Multiobjective household energy planning using evolutionary algorithms
Authors: Giovanni Colacurcio, Sergio Nesmachnow, Jamal Toutouh, Francisco Luna, Diego Rossit
Organizations: Universidad de la Rep´ublica, Uruguay · Massachusetts Institute of Technology, USA · Universidad de M´alaga, Spain · Universidad Nacional del Sur, Argentina
This article presents the advances in the design and implementation of a recommendation system for planning the use of household appliances, focused on improving energy efficiency from the point of view of both energy companies and end-users. The system proposes using historical information and data from sensors to define instances of the planning problem considering user preferences, which in turn are proposed to be solved using a multiobjective evolutionary approach, in order to minimize energy consumption and maximize quality of service offered to users. Promising results are reported on realistic instances of the problem, compared with situations where no intelligent energy planning are used (i.e., Bussiness as Usual model) and also with a greedy algorithm developed in the framework of the reference project. The proposed evolutionary approach was able to improve up to 29.0% in energy utilization and up to 65.3% in user preferences over the reference methods.
Figures & tables
Figure 1: An example of the proposed solution encoding.
Figure 2: An example of the proposed recombination operator.
Figure 3: An example of the proposed mutation operator.
#
name
users
appliances in the instance
consumption pattern
1
small.1 (s1)
2
(2,3)
weekday
2
small.2 (s2)
2
(2,3)
weekend
3
medium.1 (m1)
4
(4,4,3,2)
weekday
4
medium.2 (m2)
4
(4,4,3,2)
weekend
5
large.1 (l1)
6
(5,5,4,4,3,2)
weekday
6
large.2 (l2)
6
(5,5,4,4,3,2)
weekend
Table 1: Proposed problem instances
α
β
instance #1
instance #2
instance #3
median
IQR
median
IQR
median
IQR
0.3
0.3
15.81
1.09
179.83
0.55
132.35
6.68
0.3
0.5
11.02
1.54
167.94
2.03
110.10
4.54
0.3
1.0
4.27
6.39
128.55
28.66
54.82
13.56
0.5
0.3
34.06
0.81
311.32
28.66
253.95
5.11
0.5
0.5
26.75
1.06
298.80
23.71
224.47
12.21
Table 2: Best fitness values computed using different values of (α,β)
Figure 4: Trade-off analysis of solutions computed using different values of ( α , β )
configuration
fitness
configuration
fitness
( G , pC , pM )
median
IQR
( G , pC , pM )
median
IQR
(2500, 0.1, 0.01)
7.88
10.45
(10000, 0.5, 0.1)
40.27
7.61
(5000, 0.1, 0.01)
38.74
11.71
(2500, 0.25, 0.1)
40.73
7.01
(10000, 0.1, 0.01)
40.04
6.98
(5000, 0.25, 0.1)
41.64
6.76
(2500, 0.5, 0.01)
9.26
11.17
(10000, 0.25, 0.1)
40.72
8.67
(5000, 0.5, 0.01)
39.22
9.37
(2500, 0.1, 0.05)
39.60
6.97
Table 3: Parameter setting results for the proposed EA
instance
f (EA)
greedy
BaU
f
Δf
Δcost
Δpref
f
Δf
Δcost
Δpref
weekday
small.1
14.1
11.0
28.3%
22.6%
56.2%
-7.6
284.6%
81.1%
16.0%
medium.1
340.0
239.4
42.0%
27.2%
65.3%
69.5
388.9%
77.9%
19.4%
large.1
407.8
347.9
17.2%
20.8%
47.3%
-187.1
317.0%
70.6%
22.0%
weekend
Table 4: Experimental results: fitness values and improvements of the proposed EA over the baseline greedy algorithm and the BaU strategy.