Effects of Correlated Photovoltaic Power and Load Uncertainties On Grid-Connected Microgrid Day-Ahead Scheduling

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 8

IET Generation, Transmission & Distribution

Research Article

Effects of correlated photovoltaic power and ISSN 1751-8687


Received on 16th March 2017
Revised 20th May 2017
load uncertainties on grid-connected Accepted on 15th June 2017
E-First on 19th June 2017
microgrid day-ahead scheduling doi: 10.1049/iet-gtd.2017.0427
www.ietdl.org

Shichao Liu1,2, Peter Xiaoping Liu1,2 , Xiaoyu Wang3,4, Zhijun Wang3, Wenchao Meng3
1School of Mechanical, Electronic, and Control Engineering, Beijing Jiaotong University, Beijing 100044, People's Republic of China
2Department of Systems and Computer Engineering, Carleton University, Ottawa, ON K1S 5B6, Canada
3Department of Electronics, Carleton University, Ottawa, ON K1S 5B6, Canada
4State Key Laboratory of Power Transmission Equipment and System Security and New Technology, Chongqing University, Chongqing,

People's Republic of China


E-mail: xpliu@sce.carleton.ca

Abstract: Due to the increasing integration of photovoltaic-based distributed generators (PV-DGs), uncertainties resulted from
both PV-DG power and loads have posed a serious challenge in microgrid day-ahead scheduling and operation. In this study,
the effect of uncertainties in both PV-DG power and loads on the microgrid day-ahead scheduling is assessed. Specifically, the
correlation between the PV-DG power and load uncertainties is taken into account as this is closer to the reality. The
probabilistic optimal power flow (P-OPF) model is formulated to analyse the impact of the correlated PV-DG power and load
uncertainties. A modified Harr's two-point estimation method (MH-2PEM) is introduced to provide computation-efficient
estimation of the P-OPF solution. Results obtained by using the MH-2PEM and Monte Carlo simulation are compared in an
equivalent 44 kV distribution feeder system and the accuracy and efficiency of the MH-2PEM are verified. The variation ranges
of the microgrid day-ahead scheduling solution resulted from uncertainties in PV power and load are obtained with various
confidence levels.

 Nomenclature Due to the extensive distribution generation (DG) integration, the


distribution system operation and control becomes quite
PCi, t(QCi, t) active (reactive) power output of the ith challenging. To better manage these DGs, a microgrid system
controllable DG at tth time interval which is usually consisted of DGs, distributed storage units and
ci, t generation price of the ith controllable DG at loads has been introduced [1–3]. Among a variety of DGs in
tth time interval microgrid systems, an increasing amount of photovoltaic-based
PDi, t(QDi, t) active (reactive) power demand of the ith load distributed generators (PV-DGs) are integrated to reduce pollution
at tth time interval and improving environmental sustainability. Taking Canada as an
PNi, t(QNi, t) active (reactive) power of the ith example, in 2010, the total PV installation capacity was 281.13 
uncontrollable DG at tth time interval MW [4]. In 2013, the total Canadian PV installation capacity was
Pgrid, t active power injected from the grid at tth time significantly increased to 1210.48 MW. This rapid growth of PV-
interval DG penetration makes it difficult to obtain accurate microgrid
cgrid, t electricity price at tth time interval operation decisions [5–7]. The reason lies in the fact that both the
θi, j, t voltage angular difference between bus i and power output from PV-DGs and load are weather dependent and
bus j at tth time interval inherently variable. For example, Fig. 1 shows the variations of PV
Gi, j, t, Bi, j, t elements of the admittance matrix at tth time power and total load at 10 AM in both winter and summer days
interval during a 2-year period in the city of Ottawa. It is observed that the
SLi, j, t instantaneous power in the branch connecting PV power and load at the same hour are varying from day to day.
bus i and bus j at tth time interval Therefore, forecasted PV-DG power output and load are
V i, t voltage at the ith bus at tth time interval unavoidably accompanied with forecast errors or uncertainties [8–
θi, j, min, θi, j, max lower and upper limits of the voltage angular 10]. Based on forecasted data, the microgrid day-ahead power
difference scheduling dispatches available power production of DGs to
SLmin, SLmax minimum and maximum instantaneous power balance load demand at a minimum economic cost within the
in the branch connecting bus i and bus j system operation constraints during next 24 h. The uncertainties in
PCimin, PCimax minimum and maximum active power of PV-DG power and load will consequently result in the random
controllable DG i variation of the microgrid day-ahead scheduling decision. Without
QCimin, QCimax minimum and maximum reactive power of knowing the variation ranges of these day-ahead scheduling
controllable DG i decisions, significant power mismatches may be caused by the
PNimin, PNimax minimum and maximum active power of unexpected PV-DG and load variations.
uncontrollable DG i In terms of the microgrid day-ahead scheduling and operation, a
QNimin, QNimax minimum and maximum reactive power of variety of optimal power flow (OPF) models have been formulated
uncontrollable DG i and solved in the literature. A three-phase alternative current OPF
PGridimin, PGridimax minimum and maximum active power injected is formulated and solved by using dynamic programming method
to or absorbed from the grid for optimising overall power consumption in an microgrid with
V imin, V imax minimum and maximum voltage at bus i energy storage devices [11]. By modelling a microgrid as a linear
network, direct current OPF problems have been considered and
1 Introduction solved through a number of methods, such as linear programming
[12] and two-stage optimisation [13], multi-agent method [14],

IET Gener. Transm. Distrib., 2017, Vol. 11 Iss. 14, pp. 3620-3627 3620
© The Institution of Engineering and Technology 2017
generation from renewable-source DGs such as wind turbines and
PV generators. However, due to the random characteristic of PV
and wind power generation, forecast uncertainties are inherently
unavoidable. Therefore, it is valuable to know the variable range of
the scheduled DG power references when considering the PV and
load uncertainties. To achieve this purpose, the microgrid day-
ahead scheduling is formulated as a P-OPF problem. In specific,
the correlated PV power and load uncertainties are taken into
account.
For the analysis simplification, the following assumptions are
made in terms of this P-OPF model:

• Assumption 1: The time interval is chosen as 1 h. Within 1 h, the


load, power generation and generation price are assumed to be
constant.
• Assumption 2: We mainly focus on assessing the effect of
Fig. 1  Variations of PV power and load at 10 AM in winter and summer correlated uncertainties in both PV-DG power and loads on the
days during the 2-year period of 1 November 2003- to 31 October 2005 microgrid day-ahead scheduling through the proposed stochastic
approach. For the purpose of demonstration, the linear cost
stability-constrained optimisation [15] and so on. Although these function related to the active power scheduling is considered.
methods are very helpful in the microgrid scheduling and • Assumption 3: To model the load and PV power uncertainties,
operation, most of them are based on deterministic OPF and normal distributions are used in this work. Due to the normal
uncertainties in forecast generation and load data have been distributions are symmetric, the third moment (skewness) of the
ignored. To incorporate these uncertainties in the microgrid day- output variables is zero. Therefore, only the first two moments
ahead scheduling and operation, probabilistic methods may be (mean and variance) are calculated in the proposed ME-2PEM
more proper choices. One extensively used probabilistic method is scheme.
the Monte Carlo simulation (MCS) method. This method tries to • Assumption 4: The slack bus voltage is considered to be constant
exhaust all possible operating scenarios of the uncertain input when and the reactive power cost is not taken into account as the
analysing the power system. In [16], the MCS method is used for microgrid DG scheduling mainly considers the active power
evaluating reliability of microgrids considering the randomness of generation cost.
the DG operation. Although the MCS method can provide accurate
and reliable results, the extensively repeated study is extremely 2.1 Deterministic grid-connected microgird day-ahead
time consuming. Recently, a number of probabilistic OPF (P-OPF) scheduling formulation
models have been proposed for the day-ahead microgrid
scheduling. For instance, in [17] and [18], P-OPF models are The deterministic grid-connected microgrid day-ahead scheduling
formulated when considering the variation of load uncertainty. In is formulated as the following constrained non-linear optimisation
[19], another P-OPF model is proposed for microgrid day-ahead problem:
battery scheduling when the wind power uncertainty is taken into
N
∑t ∑i
24
account. While the above P-OPF models are very promising in min Cost = (ci, tPCi, t + cgrid, tPgrid, t)
microgrid day-ahead scheduling, the correlated PV-DG and load PCi =1 =1

uncertainties have not been considered simultaneously. In reality, constraints:


as weather conditions can affect both PV power output and load
demand, their uncertainties may be correlated with each other. PCi, t + PNi, t − PDi, t = V i, tV j, t(Gi, j, tcos(θi, j, t) + Bi, j, tsin(θi, j, t)),
According to [20], the correlation between wind and load demand QCi, t + QNi, t − QDi, t = V i, tV j, t(Gi, j, tsin(θi, j, t) − Bi, j, tcos(θi, j, t)),
variations in a distribution system has been found when conducting
probabilistic load flow. SLmin ≤ SLi, j, t ≤ SLmax,
In this paper, the impact of correlated uncertainties in PV-DG PCimin ≤ PCi, t ≤ PCimax, (1)
power and load on the microgrid day-ahead scheduling is studied.
QCimin ≤ QCi, t ≤ QCimax,
By formulating the microgrid scheduling as a P-OPF problem, a
modified Harr's two-point estimation method (MH-2PEM) is PNimin ≤ PNi, t ≤ PNimax,
introduced to approximate the mean and variation of the P-OPF QNimin ≤ QNi, t ≤ QNimax,
solution in a computation-efficient way. Then, the variation ranges
of the microgrid day-ahead scheduling solutions can be obtained. PGridmin ≤ Pgrid, t ≤ PGridmax,
An equivalent 44 kV distribution feeder system is tested to verify V imin ≤ V i, t ≤ V imax,
the effectiveness of the MH-2PEM by comparing the results with
MCS method. In addition, via the MH-2PEM, the impact of θi, j, min ≤ θi, j, t ≤ θi, j, max .
correlated PV-DG power and load uncertainties on the microgrid
day-ahead scheduling is assessed and the variation ranges of the where the notations are listed in nomenclature.
microgrid day-ahead scheduling solutions are obtained with a The objective of this microgrid day-ahead scheduling is to
variety of confidence levels. minimise the operation cost by assigning optimal power references
The rest of this paper is organised as follows. The P-OPF model PCi, t for all controllable DGs, with forecasted power production
is formulated for analysing the microgrid day-ahead scheduling PNi, t from uncontrollable DGs and load demand PDi, t. This
with correlated PV power and load uncertainties in Section 2. The deterministic DG scheduling with non-linear constraints can be
MH-2PEM is introduced in Section 3. Case studies and validations solved via the interior point method [21]. In this work, the software
are given in Section 4. Finally, conclusions are drawn in Section 5. for solving the OPF problem is Matlab R2015b with the non-linear
optimisation solver fmincon. As seen in the cost function, Pgrid, t is
2 P-OPF problem for microgrid day-ahead active power injected to or absorbed from the main grid. Its value
scheduling can be either positive or negative to ensure the equality constraints
in the OPF problem will always be satisfied. When Pgrid, t is
The objective of the microgrid day-ahead scheduling is to assign positive, it denotes that the power produced by the DGs is not
optimal hourly average power references for all controllable DGs enough to supply load and the deficit in the active power is
within the system at a minimum operation cost in the next 24 h. It supplied from the main grid. When Pgrid, t is negative, it means that
can be formulated as a OPF problem with forecasted power

IET Gener. Transm. Distrib., 2017, Vol. 11 Iss. 14, pp. 3620-3627 3621
© The Institution of Engineering and Technology 2017
the power produced by the DGs is over the need of loads, it will be are no assumptions on the probability distributions of the input
sold to the main grid. random variables. It calculates the moments of the output variables
with the statistics of the input variables. When the first few
2.2 Probabilistic models for PV power and load moments of solutions are calculated, the probability distributions
[or probability density distributions (PDFs)] of the solutions can be
As the PV power and load demand are random variables, estimated. As assumed in the literature such as [8, 18], the
forecasted PV power and load are usually used in power system uncertain PV power and load are considered to follow normal
operation. However, forecast uncertainties are inherently distributions in this work. In terms of normal distributions, two
unavoidable. As extensively used in literature, such as [8, 18], specific statistic factors including mean and STD can be used to
normal distributions are used to model the uncertain PV power and determine the characteristic of a normal distribution due to the
load. third moment (skewness) is zero (normal distributions are
The uncertain PV power at tth time interval of the operation Pt symmetric and the skewness equals to zero). The variation range of
is modelled by the following normal distribution: the random output variables could be determined by their statistic
characteristics, including mean and STD.
1 2 2
f (Pt) = 2
e(Pt − Pt, F) /2σPt (2)
2πσPt 3 MH-2PEM for solving P-OPF
The uncertainty of PV power output is highly weather dependent.
where f (Pt) is the probability density function (PDF), the mean of Meanwhile, weather conditions can also affect significantly load
Pt is set to the forecasted PV power at the tth interval Pt, F and σPt is demand. Consequently, their uncertainties could be correlated with
the standard deviation (STD) of Pt. each other. The correlation between hourly PV power output and
The following normal distribution is used to model the load demand uncertainties has been ignored when conducting the
uncertain load at tth time interval of the operation PD, t: microgrid day-ahead scheduling in the current literature. When the
uncertainties of PV output and load are identified as either
1 2 2 positively or negatively correlated, the effect of the PV power and
f (PD, t) = 2
e(PD, t − PD, t, F) /2σPD, t (3) load uncertainties on the microgrid day-ahead scheduling could be
2πσPD, t
more precise. In this work, a MH-2PEM has been introduced to
solve the P-OPF with the correlated PV power and load
where f (PD, t) is the PDF, the mean of PD, t is set to the forecasted uncertainties.
load at the tth interval PD, t, F and σPD, t is the STD of PD, t.
3.1 Correlation between PV power and load uncertainties
2.3 P-OPF formulation
The metric used in the correlation study is the correlation
According to [22, 23], a P-OPF problem refers to the OPF with coefficient. The hourly power production of a PV-DG x and hourly
consideration of uncertainty factors. As these factors randomly power demand of a load y are two random variables. Their samples
vary, the solutions of the deterministic OPF will also randomly are denoted by xi (i ∈ {1, 2, …, N}) and yi (i ∈ {1, 2, …, N}). The
change and are implicit functions of these random variables. In correlation coefficient ( ρxy) between these random variables is
order to determine the variable range of the scheduled DG power defined as
references when considering the PV and load uncertainties, the
microgrid day-ahead scheduling is formulated as a P-OPF problem Cov(x, y)
represented by the following general non-linear equation: ρxy = (5)
σ xσ y
Y = h(X) (4) where
The uncertain variable vector is denoted as N
Cov(x, y) = ∑ (xi − μx)(yi − μy) (6)
X = [PN , PD]T i=1

where PN = [PN 1, …, PNM] and PD = [PD1, …, PDM]. The output μx and μy are the mean values of x and y, σx and σy are the variance
variables are denoted as of x and y and Cov(x, y) is the covariance. The correlation
coefficient measures the strength of the related variation between
Y = [PC∗ , q∗, V ∗]T two variables. Its value ranges between −1 and 1. When the
correlation coefficient is positive, the hourly PV power and hourly
load demand vary in the same direction, either simultaneously
where PC∗ = [PC∗ 1, …, PCN

], and q∗ = [θ1,∗ 1, …, θN∗ , N ], and increasing or simultaneously decreasing. On the other hand, when
∗ ∗ ∗
V = [V 1 , …, V M ]. the correlation coefficient is negative, the hourly PV power and
As power production from uncontrollable DGs (such as wind hourly load demand vary in the opposite direction.
and PV) and load demands are inherently variable, uncertainties
exist inevitably in PNi and PDi in (1). These uncertainties existing 3.2 MH-2PEM method for solving P-OPF
in the vector X will consequently result in the random variation of
the solution vector Y . Solving the P-OPF problem is to obtain the To determine the range of output variable uncertainties, the mean
statistical metrics of the random solution vector Y with known and STD are the dominant statistical metrics [24, 25]. The
statistics or sample set of random input vector X . In practice, by procedure of solving the formulated P-OPF problem mainly
obtaining the variation ranges of the microgrid day-ahead involves computing the mean and STD of the OPF solution of the
scheduling solution, the P-OPF analysis is an efficient and reliable P-OPF problem. The steps of solving this P-OPF via the proposed
tool for decision making in microgrid energy management and MH-2PEM method are summarised as follows:
operation.
The main idea of point estimate methods [e.g. two-point • Step 1: Determine the number of random input variables n. The
estimate method (2PEM)] for solving the P-OPF is to estimate the uncertain input variable vector is X = [X1, X2, …, Xn].
first few moments [including mean (first moment), STD (second • Step 2: Calculate the mean μXi and STD σXi of each random
moment), skewness (third moment) etc.] of the output random input variable Xi, i ∈ {1, 2, …, n} based on the obtained
variables by using the known or calculated statistic parameters of historical data.
the input random variables in the OPF model. In the 2PEM, there

3622 IET Gener. Transm. Distrib., 2017, Vol. 11 Iss. 14, pp. 3620-3627
© The Institution of Engineering and Technology 2017
n 2
E(Y 2) ≃ ∑ ∑ (pk ih(Xk i) )
, ,
2
(16)
k=1i=1
• Step 11: Calculate the mean and STD

μY = E(Y), σY = E(Y 2) − μY2 (17)


• Step 12: Repeat steps 7–11 for k = k + 1 until all the random
input variables are calculated.

As seen in the MH-2PEM scheme, one random variable is


replaced by two deterministic points (concentrations) located on
each side of the mean of the probability distribution function of this
random variable. Therefore, if there are totally n random variables,
only 2n simulations of the deterministic DG scheduling are needed
in the MH-2PEM. As noticed in other 2PEMs, the MH-2PEM also
Fig. 2  Forecasted hourly load and PV power of four typical days in four is effective in large-scale systems when the number of input
seasons random variables in large systems is not too large. However, as
pointed out by the work [26, 27], the estimation accuracy of the
• Step 3: Calculate the correlation matrix RX for the input variable 2PEM decreases as the number of input random variables
vector X based on the historical dataset of each random input increases. When the number of input variables is too large, the
variable. tradeoff between the estimation accuracy and the computational
efficiency has to be considered.
r11 r12 ⋯ r1n
r21 r22 ⋯ r2n 4 Case studies and validations
RX = (7) In this section, an equivalent distribution feeder system at the level
⋮ ⋮ ⋮ ⋮
of 44 kV in the city of Ottawa is tested to evaluate impacts of
rn1 rn2 ⋯ rnn correlated PV power and load uncertainties. Based on the historical
data, the auto-regression moving average (ARMA) technique used
Each element in the matrix is calculated according to (5) and in [28] is adapted to forecast PV power and load. In the following,
(6). correlations between PV power and load variations are analysed
• Step 4: Decompose the correlation matrix RX as via sample data generated from their PDFs. Also, based on the
same set of data, the accuracy of probabilistic analysis results via
RX = VLV T (8) the proposed MH-2PEM is verified by comparing with MCS.
Finally, by using the MH-2PEM, probabilistic analysis of the grid-
where V = [V1, V2, …, Vn] is the eigenvector matrix, connected microgrid day-ahead scheduling is conducted for typical
L = diag{λ1, λ2, …, λn} is the corresponding diagonal eigenvalue days in four seasons. The variation ranges of optimal microgrid
operation results are found with different confidence levels.
matrix with λ1, λ2, …, λn being the eigenvalues.
• Step 5: Set the first two moments of the output variable vector as 4.1 PV power and load forecast
E(Y) = 0, E(Y 2) = 0.
• Step 6: Set k = 1. To forecast PV power and load in typical days at different seasons,
• Step 7: Determine locations of two concentrations the historical weather data, including solar irradiance, temperature
(ci, tPCi, t + cgrid, tPgrid, t) and ϵk, 2 and the concentration and system load data are used for the city of Ottawa, Ontario,
probabilities pk, 1 and pk, 2, according to the following equations: Canada for 2 years (from 1 November 2003, 1 : 00 to 31 October
2005, 24 : 00). The ARMA technique for forecasting wind power
proposed in [28] is used to forecast PV power and load. For more
ϵk, 1 = n (9) details on the ARMA method, readers are suggested to refer [28].
The solar irradiance and temperature data are on an hourly basis
ϵk, 2 = − n (10) and obtained from Canadian Weather Energy and Engineering
Datasets [29]. Based on the solar irradiance and temperature data,
1 the PV power is calculated as follows [30]:
pk, 1 = (11)
2n
P = Ir Aη(1 + γ(T ref − T)) (18)
1
pk, 2 = (12)
2n where Ir is the solar irradiance, T is the cell temperature, T ref is the
• Step 8: Determine the two concentration vectors Xk, 1 and Xk, 2 reference cell temperature, η is the PV generation efficiency, A is
based on the following equations: the area of PV array and γ is the temperature coefficient.
Parameters related to (18) are chosen as T ref = 25∘C, η = 0.8 and
Xk, 1 = μXk + ϵk, 1σXkV k (13) r = 0.005/ ∘C. The PV array area A is assumed as 1000 m2 without
loss of generality.
Xk, 2 = μXk + ϵk, 2σXkV k (14) Fig. 2 shows the forecasted PV power (with PV array area of
1000 m2) and total load in the city of Ottawa in typical days of four
where Vk is the kth eigenvector of the correlation matrix RX. seasons. It can be seen from this figure that the forecasted daily
• Step 9: Run the deterministic DG scheduling (1) for both load demand in winter is the highest among all the seasons, due to
concentrations Xk, i, i ∈ {1, 2}. the heating need. On the other hand, the forecasted daily PV output
• Step 10: Update E(Y) and E(Y 2), according to the following in winter is the weakest among the four seasons.
equations:
4.2 Correlation between PV power variation and load change
n 2
According to the PDF of the probabilistic models of PV power (2)
E(Y) ≃ ∑ ∑ (pk ih(Xk i) , , (15)
and load (3), random hourly PV power and load samples are
k=1i=1
generated by setting the forecasted hourly PV power and load

IET Gener. Transm. Distrib., 2017, Vol. 11 Iss. 14, pp. 3620-3627 3623
© The Institution of Engineering and Technology 2017
Fig. 5  Hourly generation price in Ontario
Fig. 3  Correlation coefficient of daily PV power and load demand
variation
To verify the accuracy of the proposed MH-2PEM, the mean
obtained above as the mean values. By using the power and load and STD of the microgrid DG scheduling solution are calculated
samples, the correlations between uncertainty hourly PV power and via both MCS method and the proposed MH-2PEM. The following
load are calculated via (5) in the four typical days. The results are two cases are considered:
shown in Fig. 3. In winter, as it is illustrated in the figure, the
hourly PV power variation is mostly in positive correlation with • Case 1 – MCS: By using generated random PV and load data as
load changes. Also, in winter, the PV power is zero after 18 : 00 the random inputs, the deterministic microgrid DG scheduling is
and thus is not be able to contribute for supporting the load demand calculated for each random input. Then, the mean and STD of
any more. In addition, in summer, the hourly PV power and hourly the DG scheduling outputs are calculated. As discussed in the
load variation are negatively correlated until 14 : 00 while they are literature, MCS is able to provide accurate results. Its mean and
positively correlated after that until 20 : 00. STD are considered as references.
• Case 2 – MH-2PEM: Based on the same dataset as previous
cases, the mean and STD of DG scheduling outputs are
4.3 Validation of the MH-2PEM via comparisons
calculated by performing the modified two-point estimation
An equivalent 44 kV distribution feeder system in the city of method.
Ottawa (see Fig. 4) is used to evaluate variation ranges of the
microgrid DG scheduling results when the correlated PV power As the MCS results (case 1) are considered as accurate, the
and load uncertainties are taken into account. The utility source is errors for the mean μ and STD σ of the DG scheduling outputs in
230 kV and the 44 kV substation is connected to the grid via a case 2 are calculated as
circuit breaker. The capacity of the substation transformer is 100 
MVA. A 20 Mvar capacitor bank is located at the substation. A PV- |(μMCS − μ)|
ϵμ = (19)
DG is connected at bus 6. A natural gas turbine and a diesel μMCS
generator are connected at bus 7 and bus 8, respectively. Variable
loads are assumed to be sited at bus 7. Each load is seen as the |(σMCS − σ)|
equivalent total load in the bus. ϵσ = (20)
σMCS
The feeder conductor parameters are R = 0.1156 and
X = 0.3708 Ω/km according to practical feeder parameters in the where μMCS and σMCS are the mean and deviation of the DG
distribution system.
scheduling solution for case 1 as the reference. The mean and STD
The random PV power profile generated by the PV power
probabilistic model (2) is used as the power output of the PV-DG at of the scheduled DG power generation PC∗ at bus 8 of the
bus 6. The load profile of the Ottawa city is scaled down by 1000 distribution feeder system in selected hours in winter and summer
times to fit the capacity level in the tested distribution feeder are listed in Tables 1 and 2.
system. However, this action will not change uncertainty As seen from Tables 1 and 2, the MH-2PEM has the very low
characteristics of load. The variable load (load 4) is located at bus estimation error for the mean value of the scheduled DG power
7. generation PC∗ with comparison to MCS. In terms of the
The forecasted hourly prices of generation in Ontario are shown computational efficiency, time spent by the MH-2PEM is 4.24 s,
in Fig. 5, based on the historical data in IESO (the ISO of Ontario) while time spent by MCS is 121.01 s. It can be seen that the
[31]. MH-2PEM significantly reduces the calculation time compared
with the MCS method.

Fig. 4  Equivalent distribution feeder system in Ottawa

3624 IET Gener. Transm. Distrib., 2017, Vol. 11 Iss. 14, pp. 3620-3627
© The Institution of Engineering and Technology 2017
Table 1 Mean and STD of PC∗ in winter day
Hour Method μ, MW σ, MW
10 : 00 MCS 3.6391 0.2515
MH-2PEM 3.6390 0.2516
ϵ 0.0027% 0.0397%
12 : 00 MCS 3.3408 0.3975
MH-2PEM 3.3405 0.4089
ϵ 0.0089% 2.8679%
14 : 00 MCS 3.4384 0.3609
MH-2PEM 3.4382 0.3641
ϵ 0.0058% 0.8866%
16 : 00 MCS 3.8909 0.1547
MH-2PEM 3.8901 0.1499
ϵ 0.0205% 3.1027%

Fig. 6  Variation ranges of PC∗ and FC in winter day


Table 2 Mean and STD of PC∗ in summer day
Hour Method μ, MW σ, MW
10 : 00 MCS 2.5984 0.5857
MH-2PEM 2.5984 0.5902
ϵ 0 0.7683%
12 : 00 MCS 2.4111 0.6316
MH-2PEM 2.4110 0.6359
ϵ 0.0041% 0.6808%
14 : 00 MCS 2.5623 0.5734
MH-2PEM 2.5623 0.5763
ϵ 0.0002% 0.5057%
16 : 00 MCS 2.9753 0.4298
MH-2PEM 2.9753 0.4288
ϵ 0 0.2326%

Fig. 7  Variation ranges of PC∗ and FC in spring day seasons. The results are shown from Figs. 6–9. As it can be seen
from these figures, means and variances of the uncertainty PV
power and load for each hour from 10 : 00 to 16 : 00 (sun hours) of
next day are calculated in the microgrid scheduling problem. By
providing the microgrid operation centre the variation ranges of the
scheduled DG power and cost, the proposed MH-2PEM can be an
efficient and reliable tool to improve the accuracy and confidence
for the decision making in microgrid energy management and
operation when PV power and load uncertainties are taken into
consideration.

4.4 Impact of ranges of controllable DG and grid active


power
To evaluate the influence of the range of decision variables on the
P-OPF model and its solution, solutions of the P-OPF are
calculated by using the same test system based on the load and PV
power data at 16 PM in a winter day. With various active power
ranges of controllable DG, the mean and STD of PC∗ are calculated
Fig. 8  Variation ranges of PC∗ and FC in summer day and shown in Table 3. As seen in Table 3, when the mean of the
optimal active power is beyond its permissible limit (e.g.
After verifying the accuracy of the MH-2PEM, it is used to 0 ≤ PC ≤ 3), the control variable is set to the limit value (e.g.
analyse variation ranges of scheduled DG power generation PC∗ PC = 2.9999). With various active power ranges of grid, the mean
when considering the correlated PV power and load uncertainties. and STD of PC∗ are calculated and listed in Table 4. As seen in this
Based on the same data generated previously, means and variances table, when the grid active power limit is reduced (e.g. from 30 to
(the first two moments) of both the scheduled DG power PC∗ and 25 MW), the solution of the P-OPF will also change (e.g. from
the minimum DG scheduling cost FC are estimated by the 3.8909 to 4.6389 MW) until the controllable DG reach its
MH-2PEM. In this work, the PDFs of the random PC∗ and FC are permissible limit (e.g. PC ≤ 5 MW). These observations are
approximated by their first and second moments (means and consistent with the conclusions made in the literature. According to
variances, respectively) and this indicates two normal distributions [32], these inequality constraints on decision variables assure that
with the calculated means and variances are considered for the these variables do not go beyond their permissible limits when the
random PC∗ and FC. As it is known, for a normal distribution, the OPF is solved by using Newton's method. If one control variable
variable falls within the range [μ − σ, μ + σ] with the probability of exceeds its limit, the control variable is set to the corresponding
68% and [μ − 2σ, μ + 2σ] with the probability of 95%. The limit. By setting the control variable limit, the final optimal
variation ranges of PC∗ and FC are found for typical days in four solution might not be reached. Usually, the DG siting and sizing
are designed by solving OPFs with many constraints when the

IET Gener. Transm. Distrib., 2017, Vol. 11 Iss. 14, pp. 3620-3627 3625
© The Institution of Engineering and Technology 2017
Table 3 Range of active power for controllable DG versus
mean and STD of PC∗
Range μ, MW σ, MW
0 ≤ PC ≤ 5 3.8909 0.1547
0 ≤ PC ≤ 4 3.8391 0.1608
0 ≤ PC ≤ 3 2.9999 1.8758 × 10−6
0 ≤ PC ≤ 2 1.9999 0
0 ≤ PC ≤ 1 0.9999 1.4901 × 10−8

Table 4 Range of active power for grid versus mean and


STD of PC∗
Range μ, MW σ, MW
−35 ≤ Pgrid ≤ 35 3.8909 0.1547

Fig. 9  Variation ranges of PC and FC in fall day −30 ≤ Pgrid ≤ 30 3.8909 0.1557
−25 ≤ Pgrid ≤ 25 4.6389 0.1960
distribution network operators started planning the installation of −20 ≤ Pgrid ≤ 20 4.9999 1.9768 × 10−7
the DG [33, 34]. Therefore, when DGs are scheduled in normal
−15 ≤ Pgrid ≤ 15 4.9999 2.5154 × 10−8
situations, there could be a feasible solution. In case of unplanned
serious contingencies, the feasibility restoration should be needed,
such as [35].
7 References
4.5 Discussions on the MH-2PEM [1] Guerrero, J., Vasquez, J., Matas, J., et al.: ‘Hierarchical control of droop-
controlled ac and dc microgrids – a general approach toward standardization’,
The proposed MH-2PEM is one of 2PEMs. As pointed out by the IEEE Trans. Ind. Electron., 2011, 58, (1), pp. 158–172
previous works [26, 27, 32], the main idea of the 2PEM is used to [2] Liu, S., Wang, X., Liu, P. X.: ‘Impact of communication delays on secondary
estimate the first few moments [including mean (first moment), frequency control in an islanded microgrid’, IEEE Trans. Ind. Electron., 2015,
62, (4), pp. 2021–2031
STD (second moment), skewness (third moment) etc.] of the output [3] Guo, X., Xu, D., Guerrero, J., et al.: ‘Space vector modulation for dc-link
random variables by using the known or calculated statistic current ripple reduction in back-to-back current-source converters for
parameters of the input random variables. In the 2PEM, no microgrid application’, IEEE Trans. Ind. Electron., 2015, 62, (10), pp. 6008–
assumptions are made regarding the probability distributions of the 6013
[4] EPIA: ‘Global market outlook: for photovoltaic 2014–2018’, 2014
input variables. It mainly calculates the moments of the output [5] Wang, Y., Zhang, P., Li, W., et al.: ‘Online overvoltage prevention control of
variables with the statistics of the input variables. As commented in photovoltaic generators in microgrids’, IEEE Trans. Smart Grid, 2012, 3, (4),
[27], while the moments of the output variables are calculated, the pp. 2071–2078
associated probability distributions are not necessary to be known. [6] Fan, B., Yang, Q., Wang, K., et al.: ‘Transient stability enhancement control
of power systems with time-varying constraints’, IET Gener. Transm. Distrib.,
Therefore, the MH-2PEM is also suitable for non-normal 2016, 10, (13), pp. 3251–3263
distribution variables. One point needs to be mentioned is that the [7] Liu, S., Liu, P. X., Wang, X.: ‘Stochastic small-signal stability analysis of
accuracy of the 2PEM will be adequate if the probability grid-connected photovoltaic systems’, IEEE Trans. Ind. Electron., 2016, 63,
distribution of the output variables is smooth, while the estimation (2), pp. 1027–1038
[8] Makarov, Y.V., Etingov, P.V., Ma, J., et al.: ‘Incorporating uncertainty of wind
error will increase when the probability distribution of the output power generation forecast into power system operation, dispatch, and unit
variables is discontinuous [25, 27]. commitment procedures’, IEEE Trans. Sustain. Energy, 2011, 2, (4), pp. 433–
442
[9] Farzan, F., Jafari, M.A., Masiello, R., et al.: ‘Toward optimal day-ahead
5 Conclusions scheduling and operation control of microgrids under uncertainty’, IEEE
Trans. Smart Grid, 2015, 6, (2), pp. 499–507
As the PV power and loads are randomly changing, forecast [10] Wu, W., Wang, K., Li, G., et al.: ‘Probabilistic load flow calculation using
uncertainties are inherent in the microgrid day-ahead scheduling cumulants and multiple integrals’, IET Gener. Transm. Distrib., 2016, 10, (7),
and operation. In this work, the influence of the correlated pp. 1703–1709
uncertainties in PV-DG power and loads on the microgrid day- [11] Levron, Y., Guerrero, J., Beck, Y.: ‘Optimal power flow in microgrids with
energy storage’, IEEE Trans. Power Syst., 2013, 28, (3), pp. 3226–3234
ahead scheduling is analysed. To achieve this objective, a P-OPF [12] Carpinelli, G., Bracale, A., Caramia, P.: ‘The great project: integer linear
model is formulated. A MH-2PEM is introduced to solve the P- programming-based day-ahead optimal scheduling of a dc microgrid’. Proc.
OPF when incorporating the correlation between PV-DG power Int. Conf. Environment and Electrical Engineering, Wroclaw, Poland, May
and load uncertainties. An equivalent 44 kV distribution feeder 2013, pp. 573–578
[13] Kanchev, H., Lu, D., Colas, F., et al.: ‘Energy management and operational
system in the Ottawa grid is used to validate the accuracy and planning of a microgrid with a PV-based active generator for smart grid
computational efficiency of the MH-2PEM. The results show the applications’, IEEE Trans. Ind. Electron., 2011, 58, (10), pp. 4583–4592
MH-2PEM can precisely estimate the means and variations of the [14] Dou, C.X., Jia, X.B., Li, H., et al.: ‘Multi-agent system based energy
microgrid day-ahead scheduling solutions within 5 s. The variation management of microgrid on day-ahead market transaction’, Electr. Power
Compon. Syst., 2016, 44, (12), pp. 1330–1344
ranges with different confidence levels are obtained for the [15] Zhang, J., Lu, C., Chung, C.Y., et al.: ‘Online re-dispatching of power
microgrid day-ahead scheduling solutions. As one of our future systems based on modal sensitivity identification’, IET Gener. Transm.
work, the MH-2PEM is going to be applied in islanded microgrids. Distrib., 2015, 9, (12), pp. 1352–1360
[16] Kennedy, S., Marden, M.: ‘Reliability of islanded microgrids with stochastic
generation and prioritized load’. Proc. Int. Conf. PowerTech, Bucharest,
6 Acknowledgments Romania, 2009, pp. 1–7
[17] Li, X., Li, Y., Zhang, S.: ‘Analysis of probabilistic optimal power flow taking
This work was partly supported by the Fundamental Research account of the variation of load power’, IEEE Trans. Power Syst., 2008, 23,
Funds for the Central Universities, Beijing Natural Science (3), pp. 992–999
Foundation under project 4172048 and Jiangxi Provincial [18] Keun Jo, B., Han, J.H., Guo, Q., et al.: ‘Probabilistic optimal power flow
analysis with undertermined loads’, J. Int. Electr. Eng., 2011, 2, (3), pp. 321–
Department of Science and Technology under grants 325
20121BBE50023, 20133BCB22002 and 612049302056. The work [19] Xu, Q., Zhang, N., Kang, C., et al.: ‘Day-ahead battery scheduling in
of Prof. Xiaoyu Wang was partly supported by the Visiting microgrid considering wind power uncertainty using ordinal optimization’.
Scholarship of State Key Laboratory of Power Transmission Proc. North American Power Symp., Pullman, Washington, 2014, pp. 1–3
Equipment and System Security and New Technology (Chongqing
University) (2007DA10512716405).
3626 IET Gener. Transm. Distrib., 2017, Vol. 11 Iss. 14, pp. 3620-3627
© The Institution of Engineering and Technology 2017
[20] Qin, Z., Li, W., Xiong, X.: ‘Incorporating multiple correlations among wind [28] Wangdee, W., Billinto, R.: ‘Considering load-carrying capability and wind
speeds, photovoltaic powers and bus loads in composite system reliability speed correlation of wecs in generation adequacy assessment’, IEEE Trans.
evaluation’, Appl. Energy, 2013, 110, (10), pp. 285–294 Energy Convers., 2006, 21, (3), pp. 734–741
[21] Wei, H., Sasaki, H., Kubokawa, J., et al.: ‘An interior point nonlinear [29] Government of Canada: ‘Canadian weather energy and engineering datasets
programming for optimal power flow problems with a novel data structure’, (CWEEDS)’, 2012
IEEE Trans. Power Syst., 1998, 13, (3), pp. 870–877 [30] NREL: ‘Pvwatts version 5 manual’, 2015
[22] Ke, D., Chung, C.Y, Sun, Y.: ‘A novel probabilistic optimal power flow [31] Independent electricity system operator. Available at http://www.ieso.ca/
model with uncertain wind power generation described by customized power-data, accessed May 2015
Gaussian mixture model’, IEEE Trans. Sustain. Energy, 2016, 7, (1), pp. 200– [32] Dommel, H.W., Tinney, W.F.: ‘Optimal power flow solutions’, IEEE Trans.
212 Power Appar. Syst., 1968, PAS-87, (10), pp. 1866–1876
[23] Zhang, H., Li, P.: ‘Probabilistic analysis for optimal power flow under [33] Keane, A., Ochoa, L.F., Borges, C.L.T., et al.: ‘State-of-the-art techniques and
uncertainty’, IET Gener. Transm. Distrib., 2010, 4, (5), pp. 553–561 challenges ahead for distributed generation planning and optimization’, IEEE
[24] Chang, C.H., Tung, Y.K., Yang, J.C.: ‘Evaluation of probability point estimate Trans. Power Syst., 2013, 28, (2), pp. 1493–1502
methods’, Appl. Math. Model., 1995, 19, (2), pp. 95–105 [34] Dent, C.J., Ochoa, L.F., Harrison, G.P.: ‘Network distributed generation
[25] Rosenblueth, E.: ‘Point estimates for probability moments’. Proc. Natl. Acad. capacity analysis using OPF with voltage step constraints’, IEEE Trans.
Sci. USA, 1975, 72, (10), pp. 3812–3814 Power Syst., 2010, 25, (1), pp. 296–304
[26] Morales, J.M., Perez-Ruiz, J.: ‘Point estimate schemes to solve the [35] Tanabe, R., Tada, Y., Chiang, H.D.: ‘A novel optimal power flow solver based
probabilistic power flow’, IEEE Trans. Power Syst., 2007, 22, (4), pp. 1594– on parameterized contingency for preventive control against insecure-
1601 contingencies’. 17th Conf. Power Syst. Computation, Stockholm, Sweden,
[27] Verbic, G., Canizares, C.A.: ‘Probabilistic optimal power flow in electricity 2011, pp. 1–7
markets based on a two-point estimate method’, IEEE Trans. Power Syst.,
2006, 21, (4), pp. 1883–1893

IET Gener. Transm. Distrib., 2017, Vol. 11 Iss. 14, pp. 3620-3627 3627
© The Institution of Engineering and Technology 2017

You might also like