electrical Power and Energy Systems 106 (2019) 392–400

Fig. 4. Flowchart of the simulation framework.

Fig. 5. Topology of the IEEE LV European test feeder. The numbered nodes are some of the nodes under observation in the implemented analysis.

Table 1 6. Implementation and simulation framework

Number of loads per phase.
Phase A 21 In our implementation, the uncertainty about active power profiles
and power factor are modeled as described in Section 3 by means of l
Phase B 19 independent Gaussian-distributed random variables r . Hence, varia-
Phase C 15 bility analysis is performed for a set of node voltages and line current

396
G. Gruosso et al. Electrical Power and Energy Systems 106 (2019) 392–400

Fig. 6. Daily time evolution of the three phases at node 898 for one of the
testing points.

magnitudes, considered here as the output variables, by using the non-

intrusive ST method described in Section 4 and subsection V-C. This is
achieved by interfacing the gPC + ST code developed at Massachusetts
Institute of Technology [17] and written in Matlab with the Load Flow
deterministic solver OpenDSS. Fig. 4 shows the qualitative flowchart of
the implemented simulation framework. To this aim, we exploit the
OpenDSS internal interface DCOM which allows bidirectional in-
formation exchange between OpenDSS and Matlab. Such an informa-
tion flow is first used to import in OpenDSS the relevant information
about the load profiles, i.e. the active power time profiles and cos( ) , at
all network nodes, generated with gPC + ST in the Matlab workspace.
Second, the information about the time waveforms of simulated net-
work voltages and currents are exported to Matlab for further proces-
sing. This includes extracting the ci expansion coefficients and using the
compact gPC model (5) to efficiently compute the detailed PDFs of the
network variables of interest.

7. Test network

The distribution network adopted in our simulations is the IEEE

European low voltage test feeder[26]. Such a test network, published
by the Test Feeders Working Group of the Distribution System Analysis
Subcommittee of the Power Systems Analysis, Computing, and Economics
Committee (PSACE), provides a valid benchmark for researchers willing
to study low voltage feeders which are common in Europe. The pre-
viously published test feeders were mainly focused on North American
style systems, which consists in radial, very wide medium voltage
networks with a large number of secondary service transformers, each
serving a couple of houses. In Europe instead the most common dis-
tribution system design is with medium voltage networks still radial, or
weekly meshed, but smaller in amplitude and serving few secondary
substations, which are larger with respect to the North American style.
Fig. 7. PDFs for the peak and minimum values of the voltage phases at node
Each substation supply a radial low voltage network. Residential and
898: subfigures (a), (b) and (c) report phases A, B, C, respectively.
small commercial loads are supplied by these networks.
The topology of the proposed network is reported in Fig. 5. The test
feeder is a radial distribution feeder with a base frequency of 50 Hz, at is neglected and just the series impedance is considered.
240 V (phase voltage)/416 V (line to line voltage), which is typical of
the Italian low voltage distribution systems. 7.1. Load data analysis
The medium voltage system supplying the substation is modeled as
a voltage source with an impedance (Thevenin equivalent). The im- The test distribution network is a 3-phase/4 wire network, with the
pedance is specified by short circuit current. The LV test feeder model is possibility of assigning the terminal powers either as 3-phase or 1-phase
composed of 906 low voltage nodes, connected by 905 branches. The loads. In this work, we assume that all of the powers are given as 1-
network is radial, with 55 load buses. The distribution lines are defined phase loads which are distributed among the three-phase lines, i.e.
by their codes and lengths. To each code specific line impedance and Phase A, Phase B and Phase C, in the numbers reported in Table 1.
shunt admittance values are assigned. Due to the short length of lines Such loads are assigned to 55 nodes and their shapes are provided
(all branches are shorter than a hundred meters) the shunt admittance by the benchmark, as 1 min time series, for a day [26]. The aim of the

397
G. Gruosso et al. Electrical Power and Energy Systems 106 (2019) 392–400

Fig. 8. Detail of the distributions of the Phase-C peak value supplied by gPC and Fig. 11. Case with 6 Uniformly-distributed statistical parameters. PDFs for the
MC (5000 samples) methods. peak values of voltage at node 898 for the three phases A, B, C.

Table 2
Simulation times.
2 variables 3 variables 6 variables

Number of simulations 10 20 84
Total time elapsed [s] 40 80 336
Time per simulation [s] 4 4 4

analysis is to study, in a probabilistic sense, the fluctuations of the node

voltages which are induced by power loads variation in order to assess
the quality of the network, therefore we focus on the Voltage Unbalance
Factor (VUF). The percentage VUF is defined as the ratio of the negative
voltage sequence component Vn to the positive voltage sequence com-
ponent Vp [28], i.e.
Vn
VFU = ·100,
Vp (15)
Fig. 9. Statistical distribution of the average value of the voltage unbalance with
factor.
VAB + a2· VBC + a ·VCA
Vn =
3 (16)

and
VAB + a ·VBC + a2· VCA
Vp = ,
3 (17)

where VAB, VBC , VCA are the phasors of the unbalanced line voltages
while a = exp(j 120°) and a2 = exp(j 240°) .

8. Numerical results

In this section, we present fresh results provided by the application

of the proposed variability analysis to the IEEE European low voltage
test feeder.
Our goal is that of investigating the effects that the variability of the
total power assigned to a given phase induces on the node voltages. To
this aim, and accordingly to load model (2), we assume that the active
powers Pn (t ) of all of the nodes assigned to a given phase line, e.g. Phase
A, are scaled by the same p Gaussian statistical parameter, e.g. Ap . As a
result, three statistically independent parameters Ap, Bp and Cp are si-
mulated. In simulations the variability degrees Ap = Bp = Cp = 0.2 are
Fig. 10. Case with 6 Gaussian-distributed statistical parameters. PDFs for the adopted.This is a simplification that we adopt in order to show the main
peak values of voltage at node 898 for the three phases A, B, C. features of the method. More complex scenarios could be implemented
where an independent statistical parameter is introduced for each load.
Assuming gPC expansion order = 3 , twenty testing points are

398
G. Gruosso et al. Electrical Power and Energy Systems 106 (2019) 392–400

generated in the statistical space and for each one of them a determi- Legendre-chaos polynomials [16].
nistic load flow analysis is performed. Fig. 6 reports the waveforms of Finally, in Table 2 we report the simulation times of the proposed
the three phase voltages at node 898, used here as the monitoring point, variability analysis for the cases l = 2, l = 3 and l = 6 statistical para-
simulated with OpenDSS in one of the testing points (i.e. for a given set meters. The simulation times are dominated by the deterministic load
of parameters ). Such waveforms exhibit sharp fluctuations over the flow simulation with OpenDSS. For the case with l = 3 parameters, one
considered time window with peaks (maxima) and valleys (minima) deterministic load flow analysis takes about 4 s and the whole varia-
that are significantly affected by the potential growth or reduction of bility analysis is completed in about 5 min. By contrast, the same ana-
the loads. The probabilistic evaluation of the achievable daily peak and lysis performed with the reference Monte Carlo method require about
minimum is thus crucial in order to assess the quality of the service 5000 load flow analyses and takes more than 5 h.
provided by the distribution infrastructure. To this aim, with the pro-
posed gPC + ST method we calculate the detailed statistical distribu- 9. Conclusion
tion of the peak and minimum values in each node. As an example,
Fig. 7 shows the statistical distribution of the peak value of voltage at In this paper, we have described an innovative simulation frame-
node 898 for the three phases. For the assumed loads uncertainty, the work for the probabilistic analysis of power distribution networks
peak value of Phase A exhibits wide variability with an almost Gaussian subject to load uncertainty. Our approach employes generalized
distribution. In fact, the peak of Phase A ranges within the interval Polynomial Chaos (gPC) algorithm and Stochastic Testing (ST) method
(254, 257) V with 90% probability. The peak values of Phase B and Phase combined with the deterministic load flow solver OpenDSS. We have
C, fluctuate within narrower intervals, i.e. about (252.5, 254) V, how- shown how the proposed method enables deriving in a very efficient
ever their distributions are nonGaussian. This is due to the nonlinear way the detailed information about the variability of a subset of electric
nature of the LF problem. variables and figure of merits that are relevant for the quality of service.
Similarly the statistical distributions of minimum voltage at the The speed up factor in computation is about 100× compared to standard
same node shows as the greatest variability is seen for the minimum Monte Carlo simulations.
value of Phase B that fluctuates into the interval (230, 243) V with 90%
probability. References
In order to check the accuracy of the gPC + ST method we compare
it with a reference MC method that uses 5000 runs (i.e. deterministic [1] Wang Y, Zhang N, Chen Q, Yang J, Kang C, Huang J. Dependent discrete con-
volution based probabilistic load flow for the active distribution system. IEEE Trans.
load flow analyses) selected with a latin-hypercube sampling. Fig. 8 Sustain. Energy 2017;8(3):1000–9.
shows the PDF for the peak value of the Phase C voltage provided by [2] Singh R, Pal BC, Jabr RA. Statistical representation of distribution system loads
proposed gPC and MC (5000 samples) methods. The two distributions using gaussian mixture model. IEEE Trans. Power Syst. 2010;25(1):29–37.
[3] Nguyen DT. Modeling load uncertainty in distribution network monitoring. IEEE
are almost superimposed, e.g. the Kullback Leibler [29] divergence Trans. Power Syst. 2015;30(5):2321–8.
between them is 0.0161, and the associated standard deviations, i.e. [4] Navarro-Espinosa A, Ochoa LF. Probabilistic impact assessment of low carbon
gPC = 0.572 V and MC = 0.565 V respectively, match within a relative technologies in lv distribution systems. IEEE Trans. Power Syst.
2016;31(3):2192–203.
accuracy of 2% . As a further check, we also use the point estimate
[5] Al H, Zivanovic R, Al-Sarawi SF. Probabilistic hosting capacity for active distribu-
scheme described in [14,30], which adopts a numerically efficient tion networks. IEEE Trans. Industr. Inf. 2017;13(5):2519–32.
samples selection method, to estimate the raw moments of the output [6] Zhang W, Xu Y, Dong Z, Wong KP. Robust security constrained-optimal power flow
variables of interest. The standard deviation predicted by the numeri- using multiple microgrids for corrective control of power systems under un-
certainty. IEEE Trans. Industr. Inf. 2017;13(4):1704–13.
cally-efficient point estimate method for the peak value of Phase-C is [7] Borges CLT, Dias JAS. A model to represent correlated time series in reliability
PEM = 0.410 V so that the relative error compared to reference MC evaluation by non-sequential monte carlo simulation. IEEE Trans. Power Syst.
analysis is about 27% . 2017;32(2):1511–9.
[8] Yang Y, Li S, Li W, Qu M. Power load probability density forecasting using gaussian
Finally, Fig. 9 shows the distribution of the average value (over the process quantile regression. Appl Energy 2018;213:499–509. https://doi.org/10.
day window) of the Voltage Unbalance Factor (VUF) figure of merit: the 1016/j.apenergy.2017.11.035.
average VUF is always smaller than 2% . [9] Xu X, Yan Z. Probabilistic load flow calculation with quasi-monte carlo and multiple
linear regression. Int J Electr Power Energy Syst 2017;88:1–12.
As a second example, we repeat variability analysis also considering [10] Mohammadi M, Shayegani A, Adaminejad H. A new approach of point estimate
the uncertainty of the power factors for the three phases. We thus add method for probabilistic load flow. Int J Electr Power Energy Syst 2013;51:54–60.
three extra statistically-independent Gaussian distributed parameters [11] Amid P, Crawford C. A cumulant-tensor based probabilistic load flow method. IEEE
Trans Power Syst 2018:1. https://doi.org/10.1109/TPWRS.2018.2811707.
A , B and C scaling power factor accordingly to (4). We assume [12] Chen P, Chen Z, Bak-Jensen B. Probabilistic load flow: a review. In: 2008 Third
nominal power factor values cos( A0 ) = cos( B0 ) = cos( C0) = 0.925 and International Conference on Electric Utility Deregulation and Restructuring and
associated scaling constants A = B = C = 0.01. Such scaling con- Power Technologies; 2008. p. 1586–91. doi: 10.1109/DRPT.2008.4523658.
[13] Giraldo JS, Castrilln JA, Mauricio GE, Castro CA. Efficient probabilistic power flow
stants meet the constraint 0.9 cos( ) 1. We perform variability for weakly-meshed distribution networks. In: 2014 IEEE PES Transmission
analysis considering the mutual effect of the six statistical parameters. Distribution Conference and Exposition – Latin America (PES T D-LA), 2014, pp.
The simulated distributions of the peak values at monitoring node are 1–6. doi: 10.1109/TDC-LA.2014.6955231.
[14] Morales JM, Perez-Ruiz J. Point estimate schemes to solve the probabilistic power
collected in Fig. 10. The distributions for the case with six parameters flow. IEEE Trans Power Syst 2007;22(4):1594–601.
are quite similar to those for three parameters shown in (7) meaning [15] Fan M, Vittal V, Heydt GT, Ayyanar R. Probabilistic power flow studies for trans-
that, for the considered scenario, the power factor uncertainty is less mission systems with photovoltaic generation using cumulants. IEEE Trans Power
Syst 2012;27(4):2251–61. https://doi.org/10.1109/TPWRS.2012.2190533.
relevant than active power variability. Even though in this paper we
[16] Xiu D, Karniadakis GE. The Wiener–Askey polynomial chaos for stochastic differ-
have focused on the case of Gaussian-distributed parameters, since this ential equations. SIAM J Sci Comput 2002;24(2):619–44.
is the most frequent case in applications, more complex scenarios may [17] Zhang Z, El-Moselhy TA, Elfadel IM, Daniel L. Stochastic testing method for tran-
sistor-level uncertainty quantification based on generalized polynomial chaos. IEEE
arise where statistical parameters are not Gaussian distributed [31,32].
Trans Comput Aided Des Integr Circ Syst 2013;32(10):1533–45.
The generalized Polynomial Chaos method allows handling statis- [18] Wu H, Zhou Y, Dong S, Song Y. Probabilistic load flow based on generalized
tical parameters with several nonGaussian statistical distributions as polynomial chaos. IEEE Trans Power Syst 2017;32(1):820–1. https://doi.org/10.
listed in [16], and complex combinations of them [33]. As an example, 1109/TPWRS.2016.2543143.
[19] Ni F, Nguyen PH, Cobben JFG. Basis-adaptive sparse polynomial chaos expansion
Fig. 11 shows the distributions of the peak values at monitoring node in for probabilistic power flow. IEEE Trans Power Syst 2017;32(1):694–704.
the case where the six parameters Ap, Bp, Cp and A , B , C scaling active [20] EPRI, Opendss. URL http://smartgrid.epri.com/SimulationTool.aspx.
powers and power factors, respectively, are all uniformly distributed [21] Dugan RC. Reference guide: the open distribution system simulator (opendss).
Electric Power Research Institute, Inc,, vol. 7 ed.; 2012.
into the interval [ 1, 1]. In this case, the expansion (5) is made of [22] Papoulis A, Pillai U. Probability, random variables and stochastic processes. 4th ed.

399
G. Gruosso et al. Electrical Power and Energy Systems 106 (2019) 392–400

McGraw-Hill; 2001. [29] Rubner Y, Tomasi C, Guibas LJ. The earth mover’s distance as a metric for image
[23] Bizzozero F, Gruosso G, Vezzini N. A time-of-use-based residential electricity de- retrieval. Int J Comput Vis 2000;40(2):99–121. https://doi.org/10.1023/
mand model for smart grid applications. In: 2016 IEEE 16th International A:1026543900054.
Conference on Environment and Electrical Engineering (EEEIC); 2016. p. 1–6. [30] Gupta N, Pant V, Das B. Probabilistic load flow incorporating generator reactive
[24] Muratori M, Rizzoni G. Residential demand response: dynamic energy management power limit violations with spline based reconstruction method. Electr Pow Syst Res
and time-varying electricity pricing. IEEE Trans Power Syst 2016;31(2):1108–17. 2014;106:203–13.
[25] Xu FY, Lai LL. Novel active time-based demand response for industrial consumers in [31] Valverde G, Saric AT, Terzija V. Probabilistic load flow with non-gaussian corre-
smart grid. IEEE Trans Industr Inf 2015;11(6):1564–73. https://doi.org/10.1109/ lated random variables using gaussian mixture models. IET Gener Transm Distrib
TII.2015.2446759. 2012;6(7):701–9. https://doi.org/10.1049/iet-gtd.2011.0545.
[26] IEEE. Distribution test feeders (may 2015). URL http://ewh.ieee.org/soc/pes/ [32] Prusty BR, Jena D. Combined cumulant and gaussian mixture approximation for
dsacom/testfeeders/index.html. correlated probabilistic load flow studies: a new approach. CSEE J Power Energy
[27] Sandu A, Sandu C, Ahmadian M. Modeling multibody systems with uncertainties. Syst 2016;2(2):71–8. https://doi.org/10.17775/CSEEJPES.2016.00024.
Part i: Theoretical and computational aspects. Multibody Sys Dyn [33] Zhang Z, El-Moselhy TA, Elfadel IM, Daniel L. Calculation of generalized poly-
2006;15(4):369–91. nomial-chaos basis functions and gauss quadrature rules in hierarchical uncertainty
[28] Pillay P, Manyage M. Definitions of voltage unbalance. IEEE Pow Eng Rev quantification. IEEE Trans Comput Aided Des Integr Circuits Syst
2001;21(5):49–51. https://doi.org/10.1109/39.920965. 2014;33(5):728–40. https://doi.org/10.1109/TCAD.2013.2295818.

400