You are on page 1of 5

Load Profiling by a Genetic Algorithm

Ovidiu Ivanov, Mihai Gavrila

Abstract. In the electricity market context, the customers


without digital meters must be categorized by the suppliers into consumer categories, using load profiling techniques. The paper proposes a new load profiling approach that determines typical load profiles (TLPs) using a genetic algorithm approach. TLPs are determined by means of a multiobjective optimization process based on two criteria: similarity between metered profiles and the TLPs and dissimilarities between different TLPs. To prove the efficiency of the new approach, TLPs are determined for 6 consumer categories with different load behaviours.

energy regulations impose that, in order to participate in the electricity market, their consumption must be monitored with electronic equipments able to meter hourly load data [1]. The Romanian electricity market is programmed to become fully open in the summer of 2007, giving the small consumers the possibility to choose their supplier. Because for these consumers, installing a digital meter is a prohibitive option, they must be categorized by their suppliers using a method which can determine a load profile by keeping the traditional monthly meter readings. Such a technique is the load profiling. Generally, a load profile describes how the consumer uses electricity. Thus, the load profiles are determined usually for consumer categories, typical weekdays, seasons, or for typical values of other parameters such as the load factor. The load profiling techniques emerged and have been used widely on many electricity markets. For instance, in the summer of 1999, when the British retail electricity was fully opened, more than 26 millions of small consumers with a peak load smaller than 100 kW entered the market [2]. This was possible particularly because load profiling techniques were used. Traditional load profiling methods classify the consumers and generate TLPs based on the metered load profiles and some other parameters. This paper presents a load profiling technique that identifies the TLPs taking into account the following aspects: The TLP associated to a consumer class must match as close as possible to the metered profiles used to generate it The generated TLPs must be as different as possible to each other

Keywords
Electricity markets, load profiling, genetic algorithms, typical load profiles.

1. Introduction
In todays electricity industry, the old vertically structured monopolies are replaced with dynamic electricity markets where the prices are determined by the economic principle of supply and demand. To participate in the electricity market, the players must provide hourly load data for the entire trading period. For small consumers, the digital meters costs are prohibitive and they must be categorized by their suppliers through load profiling techniques, which, for a given consumer, determine its typical load profiles (TLP) based on the shape of the metered load profiles. In Romania, the energy market liberalization started in the year 2000, when the wholesale electricity market operator, OPCOM, was established. Today, the market is 83.5% open, which means that any commercial and industry consumer are eligible consumers. The current
Ovidiu Ivanov is with the Technical University of Iasi, Blvd. D. Mangeron, no. 51-53, Iasi, 700050, Romania (phone +40-232278683; fax +40-232-237627; e-mail: ovidiuivanov@ee.tuiasi.ro). Mihai Gavrila is with the Technical University of Iasi, Romania (e-mail: mgavril@ee.tuiasi.ro).

This method takes advantage of the known searching and classification capabilities offered by the evolutionary algorithms. The proposed method uses a

genetic algorithm. Results are presented as TLPs for different consumer categories. The problems addressability is immediate, as beginning from summer 2007 the Romanian electricity market will be 100% opened and the market regulations impose the use of TLPs for small consumers. The authors previous works in this field include several papers presented on national and international conferences proceedings.

Dt

1 TLPt LPi N t iK
t

1 24 N ti iK

(TLP
t

24

(3)

t ,h

LPi , h )

h 1

The F2 function is defined as the overall mean square difference between the generated TLPs:

2. The Proposed Method


The basic problem can be stated as follows: For a group of consumers represented by N hourly metered load profiles, the goal is to determine a set of T typical load profiles (TLPs) that must comply with the following requirements: The TLP associated to a consumer class must match as close as possible to the metered profiles used to generate it The generated TLPs must be as different as possible to each other

F2

1 T (T 1) t

TLP
1 1 t 2 1 t 2 t1

t1

TLPt 2
(4)

1 (TLPt1 ,h TLPt 2 , h ) 24 T (T 1) h 1

24

Prior to be processed, the load profiles metered in kW are transformed in values expressed in percents from the reference periods total consume, which has been considered a month. Thus, TLPs are expressed in percents from the energy consumption of a representative consumer from that category, over the reference period. In the following, an i load profile is denoted LPi,h (i=1..N load profiles, h=1..24 hours). Based on the LPi,h curves, a number of T TLPs are generated, wich are denoted TLPt,h (where t is the current typical profile number). The set of TLPs is generated as a result of a multiobjective goal function minimizing process:

The F1 and F2 functions are complementary, in the sense that an optimal solution is a compromise between their values. Thats why in the profiling process, the inverse value of F2 has been used. The a, b weights can be chosen based on the importance of the corresponding goal function. In this paper, the two values have been set equal.

3. Genetic algorithms
Genetic algorithms (GA) are part of the so-called evolutionary algorithms, which use the natural selection mechanisms to find the optimal solution for a given problem. Their strength lies in a probabilistic and parallel search through the solutions space that of fers a better chance of finding the optimal result, compared to the traditional methods [3],[4]. GAs encode possible solutions of the problem into chains of numbers or symbols (genes) called chromosomes which represent possible solutions for the studied problem and form the search population. The degree of approach of each solution or chromosome toward the global best is measured by a fitness function. Usually, the initial population is generated randomly. Each chromosome must comply with the problem restrictions, if there are any. The optimal solution is researched by an iterative process using the genetic operators called selection, crossover and mutation. In the selection stage, the less adapted individuals, those with a low fitness function, are removed from the population and replaced by duplicating better adapted chromosomes. Several selection methods exist; in this paper the roulette rule was used. At the end of the selection stage, the number of chromosome must be the same as in the initial population. The crossover stage pairs the selected individuals, which exchange amongst them genetic information by switching chromosome parts. The child chromosomes

F a F1 b F2

(1)

where a and b are weights, F1 is an objective function associated to the similarity between the metered and typical load profiles and F2 is an objective function associated to the dissimilarity between distinct typical load profiles. The F1 function is defined as the overall mean square difference between the TLPs and the metered profiles associated with each TLP and it can be written as

F1
where

1 T

D
t 1

(2)

inherit partially the features of their parents, and represent new solutions which could be better ones. The crossover takes place with a probability called crossover rate. The mutation operator is used do preserve genetic diversity, i.e. to avoid local minima by randomly altering genes, with a probability called mutation rate. Each new chromosome generated through crossover or mutation must comply with the problem restrictions. If not, it is discarded and the parent chromosome is perpetuated instead. At this point, the resulting population replaces the initial population and the adaptive process continues in the same manner until a fixed number of iterations called generations is reached, or a stopping criterion is fulfilled. The evolutionary process often uses the elitism procedure, in which the best current chromosome is perpetuated from one generation to another in order to preserve the best found solution [5],[6]. The basic genetic algorithm can be described as follows: Step 1. Generate a random population of chromosomes and define the algorithm parameters (number of generations, crossover and mutation rate) repeat Step 2. Compute the fitness function for each chromosome Step 3. Elitism. Keep the best adapted chromosome from the current population. Step 4. Selection. Remove from the current population the poor adapted chromosomes and replace them with copies of better adapted ones. Step 5. Crossover. Create child chromosomes from the selected population, by exchanging parts between parents chromosomes. Step 6. Mutation. Randomly change gene values. Step 7. Replace the initial population with the current population until a fixed number of generations is reached or the stopping criterion is achieved Step 8. The algorithm found an optimal solution, coded in the best adapted chromosome.

The GA parameters were set as follows: number of generations: 50 crossover rate: 0.9 mutation rate: 0.1 A chromosome has a number of genes equal to the number of measured profiles for all category of consumers and each gene value represents the TLP number to which the respective metered curve is to be assigned. For instance, if the consumers metered profiles database has 60 profiles that can be assigned to maximum 4 TLPs, the chromosome will have 60 genes which can have the value 1, 2, 3 or 4. The TLPs are calculated as the average of the metered profiles associated by the algorithm to each TLP. In Fig. 1-5 are illustrated the daily metered profiles for the five studied consumer categories. For clarity reasons, only the profiles for the first week of the month are presented. In Fig. 6-10 are presented the TLPs generated by the algorithm for each consumer category.
Energy [% of monthly consumption]
0.25

0.2

0.15

0.1

0.05

0 1 Sat Wed 3 5 Sun Thu 7 9 11 13 15 17 19 21 23 Mon Fri Tue

Hour

Fig. 1. Metered load profiles for the Residential category


0.25 0.2 0.15 0.1 0.05 0
1
Sat Wed

4. Experimental results
The algorithm presented above was used for 5 consumer categories, in accordance with measurements taken in 6 MV/LV substations: Residential (substation PT270 and PT413), University campus (PT153), Supermarket (PT301), Hotel (PT206) and Hospital (PT016).

Energy [% of monthly consumption]

3
Sun Thu

9 11 13 15 17 19 21 23
Mon Fri Tue

Hour

Fig. 2. Metered load profiles for the University campus category

Energy [% of monthly consumption]

0.3 0.25 0.2 0.15 0.1 0.05 0 1 Sat Wed 3 5 7 Sun Thu 9 11 13 15 17 19 21 23 Mon Tue Hour Fri

0.25

Energy [% of monthly consumption]

0.2 0.15 0.1 0.05 0 1 Mo-Fri 3 5 7 9 11 13 15 17 19 21 23

Sat-Sun

Hour

Fig. 3. Metered load profiles for the Supermarket category


Energy consumption [% of monthly consume]
0.25 0.2 0.15 0.1 0.05 0 1 Sat Wed 3 5 Sun Thu 7 9 11 13 15 17 19 21 23 Mon Fri Tue

Fig. 6. Typical load profiles for the Residential category


0.25

Energy [% of monthly consumption]

0.2

0.15 0.1

0.05 0 1 3 5 7 9 11 13 15 17 19 21 23

Hour
Mon-Sun

Hour

Fig. 4. Metered load profiles for the Hotel category

Fig. 7. Typical load profiles for the University campus category

0.35

0.3

Energy [% of monthly consumption]

Energy [% of monthly consumption]

0.3 0.25 0.2 0.15 0.1 0.05 0 1 3 5 7 9 11 13 15 17 19 21 23

0.25 0.2 0.15 0.1 0.05 0 1 Mon-Sat 3 5 7 Sun 9 11 13 15 17 19 21 23

Sat Wed

Sun Thu

Mon Fri

Thu

Hour

Hour

Fig. 5. Metered load profiles for the Hospital category

Fig. 8. Typical load profiles for the Supermarket category

5. Conclusion
0.18

Energy [% of monthly consumption]

0.16 0.14 0.12 0.1 0.08 0.06 0.04 0.02 0 1 3 5 7 9 11 13 15 17 19 21 23

The proposed method takes advantage of GAs algorithms capability to solve multiobjective optimization problems and can be used with unlimited number of consumer categories. The results, generated for several consumer categories with different load behaviours prove the algorithms validity. They can be used to implement the load profiling techniques for the small consumers that want to participate in the electricity market.

Mo-Sun

Hour

Fig. 9. Typical load profiles for the Hotel category

References
[1] The Romanian Electricity Market Operator Website www.opcom.ro S.V. Allera S.V. and A.G. Horsburgh, Load profiling for Energy trading and settlements in the UK Electricity Markets, DistribuTECH Europe DA/DSM Conference, London, 1998 T. Back, Evolutionary Algorithms in Theory and Practice, Oxford University Press, New York, USA, 1996 A. Goldberg, A.Srinivas, Multiobjective Optimization Using Nondominated Sorting in Genetic Algorithms, Evolutionary Computation, Vol. 2, No. 3, pp 221-229, 2004 Z. Mihcalewicz, Genetic Algorithms + Data Structures = Evolution Programs, Artificial Intelligence, Springer, Berlin, 1992 J.D. Schaffer (editor), Proceedings of the 3rd International Conference on Genetic Algorithms and their Applications, Morgan Kaufmann Publishers, San Mateo, CA, 1989

0.3

Energy [% of monthly consumption]

0.25 0.2 0.15 0.1 0.05 0 1 3

[2]

[3]

Mo-Fri Sun

5 7 Sat

9 11 13 15 17 19 21 23

[4]

Hour

Fig. 10. Typical load profiles for the Hospital category

[5]

The results are consistent with the metered profiles shape for each consumer category. For example, looking at the Hospital or Supermarket categories, the profiles associated to working days and week-end are associated to different TLPs. Where the metered profiles do not differ significantly (the Hotel category, for instance) the algorithm generates only a single TLP for the entire week. The generated TLPs can be considered as valid for their respective consumers category.

[6]

You might also like