Atomic Scheduling of Appliance
Energy Consumption in Residential
Smart Grid
Kyeong Soo (Joseph) Kim
(With S. Lee, T. O. Ting@XJTLU and X.-S. Yang@Middlesex)
Department of Electrical and Electronic Engineering
Xi’an Jiaotong-Liverpool University
CeSGIC 1st International Workshop on
Smart Grid Technology and Data Processing
19 June 2015
1 / 55
Outline
Introduction
Review of Current Formulation of Appliance Energy
Consumption Scheduling
Atomic Scheduling of Appliance Energy Consumption
Starting-Time-Based Formulation
Optimal-Routing-Based Formulation
Successive Convex Relaxation
Examples
Numerical Results
Experimental Parameters
Comparison of Bounds with True Optimal Values
Cost Minimization
PAR Minimization
Conclusions
2 / 55
Next . . .
Introduction
Review of Current Formulation of Appliance Energy
Consumption Scheduling
Atomic Scheduling of Appliance Energy Consumption
Starting-Time-Based Formulation
Optimal-Routing-Based Formulation
Successive Convex Relaxation
Examples
Numerical Results
Experimental Parameters
Comparison of Bounds with True Optimal Values
Cost Minimization
PAR Minimization
Conclusions
4 / 55
Autonomous Demand-Side Management in
Smart Grid
Gateway
(Traditional)
Electricity
Grid
Greenfield
Power
Line
Bi-Directional
Communication
Links
Power Plant
5 / 55
Scheduling of Appliance Energy
Consumption
A key to autonomous DSM in
optimizing energy production and
consumption.
Based on two-way digital
communications between a utility
company and users through smart
meters at users’ premises.
Typical objectives
Peak-to-average ratio (PAR)
Total energy cost
6 / 55
Scheduling of Appliance Energy
Consumption
A key to autonomous DSM in
optimizing energy production and
consumption.
Based on two-way digital
communications between a utility
company and users through smart
meters at users’ premises.
Typical objectives
Peak-to-average ratio (PAR)
Total energy cost
6 / 55
Scheduling of Appliance Energy
Consumption
A key to autonomous DSM in
optimizing energy production and
consumption.
Based on two-way digital
communications between a utility
company and users through smart
meters at users’ premises.
Typical objectives
Peak-to-average ratio (PAR)
Total energy cost
6 / 55
Scheduling of Appliance Energy
Consumption
A key to autonomous DSM in
optimizing energy production and
consumption.
Based on two-way digital
communications between a utility
company and users through smart
meters at users’ premises.
Typical objectives
Peak-to-average ratio (PAR)
Total energy cost
6 / 55
A Question on Scheduled Energy
Consumption
Can a washing machine
successfully complete its job with
the energy consumption scheduled
as follows?
7 / 55
A Question on Scheduled Energy
Consumption
Can a washing machine
successfully complete its job with
the energy consumption scheduled
as follows?
9am 3pm12am
7 / 55
A Question on Scheduled Energy
Consumption
Can a washing machine
successfully complete its job with
the energy consumption scheduled
as follows?
9am 3pm12am
or
7 / 55
A Question on Scheduled Energy
Consumption
Can a washing machine
successfully complete its job with
the energy consumption scheduled
as follows?
9am 3pm12am
or
9am 3pm2pm10am
7 / 55
Next . . .
Introduction
Review of Current Formulation of Appliance Energy
Consumption Scheduling
Atomic Scheduling of Appliance Energy Consumption
Starting-Time-Based Formulation
Optimal-Routing-Based Formulation
Successive Convex Relaxation
Examples
Numerical Results
Experimental Parameters
Comparison of Bounds with True Optimal Values
Cost Minimization
PAR Minimization
Conclusions
8 / 55
Optimization Variables
Based on energy consumption over equally-divided time
slots of a day (typically hourly time slots) as
optimization variables, i.e.,
xn x0
n, . . . , xh
n, . . . , xH−1
n
for
User n ∈ N {1, . . ., N};
Time slot h ∈ H {0, . . ., H−1}.
Because the optimization variables take continuous
values, we can easily apply
Convex optimization [1];
Distributed algorithms through concave n-person
games [2].
9 / 55
Optimization Variables
Based on energy consumption over equally-divided time
slots of a day (typically hourly time slots) as
optimization variables, i.e.,
xn x0
n, . . . , xh
n, . . . , xH−1
n
for
User n ∈ N {1, . . ., N};
Time slot h ∈ H {0, . . ., H−1}.
Because the optimization variables take continuous
values, we can easily apply
Convex optimization [1];
Distributed algorithms through concave n-person
games [2].
9 / 55
Optimization Variables
Based on energy consumption over equally-divided time
slots of a day (typically hourly time slots) as
optimization variables, i.e.,
xn x0
n, . . . , xh
n, . . . , xH−1
n
for
User n ∈ N {1, . . ., N};
Time slot h ∈ H {0, . . ., H−1}.
Because the optimization variables take continuous
values, we can easily apply
Convex optimization [1];
Distributed algorithms through concave n-person
games [2].
9 / 55
Optimization Variables
Based on energy consumption over equally-divided time
slots of a day (typically hourly time slots) as
optimization variables, i.e.,
xn x0
n, . . . , xh
n, . . . , xH−1
n
for
User n ∈ N {1, . . ., N};
Time slot h ∈ H {0, . . ., H−1}.
Because the optimization variables take continuous
values, we can easily apply
Convex optimization [1];
Distributed algorithms through concave n-person
games [2].
9 / 55
Optimization Variables
Based on energy consumption over equally-divided time
slots of a day (typically hourly time slots) as
optimization variables, i.e.,
xn x0
n, . . . , xh
n, . . . , xH−1
n
for
User n ∈ N {1, . . ., N};
Time slot h ∈ H {0, . . ., H−1}.
Because the optimization variables take continuous
values, we can easily apply
Convex optimization [1];
Distributed algorithms through concave n-person
games [2].
9 / 55
Optimization Variables
Based on energy consumption over equally-divided time
slots of a day (typically hourly time slots) as
optimization variables, i.e.,
xn x0
n, . . . , xh
n, . . . , xH−1
n
for
User n ∈ N {1, . . ., N};
Time slot h ∈ H {0, . . ., H−1}.
Because the optimization variables take continuous
values, we can easily apply
Convex optimization [1];
Distributed algorithms through concave n-person
games [2].
9 / 55
Feasible Set
A feasible energy consumption scheduling set for user n
Xn= xn
h∈Hn
xh
n=En, γmin
n ≤xh
n≤γmax
n , ∀h∈Hn, xh
n=0, ∀h∈HHn
where
γmin
n : Minimum energy level;
γmax
n : Maximum energy level;
En: Total daily energy consumption;
Hn: Scheduling interval defined as follows:
Hn h h = i mod H, ∀i∈ αn, βn
with αn∈[0, H−1], βn∈[1, 2H−2], and 1≤βn−αn≤H−1.
10 / 55
Feasible Set
A feasible energy consumption scheduling set for user n
Xn= xn
h∈Hn
xh
n=En, γmin
n ≤xh
n≤γmax
n , ∀h∈Hn, xh
n=0, ∀h∈HHn
where
γmin
n : Minimum energy level;
γmax
n : Maximum energy level;
En: Total daily energy consumption;
Hn: Scheduling interval defined as follows:
Hn h h = i mod H, ∀i∈ αn, βn
with αn∈[0, H−1], βn∈[1, 2H−2], and 1≤βn−αn≤H−1.
10 / 55
Feasible Set
A feasible energy consumption scheduling set for user n
Xn= xn
h∈Hn
xh
n=En, γmin
n ≤xh
n≤γmax
n , ∀h∈Hn, xh
n=0, ∀h∈HHn
where
γmin
n : Minimum energy level;
γmax
n : Maximum energy level;
En: Total daily energy consumption;
Hn: Scheduling interval defined as follows:
Hn h h = i mod H, ∀i∈ αn, βn
with αn∈[0, H−1], βn∈[1, 2H−2], and 1≤βn−αn≤H−1.
10 / 55
Feasible Set
A feasible energy consumption scheduling set for user n
Xn= xn
h∈Hn
xh
n=En, γmin
n ≤xh
n≤γmax
n , ∀h∈Hn, xh
n=0, ∀h∈HHn
where
γmin
n : Minimum energy level;
γmax
n : Maximum energy level;
En: Total daily energy consumption;
Hn: Scheduling interval defined as follows:
Hn h h = i mod H, ∀i∈ αn, βn
with αn∈[0, H−1], βn∈[1, 2H−2], and 1≤βn−αn≤H−1.
10 / 55
Feasible Set
A feasible energy consumption scheduling set for user n
Xn= xn
h∈Hn
xh
n=En, γmin
n ≤xh
n≤γmax
n , ∀h∈Hn, xh
n=0, ∀h∈HHn
where
γmin
n : Minimum energy level;
γmax
n : Maximum energy level;
En: Total daily energy consumption;
Hn: Scheduling interval defined as follows:
Hn h h = i mod H, ∀i∈ αn, βn
with αn∈[0, H−1], βn∈[1, 2H−2], and 1≤βn−αn≤H−1.
10 / 55
Feasible Set
A feasible energy consumption scheduling set for user n
Xn= xn
h∈Hn
xh
n=En, γmin
n ≤xh
n≤γmax
n , ∀h∈Hn, xh
n=0, ∀h∈HHn
where
γmin
n : Minimum energy level;
γmax
n : Maximum energy level;
En: Total daily energy consumption;
Hn: Scheduling interval defined as follows:
Hn h h = i mod H, ∀i∈ αn, βn
with αn∈[0, H−1], βn∈[1, 2H−2], and 1≤βn−αn≤H−1.
10 / 55
Optimal Scheduling
The optimal scheduling is formulated as an optimization
problem for a given objective function φ(·) (e.g., total
energy cost or PAR) as follows:
minimize
xn∈Xn, ∀n∈N
φ (L(x))
where
x [x1, . . . , xN]: A vector of user energy consumption
vectors;
L(x) [L0 (x) , . . . , LH−1 (x)]: A vector of aggregate
loads across all users at each time slot, which are
defined as
Lh (x)
n∈N
xh
n.
11 / 55
Optimal Scheduling
The optimal scheduling is formulated as an optimization
problem for a given objective function φ(·) (e.g., total
energy cost or PAR) as follows:
minimize
xn∈Xn, ∀n∈N
φ (L(x))
where
x [x1, . . . , xN]: A vector of user energy consumption
vectors;
L(x) [L0 (x) , . . . , LH−1 (x)]: A vector of aggregate
loads across all users at each time slot, which are
defined as
Lh (x)
n∈N
xh
n.
11 / 55
Optimal Scheduling
The optimal scheduling is formulated as an optimization
problem for a given objective function φ(·) (e.g., total
energy cost or PAR) as follows:
minimize
xn∈Xn, ∀n∈N
φ (L(x))
where
x [x1, . . . , xN]: A vector of user energy consumption
vectors;
L(x) [L0 (x) , . . . , LH−1 (x)]: A vector of aggregate
loads across all users at each time slot, which are
defined as
Lh (x)
n∈N
xh
n.
11 / 55
Optimal Scheduling
The optimal scheduling is formulated as an optimization
problem for a given objective function φ(·) (e.g., total
energy cost or PAR) as follows:
minimize
xn∈Xn, ∀n∈N
φ (L(x))
where
x [x1, . . . , xN]: A vector of user energy consumption
vectors;
L(x) [L0 (x) , . . . , LH−1 (x)]: A vector of aggregate
loads across all users at each time slot, which are
defined as
Lh (x)
n∈N
xh
n.
11 / 55
Objective Functions
For energy cost minimization:
φ (L(x)) =
h∈H
Ch (Lh(x))
where
Ch(·): A cost function for generating or distributing
electricity energy at a time slot h.
For PAR minimization:
φ (L(x)) =
H max
h∈H
Lh(x)
n∈N
En
12 / 55
Objective Functions
For energy cost minimization:
φ (L(x)) =
h∈H
Ch (Lh(x))
where
Ch(·): A cost function for generating or distributing
electricity energy at a time slot h.
For PAR minimization:
φ (L(x)) =
H max
h∈H
Lh(x)
n∈N
En
12 / 55
Atomic vs. Non-Atomic Scheduling
αn,a βn,a
αn,a βn,a
Gap Gap
γmin
n
γmax
n
(a)
γop
n (·)
γmin
n
γmax
n
(b)
Examples of (a) non-atomic and (b) atomic scheduling.
γmin
n : Minimum energy level
γmax
n : Maximum energy level
γ
op
n (·): Operating energy level
13 / 55
Non-Atomic Scheduling Example
Appliance 1
Appliance 2
Scheduled
Consumption
?
14 / 55
Non-Atomic Scheduling Example: Case 1
Appliance 1
Appliance 2
Scheduled
Consumption
15 / 55
Non-Atomic Scheduling Example: Case 2
Appliance 1
Appliance 2
Scheduled
Consumption
16 / 55
Non-Atomic Scheduling Example: Case 3
Appliance 1
Appliance 2
Scheduled
Consumption
17 / 55
Next . . .
Introduction
Review of Current Formulation of Appliance Energy
Consumption Scheduling
Atomic Scheduling of Appliance Energy Consumption
Starting-Time-Based Formulation
Optimal-Routing-Based Formulation
Successive Convex Relaxation
Examples
Numerical Results
Experimental Parameters
Comparison of Bounds with True Optimal Values
Cost Minimization
PAR Minimization
Conclusions
18 / 55
Overview of Atomic Scheduling Problem
Formulation & Solution
Starting-Time-Based
Formulation
Optimal-Routing-Based
Formulation
Convex Relaxation
Successive Convex
Relaxation with
Fractional-Value
Dropping
Convex
Optimization
(Feasible Upper Bound)
Combinatorial
Optimization
Boolean-Convex
Optmization
Convex
Optmization
(Lower Bound)
19 / 55
Next . . .
Introduction
Review of Current Formulation of Appliance Energy
Consumption Scheduling
Atomic Scheduling of Appliance Energy Consumption
Starting-Time-Based Formulation
Optimal-Routing-Based Formulation
Successive Convex Relaxation
Examples
Numerical Results
Experimental Parameters
Comparison of Bounds with True Optimal Values
Cost Minimization
PAR Minimization
Conclusions
20 / 55
Starting-Time-Based Formulation
Optimization variables:
s [s1, . . . , sN]
A feasible set for user n:
Sn sn sn = i mod H, ∀i∈[αn, βn−δn+1]
Aggregate load across all users at each time slot h:
Lh(s)
n∈N
γ
op
n ((h − sn) mod H) IRn(sn)(h)
where
IRn(sn)(h): An indicator function for a set Rn(sn).
Rn(sn): A range of user n’s appliance operation for sn
defined as follows:
Rn(sn) h h = i mod H, ∀i∈ [sn, sn + δn − 1]
21 / 55
Starting-Time-Based Formulation
Optimization variables:
s [s1, . . . , sN]
A feasible set for user n:
Sn sn sn = i mod H, ∀i∈[αn, βn−δn+1]
Aggregate load across all users at each time slot h:
Lh(s)
n∈N
γ
op
n ((h − sn) mod H) IRn(sn)(h)
where
IRn(sn)(h): An indicator function for a set Rn(sn).
Rn(sn): A range of user n’s appliance operation for sn
defined as follows:
Rn(sn) h h = i mod H, ∀i∈ [sn, sn + δn − 1]
21 / 55
Starting-Time-Based Formulation
Optimization variables:
s [s1, . . . , sN]
A feasible set for user n:
Sn sn sn = i mod H, ∀i∈[αn, βn−δn+1]
Aggregate load across all users at each time slot h:
Lh(s)
n∈N
γ
op
n ((h − sn) mod H) IRn(sn)(h)
where
IRn(sn)(h): An indicator function for a set Rn(sn).
Rn(sn): A range of user n’s appliance operation for sn
defined as follows:
Rn(sn) h h = i mod H, ∀i∈ [sn, sn + δn − 1]
21 / 55
Starting-Time-Based Formulation
Optimization variables:
s [s1, . . . , sN]
A feasible set for user n:
Sn sn sn = i mod H, ∀i∈[αn, βn−δn+1]
Aggregate load across all users at each time slot h:
Lh(s)
n∈N
γ
op
n ((h − sn) mod H) IRn(sn)(h)
where
IRn(sn)(h): An indicator function for a set Rn(sn).
Rn(sn): A range of user n’s appliance operation for sn
defined as follows:
Rn(sn) h h = i mod H, ∀i∈ [sn, sn + δn − 1]
21 / 55
Starting-Time-Based Formulation
Optimization variables:
s [s1, . . . , sN]
A feasible set for user n:
Sn sn sn = i mod H, ∀i∈[αn, βn−δn+1]
Aggregate load across all users at each time slot h:
Lh(s)
n∈N
γ
op
n ((h − sn) mod H) IRn(sn)(h)
where
IRn(sn)(h): An indicator function for a set Rn(sn).
Rn(sn): A range of user n’s appliance operation for sn
defined as follows:
Rn(sn) h h = i mod H, ∀i∈ [sn, sn + δn − 1]
21 / 55
Starting-Time-Based Formulation
Optimization variables:
s [s1, . . . , sN]
A feasible set for user n:
Sn sn sn = i mod H, ∀i∈[αn, βn−δn+1]
Aggregate load across all users at each time slot h:
Lh(s)
n∈N
γ
op
n ((h − sn) mod H) IRn(sn)(h)
where
IRn(sn)(h): An indicator function for a set Rn(sn).
Rn(sn): A range of user n’s appliance operation for sn
defined as follows:
Rn(sn) h h = i mod H, ∀i∈ [sn, sn + δn − 1]
21 / 55
Issues with Starting-Time-Based Formulation
Because the feasible set is now discrete, we have to
evaluate the objective function for all the elements in the
feasible set.
The optimization by direct enumeration becomes
impractical for large N and H.
When N=100 and H=24 with the worst case scenario
of αn=0, βn=23, and δn=1 for all n∈N, we need to
evaluate the objective function 24100 times, which is
on the order of 10138 times!
22 / 55
Next . . .
Introduction
Review of Current Formulation of Appliance Energy
Consumption Scheduling
Atomic Scheduling of Appliance Energy Consumption
Starting-Time-Based Formulation
Optimal-Routing-Based Formulation
Successive Convex Relaxation
Examples
Numerical Results
Experimental Parameters
Comparison of Bounds with True Optimal Values
Cost Minimization
PAR Minimization
Conclusions
23 / 55
Why London Eye?
Does the London Eye
have something to
do with the atomic
scheduling?
24 / 55
A Network, A Path, and Links
S
D
0 1
2
3
4
5
6
7
8
9
10
111213
14
15
16
17
18
19
20
21
22
23
l9,10
l10,11l11,12
p9,3
A network connecting
the source (S) and the
destination (D)
through 24
intermediate nodes
with a path (p9,3
) and
its constituent links
(l9,10
, l10,11
, and l11,12
).
25 / 55
Mapping of Atomic Operations to Flows
0 1
2
3
4
5
6
7
8
9
10
111213
14
15
16
17
18
19
20
21
22
23
f0
1
f1
1
f2
1
f3
1
f4
1
f9
2
f10
2
f11
2f12
2
S
D
Mapping of all
possible atomic
operations of two
appliances into two
groups of flows
(f0
1
, . . . , f4
1
and
f9
2
, . . . , f12
2
) over
multiple paths on the
network.
26 / 55
Optimization Variables and Feasible Set
Optimization variables: Flow configurations of all users
defined as
f f1, . . . , fn, . . . , fN
where
fn f0
n , . . . , fH−1
n .
A feasible atomic energy consumption scheduling set
for user n:
Fn = fn
s∈Sn
fs
n=1, fs
n∈ {0, 1} , ∀s∈Sn, fs
n=0, ∀s∈HSn
where Sn is the feasible set of starting times for user n
that is already defined in starting-time-based
formulation.
27 / 55
Optimization Variables and Feasible Set
Optimization variables: Flow configurations of all users
defined as
f f1, . . . , fn, . . . , fN
where
fn f0
n , . . . , fH−1
n .
A feasible atomic energy consumption scheduling set
for user n:
Fn = fn
s∈Sn
fs
n=1, fs
n∈ {0, 1} , ∀s∈Sn, fs
n=0, ∀s∈HSn
where Sn is the feasible set of starting times for user n
that is already defined in starting-time-based
formulation.
27 / 55
Optimization Variables and Feasible Set
Optimization variables: Flow configurations of all users
defined as
f f1, . . . , fn, . . . , fN
where
fn f0
n , . . . , fH−1
n .
A feasible atomic energy consumption scheduling set
for user n:
Fn = fn
s∈Sn
fs
n=1, fs
n∈ {0, 1} , ∀s∈Sn, fs
n=0, ∀s∈HSn
where Sn is the feasible set of starting times for user n
that is already defined in starting-time-based
formulation.
27 / 55
Optimization Variables and Feasible Set
Optimization variables: Flow configurations of all users
defined as
f f1, . . . , fn, . . . , fN
where
fn f0
n , . . . , fH−1
n .
A feasible atomic energy consumption scheduling set
for user n:
Fn = fn
s∈Sn
fs
n=1, fs
n∈ {0, 1} , ∀s∈Sn, fs
n=0, ∀s∈HSn
where Sn is the feasible set of starting times for user n
that is already defined in starting-time-based
formulation.
27 / 55
Atomic Optimal Scheduling
Atomic optimal scheduling for an objective function
of φ(·) is formulated as follows:
minimize
fn∈Fn,∀n∈N
φ (L (f)) .
where
L (f) [L0 (f) , . . . , LH−1 (f)]: A vector of aggregate loads
across all flows at each time slot, which are defined as
Lh(f)
n∈N
γ
op
n ((h−s) modH)


s∈Sn
fs
nIRn(s)(h)


.
Note that for a convex objective function, this
problem becomes a Boolean-convex problem, since the
optimization variable fs
n is restricted to only 0 or 1.
28 / 55
Atomic Optimal Scheduling
Atomic optimal scheduling for an objective function
of φ(·) is formulated as follows:
minimize
fn∈Fn,∀n∈N
φ (L (f)) .
where
L (f) [L0 (f) , . . . , LH−1 (f)]: A vector of aggregate loads
across all flows at each time slot, which are defined as
Lh(f)
n∈N
γ
op
n ((h−s) modH)


s∈Sn
fs
nIRn(s)(h)


.
Note that for a convex objective function, this
problem becomes a Boolean-convex problem, since the
optimization variable fs
n is restricted to only 0 or 1.
28 / 55
Atomic Optimal Scheduling
Atomic optimal scheduling for an objective function
of φ(·) is formulated as follows:
minimize
fn∈Fn,∀n∈N
φ (L (f)) .
where
L (f) [L0 (f) , . . . , LH−1 (f)]: A vector of aggregate loads
across all flows at each time slot, which are defined as
Lh(f)
n∈N
γ
op
n ((h−s) modH)


s∈Sn
fs
nIRn(s)(h)


.
Note that for a convex objective function, this
problem becomes a Boolean-convex problem, since the
optimization variable fs
n is restricted to only 0 or 1.
28 / 55
Atomic Optimal Scheduling
Atomic optimal scheduling for an objective function
of φ(·) is formulated as follows:
minimize
fn∈Fn,∀n∈N
φ (L (f)) .
where
L (f) [L0 (f) , . . . , LH−1 (f)]: A vector of aggregate loads
across all flows at each time slot, which are defined as
Lh(f)
n∈N
γ
op
n ((h−s) modH)


s∈Sn
fs
nIRn(s)(h)


.
Note that for a convex objective function, this
problem becomes a Boolean-convex problem, since the
optimization variable fs
n is restricted to only 0 or 1.
28 / 55
Next . . .
Introduction
Review of Current Formulation of Appliance Energy
Consumption Scheduling
Atomic Scheduling of Appliance Energy Consumption
Starting-Time-Based Formulation
Optimal-Routing-Based Formulation
Successive Convex Relaxation
Examples
Numerical Results
Experimental Parameters
Comparison of Bounds with True Optimal Values
Cost Minimization
PAR Minimization
Conclusions
29 / 55
Relaxed Atomic Optimal Scheduling
We can relax the atomic optimal scheduling problem
by replacing fs
n∈ {0, 1} with 0≤fs
n≤1 in constraints as
follows:
minimize
fn∈ ˆFn,∀n∈N
φ (L (f))
where
ˆFn = fn
s∈Sn
fs
n=1, 0 ≤ fs
n ≤ 1, ∀s∈Sn, fs
n=0, ∀s∈HSn .
For a convex objective function, this problem
becomes convex because ˆFn is now a convex set. It
can be solved efficiently, for instance, using the
well-known interior-point method [1].
30 / 55
Relaxed Atomic Optimal Scheduling
We can relax the atomic optimal scheduling problem
by replacing fs
n∈ {0, 1} with 0≤fs
n≤1 in constraints as
follows:
minimize
fn∈ ˆFn,∀n∈N
φ (L (f))
where
ˆFn = fn
s∈Sn
fs
n=1, 0 ≤ fs
n ≤ 1, ∀s∈Sn, fs
n=0, ∀s∈HSn .
For a convex objective function, this problem
becomes convex because ˆFn is now a convex set. It
can be solved efficiently, for instance, using the
well-known interior-point method [1].
30 / 55
Relaxed Atomic Optimal Scheduling
We can relax the atomic optimal scheduling problem
by replacing fs
n∈ {0, 1} with 0≤fs
n≤1 in constraints as
follows:
minimize
fn∈ ˆFn,∀n∈N
φ (L (f))
where
ˆFn = fn
s∈Sn
fs
n=1, 0 ≤ fs
n ≤ 1, ∀s∈Sn, fs
n=0, ∀s∈HSn .
For a convex objective function, this problem
becomes convex because ˆFn is now a convex set. It
can be solved efficiently, for instance, using the
well-known interior-point method [1].
30 / 55
Relaxed vs. Original Scheduling Problems
The relaxed atomic optimal scheduling problem is
not equivalent to the original problem.
The elements of the optimal solution from the relaxed
problem can take fractional values (e.g., 0.75).
The optimal solution of the relaxed problem,
however, provides a lower bound on the optimal
solution of the original problem.
The feasible set for the relaxed problem contains the
feasible set for the original problem.
31 / 55
Relaxed vs. Original Scheduling Problems
The relaxed atomic optimal scheduling problem is
not equivalent to the original problem.
The elements of the optimal solution from the relaxed
problem can take fractional values (e.g., 0.75).
The optimal solution of the relaxed problem,
however, provides a lower bound on the optimal
solution of the original problem.
The feasible set for the relaxed problem contains the
feasible set for the original problem.
31 / 55
Relaxed vs. Original Scheduling Problems
The relaxed atomic optimal scheduling problem is
not equivalent to the original problem.
The elements of the optimal solution from the relaxed
problem can take fractional values (e.g., 0.75).
The optimal solution of the relaxed problem,
however, provides a lower bound on the optimal
solution of the original problem.
The feasible set for the relaxed problem contains the
feasible set for the original problem.
31 / 55
Relaxed vs. Original Scheduling Problems
The relaxed atomic optimal scheduling problem is
not equivalent to the original problem.
The elements of the optimal solution from the relaxed
problem can take fractional values (e.g., 0.75).
The optimal solution of the relaxed problem,
however, provides a lower bound on the optimal
solution of the original problem.
The feasible set for the relaxed problem contains the
feasible set for the original problem.
31 / 55
Successive Convex Relaxation
First, we solve the relaxed convex optimization problem.
Then, carry out the following procedures:
1. Identify the maximum element of each user flow
configuration vector (i.e., corresponding to fn) and
exclude them in the following procedures.
2. Arrange in ascending order the remaining elements
that are less than 1.
3. Drop the smallest element and add a zero constraint
for it.
32 / 55
Successive Convex Relaxation
First, we solve the relaxed convex optimization problem.
Then, carry out the following procedures:
1. Identify the maximum element of each user flow
configuration vector (i.e., corresponding to fn) and
exclude them in the following procedures.
2. Arrange in ascending order the remaining elements
that are less than 1.
3. Drop the smallest element and add a zero constraint
for it.
32 / 55
Successive Convex Relaxation
First, we solve the relaxed convex optimization problem.
Then, carry out the following procedures:
1. Identify the maximum element of each user flow
configuration vector (i.e., corresponding to fn) and
exclude them in the following procedures.
2. Arrange in ascending order the remaining elements
that are less than 1.
3. Drop the smallest element and add a zero constraint
for it.
32 / 55
Successive Convex Relaxation
First, we solve the relaxed convex optimization problem.
Then, carry out the following procedures:
1. Identify the maximum element of each user flow
configuration vector (i.e., corresponding to fn) and
exclude them in the following procedures.
2. Arrange in ascending order the remaining elements
that are less than 1.
3. Drop the smallest element and add a zero constraint
for it.
32 / 55
Successive Convex Relaxation (Cont.)
4. For the rest of the elements, drop them and add zero
constraints from the smallest element up to ND
elements in total (including the one in step 3) as far as
the element is less than a dropping threshold (θD);
otherwise, stop dropping and go to the next step.
5. If there remains only one nonzero element per user
flow configuration vector, stop here (a solution
found); otherwise, solve a new relaxed convex
optimization problem with augmented constraints
and repeat the whole procedure from step 1.
33 / 55
Successive Convex Relaxation (Cont.)
4. For the rest of the elements, drop them and add zero
constraints from the smallest element up to ND
elements in total (including the one in step 3) as far as
the element is less than a dropping threshold (θD);
otherwise, stop dropping and go to the next step.
5. If there remains only one nonzero element per user
flow configuration vector, stop here (a solution
found); otherwise, solve a new relaxed convex
optimization problem with augmented constraints
and repeat the whole procedure from step 1.
33 / 55
Successive Convex Relaxation: An Example
Consider a simple case of N=2, H=4, and ND = 1.
Initial condition:
f = [ 0.25 0.25 0.25 0.25 | 0.3 0.2 0.25 0.25 ]
After 1st step:
f = [ 0.0 0.2 0.3 0.5 | 0.7 0.0 0.2 0.1 ]
After 2nd step:
f = [ 0.0 0.0 0.2 0.8 | 0.9 0.0 0.1 0.0 ]
Stop here. Solution found!
34 / 55
Successive Convex Relaxation: An Example
Consider a simple case of N=2, H=4, and ND = 1.
Initial condition:
f = [ 0.25 0.25 0.25 0.25 | 0.3 0.2 0.25 0.25 ]
After 1st step:
f = [ 0.0 0.2 0.3 0.5 | 0.7 0.0 0.2 0.1 ]
After 2nd step:
f = [ 0.0 0.0 0.2 0.8 | 0.9 0.0 0.1 0.0 ]
Stop here. Solution found!
34 / 55
Successive Convex Relaxation: An Example
Consider a simple case of N=2, H=4, and ND = 1.
Initial condition:
f = [ 0.25 0.25 0.25 0.25 | 0.3 0.2 0.25 0.25 ]
After 1st step:
f = [ 0.0 0.2 0.3 0.5 | 0.7 0.0 0.2 0.1 ]
After 2nd step:
f = [ 0.0 0.0 0.2 0.8 | 0.9 0.0 0.1 0.0 ]
Stop here. Solution found!
34 / 55
Successive Convex Relaxation: An Example
Consider a simple case of N=2, H=4, and ND = 1.
Initial condition:
f = [ 0.25 0.25 0.25 0.25 | 0.3 0.2 0.25 0.25 ]
After 1st step:
f = [ 0.0 0.2 0.3 0.5 | 0.7 0.0 0.2 0.1 ]
After 2nd step:
f = [ 0.0 0.0 0.2 0.8 | 0.9 0.0 0.1 0.0 ]
Stop here. Solution found!
34 / 55
Next . . .
Introduction
Review of Current Formulation of Appliance Energy
Consumption Scheduling
Atomic Scheduling of Appliance Energy Consumption
Starting-Time-Based Formulation
Optimal-Routing-Based Formulation
Successive Convex Relaxation
Examples
Numerical Results
Experimental Parameters
Comparison of Bounds with True Optimal Values
Cost Minimization
PAR Minimization
Conclusions
35 / 55
Energy Cost and PAR Minimization
Energy cost minization
minimize
fn∈ ˆFn,∀n∈N
h∈H
Ch (Lh(f)) .
PAR minimization
minimize
Γ,fn∈ ˆFn,∀n∈N
Γ
subject to Γ ≥ Lh(f), ∀h ∈ H.
Note that PAR minimization is formulated as a relaxed
linear program by introducing a new auxiliary variable
Γ.
36 / 55
Energy Cost and PAR Minimization
Energy cost minization
minimize
fn∈ ˆFn,∀n∈N
h∈H
Ch (Lh(f)) .
PAR minimization
minimize
Γ,fn∈ ˆFn,∀n∈N
Γ
subject to Γ ≥ Lh(f), ∀h ∈ H.
Note that PAR minimization is formulated as a relaxed
linear program by introducing a new auxiliary variable
Γ.
36 / 55
Energy Cost and PAR Minimization
Energy cost minization
minimize
fn∈ ˆFn,∀n∈N
h∈H
Ch (Lh(f)) .
PAR minimization
minimize
Γ,fn∈ ˆFn,∀n∈N
Γ
subject to Γ ≥ Lh(f), ∀h ∈ H.
Note that PAR minimization is formulated as a relaxed
linear program by introducing a new auxiliary variable
Γ.
36 / 55
Next . . .
Introduction
Review of Current Formulation of Appliance Energy
Consumption Scheduling
Atomic Scheduling of Appliance Energy Consumption
Starting-Time-Based Formulation
Optimal-Routing-Based Formulation
Successive Convex Relaxation
Examples
Numerical Results
Experimental Parameters
Comparison of Bounds with True Optimal Values
Cost Minimization
PAR Minimization
Conclusions
37 / 55
Next . . .
Introduction
Review of Current Formulation of Appliance Energy
Consumption Scheduling
Atomic Scheduling of Appliance Energy Consumption
Starting-Time-Based Formulation
Optimal-Routing-Based Formulation
Successive Convex Relaxation
Examples
Numerical Results
Experimental Parameters
Comparison of Bounds with True Optimal Values
Cost Minimization
PAR Minimization
Conclusions
38 / 55
Appliance Energy Consumption
Requirements
Appliance
Parameters
α [h] β [h] γop
[kWh] δ [h]
Dish Washer 0 23 0.7200 2
Washing Machine
0 23 0.4967 3
(Energy Star)
Washing Machine
0 23 0.6467 3
(Regular)
Clothes Dryer 0 23 0.6250 4
PHEV1
222
292
3.3000 3
1
Plug-in hybrid electric vehicle.
2
Scheduling interval of 10 PM–5 AM.
An appliance is randomly selected for each user.
Constant operating energy levels (γop
) are assumed. 39 / 55
Hourly Cost Function
We assume a simple quadratic hourly cost function as in
[3], i.e.,
Ch (Lh) = ahL2
h [cent]
where
ah =
0.2 if h ∈ [0, 7],
0.3 if h ∈ [8, 23].
40 / 55
Hourly Cost Function
We assume a simple quadratic hourly cost function as in
[3], i.e.,
Ch (Lh) = ahL2
h [cent]
where
ah =
0.2 if h ∈ [0, 7],
0.3 if h ∈ [8, 23].
40 / 55
Performance Measures and Parameter Values
Performance measures
Lower bound: LB
Upper bound with ND: UB (ND)
Gap: G UB(D) − LB
Number of iterations
Parameter values for successive convex relaxation
Dropping threshold (θD): 0.1
Maximum number of fractional-valued elements that
can be dropped per iteration (ND): 1, 2, 5, 10
41 / 55
Performance Measures and Parameter Values
Performance measures
Lower bound: LB
Upper bound with ND: UB (ND)
Gap: G UB(D) − LB
Number of iterations
Parameter values for successive convex relaxation
Dropping threshold (θD): 0.1
Maximum number of fractional-valued elements that
can be dropped per iteration (ND): 1, 2, 5, 10
41 / 55
Performance Measures and Parameter Values
Performance measures
Lower bound: LB
Upper bound with ND: UB (ND)
Gap: G UB(D) − LB
Number of iterations
Parameter values for successive convex relaxation
Dropping threshold (θD): 0.1
Maximum number of fractional-valued elements that
can be dropped per iteration (ND): 1, 2, 5, 10
41 / 55
Performance Measures and Parameter Values
Performance measures
Lower bound: LB
Upper bound with ND: UB (ND)
Gap: G UB(D) − LB
Number of iterations
Parameter values for successive convex relaxation
Dropping threshold (θD): 0.1
Maximum number of fractional-valued elements that
can be dropped per iteration (ND): 1, 2, 5, 10
41 / 55
Performance Measures and Parameter Values
Performance measures
Lower bound: LB
Upper bound with ND: UB (ND)
Gap: G UB(D) − LB
Number of iterations
Parameter values for successive convex relaxation
Dropping threshold (θD): 0.1
Maximum number of fractional-valued elements that
can be dropped per iteration (ND): 1, 2, 5, 10
41 / 55
Performance Measures and Parameter Values
Performance measures
Lower bound: LB
Upper bound with ND: UB (ND)
Gap: G UB(D) − LB
Number of iterations
Parameter values for successive convex relaxation
Dropping threshold (θD): 0.1
Maximum number of fractional-valued elements that
can be dropped per iteration (ND): 1, 2, 5, 10
41 / 55
Performance Measures and Parameter Values
Performance measures
Lower bound: LB
Upper bound with ND: UB (ND)
Gap: G UB(D) − LB
Number of iterations
Parameter values for successive convex relaxation
Dropping threshold (θD): 0.1
Maximum number of fractional-valued elements that
can be dropped per iteration (ND): 1, 2, 5, 10
41 / 55
Performance Measures and Parameter Values
Performance measures
Lower bound: LB
Upper bound with ND: UB (ND)
Gap: G UB(D) − LB
Number of iterations
Parameter values for successive convex relaxation
Dropping threshold (θD): 0.1
Maximum number of fractional-valued elements that
can be dropped per iteration (ND): 1, 2, 5, 10
41 / 55
Next . . .
Introduction
Review of Current Formulation of Appliance Energy
Consumption Scheduling
Atomic Scheduling of Appliance Energy Consumption
Starting-Time-Based Formulation
Optimal-Routing-Based Formulation
Successive Convex Relaxation
Examples
Numerical Results
Experimental Parameters
Comparison of Bounds with True Optimal Values
Cost Minimization
PAR Minimization
Conclusions
42 / 55
Upper/Lower Bounds vs. True Optimal
Values: Cost Minimization
2 3 4 5 6 7 8 9 10
0
5 · 10−2
0.1
0.15
0.2
0.25
0.3
0.35
N
EnergyCost[USD] LB
GO
UB(1)
UB(2)
UB(5)
UB(10)
43 / 55
Upper/Lower Bounds vs. True Optimal
Values: PAR Minimization
2 3 4 5 6 7 8 9 10
0
1
2
3
4
5
6
7
8
9
10
N
PARinAggregatedLoad LB
GO
UB(1)
UB(2)
UB(5)
UB(10)
44 / 55
Next . . .
Introduction
Review of Current Formulation of Appliance Energy
Consumption Scheduling
Atomic Scheduling of Appliance Energy Consumption
Starting-Time-Based Formulation
Optimal-Routing-Based Formulation
Successive Convex Relaxation
Examples
Numerical Results
Experimental Parameters
Comparison of Bounds with True Optimal Values
Cost Minimization
PAR Minimization
Conclusions
45 / 55
Cost Minimization: Upper/Lower Bounds
2 10 20 30 40 50
0
1
2
3
4
5
N
EnergyCost[USD]
LB
UB(1)
UB(2)
UB(5)
UB(10)
46 / 55
Cost Minimization: Gaps
2 10 20 30 40 50
0
1
2
3
4
5
6
7
×10−2
N
Gap[USD]
UB(1)−LB
UB(2)−LB
UB(5)−LB
UB(10)−LB
47 / 55
Cost Minimization: Number of Iterations
2 10 20 30 40 50
0
200
400
600
800
1,000
N
NumberofIterations
UB(1)
UB(2)
UB(5)
UB(10)
48 / 55
Next . . .
Introduction
Review of Current Formulation of Appliance Energy
Consumption Scheduling
Atomic Scheduling of Appliance Energy Consumption
Starting-Time-Based Formulation
Optimal-Routing-Based Formulation
Successive Convex Relaxation
Examples
Numerical Results
Experimental Parameters
Comparison of Bounds with True Optimal Values
Cost Minimization
PAR Minimization
Conclusions
49 / 55
Par Minimization: Upper/Lower Bounds
2 10 20 30 40 50
0
5
10
15
20
25
N
PARinAggregatedLoad
LB
UB(1)
UB(2)
UB(5)
UB(10)
50 / 55
Par Minimization: Gaps
2 10 20 30 40 50
0
1
2
3
N
Gap
UB(1)−LB
UB(2)−LB
UB(5)−LB
UB(10)−LB
51 / 55
Par Minimization: Number of Iterations
2 10 20 30 40 50
0
200
400
600
800
1,000
N
NumberofIterations
UB(1)
UB(2)
UB(5)
UB(10)
52 / 55
Next . . .
Introduction
Review of Current Formulation of Appliance Energy
Consumption Scheduling
Atomic Scheduling of Appliance Energy Consumption
Starting-Time-Based Formulation
Optimal-Routing-Based Formulation
Successive Convex Relaxation
Examples
Numerical Results
Experimental Parameters
Comparison of Bounds with True Optimal Values
Cost Minimization
PAR Minimization
Conclusions
53 / 55
Conclusions
We have provided a new formulation of appliance energy
consumption scheduling based on the optimal routing
framework.
It guarantees the atomicity of resulting scheduled
energy consumption.
We have also provided an efficient solution technique
based on successive convex relaxation for a
Boolean-convex problem resulting from a convex
objective function.
It enables us to carry out systematic analysis of the
original problem with both upper and lower bounds.
Possible extensions to the current work include
Distributed atomic energy consumption scheduling;
Advanced techniques refining the feasible region at
each iterative step to reduce the total number of
iterations.
54 / 55
Conclusions
We have provided a new formulation of appliance energy
consumption scheduling based on the optimal routing
framework.
It guarantees the atomicity of resulting scheduled
energy consumption.
We have also provided an efficient solution technique
based on successive convex relaxation for a
Boolean-convex problem resulting from a convex
objective function.
It enables us to carry out systematic analysis of the
original problem with both upper and lower bounds.
Possible extensions to the current work include
Distributed atomic energy consumption scheduling;
Advanced techniques refining the feasible region at
each iterative step to reduce the total number of
iterations.
54 / 55
Conclusions
We have provided a new formulation of appliance energy
consumption scheduling based on the optimal routing
framework.
It guarantees the atomicity of resulting scheduled
energy consumption.
We have also provided an efficient solution technique
based on successive convex relaxation for a
Boolean-convex problem resulting from a convex
objective function.
It enables us to carry out systematic analysis of the
original problem with both upper and lower bounds.
Possible extensions to the current work include
Distributed atomic energy consumption scheduling;
Advanced techniques refining the feasible region at
each iterative step to reduce the total number of
iterations.
54 / 55
Conclusions
We have provided a new formulation of appliance energy
consumption scheduling based on the optimal routing
framework.
It guarantees the atomicity of resulting scheduled
energy consumption.
We have also provided an efficient solution technique
based on successive convex relaxation for a
Boolean-convex problem resulting from a convex
objective function.
It enables us to carry out systematic analysis of the
original problem with both upper and lower bounds.
Possible extensions to the current work include
Distributed atomic energy consumption scheduling;
Advanced techniques refining the feasible region at
each iterative step to reduce the total number of
iterations.
54 / 55
Conclusions
We have provided a new formulation of appliance energy
consumption scheduling based on the optimal routing
framework.
It guarantees the atomicity of resulting scheduled
energy consumption.
We have also provided an efficient solution technique
based on successive convex relaxation for a
Boolean-convex problem resulting from a convex
objective function.
It enables us to carry out systematic analysis of the
original problem with both upper and lower bounds.
Possible extensions to the current work include
Distributed atomic energy consumption scheduling;
Advanced techniques refining the feasible region at
each iterative step to reduce the total number of
iterations.
54 / 55
Conclusions
We have provided a new formulation of appliance energy
consumption scheduling based on the optimal routing
framework.
It guarantees the atomicity of resulting scheduled
energy consumption.
We have also provided an efficient solution technique
based on successive convex relaxation for a
Boolean-convex problem resulting from a convex
objective function.
It enables us to carry out systematic analysis of the
original problem with both upper and lower bounds.
Possible extensions to the current work include
Distributed atomic energy consumption scheduling;
Advanced techniques refining the feasible region at
each iterative step to reduce the total number of
iterations.
54 / 55
Conclusions
We have provided a new formulation of appliance energy
consumption scheduling based on the optimal routing
framework.
It guarantees the atomicity of resulting scheduled
energy consumption.
We have also provided an efficient solution technique
based on successive convex relaxation for a
Boolean-convex problem resulting from a convex
objective function.
It enables us to carry out systematic analysis of the
original problem with both upper and lower bounds.
Possible extensions to the current work include
Distributed atomic energy consumption scheduling;
Advanced techniques refining the feasible region at
each iterative step to reduce the total number of
iterations.
54 / 55
References I
S. Boyd and L. Vandenberghe, Convex optimization.
Cambridge, U.K.: Cambridge Universtiy Press, 2004.
J. B. Rosen, “Existence and uniqueness of equilibrium
for concave N-person games,” Econometrica, vol. 33,
no. 3, pp. 520–534, Jul. 1965.
A.-H. Mohsenian-Rad, V. W. S. Wong, J. Jatskevich,
R. Schober, and A. Leon-Garcia, “Autonomous
demand-side management based on game-theoretic
energy consumption scheduling for the future smart
grid,” IEEE Trans. Smart Grid, vol. 1, no. 3, pp.
320–331, Dec. 2010.
55 / 55

Atomic Scheduling of Appliance Energy Consumption in Residential Smart Grid

  • 1.
    Atomic Scheduling ofAppliance Energy Consumption in Residential Smart Grid Kyeong Soo (Joseph) Kim (With S. Lee, T. O. Ting@XJTLU and X.-S. Yang@Middlesex) Department of Electrical and Electronic Engineering Xi’an Jiaotong-Liverpool University CeSGIC 1st International Workshop on Smart Grid Technology and Data Processing 19 June 2015 1 / 55
  • 2.
    Outline Introduction Review of CurrentFormulation of Appliance Energy Consumption Scheduling Atomic Scheduling of Appliance Energy Consumption Starting-Time-Based Formulation Optimal-Routing-Based Formulation Successive Convex Relaxation Examples Numerical Results Experimental Parameters Comparison of Bounds with True Optimal Values Cost Minimization PAR Minimization Conclusions 2 / 55
  • 4.
    Next . .. Introduction Review of Current Formulation of Appliance Energy Consumption Scheduling Atomic Scheduling of Appliance Energy Consumption Starting-Time-Based Formulation Optimal-Routing-Based Formulation Successive Convex Relaxation Examples Numerical Results Experimental Parameters Comparison of Bounds with True Optimal Values Cost Minimization PAR Minimization Conclusions 4 / 55
  • 5.
    Autonomous Demand-Side Managementin Smart Grid Gateway (Traditional) Electricity Grid Greenfield Power Line Bi-Directional Communication Links Power Plant 5 / 55
  • 6.
    Scheduling of ApplianceEnergy Consumption A key to autonomous DSM in optimizing energy production and consumption. Based on two-way digital communications between a utility company and users through smart meters at users’ premises. Typical objectives Peak-to-average ratio (PAR) Total energy cost 6 / 55
  • 7.
    Scheduling of ApplianceEnergy Consumption A key to autonomous DSM in optimizing energy production and consumption. Based on two-way digital communications between a utility company and users through smart meters at users’ premises. Typical objectives Peak-to-average ratio (PAR) Total energy cost 6 / 55
  • 8.
    Scheduling of ApplianceEnergy Consumption A key to autonomous DSM in optimizing energy production and consumption. Based on two-way digital communications between a utility company and users through smart meters at users’ premises. Typical objectives Peak-to-average ratio (PAR) Total energy cost 6 / 55
  • 9.
    Scheduling of ApplianceEnergy Consumption A key to autonomous DSM in optimizing energy production and consumption. Based on two-way digital communications between a utility company and users through smart meters at users’ premises. Typical objectives Peak-to-average ratio (PAR) Total energy cost 6 / 55
  • 10.
    A Question onScheduled Energy Consumption Can a washing machine successfully complete its job with the energy consumption scheduled as follows? 7 / 55
  • 11.
    A Question onScheduled Energy Consumption Can a washing machine successfully complete its job with the energy consumption scheduled as follows? 9am 3pm12am 7 / 55
  • 12.
    A Question onScheduled Energy Consumption Can a washing machine successfully complete its job with the energy consumption scheduled as follows? 9am 3pm12am or 7 / 55
  • 13.
    A Question onScheduled Energy Consumption Can a washing machine successfully complete its job with the energy consumption scheduled as follows? 9am 3pm12am or 9am 3pm2pm10am 7 / 55
  • 14.
    Next . .. Introduction Review of Current Formulation of Appliance Energy Consumption Scheduling Atomic Scheduling of Appliance Energy Consumption Starting-Time-Based Formulation Optimal-Routing-Based Formulation Successive Convex Relaxation Examples Numerical Results Experimental Parameters Comparison of Bounds with True Optimal Values Cost Minimization PAR Minimization Conclusions 8 / 55
  • 15.
    Optimization Variables Based onenergy consumption over equally-divided time slots of a day (typically hourly time slots) as optimization variables, i.e., xn x0 n, . . . , xh n, . . . , xH−1 n for User n ∈ N {1, . . ., N}; Time slot h ∈ H {0, . . ., H−1}. Because the optimization variables take continuous values, we can easily apply Convex optimization [1]; Distributed algorithms through concave n-person games [2]. 9 / 55
  • 16.
    Optimization Variables Based onenergy consumption over equally-divided time slots of a day (typically hourly time slots) as optimization variables, i.e., xn x0 n, . . . , xh n, . . . , xH−1 n for User n ∈ N {1, . . ., N}; Time slot h ∈ H {0, . . ., H−1}. Because the optimization variables take continuous values, we can easily apply Convex optimization [1]; Distributed algorithms through concave n-person games [2]. 9 / 55
  • 17.
    Optimization Variables Based onenergy consumption over equally-divided time slots of a day (typically hourly time slots) as optimization variables, i.e., xn x0 n, . . . , xh n, . . . , xH−1 n for User n ∈ N {1, . . ., N}; Time slot h ∈ H {0, . . ., H−1}. Because the optimization variables take continuous values, we can easily apply Convex optimization [1]; Distributed algorithms through concave n-person games [2]. 9 / 55
  • 18.
    Optimization Variables Based onenergy consumption over equally-divided time slots of a day (typically hourly time slots) as optimization variables, i.e., xn x0 n, . . . , xh n, . . . , xH−1 n for User n ∈ N {1, . . ., N}; Time slot h ∈ H {0, . . ., H−1}. Because the optimization variables take continuous values, we can easily apply Convex optimization [1]; Distributed algorithms through concave n-person games [2]. 9 / 55
  • 19.
    Optimization Variables Based onenergy consumption over equally-divided time slots of a day (typically hourly time slots) as optimization variables, i.e., xn x0 n, . . . , xh n, . . . , xH−1 n for User n ∈ N {1, . . ., N}; Time slot h ∈ H {0, . . ., H−1}. Because the optimization variables take continuous values, we can easily apply Convex optimization [1]; Distributed algorithms through concave n-person games [2]. 9 / 55
  • 20.
    Optimization Variables Based onenergy consumption over equally-divided time slots of a day (typically hourly time slots) as optimization variables, i.e., xn x0 n, . . . , xh n, . . . , xH−1 n for User n ∈ N {1, . . ., N}; Time slot h ∈ H {0, . . ., H−1}. Because the optimization variables take continuous values, we can easily apply Convex optimization [1]; Distributed algorithms through concave n-person games [2]. 9 / 55
  • 21.
    Feasible Set A feasibleenergy consumption scheduling set for user n Xn= xn h∈Hn xh n=En, γmin n ≤xh n≤γmax n , ∀h∈Hn, xh n=0, ∀h∈HHn where γmin n : Minimum energy level; γmax n : Maximum energy level; En: Total daily energy consumption; Hn: Scheduling interval defined as follows: Hn h h = i mod H, ∀i∈ αn, βn with αn∈[0, H−1], βn∈[1, 2H−2], and 1≤βn−αn≤H−1. 10 / 55
  • 22.
    Feasible Set A feasibleenergy consumption scheduling set for user n Xn= xn h∈Hn xh n=En, γmin n ≤xh n≤γmax n , ∀h∈Hn, xh n=0, ∀h∈HHn where γmin n : Minimum energy level; γmax n : Maximum energy level; En: Total daily energy consumption; Hn: Scheduling interval defined as follows: Hn h h = i mod H, ∀i∈ αn, βn with αn∈[0, H−1], βn∈[1, 2H−2], and 1≤βn−αn≤H−1. 10 / 55
  • 23.
    Feasible Set A feasibleenergy consumption scheduling set for user n Xn= xn h∈Hn xh n=En, γmin n ≤xh n≤γmax n , ∀h∈Hn, xh n=0, ∀h∈HHn where γmin n : Minimum energy level; γmax n : Maximum energy level; En: Total daily energy consumption; Hn: Scheduling interval defined as follows: Hn h h = i mod H, ∀i∈ αn, βn with αn∈[0, H−1], βn∈[1, 2H−2], and 1≤βn−αn≤H−1. 10 / 55
  • 24.
    Feasible Set A feasibleenergy consumption scheduling set for user n Xn= xn h∈Hn xh n=En, γmin n ≤xh n≤γmax n , ∀h∈Hn, xh n=0, ∀h∈HHn where γmin n : Minimum energy level; γmax n : Maximum energy level; En: Total daily energy consumption; Hn: Scheduling interval defined as follows: Hn h h = i mod H, ∀i∈ αn, βn with αn∈[0, H−1], βn∈[1, 2H−2], and 1≤βn−αn≤H−1. 10 / 55
  • 25.
    Feasible Set A feasibleenergy consumption scheduling set for user n Xn= xn h∈Hn xh n=En, γmin n ≤xh n≤γmax n , ∀h∈Hn, xh n=0, ∀h∈HHn where γmin n : Minimum energy level; γmax n : Maximum energy level; En: Total daily energy consumption; Hn: Scheduling interval defined as follows: Hn h h = i mod H, ∀i∈ αn, βn with αn∈[0, H−1], βn∈[1, 2H−2], and 1≤βn−αn≤H−1. 10 / 55
  • 26.
    Feasible Set A feasibleenergy consumption scheduling set for user n Xn= xn h∈Hn xh n=En, γmin n ≤xh n≤γmax n , ∀h∈Hn, xh n=0, ∀h∈HHn where γmin n : Minimum energy level; γmax n : Maximum energy level; En: Total daily energy consumption; Hn: Scheduling interval defined as follows: Hn h h = i mod H, ∀i∈ αn, βn with αn∈[0, H−1], βn∈[1, 2H−2], and 1≤βn−αn≤H−1. 10 / 55
  • 27.
    Optimal Scheduling The optimalscheduling is formulated as an optimization problem for a given objective function φ(·) (e.g., total energy cost or PAR) as follows: minimize xn∈Xn, ∀n∈N φ (L(x)) where x [x1, . . . , xN]: A vector of user energy consumption vectors; L(x) [L0 (x) , . . . , LH−1 (x)]: A vector of aggregate loads across all users at each time slot, which are defined as Lh (x) n∈N xh n. 11 / 55
  • 28.
    Optimal Scheduling The optimalscheduling is formulated as an optimization problem for a given objective function φ(·) (e.g., total energy cost or PAR) as follows: minimize xn∈Xn, ∀n∈N φ (L(x)) where x [x1, . . . , xN]: A vector of user energy consumption vectors; L(x) [L0 (x) , . . . , LH−1 (x)]: A vector of aggregate loads across all users at each time slot, which are defined as Lh (x) n∈N xh n. 11 / 55
  • 29.
    Optimal Scheduling The optimalscheduling is formulated as an optimization problem for a given objective function φ(·) (e.g., total energy cost or PAR) as follows: minimize xn∈Xn, ∀n∈N φ (L(x)) where x [x1, . . . , xN]: A vector of user energy consumption vectors; L(x) [L0 (x) , . . . , LH−1 (x)]: A vector of aggregate loads across all users at each time slot, which are defined as Lh (x) n∈N xh n. 11 / 55
  • 30.
    Optimal Scheduling The optimalscheduling is formulated as an optimization problem for a given objective function φ(·) (e.g., total energy cost or PAR) as follows: minimize xn∈Xn, ∀n∈N φ (L(x)) where x [x1, . . . , xN]: A vector of user energy consumption vectors; L(x) [L0 (x) , . . . , LH−1 (x)]: A vector of aggregate loads across all users at each time slot, which are defined as Lh (x) n∈N xh n. 11 / 55
  • 31.
    Objective Functions For energycost minimization: φ (L(x)) = h∈H Ch (Lh(x)) where Ch(·): A cost function for generating or distributing electricity energy at a time slot h. For PAR minimization: φ (L(x)) = H max h∈H Lh(x) n∈N En 12 / 55
  • 32.
    Objective Functions For energycost minimization: φ (L(x)) = h∈H Ch (Lh(x)) where Ch(·): A cost function for generating or distributing electricity energy at a time slot h. For PAR minimization: φ (L(x)) = H max h∈H Lh(x) n∈N En 12 / 55
  • 33.
    Atomic vs. Non-AtomicScheduling αn,a βn,a αn,a βn,a Gap Gap γmin n γmax n (a) γop n (·) γmin n γmax n (b) Examples of (a) non-atomic and (b) atomic scheduling. γmin n : Minimum energy level γmax n : Maximum energy level γ op n (·): Operating energy level 13 / 55
  • 34.
    Non-Atomic Scheduling Example Appliance1 Appliance 2 Scheduled Consumption ? 14 / 55
  • 35.
    Non-Atomic Scheduling Example:Case 1 Appliance 1 Appliance 2 Scheduled Consumption 15 / 55
  • 36.
    Non-Atomic Scheduling Example:Case 2 Appliance 1 Appliance 2 Scheduled Consumption 16 / 55
  • 37.
    Non-Atomic Scheduling Example:Case 3 Appliance 1 Appliance 2 Scheduled Consumption 17 / 55
  • 38.
    Next . .. Introduction Review of Current Formulation of Appliance Energy Consumption Scheduling Atomic Scheduling of Appliance Energy Consumption Starting-Time-Based Formulation Optimal-Routing-Based Formulation Successive Convex Relaxation Examples Numerical Results Experimental Parameters Comparison of Bounds with True Optimal Values Cost Minimization PAR Minimization Conclusions 18 / 55
  • 39.
    Overview of AtomicScheduling Problem Formulation & Solution Starting-Time-Based Formulation Optimal-Routing-Based Formulation Convex Relaxation Successive Convex Relaxation with Fractional-Value Dropping Convex Optimization (Feasible Upper Bound) Combinatorial Optimization Boolean-Convex Optmization Convex Optmization (Lower Bound) 19 / 55
  • 40.
    Next . .. Introduction Review of Current Formulation of Appliance Energy Consumption Scheduling Atomic Scheduling of Appliance Energy Consumption Starting-Time-Based Formulation Optimal-Routing-Based Formulation Successive Convex Relaxation Examples Numerical Results Experimental Parameters Comparison of Bounds with True Optimal Values Cost Minimization PAR Minimization Conclusions 20 / 55
  • 41.
    Starting-Time-Based Formulation Optimization variables: s[s1, . . . , sN] A feasible set for user n: Sn sn sn = i mod H, ∀i∈[αn, βn−δn+1] Aggregate load across all users at each time slot h: Lh(s) n∈N γ op n ((h − sn) mod H) IRn(sn)(h) where IRn(sn)(h): An indicator function for a set Rn(sn). Rn(sn): A range of user n’s appliance operation for sn defined as follows: Rn(sn) h h = i mod H, ∀i∈ [sn, sn + δn − 1] 21 / 55
  • 42.
    Starting-Time-Based Formulation Optimization variables: s[s1, . . . , sN] A feasible set for user n: Sn sn sn = i mod H, ∀i∈[αn, βn−δn+1] Aggregate load across all users at each time slot h: Lh(s) n∈N γ op n ((h − sn) mod H) IRn(sn)(h) where IRn(sn)(h): An indicator function for a set Rn(sn). Rn(sn): A range of user n’s appliance operation for sn defined as follows: Rn(sn) h h = i mod H, ∀i∈ [sn, sn + δn − 1] 21 / 55
  • 43.
    Starting-Time-Based Formulation Optimization variables: s[s1, . . . , sN] A feasible set for user n: Sn sn sn = i mod H, ∀i∈[αn, βn−δn+1] Aggregate load across all users at each time slot h: Lh(s) n∈N γ op n ((h − sn) mod H) IRn(sn)(h) where IRn(sn)(h): An indicator function for a set Rn(sn). Rn(sn): A range of user n’s appliance operation for sn defined as follows: Rn(sn) h h = i mod H, ∀i∈ [sn, sn + δn − 1] 21 / 55
  • 44.
    Starting-Time-Based Formulation Optimization variables: s[s1, . . . , sN] A feasible set for user n: Sn sn sn = i mod H, ∀i∈[αn, βn−δn+1] Aggregate load across all users at each time slot h: Lh(s) n∈N γ op n ((h − sn) mod H) IRn(sn)(h) where IRn(sn)(h): An indicator function for a set Rn(sn). Rn(sn): A range of user n’s appliance operation for sn defined as follows: Rn(sn) h h = i mod H, ∀i∈ [sn, sn + δn − 1] 21 / 55
  • 45.
    Starting-Time-Based Formulation Optimization variables: s[s1, . . . , sN] A feasible set for user n: Sn sn sn = i mod H, ∀i∈[αn, βn−δn+1] Aggregate load across all users at each time slot h: Lh(s) n∈N γ op n ((h − sn) mod H) IRn(sn)(h) where IRn(sn)(h): An indicator function for a set Rn(sn). Rn(sn): A range of user n’s appliance operation for sn defined as follows: Rn(sn) h h = i mod H, ∀i∈ [sn, sn + δn − 1] 21 / 55
  • 46.
    Starting-Time-Based Formulation Optimization variables: s[s1, . . . , sN] A feasible set for user n: Sn sn sn = i mod H, ∀i∈[αn, βn−δn+1] Aggregate load across all users at each time slot h: Lh(s) n∈N γ op n ((h − sn) mod H) IRn(sn)(h) where IRn(sn)(h): An indicator function for a set Rn(sn). Rn(sn): A range of user n’s appliance operation for sn defined as follows: Rn(sn) h h = i mod H, ∀i∈ [sn, sn + δn − 1] 21 / 55
  • 47.
    Issues with Starting-Time-BasedFormulation Because the feasible set is now discrete, we have to evaluate the objective function for all the elements in the feasible set. The optimization by direct enumeration becomes impractical for large N and H. When N=100 and H=24 with the worst case scenario of αn=0, βn=23, and δn=1 for all n∈N, we need to evaluate the objective function 24100 times, which is on the order of 10138 times! 22 / 55
  • 48.
    Next . .. Introduction Review of Current Formulation of Appliance Energy Consumption Scheduling Atomic Scheduling of Appliance Energy Consumption Starting-Time-Based Formulation Optimal-Routing-Based Formulation Successive Convex Relaxation Examples Numerical Results Experimental Parameters Comparison of Bounds with True Optimal Values Cost Minimization PAR Minimization Conclusions 23 / 55
  • 49.
    Why London Eye? Doesthe London Eye have something to do with the atomic scheduling? 24 / 55
  • 50.
    A Network, APath, and Links S D 0 1 2 3 4 5 6 7 8 9 10 111213 14 15 16 17 18 19 20 21 22 23 l9,10 l10,11l11,12 p9,3 A network connecting the source (S) and the destination (D) through 24 intermediate nodes with a path (p9,3 ) and its constituent links (l9,10 , l10,11 , and l11,12 ). 25 / 55
  • 51.
    Mapping of AtomicOperations to Flows 0 1 2 3 4 5 6 7 8 9 10 111213 14 15 16 17 18 19 20 21 22 23 f0 1 f1 1 f2 1 f3 1 f4 1 f9 2 f10 2 f11 2f12 2 S D Mapping of all possible atomic operations of two appliances into two groups of flows (f0 1 , . . . , f4 1 and f9 2 , . . . , f12 2 ) over multiple paths on the network. 26 / 55
  • 52.
    Optimization Variables andFeasible Set Optimization variables: Flow configurations of all users defined as f f1, . . . , fn, . . . , fN where fn f0 n , . . . , fH−1 n . A feasible atomic energy consumption scheduling set for user n: Fn = fn s∈Sn fs n=1, fs n∈ {0, 1} , ∀s∈Sn, fs n=0, ∀s∈HSn where Sn is the feasible set of starting times for user n that is already defined in starting-time-based formulation. 27 / 55
  • 53.
    Optimization Variables andFeasible Set Optimization variables: Flow configurations of all users defined as f f1, . . . , fn, . . . , fN where fn f0 n , . . . , fH−1 n . A feasible atomic energy consumption scheduling set for user n: Fn = fn s∈Sn fs n=1, fs n∈ {0, 1} , ∀s∈Sn, fs n=0, ∀s∈HSn where Sn is the feasible set of starting times for user n that is already defined in starting-time-based formulation. 27 / 55
  • 54.
    Optimization Variables andFeasible Set Optimization variables: Flow configurations of all users defined as f f1, . . . , fn, . . . , fN where fn f0 n , . . . , fH−1 n . A feasible atomic energy consumption scheduling set for user n: Fn = fn s∈Sn fs n=1, fs n∈ {0, 1} , ∀s∈Sn, fs n=0, ∀s∈HSn where Sn is the feasible set of starting times for user n that is already defined in starting-time-based formulation. 27 / 55
  • 55.
    Optimization Variables andFeasible Set Optimization variables: Flow configurations of all users defined as f f1, . . . , fn, . . . , fN where fn f0 n , . . . , fH−1 n . A feasible atomic energy consumption scheduling set for user n: Fn = fn s∈Sn fs n=1, fs n∈ {0, 1} , ∀s∈Sn, fs n=0, ∀s∈HSn where Sn is the feasible set of starting times for user n that is already defined in starting-time-based formulation. 27 / 55
  • 56.
    Atomic Optimal Scheduling Atomicoptimal scheduling for an objective function of φ(·) is formulated as follows: minimize fn∈Fn,∀n∈N φ (L (f)) . where L (f) [L0 (f) , . . . , LH−1 (f)]: A vector of aggregate loads across all flows at each time slot, which are defined as Lh(f) n∈N γ op n ((h−s) modH)   s∈Sn fs nIRn(s)(h)   . Note that for a convex objective function, this problem becomes a Boolean-convex problem, since the optimization variable fs n is restricted to only 0 or 1. 28 / 55
  • 57.
    Atomic Optimal Scheduling Atomicoptimal scheduling for an objective function of φ(·) is formulated as follows: minimize fn∈Fn,∀n∈N φ (L (f)) . where L (f) [L0 (f) , . . . , LH−1 (f)]: A vector of aggregate loads across all flows at each time slot, which are defined as Lh(f) n∈N γ op n ((h−s) modH)   s∈Sn fs nIRn(s)(h)   . Note that for a convex objective function, this problem becomes a Boolean-convex problem, since the optimization variable fs n is restricted to only 0 or 1. 28 / 55
  • 58.
    Atomic Optimal Scheduling Atomicoptimal scheduling for an objective function of φ(·) is formulated as follows: minimize fn∈Fn,∀n∈N φ (L (f)) . where L (f) [L0 (f) , . . . , LH−1 (f)]: A vector of aggregate loads across all flows at each time slot, which are defined as Lh(f) n∈N γ op n ((h−s) modH)   s∈Sn fs nIRn(s)(h)   . Note that for a convex objective function, this problem becomes a Boolean-convex problem, since the optimization variable fs n is restricted to only 0 or 1. 28 / 55
  • 59.
    Atomic Optimal Scheduling Atomicoptimal scheduling for an objective function of φ(·) is formulated as follows: minimize fn∈Fn,∀n∈N φ (L (f)) . where L (f) [L0 (f) , . . . , LH−1 (f)]: A vector of aggregate loads across all flows at each time slot, which are defined as Lh(f) n∈N γ op n ((h−s) modH)   s∈Sn fs nIRn(s)(h)   . Note that for a convex objective function, this problem becomes a Boolean-convex problem, since the optimization variable fs n is restricted to only 0 or 1. 28 / 55
  • 60.
    Next . .. Introduction Review of Current Formulation of Appliance Energy Consumption Scheduling Atomic Scheduling of Appliance Energy Consumption Starting-Time-Based Formulation Optimal-Routing-Based Formulation Successive Convex Relaxation Examples Numerical Results Experimental Parameters Comparison of Bounds with True Optimal Values Cost Minimization PAR Minimization Conclusions 29 / 55
  • 61.
    Relaxed Atomic OptimalScheduling We can relax the atomic optimal scheduling problem by replacing fs n∈ {0, 1} with 0≤fs n≤1 in constraints as follows: minimize fn∈ ˆFn,∀n∈N φ (L (f)) where ˆFn = fn s∈Sn fs n=1, 0 ≤ fs n ≤ 1, ∀s∈Sn, fs n=0, ∀s∈HSn . For a convex objective function, this problem becomes convex because ˆFn is now a convex set. It can be solved efficiently, for instance, using the well-known interior-point method [1]. 30 / 55
  • 62.
    Relaxed Atomic OptimalScheduling We can relax the atomic optimal scheduling problem by replacing fs n∈ {0, 1} with 0≤fs n≤1 in constraints as follows: minimize fn∈ ˆFn,∀n∈N φ (L (f)) where ˆFn = fn s∈Sn fs n=1, 0 ≤ fs n ≤ 1, ∀s∈Sn, fs n=0, ∀s∈HSn . For a convex objective function, this problem becomes convex because ˆFn is now a convex set. It can be solved efficiently, for instance, using the well-known interior-point method [1]. 30 / 55
  • 63.
    Relaxed Atomic OptimalScheduling We can relax the atomic optimal scheduling problem by replacing fs n∈ {0, 1} with 0≤fs n≤1 in constraints as follows: minimize fn∈ ˆFn,∀n∈N φ (L (f)) where ˆFn = fn s∈Sn fs n=1, 0 ≤ fs n ≤ 1, ∀s∈Sn, fs n=0, ∀s∈HSn . For a convex objective function, this problem becomes convex because ˆFn is now a convex set. It can be solved efficiently, for instance, using the well-known interior-point method [1]. 30 / 55
  • 64.
    Relaxed vs. OriginalScheduling Problems The relaxed atomic optimal scheduling problem is not equivalent to the original problem. The elements of the optimal solution from the relaxed problem can take fractional values (e.g., 0.75). The optimal solution of the relaxed problem, however, provides a lower bound on the optimal solution of the original problem. The feasible set for the relaxed problem contains the feasible set for the original problem. 31 / 55
  • 65.
    Relaxed vs. OriginalScheduling Problems The relaxed atomic optimal scheduling problem is not equivalent to the original problem. The elements of the optimal solution from the relaxed problem can take fractional values (e.g., 0.75). The optimal solution of the relaxed problem, however, provides a lower bound on the optimal solution of the original problem. The feasible set for the relaxed problem contains the feasible set for the original problem. 31 / 55
  • 66.
    Relaxed vs. OriginalScheduling Problems The relaxed atomic optimal scheduling problem is not equivalent to the original problem. The elements of the optimal solution from the relaxed problem can take fractional values (e.g., 0.75). The optimal solution of the relaxed problem, however, provides a lower bound on the optimal solution of the original problem. The feasible set for the relaxed problem contains the feasible set for the original problem. 31 / 55
  • 67.
    Relaxed vs. OriginalScheduling Problems The relaxed atomic optimal scheduling problem is not equivalent to the original problem. The elements of the optimal solution from the relaxed problem can take fractional values (e.g., 0.75). The optimal solution of the relaxed problem, however, provides a lower bound on the optimal solution of the original problem. The feasible set for the relaxed problem contains the feasible set for the original problem. 31 / 55
  • 68.
    Successive Convex Relaxation First,we solve the relaxed convex optimization problem. Then, carry out the following procedures: 1. Identify the maximum element of each user flow configuration vector (i.e., corresponding to fn) and exclude them in the following procedures. 2. Arrange in ascending order the remaining elements that are less than 1. 3. Drop the smallest element and add a zero constraint for it. 32 / 55
  • 69.
    Successive Convex Relaxation First,we solve the relaxed convex optimization problem. Then, carry out the following procedures: 1. Identify the maximum element of each user flow configuration vector (i.e., corresponding to fn) and exclude them in the following procedures. 2. Arrange in ascending order the remaining elements that are less than 1. 3. Drop the smallest element and add a zero constraint for it. 32 / 55
  • 70.
    Successive Convex Relaxation First,we solve the relaxed convex optimization problem. Then, carry out the following procedures: 1. Identify the maximum element of each user flow configuration vector (i.e., corresponding to fn) and exclude them in the following procedures. 2. Arrange in ascending order the remaining elements that are less than 1. 3. Drop the smallest element and add a zero constraint for it. 32 / 55
  • 71.
    Successive Convex Relaxation First,we solve the relaxed convex optimization problem. Then, carry out the following procedures: 1. Identify the maximum element of each user flow configuration vector (i.e., corresponding to fn) and exclude them in the following procedures. 2. Arrange in ascending order the remaining elements that are less than 1. 3. Drop the smallest element and add a zero constraint for it. 32 / 55
  • 72.
    Successive Convex Relaxation(Cont.) 4. For the rest of the elements, drop them and add zero constraints from the smallest element up to ND elements in total (including the one in step 3) as far as the element is less than a dropping threshold (θD); otherwise, stop dropping and go to the next step. 5. If there remains only one nonzero element per user flow configuration vector, stop here (a solution found); otherwise, solve a new relaxed convex optimization problem with augmented constraints and repeat the whole procedure from step 1. 33 / 55
  • 73.
    Successive Convex Relaxation(Cont.) 4. For the rest of the elements, drop them and add zero constraints from the smallest element up to ND elements in total (including the one in step 3) as far as the element is less than a dropping threshold (θD); otherwise, stop dropping and go to the next step. 5. If there remains only one nonzero element per user flow configuration vector, stop here (a solution found); otherwise, solve a new relaxed convex optimization problem with augmented constraints and repeat the whole procedure from step 1. 33 / 55
  • 74.
    Successive Convex Relaxation:An Example Consider a simple case of N=2, H=4, and ND = 1. Initial condition: f = [ 0.25 0.25 0.25 0.25 | 0.3 0.2 0.25 0.25 ] After 1st step: f = [ 0.0 0.2 0.3 0.5 | 0.7 0.0 0.2 0.1 ] After 2nd step: f = [ 0.0 0.0 0.2 0.8 | 0.9 0.0 0.1 0.0 ] Stop here. Solution found! 34 / 55
  • 75.
    Successive Convex Relaxation:An Example Consider a simple case of N=2, H=4, and ND = 1. Initial condition: f = [ 0.25 0.25 0.25 0.25 | 0.3 0.2 0.25 0.25 ] After 1st step: f = [ 0.0 0.2 0.3 0.5 | 0.7 0.0 0.2 0.1 ] After 2nd step: f = [ 0.0 0.0 0.2 0.8 | 0.9 0.0 0.1 0.0 ] Stop here. Solution found! 34 / 55
  • 76.
    Successive Convex Relaxation:An Example Consider a simple case of N=2, H=4, and ND = 1. Initial condition: f = [ 0.25 0.25 0.25 0.25 | 0.3 0.2 0.25 0.25 ] After 1st step: f = [ 0.0 0.2 0.3 0.5 | 0.7 0.0 0.2 0.1 ] After 2nd step: f = [ 0.0 0.0 0.2 0.8 | 0.9 0.0 0.1 0.0 ] Stop here. Solution found! 34 / 55
  • 77.
    Successive Convex Relaxation:An Example Consider a simple case of N=2, H=4, and ND = 1. Initial condition: f = [ 0.25 0.25 0.25 0.25 | 0.3 0.2 0.25 0.25 ] After 1st step: f = [ 0.0 0.2 0.3 0.5 | 0.7 0.0 0.2 0.1 ] After 2nd step: f = [ 0.0 0.0 0.2 0.8 | 0.9 0.0 0.1 0.0 ] Stop here. Solution found! 34 / 55
  • 78.
    Next . .. Introduction Review of Current Formulation of Appliance Energy Consumption Scheduling Atomic Scheduling of Appliance Energy Consumption Starting-Time-Based Formulation Optimal-Routing-Based Formulation Successive Convex Relaxation Examples Numerical Results Experimental Parameters Comparison of Bounds with True Optimal Values Cost Minimization PAR Minimization Conclusions 35 / 55
  • 79.
    Energy Cost andPAR Minimization Energy cost minization minimize fn∈ ˆFn,∀n∈N h∈H Ch (Lh(f)) . PAR minimization minimize Γ,fn∈ ˆFn,∀n∈N Γ subject to Γ ≥ Lh(f), ∀h ∈ H. Note that PAR minimization is formulated as a relaxed linear program by introducing a new auxiliary variable Γ. 36 / 55
  • 80.
    Energy Cost andPAR Minimization Energy cost minization minimize fn∈ ˆFn,∀n∈N h∈H Ch (Lh(f)) . PAR minimization minimize Γ,fn∈ ˆFn,∀n∈N Γ subject to Γ ≥ Lh(f), ∀h ∈ H. Note that PAR minimization is formulated as a relaxed linear program by introducing a new auxiliary variable Γ. 36 / 55
  • 81.
    Energy Cost andPAR Minimization Energy cost minization minimize fn∈ ˆFn,∀n∈N h∈H Ch (Lh(f)) . PAR minimization minimize Γ,fn∈ ˆFn,∀n∈N Γ subject to Γ ≥ Lh(f), ∀h ∈ H. Note that PAR minimization is formulated as a relaxed linear program by introducing a new auxiliary variable Γ. 36 / 55
  • 82.
    Next . .. Introduction Review of Current Formulation of Appliance Energy Consumption Scheduling Atomic Scheduling of Appliance Energy Consumption Starting-Time-Based Formulation Optimal-Routing-Based Formulation Successive Convex Relaxation Examples Numerical Results Experimental Parameters Comparison of Bounds with True Optimal Values Cost Minimization PAR Minimization Conclusions 37 / 55
  • 83.
    Next . .. Introduction Review of Current Formulation of Appliance Energy Consumption Scheduling Atomic Scheduling of Appliance Energy Consumption Starting-Time-Based Formulation Optimal-Routing-Based Formulation Successive Convex Relaxation Examples Numerical Results Experimental Parameters Comparison of Bounds with True Optimal Values Cost Minimization PAR Minimization Conclusions 38 / 55
  • 84.
    Appliance Energy Consumption Requirements Appliance Parameters α[h] β [h] γop [kWh] δ [h] Dish Washer 0 23 0.7200 2 Washing Machine 0 23 0.4967 3 (Energy Star) Washing Machine 0 23 0.6467 3 (Regular) Clothes Dryer 0 23 0.6250 4 PHEV1 222 292 3.3000 3 1 Plug-in hybrid electric vehicle. 2 Scheduling interval of 10 PM–5 AM. An appliance is randomly selected for each user. Constant operating energy levels (γop ) are assumed. 39 / 55
  • 85.
    Hourly Cost Function Weassume a simple quadratic hourly cost function as in [3], i.e., Ch (Lh) = ahL2 h [cent] where ah = 0.2 if h ∈ [0, 7], 0.3 if h ∈ [8, 23]. 40 / 55
  • 86.
    Hourly Cost Function Weassume a simple quadratic hourly cost function as in [3], i.e., Ch (Lh) = ahL2 h [cent] where ah = 0.2 if h ∈ [0, 7], 0.3 if h ∈ [8, 23]. 40 / 55
  • 87.
    Performance Measures andParameter Values Performance measures Lower bound: LB Upper bound with ND: UB (ND) Gap: G UB(D) − LB Number of iterations Parameter values for successive convex relaxation Dropping threshold (θD): 0.1 Maximum number of fractional-valued elements that can be dropped per iteration (ND): 1, 2, 5, 10 41 / 55
  • 88.
    Performance Measures andParameter Values Performance measures Lower bound: LB Upper bound with ND: UB (ND) Gap: G UB(D) − LB Number of iterations Parameter values for successive convex relaxation Dropping threshold (θD): 0.1 Maximum number of fractional-valued elements that can be dropped per iteration (ND): 1, 2, 5, 10 41 / 55
  • 89.
    Performance Measures andParameter Values Performance measures Lower bound: LB Upper bound with ND: UB (ND) Gap: G UB(D) − LB Number of iterations Parameter values for successive convex relaxation Dropping threshold (θD): 0.1 Maximum number of fractional-valued elements that can be dropped per iteration (ND): 1, 2, 5, 10 41 / 55
  • 90.
    Performance Measures andParameter Values Performance measures Lower bound: LB Upper bound with ND: UB (ND) Gap: G UB(D) − LB Number of iterations Parameter values for successive convex relaxation Dropping threshold (θD): 0.1 Maximum number of fractional-valued elements that can be dropped per iteration (ND): 1, 2, 5, 10 41 / 55
  • 91.
    Performance Measures andParameter Values Performance measures Lower bound: LB Upper bound with ND: UB (ND) Gap: G UB(D) − LB Number of iterations Parameter values for successive convex relaxation Dropping threshold (θD): 0.1 Maximum number of fractional-valued elements that can be dropped per iteration (ND): 1, 2, 5, 10 41 / 55
  • 92.
    Performance Measures andParameter Values Performance measures Lower bound: LB Upper bound with ND: UB (ND) Gap: G UB(D) − LB Number of iterations Parameter values for successive convex relaxation Dropping threshold (θD): 0.1 Maximum number of fractional-valued elements that can be dropped per iteration (ND): 1, 2, 5, 10 41 / 55
  • 93.
    Performance Measures andParameter Values Performance measures Lower bound: LB Upper bound with ND: UB (ND) Gap: G UB(D) − LB Number of iterations Parameter values for successive convex relaxation Dropping threshold (θD): 0.1 Maximum number of fractional-valued elements that can be dropped per iteration (ND): 1, 2, 5, 10 41 / 55
  • 94.
    Performance Measures andParameter Values Performance measures Lower bound: LB Upper bound with ND: UB (ND) Gap: G UB(D) − LB Number of iterations Parameter values for successive convex relaxation Dropping threshold (θD): 0.1 Maximum number of fractional-valued elements that can be dropped per iteration (ND): 1, 2, 5, 10 41 / 55
  • 95.
    Next . .. Introduction Review of Current Formulation of Appliance Energy Consumption Scheduling Atomic Scheduling of Appliance Energy Consumption Starting-Time-Based Formulation Optimal-Routing-Based Formulation Successive Convex Relaxation Examples Numerical Results Experimental Parameters Comparison of Bounds with True Optimal Values Cost Minimization PAR Minimization Conclusions 42 / 55
  • 96.
    Upper/Lower Bounds vs.True Optimal Values: Cost Minimization 2 3 4 5 6 7 8 9 10 0 5 · 10−2 0.1 0.15 0.2 0.25 0.3 0.35 N EnergyCost[USD] LB GO UB(1) UB(2) UB(5) UB(10) 43 / 55
  • 97.
    Upper/Lower Bounds vs.True Optimal Values: PAR Minimization 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10 N PARinAggregatedLoad LB GO UB(1) UB(2) UB(5) UB(10) 44 / 55
  • 98.
    Next . .. Introduction Review of Current Formulation of Appliance Energy Consumption Scheduling Atomic Scheduling of Appliance Energy Consumption Starting-Time-Based Formulation Optimal-Routing-Based Formulation Successive Convex Relaxation Examples Numerical Results Experimental Parameters Comparison of Bounds with True Optimal Values Cost Minimization PAR Minimization Conclusions 45 / 55
  • 99.
    Cost Minimization: Upper/LowerBounds 2 10 20 30 40 50 0 1 2 3 4 5 N EnergyCost[USD] LB UB(1) UB(2) UB(5) UB(10) 46 / 55
  • 100.
    Cost Minimization: Gaps 210 20 30 40 50 0 1 2 3 4 5 6 7 ×10−2 N Gap[USD] UB(1)−LB UB(2)−LB UB(5)−LB UB(10)−LB 47 / 55
  • 101.
    Cost Minimization: Numberof Iterations 2 10 20 30 40 50 0 200 400 600 800 1,000 N NumberofIterations UB(1) UB(2) UB(5) UB(10) 48 / 55
  • 102.
    Next . .. Introduction Review of Current Formulation of Appliance Energy Consumption Scheduling Atomic Scheduling of Appliance Energy Consumption Starting-Time-Based Formulation Optimal-Routing-Based Formulation Successive Convex Relaxation Examples Numerical Results Experimental Parameters Comparison of Bounds with True Optimal Values Cost Minimization PAR Minimization Conclusions 49 / 55
  • 103.
    Par Minimization: Upper/LowerBounds 2 10 20 30 40 50 0 5 10 15 20 25 N PARinAggregatedLoad LB UB(1) UB(2) UB(5) UB(10) 50 / 55
  • 104.
    Par Minimization: Gaps 210 20 30 40 50 0 1 2 3 N Gap UB(1)−LB UB(2)−LB UB(5)−LB UB(10)−LB 51 / 55
  • 105.
    Par Minimization: Numberof Iterations 2 10 20 30 40 50 0 200 400 600 800 1,000 N NumberofIterations UB(1) UB(2) UB(5) UB(10) 52 / 55
  • 106.
    Next . .. Introduction Review of Current Formulation of Appliance Energy Consumption Scheduling Atomic Scheduling of Appliance Energy Consumption Starting-Time-Based Formulation Optimal-Routing-Based Formulation Successive Convex Relaxation Examples Numerical Results Experimental Parameters Comparison of Bounds with True Optimal Values Cost Minimization PAR Minimization Conclusions 53 / 55
  • 107.
    Conclusions We have provideda new formulation of appliance energy consumption scheduling based on the optimal routing framework. It guarantees the atomicity of resulting scheduled energy consumption. We have also provided an efficient solution technique based on successive convex relaxation for a Boolean-convex problem resulting from a convex objective function. It enables us to carry out systematic analysis of the original problem with both upper and lower bounds. Possible extensions to the current work include Distributed atomic energy consumption scheduling; Advanced techniques refining the feasible region at each iterative step to reduce the total number of iterations. 54 / 55
  • 108.
    Conclusions We have provideda new formulation of appliance energy consumption scheduling based on the optimal routing framework. It guarantees the atomicity of resulting scheduled energy consumption. We have also provided an efficient solution technique based on successive convex relaxation for a Boolean-convex problem resulting from a convex objective function. It enables us to carry out systematic analysis of the original problem with both upper and lower bounds. Possible extensions to the current work include Distributed atomic energy consumption scheduling; Advanced techniques refining the feasible region at each iterative step to reduce the total number of iterations. 54 / 55
  • 109.
    Conclusions We have provideda new formulation of appliance energy consumption scheduling based on the optimal routing framework. It guarantees the atomicity of resulting scheduled energy consumption. We have also provided an efficient solution technique based on successive convex relaxation for a Boolean-convex problem resulting from a convex objective function. It enables us to carry out systematic analysis of the original problem with both upper and lower bounds. Possible extensions to the current work include Distributed atomic energy consumption scheduling; Advanced techniques refining the feasible region at each iterative step to reduce the total number of iterations. 54 / 55
  • 110.
    Conclusions We have provideda new formulation of appliance energy consumption scheduling based on the optimal routing framework. It guarantees the atomicity of resulting scheduled energy consumption. We have also provided an efficient solution technique based on successive convex relaxation for a Boolean-convex problem resulting from a convex objective function. It enables us to carry out systematic analysis of the original problem with both upper and lower bounds. Possible extensions to the current work include Distributed atomic energy consumption scheduling; Advanced techniques refining the feasible region at each iterative step to reduce the total number of iterations. 54 / 55
  • 111.
    Conclusions We have provideda new formulation of appliance energy consumption scheduling based on the optimal routing framework. It guarantees the atomicity of resulting scheduled energy consumption. We have also provided an efficient solution technique based on successive convex relaxation for a Boolean-convex problem resulting from a convex objective function. It enables us to carry out systematic analysis of the original problem with both upper and lower bounds. Possible extensions to the current work include Distributed atomic energy consumption scheduling; Advanced techniques refining the feasible region at each iterative step to reduce the total number of iterations. 54 / 55
  • 112.
    Conclusions We have provideda new formulation of appliance energy consumption scheduling based on the optimal routing framework. It guarantees the atomicity of resulting scheduled energy consumption. We have also provided an efficient solution technique based on successive convex relaxation for a Boolean-convex problem resulting from a convex objective function. It enables us to carry out systematic analysis of the original problem with both upper and lower bounds. Possible extensions to the current work include Distributed atomic energy consumption scheduling; Advanced techniques refining the feasible region at each iterative step to reduce the total number of iterations. 54 / 55
  • 113.
    Conclusions We have provideda new formulation of appliance energy consumption scheduling based on the optimal routing framework. It guarantees the atomicity of resulting scheduled energy consumption. We have also provided an efficient solution technique based on successive convex relaxation for a Boolean-convex problem resulting from a convex objective function. It enables us to carry out systematic analysis of the original problem with both upper and lower bounds. Possible extensions to the current work include Distributed atomic energy consumption scheduling; Advanced techniques refining the feasible region at each iterative step to reduce the total number of iterations. 54 / 55
  • 114.
    References I S. Boydand L. Vandenberghe, Convex optimization. Cambridge, U.K.: Cambridge Universtiy Press, 2004. J. B. Rosen, “Existence and uniqueness of equilibrium for concave N-person games,” Econometrica, vol. 33, no. 3, pp. 520–534, Jul. 1965. A.-H. Mohsenian-Rad, V. W. S. Wong, J. Jatskevich, R. Schober, and A. Leon-Garcia, “Autonomous demand-side management based on game-theoretic energy consumption scheduling for the future smart grid,” IEEE Trans. Smart Grid, vol. 1, no. 3, pp. 320–331, Dec. 2010. 55 / 55