Towards Building an Optimal Demand Response

Framework for DC Distribution Networks

Hamed Mohsenian-Rad, Member, IEEE and Ali Davoudi, Member, IEEE

Abstract— Direct current (DC) power systems have recently Moreover, many of the emerging new loads are electronic DC
been proposed as a promising technology for distribution net- loads (e.g., in data centers). Even some of the traditionally AC
works and microgrids. By eliminating unnecessary conversion loads, such as induction machines, can appear as DC devices
stages, DC distribution systems can enable seamless integration
of natively DC devices such as photovoltaic cells and batteries. when controlled by drive systems [16]. Hence, DC distribution
Moreover, DC technologies can overcome several disadvantages systems improve conversion efficiency by eliminating the AC-
of alternating current (AC) distribution systems, such as synchro- DC conversion stage. DC microgrids are also shown to have
nization requirements and reactive power compensation. There- about two orders-of-magnitude more availability compared
fore, in this paper, the first steps are taken towards designing de- to their AC counterparts, thus making them ideal candidates
mand response programs for DC distribution networks. The idea
is to adjust the internal parameters of power electronics loads for mission critical applications [17], [18]. Finally, using DC
to ensure reliable and efficient operation of the DC distribution technologies in distribution networks can overcome several
system. In this regard, first, an optimization-based foundation is disadvantages of AC technologies, including synchronization
proposed for demand response in DC distribution networks in requirements, reactive power flow, and harmonics [15].
presence of distributed generators. Then, the formulated problem Given the benefits of DC distribution networks and the
is solved using both centralized and decentralized approaches,
where the latter requires devising a pricing mechanism. Finally, important role of demand response in the future smart grid,
simulation results are presented to assess the performance and in this paper, the first steps are taken towards designing
to gain insights into the proposed demand-response paradigm. demand response programs for DC distribution systems and
DC microgrids. In this regard, the recent advancements in
Keywords: DC distribution networks, power electronics load,
demand response, distributed generation, convex optimization.
power electronics [19]–[21] are combined with techniques
from convex optimization and dual decomposition [22] to
build a new foundation for optimal demand response in DC
I. I NTRODUCTION distribution networks and DC microgrids. Both centralized
and decentralized design approaches are investigated, where
Demand response (DR) programs are designed to control
the latter requires devising a pricing mechanism. The main
the consumer resources in response to changes in the grid’s
contributions in this paper can be summarized as follow:
operating conditions [1], [2]. The majority of prior studies on
• A New Demand Response Paradigm: While most prior
demand response do not take into account the characteristics
of the underlying physical power system. Instead, they mainly studies on demand response either do not consider the
focus on balancing the load across time to reduce the load characteristics of the underlying physical power system
at peak hours in large networks, e.g., see [3]–[7]. However, or they assume that the underlying power system is AC, in
there are also some recent studies that have incorporated the this paper, a new demand response paradigm is proposed
impact of grid topology, power flow equations, and voltage that is designed specifically for DC distribution networks.
control into the design of DR programs to tackle the challenges • Centralized Design: A new optimization-based frame-

at the distribution level. So far, the focus has been only on work is developed to select the internal parameters of
alternating current (AC) distribution networks, e.g., see [8], power electronics loads to assure desirable operation
[9]. In contrast, in this paper, demand response programs for of the DC distribution network. This design takes into
direct current (DC) distribution networks are investigated. account the power flow equations, power draw limits
Traditionally, DC power systems have been used in telecom- of DC sources, DC-DC conversion, solar panels, solar
munications, naval ships, and industrial systems [10], [11]. irradiance, and proportional fairness. The formulated op-
More recently, they are proposed also for microgrids [12] timization problem is shown to be convex and tractable.
and distribution networks [13]–[15]. DC distribution networks • Decentralized Design and Pricing: An algorithm and a
can offer several important advantages. For example, they can pricing method are developed to implement the proposed
enable more efficient interconnection of most energy storage demand response design in a decentralized fashion.
units and renewable energy sources, such as photovoltaic (PV) • Computer Simulations: Various computer simulations are
systems, batteries, and fuel cells, that are natively DC sources. conducted to evaluate the designs and to gain insights.
H. Mohsenian-Rad is with the Department of Electrical Engineering at the Compared to the conference paper in [23], the following
University of California, Riverside, CA, USA, e-mail: [email protected]. A. aspects are new in this journal version. First, a detailed model
Davoudi is with the Department of Electrical Engineering at the University for solar panels is incorporated in the problem formulation.
of Texas, Arlington, TX, USA, e-mail: [email protected]. H. Mohsenian-Rad
is the corresponding author for this paper. This work was supported by the Second, the impact of permissible duty cycles of DC-DC
National Science Foundation (NSF) Grants 1253516, 1137354, and 1319798. converters is considered in order to make the new design
 where N \L denotes the set of all non-load buses, Nk ⊂ N
denotes the set of all neighboring buses of bus k, and Rik de-
notes the line resistance between buses i and k. The equalities
in (2) and (3) are the fundamental equations to understand DC
distribution systems and DC microgrids. Note that, in general,
(a) (b) calculating the line resistance Rik can be difficult in practice.
Interested readers can refer to [25] for more details on how
Fig. 1. A DC-DC boost converter can be seen as a power electronics load
with variable resistor. The effective external resistor Ri is controlled by setting to calculate the line resistance in distribution networks using
the duty cycle for switch q [21]. A buck converter can be modeled similarly. catalogue data and synchronous measurements. Other papers
that similarly consider line resistance in the context of DC
more appropriate for a practical implementation. Third, new distribution networks include [26] and [27].
simulation results, in particular in Sections V-D and V-E, are From (2), at each load bus i ∈ L we have
added to further strengthen the analysis.
Vi V s − Vi X Vk − Vi
The rest of this paper is organized as follows. The funda- = i s + .
mental system model is explained in Section II. The optimal Ri Ri Rik
k∈Ni
demand response problem is formulated in Section III. Decen- . Vs−V
! (4)
i
X Vk − Vi
i
tralized design is discussed in Section IV. Simulation results ⇒ Ri = Vi + .
are given in Section V. The paper is concluded in Section VI. Ris Rik
k∈Ni

II. S YSTEM M ODEL Also from (3), at each non-load bus j ∈ N \L, we have
With the increasing penetration of power electronics loads, Vj − Vjs X Vj X Vk X Vk
+ − = . (5)
i.e., loads that are supplied with power electronics converters, Rjs Rjk Rjk Rjk
k∈Nj k∈Nj ,k∈L
/ k∈Nj ,k∈L
the internal load characteristics can be controlled to reflect a
desirable effective impedance [19], [20]. In particular, a power The left hand side in (5) is a linear combination of variables
electronics load with a switch-mode converter can be seen by Vj for all j ∈ N \L; while the right hand side in (5) is a
the rest of the grid as a variable resistor load, as shown in linear combination of variables Vi for all i ∈ L. Therefore,
Fig. 1 [21]. For the boost converter example in this figure, the system of |N | − |L| equations in (5) can be solved to
the relationship between the internal resistance Ro and the obtain the voltages Vj for all j ∈ N \L in terms of voltages
effective external resistance Ri can be written as Vi for all i ∈ L. More specifically, from (5) we can derive:
Ri = (1 − D)2 Ro ,
X
(1) Vj = aji Vi + bj , ∀ j ∈ N \L, (6)
i∈L
where D ∈ [0, 1) denotes the duty cycle of the active switch
component q. A similar variable resistor load model can be where parameters aji and bj are constant. For example,
found for pulse-width modulation (PWM) rectifiers that feed consider the four-bus DC power system in Fig. 2(b). Here,
DC distribution systems from AC sources [24, Chapter 18]. L = {1, 3} and N \L = {2, 4}. It can be shown that
By controlling the variable resistors for all power electronics
loads, we can affect the power delivered to each user, the a21 , 1/(1 + R12 /R23 + R12 /R2s ),
power drawn from each source, the power loss on each line, a23 , 1/(1 + R23 /R12 + R23 /R2s ), (7)
and the voltage level at each bus. Therefore, this system-level b2 , V2s /(1 + R2s /R12 + R2s /R23 ).
formulation can help define a demand-response paradigm.
Parameters a41 , a4,3 , and b4 can be derived similarly.
A. Formulation of a DC Power System
Consider a DC power system such as the examples that B. Power Electronics Loads
are shown in Fig. 2. Let N denote the set of all DC buses. By replacing (6) in (4), at each load bus i ∈ L, we have
A DC power source at bus k is modeled using its Thevenin’s 
representation with a fixed voltage Vks , a fixed internal resistor . Vs−V X Vi Vk
i
X
Rks , and a maximum power limit Pkmax . If there does not exist Ri = V i  i s − +
Ri Rik Rik
any source at bus k, then we simply assume that Rks = ∞. k∈Ni k∈Ni ,k∈L
(8)
Next, let L ⊂ N denote the set of load buses. Each load bus ! 
X X
i ∈ L serves a power electronics load with an effective variable

+ akl Vl + bk Rik  .
resistor Ri . Using the Kirchhoff’s current law, we have k∈Ni ,k∈L
/ l∈L
Vi Vi − Vis X Vi − Vk
+ s + = 0, ∀ i ∈ L, (2) The above equation expresses the variable resistor of each
Ri Ri Rik power electronics load in terms of the voltages at all load
k∈Ni

and buses. Therefore, using (8), together with the rest of the system
model in this section, we can formulate different objective
Vj − Vjs X Vj − Vk
s + = 0, ∀ j ∈ N \L, (3) functions in terms of bus voltages, which can be adjusted by
Rj Rjk changing the effective resistance of converters / rectifiers.
k∈Nj
Fig. 2. Two examples for DC power systems: (a) A five-bus radial power distribution network with four power electronics loads. (b) A four-bus DC microgrid
with two power electronics loads. Demand response is done by changing the variable resistors for the participating power electronics loads.

(a)
55 1000

Solar Irradiance (W/m2)

April 30, 2003
50 1000 W/m2 800
600 W/m2
45 2
600
300 W/m
40 400

35 200

0
Power (W)

30 0 4 8 12 16 20 24

25 (b)
1000
Solar Irradiance (W/m2)

20 Feb 14, 2003

800
15
600
10
400
5
200
0 0
0 5 10 15 20 25 0 4 8 12 16 20 24
Voltage (V) Hour of Day

Fig. 3. The characteristics curves of a solar panel for three irradiance levels. Fig. 4. Two examples for daily solar irradiance in Abilene, TX based on
The MPPT solution is shown by a solid point on each curve [29]. measurements every 15 minutes: (a) A mostly sunny day. (b) A cloudy day.

C. Solar Panels III. O PTIMAL D EMAND R ESPONSE

The focus of the proposed demand response paradigm
The system model explained above can support distributed
in DC distribution networks is to treat power converter as
generation (DG), where PV panels are connected to the DC
power buffer and adjust the variable effective resistance of
distribution network or DC microgrid. Each PV panel is
power electronics loads to assure DC power system reliability
equipped with a constant voltage (CV) control mechanism [28]
and efficiency. The three key intended characteristics of this
to achieve maximum power point tracking (MPPT). Therefore,
demand response paradigm are as follows:
like a conventional generator, a PV panel at bus i is modeled
• The amount of power drawn from the DC power source
as a constant voltage DC source. However, the values of Vis
and Pimax must be updated based on the solar irradiance level. at each source bus should not exceed its supply limit.
That is, given the irradiance level, the CV-based MPPT control • The power electronics loads should support their intended
mechanism adjusts the values of Vis and Pimax according to demand response solutions based on their characteristics.
the voltage-power characteristics curves of the PV panel [29], • The power should be delivered fairly among the loads at
as shown in Fig. 3. For example, when the irradiance level different locations of the DC distribution network.
is 1000 W/ms then Vis = 17.8 V and Pimax = 47.4 W . First, consider the power draw limits from DC sources. For
Once the proposed optimal demand response method is im- the DC power source at each source bus k, the amount of
plemented, whenever the irradiance level changes during the current drawn can be calculated as (Vks − Vk )/Rks . Therefore,
day, the users’ power electronics loads are adjusted to control the power limit requirement can be formulated as Vks (Vks −
their power consumption in order to achieve the best overall Vk )/Rks ≤ Pkmax . In other words, it is required to have
system performance across the DC distribution network. In
practice, such adjustment can be done on a periodic basis. Vk ≥ Vks − Pkmax Rks /Vks , ∀ k ∈ N. (9)
Two examples for daily changes in the solar irradiance are
From (6), one can rewrite the above constraints as
shown in Fig. 4 based on the experimental measurements in
[30]. Vi ≥ ei , ∀ i ∈ L, (10)
and X For each i ∈ L, we have
aji Vi ≥ ej , ∀ j ∈ N \L, (11) Rimin s X Rimin
i∈L fi = V + bk , (26)
Ris i Rik
k∈Ni ,k∈L
/
where for each i ∈ L, we have ei = Vis − Pimax Ris /Vis and
for each j ∈ N \L, we have ej = Vjs − Pjmax Rjs /Vjs − bj . Rimax s X Rimax
gi = V + bk . (27)
Next, using (8), at each load bus i ∈ L, the power delivered Ris i Rik
k∈Ni ,k∈L
/
to the power electronics load is calculated as
  Depending on the load characteristics, the constraints in (20)
Vi 2 X and (21) may sometimes be relaxed. For example, if Dimax ≈
Pi = = Vi  cij Vj + di  , (12) 1, then Rimin = 0 and constraint (20) becomes Vi ≥ 0.
Ri
j∈L The optimal DR problem can now be formulated for DC
where distribution networks. Here, the notion of proportional fair-
X aki X 1 1 ness from utility theory [31] is used. Proportional fairness is
cii = − − s, (13)
Rik Rik Ri achieved in power delivered to users if one can maximize:
k∈Ni ,k∈L
/ k∈Ni Y
Pi . (28)
Vis X bk
di = + , (14) i∈L
Ris Rik
k∈Ni ,k∈L
/ The expression in (28) is directly linked to the concept of Nash
and for each load bus k ∈ L\{i}, we have bargaining solution in Cooperative Game Theory [32]. Intu-
itively, any Pi that is too low can significantly lower the overall
X alk 1k∈Ni
cik = + . (15) product in (28). Therefore, by maximizing (28), no single load
Rik Rik suffers from a power draw that is too low, assuring fairness.
l∈Ni ,l∈L
/
Since the logarithm function is a monotonically increasing
In (15), 1k∈Ni is an indicator function. That is, if k ∈ Ni , function, maximizing (28) is equivalent to maximizing [33]:
then 1k∈Ni = 1; otherwise, 1k∈Ni = 0. !
Whether a power electronics load can support a demand Y X
log Pi = log (Pi ) . (29)
response solution depends on the characteristics of its DC-DC
i∈L i∈L
converter. Let Dimin and Dimax denote the minimum and the
maximum permissible duty cycles that the DC-DC converter From this, together with (10)-(12), to assure fair power deliv-
at load bus i can support. It is required that ery to users while observing the power limits of DC sources,
one must solve the following optimization problem:
Dimin ≤ Di ≤ Dimax , ∀ i ∈ L. (16)  
X X X
From (1), one can rewrite the above constraints as maximize log (Vi ) + log  cij Vj + di 
Vi ≥ei , ∀i∈L
i∈L i∈L j∈L
Rimin ≤ Ri ≤ Rimax , ∀ i ∈ L, (17) X
subject to aji Vi ≥ ej , ∀ j ∈ N \L,
where i∈L (30)
Rimin = (1 − Dimax )2 Rio ,
X
(18) mik Vk ≥ fi , ∀ i ∈ L,
k∈L
Rimax = (1 − Dimin )2 Rio . (19) X
nik Vk ≤ gi , ∀ i ∈ L.
Here, Rio denotes the internal resistance of the actual end load k∈L
at bus i. From (8) and (17) and after reordering the terms, the The following theorem summarizes the key characteristics
following constraints must hold on the voltages at load buses: and the importance of the above optimization problem.
X
mik Vk ≥ fi , ∀ i ∈ L, (20) Theorem 1: (a) Problem (30) is a convex optimization prob-
k∈L lem. (b) Given Vi∗ for all i ∈ L as the optimal solution of (30),
the optimal values for variable resistors are obtained using (8).
and X The optimal duty cycles for the switch-mode converters can
nik Vk ≤ gi , ∀ i ∈ L, (21)
be selected accordingly, e.g., using (1) for a boost converter.
k∈L
Proof: To prove part (a), it is noted that since logarithm is
where
a concave function and the expression inside the parenthesis
mii = 1 − cii Rimin , (22) for the second logarithmic term in the objective function in
nii = 1 − cii Rimax , (23) (30) is affine, the objective function in optimization problem
(30) is concave. From this, together with the fact that all
and for each k ∈ L\{i}, we have constraints are linear, the maximization problem in (30) is a
convex program. The proof of part (b) is evident.
mik = cik Rimin , (24)
From Theorem 1, if direct load control (DLC) is possible,
nik = cik Rimax . (25) then optimal demand response is achieved in four steps. First,
collect all needed data from each load, bus, and link in the DC Algorithm 1 : Optimal Demand Response in DC Networks.
distribution network and send them to an aggregator. Second, 1: Initialize V L and Lagrange multipliers λN \L , µL , ρL .
solve problem (30) using standard convex programming tools. 2: Repeat
Third, announce the optimal effective resistance to each load 3: Repeat
that participates in demand response. Four, each load will 4: Each load bus i ∈ L updates Vi by solving (34).
adjust the duty cycle for its DC-DC converter accordingly. 5: Broadcast Vi to all other buses.
6: Until no entry of vector V L changes.
7: Each bus j ∈ N \L updates λj using (35).
IV. D ECENTRALIZED I MPLEMENTATION 8: Broadcast λj to all other buses.
If DLC is not feasible, then optimal demand response is 9: Each bus i ∈ L updates µj and ρj using (36) and (37).
done via pricing and a decentralized algorithm as we see next. 10: Broadcast µi and ρi to all other buses.
11: Until no entry of vectors λN \L , µL , or ρL changes.
12: Each load bus i ∈ L updates Ri using (8).
A. Dual Decomposition
From duality theory [22, Chapter 5], the Lagrangian asso-
ciated with the primal optimization problem (30) becomes: In order to solve problem (32) using the coordinate ascent
X method, one can fix all variables Vk for k ∈ L\{i}, and then
L(V L , λN \L , µL , ρL ) = log (Vi ) solve the optimization problem with respect to Vi as follows:
i∈L !
  ! X X X
X X X X maximize log (Vi ) + log cik Vk + di
+ log  cij Vj + di  − λj ej − ajk Vk Vi ≥ei
i∈L i∈L k∈L
i∈L j∈L j∈N \L k∈L  
! ! X X X
X X X X +  λj aji + µk mik − ρk nik  Vi .
− µi fi − mik Vk + ρi gi − nik Vk ,
i∈L j∈N \L k∈L
i∈L k∈L i∈L k∈L
(31) (34)

where λj ≥ 0, µi ≥ 0, and ρi ≥ 0 are the Lagrange multipliers This procedure is repeated for all i ∈ L, leading to an iterative
associated with the inequality constraints in (10), (20), and algorithm. Since problem (32) is convex, if the iterations are
(21), respectively. We define the Lagrange dual function as implemented in the form of a Gauss-Seidel algorithm, where
users take turns, then the iterations are guaranteed to converge
X X X
g(λN \L , µL , ρL ) = − λj ej − µi fi + ρi gi to the optimal solution of problem (32) [34, Proposition 2.5].
j∈N \L i∈L i∈L In order to solve problem (33) using the gradient method, at
X each iteration, we update the Lagrange multipliers as follows:
+ maximum log (Vi )
Vi ≥ei , ∀i∈L " !#+
i∈L X
λ j ← λ j + γ ej − , j ∈ N \L, (35)
 
X X ajk Vk
+ log  cij Vj + di  k∈L
i∈L j∈L " !#+
  X
X X X µi ← µj + σ fi − mik Vk , i ∈ L, (36)
+  λj aji + µk mik − ρk nik  Vi . k∈L
i∈L j∈N \L k∈L
(32) "
X
!#+
ρi ← ρj − ξ gi − nik Vk , i ∈ L, (37)
Finally, one can define the dual optimization problem as
k∈L

minimize g(λN \L ). (33) where γ > 0, σ > 0, and ξ > 0 are stepsize parameters
λj ≥0, ∀j∈N \L
and [z]+ = max{z, 0}. If the stepsize parameters are small
Recall from Theorem 1 that the primal optimization problem enough or diminishing, then the convergence of the iterations
(30) is a convex program. From this, together with the fact that in (35)-(37) to the solution of problem (33) is guaranteed [35].
the linear constraints in (30) satisfy the Slater’s condition, We are now ready to introduce our proposed decentralized
strong duality holds, and the duality gap is zero, and the demand response scheme in Algorithm 1. There are two loops
Lagrange multipliers always exist [22, Section 5.2.3]. There- in this algorithm. The inner-loop solves problem (32) using the
fore, if one can iteratively solve the maximization in (32) and coordinated ascent method. The outer loop solves problem (33)
the minimization in (33), then after convergence, the global using the gradient method. As we explained earlier, since the
optimal solution of the primal problem (30) will be achieved. duality gap is zero, Algorithm 1 will converge to the optimal
In this regard, the coordinate ascent method [34, Section 3.2.4] solution of problem (30), as long as the load bus voltages in
can be used to solve problem (32) and the gradient method Line 4 are updated sequentially and the step size to update the
[22, Section 9.3] can be used to solve problem (33). Lagrange multipliers in Line 7 is small enough or diminishing.
B. Pricing Interpretation (a) (b)
2.5 2.5
Consider the local problem (34) that each load bus (or user)
must solve when Algorithm 1 is implemented. From (12), and
2 1.87 2
after reordering the terms, it can be shown that problem (34)
is equivalent [22, Section 4.1.3] to the following problem:
1.48 1.43

Power Delivery

Power Delivery
X 1.5 1.5
maximize log (Pi ) + αi Vi + log (cki Vi + βik ) (38) 1.28
Vi ≥ei
k∈L\{i}
1.01
where 1 1 0.88 0.90 0.88
X X X
αi = λj aji + µk mik − ρk nik , (39)
0.5 0.5
j∈N \L k∈L k∈L

and for each k ∈ N \{i}, we have

X 0 0
1 3 4 5 1 3 4 5
βik = ckl Vl + dk . (40) Load Bus Number Load Bus Number
l∈L\{i}

As far as solving problem (38) is concerned, αi and βik are Fig. 5. Power delivery to end users based on the solution of problem (30) in
the radial power distribution network of Fig. 2(a) with four power electronics
constant. They can take both positive and negative values. loads: (a) With distributed generation. (b) Without distributed generation.
The objective function in (38) can be interpreted as follows.
The first term, log(Pi ), can be seen as user i’s monotonically-
increasing and concave utility function that quantifies user i’s and P3max . Next, assume that the distributed generator at bus
level of satisfaction when it draws power Pi , c.f. [36], [37]. 3 is disconnected from the network. The optimal values for
The second term, αi Vi , is for voltage regulation. If αi < 0, the variable resistors are updated accordingly and we have
then user i is encouraged to reduce its voltage. If αi > 0, R1∗ = 0.9159, R3∗ = 0.9192, R4∗ = 0.9189, R5∗ = 0.9190.
then user i Pis encouraged to increase its voltage. Finally, the The amount of power delivered to each load is shown in Fig.
third term, k∈L\{i} log(cki Vi + βik ), can enforce fairness. 5(b). One can see that the proposed design can maintain both
The higher the value of βik , the stronger user i is encouraged efficiency and fairness under different grid conditions.
to regulate its voltage and allow more power delivery to user
k. B. Different Random Scenarios
Based on the discussion above, one can interpret αi as a
voltage regulating price and βik as a fairness enforcement In order to have a base for performance comparison, in this
price. From (39), the voltage regulating price depends on λj , section, we consider the utility-based demand response design
i.e., whether any DC source at a non-load bus j has reached in [36]. Note that, if the utility functions are logarithmic, then
its power delivery limit, and also on aji , i.e., the way that the objective function of the utility maximization problem
the operation of the variable resistor load at load bus i may in [36] becomes identical to the objective function of the
affect the voltage at the terminal of a DC source at a non- optimization problem in equation (30) of this paper. However,
load bus j. From (40), the fairness enforcement price depends since the design in [36] does not take into account the details of
on the amount of current that user k can draw from the DC the underlying physical power system, if the design guidelines
distribution network if Ri → 0 and, accordingly, Vi → 0. in [36] are followed, then all users that are connected to the
same feeder and have the same utility functions will have the
V. P ERFORMANCE E VALUATION same effective resistance in their loads. For example, for the
radial DC distribution network in Fig. 2(a), [36] results in
In this section, the performance of the proposed optimal
demand response framework for DC distribution networks R1 = R3 = R4 = R5 . (41)
is assessed. We examine both centralized and decentralized
The proposed method here is compared with the design
designs. All parameters are for a per unit system with base
approach in [36] based on (41), and the results are shown
voltage of 380 V DC and base power of 10 kW [38]. Unless
in Fig. 6(a). Here, 100 different scenarios are considered by
stated otherwise, it is assumed that Rimin = 0 and Rimax  1.
randomly choosing parameters R12 , R23 , R24 , R45 , R1s , R3s ,
P1max , and P3max . The resulting proportional fairness objective
A. Achieving Efficiency and Fairness values are compared for each scenario. To ensure a fair
First, consider the radial distribution network in Fig. 2(a) comparison, the best possible performance when (41) holds
with R12 = 0.01, R23 = 0.03, R24 = 0.01, R45 = 0.01, is considered for each scenario. This gives an upper bound
R1s = 0.01, R3s = 0.01, V1s = 1, V3s = 1, P1max = 4, and for the design performance in [36]. Such upper bound is then
P3max = 2.5. In presence of the distributed generator at bus 3, compared with the proposed design in this paper. The details
the optimal values for the variable resistors are R1∗ = 0.4921, of such comparison for scenario number 1 are shown in Fig.
R3∗ = 0.7404, R4∗ = 0.5563, and R5∗ = 0.5565. The amount 6(b), as an example. On average, and across all 100 random
of power delivered to each load is shown in Fig. 5(a). At scenarios, the proportional fairness objective value increases
optimality, both DC sources reach their power limits P1max by 11.2% when the proposed design is implemented compared
 (a) (a)
3 0.938
Optimal Design Optimal Solution (Centralized Design)
Best Performance when R1 = R3 = R4 = R5 0.937
2

Voltage V1
0.936
Proportional Fairness Objective Value

1
0.935
0 0.934

−1 0.933
0 50 100 150 200 250

−2
0.953
−3 0.952
Optimal Solution (Centralized Design)

Voltage V3
−4 0.951

0.95
−5
0.949
−6 0.948
0 20 40 60 80 100 0 50 100 150 200 250
Scenario Number Iteration Number

(b) (b)
−1.9 10.5

Optimal Objective Value (Our Design) 10

−2

Price α 1
λ2 = 0
Proportional Fairness Objective Value

9.5
−2.1
10.8% 9

−2.2 8.5
0 50 100 150 200 250

−2.3
28

−2.4 27
Price α 3

26
−2.5 Infeasible Objective Value when R1 = R3 = R4 = R5 = R
λ2 = 0
25

−2.6 24
0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0 50 100 150 200 250
R Iteration Number

Fig. 6. Performance comparison between the proposed optimal design and the Fig. 7. Performance evaluation of distributed Algorithm 1 for the DC
case where R1 = R3 = R4 = R5 in the radial power distribution network microgrid in Fig. 2(b): (a) The trends of the voltages at load buses 1 and
of Fig. 2(a) with four power electronics loads: (a) Comparison across 100 3. (b) The trends of the voltage regulation prices at load buses 1 and 3.
random scenarios. (b) Detailed comparison based on scenario number 1.

The simulation results when we run Algorithm 1 are shown

with the best performance possible when (41) holds. Given in Fig. 7. We can see that the distributed design results in
that the proposed method in this paper and the method in the exact optimal performance as in the centralized design.
[36] have identical objective functions for logarithmic utilities, In particular, both V1 and V3 , i.e., the voltages at the load
the advantage of our proposed design is evident. Of course, buses, converge to their optimal values as shown in Fig. 7(a).
we must note that, while the design in [36] is general and The voltage regulation price signals are shown in Fig. 7(b). At
can accommodate different choices of utility functions beyond optimality, only the DC source at bus 4 reaches its power limit
logarithmic utilities, the proposed method in this paper, how- P4max . Therefore, at steady state, we have λ4 > 0 while λ2 =
ever, is specifically designed for logarithmic utility functions 0. Note that, only the iterations in the outer loop of Algorithm
towards achieving proportional fairness. 1 are shown here. Typically, the inner loop converges within
3 or 4 iterations. Constraints (20) and (21) are not binding.
C. Decentralized Design and Pricing It is interesting to compare Fig. 7 with the similar results
Consider the DC microgrid in Fig. 2(b), where R12 = 0.01, in Fig. 5 of the conference version of this work in [23]. Here,
R14 = 0.03, R23 = 0.01, R34 = 0.01, R2s = 0.01, R4s = 0.01, we have chosen a large step size γ = 100 for the Lagrange
V2s = 1, V4s = 1, P2max = 4, and P4max = 3. Note that, the multiplier update equation in (35). This has resulted in an
two DC power sources are connected to non-load buses. The order of magnitude faster convergence in Fig. 7, compared to
power draw constraints in this example are Fig. 5 in [23].

0.3333 V1 + 0.3333 V3 ≥ 0.6267, (42) D. Impact of Time-Varying Renewable Generation

and Next, the variable resistors are adjusted in response to
0.1429 V1 + 0.4286 V3 ≥ 0.5414. (43) changes in solar irradiance. The system setup is the same as
 (a)
3 1
Proportional Fairness

Objective Value Sunny Day

Objective Value

2 0.8 Cloudy Day

No DG

Proportional Fairness Objective Value

1 0.6

0 0.4

−1
0 4 8 12 16 20 24 0.2

(b) 0
3
Power Electronics Loads

User 1
2.5 −0.2
Power Draw by

User 3
2 User 4
User 5 −0.4
1.5
−0.6
1

0.5 −0.8
0 10 20 30 40 50 60 70 80 90 0.3 0.4 0.5 0.6 0.7 0.8
Hour of Day Dmax
1
= Dmax
3
= Dmax
4
= Dmax
5

Fig. 8. Simulation results when the variable resistors in Fig. 2(a) are updated Fig. 9. The impact of changing permissible range for duty cycles of power
every 15 minutes based on the irradiance data in Fig. 4(a): (a) Proportional electronics loads. We have Dimin = 0.2 and Dimax varies from 0.3 to 0.8.
fairness objective value. (b) Power draw by power electronics loads. As expected, a wider range for the duty cycle improves the objective value.

that in Section V-A. However, the power and voltage at the PV an optimization-based DC demand response foundation to
panel are updated every 15 minutes using the data in Fig. 4(a). ensure efficient and fair power system operation. It was
At 600 W/m2 solar irradiance, the delivered power ad voltage shown that the formulated optimization problem is convex
are P3max = 2.5 MW and V3s = 380 V DC, respectively. and, therefore, tractable. Both centralized and decentralized
The values of P3max and V3s at other irradiance levels are designs were examined. Furthermore, a pricing mechanism
adjusted based on the curves in Fig. 3, with interpolation when was developed to enforce optimal DC demand response in a
necessary. The results are shown in Fig. 8. We can see that the distributed fashion. Various simulation results showed that the
proportional fairness objective value in Fig. 8(a) resembles the proposed design can maximize the efficiency of renewable and
daily irradiance curve in Fig. 4(a). At each point, the optimal distributed generation resources such as photovoltaic panels.
objective value of problem (30) is achieved based on the PV The results in this paper can be extended in several di-
panel conditions at the corresponding time of day. The amount rections. For example, design objective functions other than
of power draw by each end user is also adjusted accordingly proportional fairness can be examined. Other constraints, e.g.,
at each point as we can see in Fig. 8(b). with respect to DC distribution loss, can also be considered.
Other MPPT models such as constant power systems can be
considered for PV panels. Data collection and communica-
E. Impact of DC-DC Converter Duty Cycle Limitations tions among power converters can be realized using different
So far, all DC-DC converters that serve the power electron- networking protocols depending on missions and applications.
ics loads are assumed ideal and support any duty cycle. In In particular, the impact of using different data transmission
other words, at each load bus i ∈ L, we have Rimin = 0 techniques such as wireless mesh networks, cellular, and
and Rimin  1. Next, we relax this assumption and study the power lines carriers can be investigated. The proposed design
impact of changing the permissible range for the duty cycles may also be evaluated using hardware implementation.
of the DC-DC converters. We assume that at each load bus
i ∈ L, we have Dimin = 0.2 and Dimax varies from 0.3 to
0.7. The results for the proportional fairness objective value R EFERENCES
are shown in Fig. 9. In all cases, the objective value is much [1] K. Hamilton and N. Gulhar, “Taking demand response to the next level,”
higher when the duty cycles are less restricted. More flexible IEEE Power and Energy Magazine, vol. 8, no. 3, pp. 60–65, May 2010.
[2] G. Strbac, “Demand side management: Benefits and challenges,” Energy
DC-DC converters are particularly beneficial in presence of Policy, vol. 36, no. 12, pp. 4419–4426, Dec. 2008.
PV panels. Here, the solar irradiance data for the sunny day [3] H. Mohsenian-Rad and A. Leon-Garcia, “Optimal Residential Load
and the cloudy day come from Figs. 4(a) and (b), respectively. Control with Price Prediction in Real-Time Electricity Pricing Environ-
ments,” IEEE Trans. on Smart Grid, vol. 1, pp. 120–133, Sept. 2010.
[4] A. Molderink, V. Bakker, J. L. Hurink, and G. J. M. Smit, “Comparing
demand side management approaches,” in Proc. of IEEE PES ISGT’12,
VI. C ONCLUSIONS AND F UTURE W ORK Washington, DC, Jan. 2010.
Given the great benefits of DC distribution networks and be- [5] H. Mohsenian-Rad, V. Wong, J. Jatskevich, R. Schober, and A. Leon-
Garcia, “Autonomous Demand Side Management Based on Game-
cause of the importance of demand response in the future smart Theoretic Energy Consumption Scheduling for the Future Smart Grid,”
grid, in this paper, the first steps were taken towards designing IEEE Trans. on Smart Grid, vol. 1, no. 3, pp. 320–331, Dec. 2010.
demand response programs for DC distribution systems. In [6] Z. Baharlouei, M. Hashemi, H. Narimani, and H. Mohsenian-Rad,
“Achieving Optimality and Fairness in Autonomous Demand Response:
this regard, the recent advancements in power electronics were Benchmarks and Billing Mechanisms,” IEEE Trans. on Smart Grid,
combined with techniques from optimization theory to develop vol. 4, no. 2, pp. 968–975, June 2013.
 [7] J. Medina, N. Muller, and I. Roytelman, “Demand Response and Dis- [33] L. Massoulie, “Structural properties of proportional fairness: Stability
tribution Grid Operations: Opportunities and Challenges,” IEEE Trans. and insensitivity,” Annals of Applied Probability, vol. 17, no. 3, pp.
on Smart Grid, vol. 1, no. 2, pp. 193–198, Sept. 2010. 809–839, Mar. 2007.
[8] N. Li, L. Chen, and S. Low, “Demand response in radial distribution [34] D. P. Bertsekas and J. N. Tsitsiklis, Parallel and Distributed Computa-
networks,” in Proc. of the IEEE Systems and Computers Asilomar tion: Numerical Methods. Belmont, MA: Athena Scientific, 1997.
Conference on Signals, Pacific Grove, CA, Nov. 2012. [35] D. P. Palomar and M. Chiang, “A Tutorial on Decomposition Methods
[9] S. Mohagheghi, J. Stoupis, Z. Wang, Z. i, and H. Kazemzadeh, “De- for Network Utility Maximization,” IEEE Journal On Selected Areas in
mand response architecture: Integration into the distribution management Communications, vol. 24, no. 8, pp. 1439–1451, Aug. 2006.
system,” in Proc. of IEEE International Conference on Smart Grid [36] P. Samadi, H. Mohsenian-Rad, R. Schober, V. Wong, and J. Jatskevich,
Communications, Gaithersburg, MD, Oct. 2010. “Optimal real-time pricing algorithm based on utility maximization for
[10] E. L. Zivi, “Integrated shipboard power and automation control challenge smart grids,” in Proc of IEEE International Conference on Smart Grid
problem,” in Proc. of the IEEE Power Engineering Society Summer Communications, Gaithersburg, MD, Oct. 2010.
Meeting, Chicago, IL, July 2002. [37] P. Samadi, H. Mohsenian-Rad, R. Schober, and V. Wong, “Advanced
[11] M. E. Baran and N. R. Mahajan, “DC distribution for industrial systems: demand side management for the future smart grid using mechanism
Opportunities and challenges,” IEEE Trans. on Industrial Applications, design,” IEEE Trans. on Smart Grid, vol. 3, no. 3, pp. 1170–1180, 2012.
vol. 39, no. 6, pp. 1596–1601, Nov. 2003. [38] D. Dong, F. Lou, X. Zhang, D. Boroyevich, and P. Mattavelli, “Grid-
[12] H. Kakigano, Y. Miura, and T. Ise, “Low voltage bipolar type DC interface bidirectional converter for residential dc distribution systems-
microgrid for super high quality distribution,” IEEE Trans. on Power part 2: Ac and dc interface design with passive components minimiza-
Electronics, vol. 25, no. 12, pp. 3066–3075, Nov. 2010. tion,” IEEE Trans Power Electronics, vol. 28, pp. 1667–1679, Apr. 2013.
[13] M. Brenna, E. Tironi, and G. Ubezio, “Proposal of a local dc distribution
network with distributed energy resources,” in Proc. of the IEEE ICHQP,
Lake Placid, NY, 2004.
[14] D. Salomonsson and A. Sannino, “Low-Voltage DC Distribution System
for Commercial Power Systems With Sensitive Electronic Loads,” IEEE
Trans. on Power Delivery, vol. 22, no. 3, pp. 1620–1627, July 2007.
[15] R. S. Balog, W. Weaver, and P. T. Krein, “The Load as an Energy Asset
in a Distributed DC SmartGrid Architecture,” IEEE Trans. on Smart
Grid, vol. 3, no. 1, pp. 253–260, Mar. 2012.
[16] A. Kwasinski and C. N. Onwuchekwa, “Dynamic behaviour and sta-
blization of dc microgrids with instantaneous constant-power loads,”
IEEE Trans Power Electronics, vol. 26, pp. 822–834, Mar. 2011. Hamed Mohsenian-Rad (S’04-M’09) received the
[17] H. Ikebe, “Power systems for telecommunications in the IT age,” in Ph.D. degree in Electrical and Computer Engineer-
Proc. of IEEE INTELEC, Yokohama, Japan, Oct. 2003. ing from the University of British Columbia in Van-
[18] A. Kwasinski and C. N. Onwuchekwa, “Quantitate evaluation of dc couver, Canada in 2008. Currently, he is an Assistant
microgrids availability: effects of system architecture and converter Professor of Electrical Engineering at the University
topology design choices,” IEEE Trans Power Electronics, vol. 26, pp. of California at Riverside. Dr. Mohsenian-Rad is
835–851, Mar. 2011. the recipient of the NSF CAREER Award, the Best
[19] W. Weaver, “Dynamic Energy Resource Control of Power Electronics Paper Award from the IEEE International Confer-
in Local Area Power Networks,” IEEE Trans. on Power Electronics, ence on Smart Grid Communications 2012, and the
vol. 26, no. 3, pp. 852–859, Mar. 2011. Best Paper Award from the IEEE Power and Energy
[20] A. Mohamed and O. Mohammed, “Power flow control in dc distribution Society General Meeting 2013. He serves as Editor
systems,” in Proc. of the IEEE NAPS, Arlington, TX, 2010. for the IEEE Trans. on Smart Grid, the IEEE Communications Surveys and
[21] W. Weaver and P. T. Krein, “Game-Theoretic Control of Small-Scale Tutorials, and the IEEE Communications Letters. His research interests are
Power Systems,” IEEE Trans. on Power Delivery, vol. 24, no. 3, pp. the design, optimization, and game-theoretic analysis of power systems.
1560–1567, July 2009.
[22] S. Boyd and L. Vandenberghe, Convex Optimization. New York, NY:
Cambridge University Press, 2004.
[23] H. Mohsenian-Rad and A. Davoudi, “Optimal demand response in dc
distribution networks,” in Proc. of IEEE International Conference on
Smart Grid Communications, Vancouver, BC, Oct. 2013.
[24] R. W. Erickson and D. Maksimovic, Fundamentals of Power Electronics.
New York, NY: Kluwer Academic Publishers, 2004.
[25] A. M. Dumitrescu, I. Florea, M. Naumof, and M. Calin, “Model valida-
tion of distribution feeders using synchronous measurements,” in Proc.
of IEEE International Instrumentation and Measurement Technology
Conference, Graz, Austria, May 2012.
Ali Davoudi (S’04-M’11) received the Ph.D. degree
[26] S. Grillo, V. Musolino, L. Piegari, and E. Tironi, “A control strategy for
in Electrical and Computer Engineering from the
optimizing the power flows supplied by two different storage units,” in
University of Illinois, Urbana, IL, USA, in 2010. He
Proc. of IEEE PowerTech, Trondheim , Norway, June 2011.
is currently an Assistant Professor at the Electrical
[27] A. Mohamed and O. Mohammed, “Connectivity of dc microgrids involv-
Engineering Department of the University of Texas,
ing sustainable energy sources,” in Proc. of IEEE Industry Applications
Arlington, TX, USA. He worked for Solar Bridge
Society Annual Meeting, Orlando, FL, Oct. 2011.
Technologies, Texas Instruments Inc., and Royal
[28] G. J. Yu, Y. S. Jung, J. Y. Choi, I. Choy, J. H. Song, and G. S. Philips Electronics. He has served as the editor or
Kim, “A novel two-mode MPPT control algorithm based on comparative guest editor of IEEE Transactions on Power Elec-
study of existing algorithms,” in Proc. of IEEE Photovoltaic Specialists tronics, Energy Conversion, Smart Grids, Industry
Conference, New Orleans, LA, May 2002. Applications, Industrial Informatics, and Vehicular
[29] T. Esram and P. L. Chapman, “Comparison of photovoltaic array Technology. His research interests are various control aspects of power
maximum power point tracking techniques,” IEEE Trans. on Energy electronics, energy conversion, and finite-inertia power systems.
Conversion, vol. 22, no. 2, pp. 439–449, June 2007.
[30] Solar Energy Laboratory at The University of
Texas at Austin, “Texas Solar Radiation Database,”
http://www.me.utexas.edu/∼solarlab/tsrdb/tsrdb.html, July 2013.
[31] F. Kelly, A. Maulloo, and D. Tan, “Rate control for communication
networks: Shadow price proportional fairness and stability,” Journal of
Operations Research, vol. 49, pp. 237–252, Jan. 1998.
[32] D. Fudenberg and J. Tirole, Game Theory. The MIT Press, 1991.