Anda di halaman 1dari 13

Energy Conversion and Management 52 (2011) 199211

Contents lists available at ScienceDirect

Energy Conversion and Management


journal homepage: www.elsevier.com/locate/enconman

A novel modeling approach for hourly forecasting of long-term electric energy demand
mmhan Basaran Filik *, mer Nezih Gerek, Mehmet Kurban
Department of Electrical and Electronics Engineering, Anadolu University, Eskisehir, Turkey

a r t i c l e

i n f o

a b s t r a c t
In this study, a novel mathematical method is proposed for modeling and forecasting electric energy demand. The method is capable of making long-term forecasts. However, unlike other long-term forecasting models, the proposed method produces hourly results with improved accuracy. The model is constructed and veried using 26-year-long real-life load data (4 years with hourly resolution) obtained from the Turkish Electric Power Company. The overall method consists of a nested combination of three subsections for modeling. The rst section is the coarse level for modeling variations of yearly average loads. The second section renes this structure by modeling weekly residual load variations within a year. The nal section reaches to the hourly resolution by modeling variations within a week, using a novel 2-D mathematical representation at this resolution. The adoptions of such nested forecasting methodology together with the proposed 2-D representation for hourly load constitute the novelties of this work. The major advantage of the proposed approach is that it enables the possibility of making short-, medium-, and long-term hourly load forecasting within a single framework. Several mathematical functions are applied as models at each level of the nested system for achieving the minimal forecasting error. Proposed model functions with their corresponding forecasting accuracies are presented in terms of root mean square error (RMSE) and mean absolute percentage error (MAPE). 2010 Elsevier Ltd. All rights reserved.

Article history: Received 28 October 2009 Accepted 22 June 2010 Available online 17 July 2010 Keywords: Forecasting Mathematical modeling Surface tting Energy demand

1. Introduction Forecasting of electric energy demand is essential for power planning and operation. The primal problem of the planning is the forecasting of energy demand values in the future. This is very important for companies in producing, delivering and reselling electric energy. Correct load forecasting improves the reliability of power system. Economic dispatch, fuel scheduling, and hydrothermal unit commitment can be performed efciently with a precise forecast. Traditionally, load forecasting can be considered three categories [1]:  Short-term forecasts are usually applied to intervals ranging from one hour to 1 week,  Medium-term forecasts are usually for a week to a year, and  Long-term forecasts are for ranges that are longer than a year. Owing to the importance of load forecasting, various models have been proposed for the short-term load forecasting in the last decades, such as regression-based methods [24], Box Jenkins model [5], time-series approaches [69], Kalman lters [10], expert
* Corresponding author. Tel.: +90 5053869556; fax: +90 2223239501. E-mail address: ubasaran@anadolu.edu.tr (.Bas. Filik). 0196-8904/$ - see front matter 2010 Elsevier Ltd. All rights reserved. doi:10.1016/j.enconman.2010.06.059

system techniques [11,12], neural network models [1321], fuzzy logic [22,23], and fuzzy-neural network structures [24,25]. Recently, applications of hybrid ANNs model with statistical methods or other intelligent approaches have received attentions. Examples of such systems are hybrids with Bayesian inference [26,27], self-organizing map [28], wavelet transform [29,30], and particle swarm optimization [31]. Short-term load forecasting has attracted the attention of most researchers because of its inuence in the daily operation of generation and distribution systems. This forecasting process tries to estimate the load or energy that will be demanded in the next hours or days. The other time horizons are also important for power system planning. The forecasting of energy that will be demanded by the consumers in coming weeks or months is very useful for the maintenance planning of grids and for market researching for electricity producers and resellers. This kind of prediction is known as medium-term forecasting. There are several methods of medium-term load forecasting such as time-series approaches [32,33], neural network models [34,35], and Fourier series approach [36]. Long-term forecasting pursues the prediction of the annual load peak or global energy that consumers will demand up to 20 years ahead, in order to schedule expansion planning strategies for production and distribution systems [36]. The major methods for long-term load forecasting are time-series approach [37],

200

.Bas. Filik et al. / Energy Conversion and Management 52 (2011) 199211

intelligent methods [38], neuro-fuzzy approach [39], dynamic simulation theory [40], hierarchical neural model [41], and support vector machines [42]. The objective of this study is to develop a novel method for precise load forecasting within the horizons of short-, medium-, and long-terms, all in hourly accuracy. The proposed work is unique in the sense that it attempts to make the long-term forecasting with an hourly accuracy. Possession of such a resolution would provide greater accuracy in energy strategy planning which enables almost instantaneous dispatching of electric energy. Normally, long-term forecasting is only performed on a yearly average basis [3742], whereas hourly accuracy is used for short term prediction [231]. This work provides hourly accuracy despite its availability in dramatically long prediction horizons, ranging from short to long terms. Using the method proposed here, it is now possible to forecast the load of a particular hour of a given year/month/day in the (reasonably near, but

still in long-term range) future. Similar to the previous works in the forecasting eld, the proposed methodology utilizes only past data of the demand time series to forecast future values. The remaining parts of the paper are organized as follows. In Section 2, the quasi-periodic load characteristics, its mathematical visualization, and advantages of the 2-D representation of the hourly load are briey discussed. In Section 3, the idea that enables hourly forecasting in long-term ranges is introduced: the threelevel nested system. The levels that rene the resolution in yearly, weekly, and hourly variations are explained by model determination for electric energy demand for each level of the nested system. Applications of the model and forecasting accuracy results are presented and the results of utilization of different mathematical models for each level are compared in the sense of RMSE and MAPE in Section 4. Finally, in Section 5, the main contributions of this paper are highlighted and possible research directions are discussed.

2.9

x 10

2002 3 2.9 2.8 2.7 2.6 2.5 2.4 2.3

x 10

2003

Total Energy, MW

2.8 2.7 2.6 2.5 2.4 2.3 2.2 10 20 30 40 50

Total Energy, MW

10

20

30

40

50

Weeks
x 10
6

Weeks
x 10
6

2004 3.4 3.2 3 2.8

2005

Total Energy, MW

3.2 3 2.8 2.6 10 20 30 40 50

Total Energy, MW

10

20

30

40

50

Weeks

Weeks

Fig. 1. Total energy values corresponding to each week of years from 2002 to 2005.

2.5 x 10

MW
1.5 1 0

0.5

1.5

2.5

Hours
Fig. 2. 1-D representation of hourly consecutively load data of years from 2002 to 2005.

3.5 4 x 10

.Bas. Filik et al. / Energy Conversion and Management 52 (2011) 199211

201

2. Load characteristics and its mathematical representation In this study, Turkeys hourly actual (real) loads for years from 2002 to 2005 and annual energy consumption values for years from 1982 to 2007 are used. Efcient utilization of these data requires a thorough analysis to understand the underlying dynamics. The load prole is a dynamic process that exhibits an hourly oscillatory behavior with a quasi-period of 1 day. Another observation is that, there are also two other underlying oscillations with periods of 1 week, and 1 year, respectively. The periodicities of the above processes are imperfect due to irregular behavior of electric equipments and several exceptions such as holidays, failures on power plants, weather condition or other random disturbances changing effect the load values. Nevertheless, deterministic (non-random) parts, i.e., the very general behavior of these oscillations can be modeled using waveform samples or mathematical models. Apart from the mentioned oscillatory behaviors, there is also a pronounced increase in the average load value every year, which could also be modeled using increasing functions in polynomial or exponential forms. In this section, we will start analyzing the periodicities within the range of 1-week. The incorporation of the weekly model into longer periods will be explained in Section 3. Weekly total energy values corresponding to weeks of years from 2002 to 2005 are shown in Fig. 1. In this gure, outlier effects

such as public holidays are eliminated and smoothed using neighbor-averaging interpolation. Seasonal variations (such as the increased demand in extreme cold months of winter time and extreme hot months of summer) can be easily observed from these gures. Spring seasons, especially May (between 17 and 20 weeks) and June (between 21 and 24 weeks) have lower demand. Another observation is that, the total energy demand value is increasing by year. For example, the peak value of the total energy is 2,885,127 MW in 2002; but, this value reaches 3,387,354 MW in 2005. In this section, a 2-D surface rendition approach is proposed for the compact visualization of all these changes and increments. Despite the ease and availability of the mesh-plot forms, the 2-D representation and rendition for the load demand values was never utilized in the literature. Due to its compact visualization property, the 2-D representation is used to exhibit load variations. The 4 years (from 2002 to 2005) hourly load data are drawn in classical 1-D form in Fig. 2. In contrast, the same data is drawn in 2-D dimension in Fig. 3. Due to the correlative behavior at extreme lags, mathematical modeling of the plot in Fig. 2 is overly complicated. On the other hand, if the hours within a day are separated to a single axis to form a 2-D matrix representation, the sizes of the 2-D gure become relatively smaller, enabling the possibility of achieving less complex models. Besides, the 2-D representation visually provides a better insight and information about how the electric load varies (Fig. 3).

Fig. 3. 2-D representation of hourly consecutively load data of years from 2002 to 2005.

(a)
x 10
4

(b)
x 10
4

2 1.8

2 1.8

MW

1.6 1.4

1.2 30 1 20 15 20 15 10 5 5 25 1 20 20 15 30

MW

1.6 1.4 1.2

Days

Hours

10

10 5

Days

Hours 10

Fig. 4. (a) Hourly load data in 2-D time plot for January 2002. (b) Hourly load data in 2-D time plot for January 2002 (the day assumed to start 5:00 AM).

202

.Bas. Filik et al. / Energy Conversion and Management 52 (2011) 199211

A detailed (zoomed) 2-D plot of 1 month period (January 2002) is shown in Fig. 4. When the whole weeks are examined in point of the energy consumption, it can be seen that weekend days exhibit less demand than the remaining 5 week days. It must be noted that for a lower complexity model, it is better to take the starting hour of the day as 5:00 AM, which typically corresponds to the minimum demand hour. Comparing Fig. 4a and b, one can see that the weekly surface blob in Fig. 4b is more useful in terms of pro-

viding a simpler mathematical model. This strategy was already applied in presenting the mesh plot of Fig. 3.

3. Modeling of electric energy demand Once the quasi-oscillatory behavior of the energy demand is illustrated, the next step is to determine how each periodicity

(a)
MW

4 3.5 3 2.5 2 1.5 1 0.5 0

x 10

200

400

600

800

1000

1200

Weeks

(b)

1.1 1.05 1 0.95 0.9 0.85

200

400

600

800

1000

1200

Weeks

(c)

7.5 7 6.5 6 5.5 5 4.5

x 10

-3

x 10

-3

Energy

6 5 4
20 15

Hours 10 5
20 40 60 80 100 120 140 160
1

Hours

Days

(i) =
4

(ii)

(d)

2.5 x 10

MW
1.5 1 0 0.5 1 1.5 2 2.5 3 3.5

Hours

x 10

Fig. 5. The combination of three nested sub-parts of the proposed method: (a) yearly (coarse) load variations on an axis of week, (b) weekly residual load variations on an axis of week, (c) (i) normalized hourly variations visualized in 1-D within a week, or (ii) normalized hourly variations visualized in 2-D within in a week, (d) overall load data as a summation of (a), (b), and (c) (shown for 4 years of data).

.Bas. Filik et al. / Energy Conversion and Management 52 (2011) 199211

203

(together with the overall increase) can be modeled. The proposed method consists of a nested combination of three sub-parts for modeling as explained in detail in this section. The rst part is the coarse level for modeling the increase in yearly average loads. The second part renes this structure by modeling weekly residual load variations within a year. The third part renes further hourly variations and its modeling within a week. Finally, the overall model is constructed by the combination of these parts. The nested combination of three sub-parts is illustrated in Fig. 5. 3.1. Level 1: yearly coarse modeling The rst part of the model consists of the increase in yearly average loads. In order to observe dependence on yearly- and within a year-changes and to obtain a mathematical model, actual hourly load values for years from 2002 to 2005, and total energy consumption values for the years 19822007 are used. This was due to the fact that the 20022005 data were hourly, whereas the data since 1982 were only yearly. As a result, yearly load variation from 1982 to 2007 is shown in Fig. 6. The yearly data is, then, interpolated to weekly resolution. The interpolation was carried out using a template obtained by actual weekly data of 2002 to 2005. This variation template was scaled to the yearly averages for years from 1982 to 2007, and displayed over a weekly axis in Fig. 7. In essence, the plot in Fig. 6 can be considered as a lower resolution version of Fig. 7, or, conversely,

Fig. 7 can be considered as a ner (weekly) resolution version of Fig. 6. The increase in electric energy demand in years is clear in both gures. Total energy demand for the rst week of 1982 is 523,120 MW h whereas it is 4,140,000 MW h for the last week of 2007. The task is, therefore, to nd an accurate week template to construct Fig. 7. If a correct model is found for Fig. 7, desired load demand values for any week within any year can be achieved, but the explanations of that modeling method is left to Section 3.2. When the total energy graph for the years from 1982 to 2007 is examined, both the increase with respect to years and changes within a year can be seen. The waveform in Fig. 7 is fairly wideband, so it is still very difcult to model this graph directly with a reasonable order of model complexity. Correct modeling can be done by splitting the graph into two parts: yearly average loads f(w) (i.e., the function in Fig. 6), and weekly residual load variations g(w) within a year (i.e., the functional difference between Figs. 7 and 6), both of which are functions of week, notated by w. The averages are obtained by directly taking the means of hourly data within a time window of 1 week. The variations are modeled using m curve tting tools to obtain modeled functions fj w and g m w, i where m indicates that these are the model functions of actual load, and i and j indicate indexes for various types of analytical models. The function plotted in Fig. 7 is notated as H(w) in the study, and is dened as follows:

Hw f w gw;

w 1; . . . ; 1300

2 1.8 1.6 1.4

x 10

MWh

1.2 1 0.8 0.6 0.4 0.2 5 10 15 20 25

Years
Fig. 6. Yearly load changes from 1982 to 2007.

x 10
4.5 4 3.5 3

MW

2.5 2 1.5 1 0.5 0 200 400 600 800 1000 1200

Weeks
Fig. 7. Total energy model from 1982 to 2007 (20022005 data are from original hourly data).

204

.Bas. Filik et al. / Energy Conversion and Management 52 (2011) 199211

Consequently, a modeled version Hm w can be dened as i;j follows:


m Hm w fj w g m w; i i;j

i 1; . . . ; 4 & j 1; 2

It is possible to model the f(w) and g(w) in different ways. In order to dene different models Hm w notation is used where the i;j index in parenthesis corresponds to the model number. Modeling of f(w): The increasing part of H(w), which is f(1)(w), m can be modeled with a second order polynomial function, f2 w. The actual and model versions of the increasing part are plotted on the same graph in Fig. 8. m The 2nd order polynomial function of yearly coarse model f2 is:
m f2 w 1:591w2 2:397 102 w 5:027 105

g m w and g m ; respectively. Due to their relatively oscillatory a b behavior, polynomial models were not adopted here. Instead, these functions are modeled by sums of sinusoidal functions, say g m w, and Fourier series coefcients, g m asinus aFourier w. The rst residual data type, g m w, is found by dividing load demand to yearly avera age data (in Section 3.1). The other one, g m w, is found by dividing b load demand to a tted function of the yearly coarse model. The considered residual weekly variations graphs of ga(w) and gb are presented in Fig. 9. The graphs of model approximations of two waveforms in Fig. 9a using eight sinusoidal harmonics (corresponding to 16 Fourier series (F.S.) coefcients) can be seen in Fig. 10. The function of g m is obtained as follows: asinus

m where the numerical values of coefcients of f2 w are always within 95% accuracy.

g m w g m w 1 asinus 1:648 sin1:1942 102 w 0:3196 0:729 sin3:252 103 w 2:624 3:651 102 sin8:978 103 w 2:176 5:943 102 sin0:241w 0:887 9:16 103 sin1:374 102 w 1:72 4:054 102 sin0:121w 1:343 3:392 103 sin2:162 102 w 0:8948 2:942 102 sin0:362w 2:662 4

3.2. Level 2: weekly residual modeling The second part governs the modeling of weekly residual load variations within a year. Modeling of g(w): Prior to modeling, two approaches are adopted to extract and recall the weekly residual. These two residual types are notated by ga(w) and gb(w). The models corresponding to these two functions, ga(w) and gb, are notated by

4 3.5 3 2.5

x 10

MW

2 1.5 1 0.5 0 200 400 600 800


f(1) (w), increasing data
m f(2) (w), fitted model

1000

1200

Weeks
Fig. 8. The graphs of yearly coarse model and its tted (2nd order polynomial) model.

(a)

1.1 1.05 1 0.95 0.9 0.85


ga(w)

200

400

600

800

1000

1200

Weeks

(b)

1.1 1 0.9 0.8


gb(w)

200

400

600

800

1000

1200

Weeks
Fig. 9. Normalized residual weekly variations: (a) ga(w), (b) gb(w).

.Bas. Filik et al. / Energy Conversion and Management 52 (2011) 199211

205

The coefcients of g m w have 95% condence bounds. asinus Next, the function with Fourier series (F.S.) model is considered. The difference between direct F.S. and plain sinusoidal models is that F.S. coefcients consider sine and cosine terms without phase factors. The graphs of model approximation of data in Fig. 9a using eight F.S. coefcients can be seen in Fig. 11. The function of the g m aFourier w constructed by the sum of eight Fourier terms is given as follows:

g m w g m w 2 aFourier 0:948 4 102 cos0:1208w 5:007 103 sin0:1208w 4:441 102 cos2 0:1208w 3:973 102 sin2 0:1208w 1:142 102 cos3 0:1208w 2:70 102 sin3 0:1208w 9:195 103 cos4 0:1208w 6:313 103 sin4 0:1208w 5

1.15 1.1 1.05 1 0.95 0.9 0.85 0.8 200 400 600 800
ga(w), normalized data gasinus (w), sum of sinus functions model
m

1000

1200

Weeks
Fig. 10. Model approximation of data in Fig. 9a using eight sinusoidal harmonics (corresponding to 16 F.S. coefcients).

1.1

1.05

0.95

0.9
ga(w), normalized data gaFourierm(w), Fourier series model

0.85

200

400

600

800

1000

1200

Weeks
Fig. 11. Model approximation of data in Fig. 9a using eight F.S. coefcients.

1.15 1.1 1.05 1 0.95 0.9 0.85 0.8 200 400 600 800
gb(w), normalized data gbsinus (w), sum of sinus functions model
m

1000

1200

Weeks
Fig. 12. Model approximation of data in Fig. 9b using eight sinusoidal harmonics (corresponding to 16 F.S. coefcients).

206

.Bas. Filik et al. / Energy Conversion and Management 52 (2011) 199211

Again, the coefcients of g m aFourier w have 95% condence bounds. Similarly, the function of g m is modeled by using sum of sinus funcb tions g m w and F.S. coefcients to construct g m bsinus bFourier w.

The graphs of model approximation results of data in Fig. 9b using eight sinusoidal harmonics (corresponding to 16 F.S. coefcients) can be seen in Fig. 12. The function of g m w is found as follows: bsinus

1.15 1.1 1.05 1 0.95 0.9 0.85 0.8


gb(w), normalized data gbFourier (w), Fourier series model
m

200

400

600

800

1000

1200

Weeks
Fig. 13. Model approximation of data in Fig. 9b using eight F.S. coefcients.

(a)
Energy

x 10
7 6

-3

(b)

7.5 7 6.5

x 10

-3

=
4 20 15

6 5.5 5

Hours 10
5 1 2 3 4

4.5

Days

20

40

60

80

100

120

140

160

Hours

Fig. 14. (a) Normalized hourly variations visualized in 2-D in a week. (b) Normalized hourly variations visualized in 1-D in a week.

x 10
7.5 7

-3

Model of week template, T(2)(h,d) Week template, T(1)(h,d)

Energy

6.5 6 5.5 5 4.5 20

15

Hours

10 5 1 2 3

Days

Fig. 15. The week template and its modeling surface.

.Bas. Filik et al. / Energy Conversion and Management 52 (2011) 199211

207

g m w g m w 3 bsinus 1:873 sin2:27 10 w 6:79 10 1:162 sin4:06 103 w 1:79 5:736 101 sin7:24 10 w 2:78 3:36 10
3 1 3 2

to know only seven coefcients and their coordinates in order to generate the week template data. As a result, the function of the week template model (notated by T(2)(h, d) is found as:
h i h i T 2 h; d 1:17 102 cos p2h1 cos p2d1 2:06 104 cos 2p2h1 cos p2d1 48 14 48 14 h i h i 9:86 104 cos 3p2h1 cos p2d1 5:71 104 cos 5p2h1 cos p2d1 48 14 48 14 h i h i 1:38 104 cos 9p2h1 cos p2d1 5:40 104 cos p2h1 cos 3p2d1 48 14 48 14 h i 1:85 104 cos 3p2h1 cos 3p2d1 48 14

sin8:31

103 w 5:21 5:91 sin2:41 101 w 0:73 3:89 102 sin0:121w 1:05 6:67 103 sin2:33 102 w 0:89 2:91 102 sin0:362w 2:57 6

8
The week template structure and its modeling surface are rendered in an overlaid plot in Fig. 15. 3.4. Determination of the total model output The above candidates of three functions for modeling the total yearly variations, marginal weekly variations, and nally, hourly residual variations must be combined to construct the nal model. It must be noted that the polynomial function for modeling yearly variations provide all the necessary scaling factor for any given year for forecasting. Then, the scaling factor (or, overall value of the year) multiplies the weekly variation model. The result is a load demand function of a particular year with weekly resolution. In the last step, each week value provides the necessary scaling factor for the hourly resolution data (which was constructed to have unit energy, rendered in 1-D or 2-D form). Therefore, when a week value is multiplied by the hourly week template, the hourly resolution is achieved. This enables one to predict the desired load demand value for a particular hour of a particular week of a particular year, which can be considered as a system (with inputs: hour, day, week, and year) that is illustrated in Fig. 16. Model output value for different functions is found using Eq. (9). As usual, h shows hour, d shows day, w shows week, and y shows year.

The coefcients of g m w have 95% condence bounds. The bsinus graphs of model approximation of data in Fig. 9b using eight F.S. coefcients can be seen in Fig. 13. The function of the g m bFourier w constructed by sum of eight Fourier terms is given as follows:

g m w g m w 4 bFourier 0:973 3:55 102 cos0:1208w 1:48 102 sin0:1208w 3:77 102 cos2 0:1208w 4:49 102 sin2 0:1208w 1:31 102 cos3 0:1208w 2:63 102 sin3 0:1208w 8:92 103 cos4 0:1208w 6:65 103 sin4 0:1208w
The coefcients of gm bFourier w have 95% condence bounds.

3.3. Level 3: hourly modeling Hourly modeling is the third part of the nested combination method, providing the highest resolution. To nd a proper hourly-based model for a year, an hourly week template is created using the hourly values for the year 2002. The template is constructed for 1-week-long hourly data. In order to nd the week template, the average hourly shape of 52 weeks in year 2002 is found. Clearly, most of the time, the hourly graph within a week (with length = 24 7 = 168) remains similar throughout the year. The arithmetic average of 52 of these graphs provides a nominal week template. Prior to averaging, in order to normalize the effective energy demand and avoid effects of overall increase or decrease within a year, the total energies of the hourly curves for each week are normalized to 1. Consequently, the total energy of the constructed template also becomes one. Normalized hourly variations visualized in 1-D and 2-D (which is the proposed hourly model approach) are presented in Fig. 14a and b. The week template structure is called T(1)(h, d) in the study. The week template structure rendered in 2-D is then modeled using 2-D discrete cosine transform (2D-DCT). The model consists of the seven largest (in magnitude) 2D-DCT coefcients and their coordinates along the 7 24 2-D grid. The model output can, therefore, be considered as the inverse 2D-DCT transform of the DCT domain with seven samples taken from the model, and all other transform samples taken as zero. In this case it is required

Lm h; d; w; y Hm y 1 1983 52 w T j h; d i;j i;j


Hm i;j

where is dened as in (2), f(j)(w) is shown in Fig. 8, g m w is i dened in Eqs. (4)(7) for i = 1, . . ., 4 respectively, and T j h; d is shown in Fig. 15. The actual load is called L(h, d, w, y) in the study. Since the desired load demand values can be extracted for any year within the reasonable range of the polynomial model, the model can be considered to work on long-term prediction. On the other hand, the resolution is hourly. 4. Application and results The accuracy of the proposed method is tested using hourly actual load values for the years 20022005. The forecasting results are obtained for the proposed model variations and different years

Table 1 MAPE, RMSE, and normalized RMSE values for hourly 4 years (20022005) and different models. Lm i;j MAPE (%) j 1 1 1 1 2 2 2 2 6.13 6.39 5.74 5.77 7.70 8.01 7.28 7.29 1411.1 1455.0 1373.3 1383.4 1624.3 1687.7 1550.5 1558.4 58.795 60.625 57.220 57.625 67.679 70.321 64.604 64.933 RMSE Normalized RMSE

h d Model (h,d,w,y) w y
Fig. 16. Basic model structure of the proposed method.

i 1 2 3 4 1 2 3 4

L(output)

208

.Bas. Filik et al. / Energy Conversion and Management 52 (2011) 199211

Table 2 MAPE and RMSE values for training and test years. Training data 2002 MAPE (%) RMSE Test data 2002 3.03 613.8 Test data 2003 5.74 1274.3 Test data 2004 6.90 1560 Test data 2005 7.30 1763.5

in terms of root mean square error (RMSE) and mean absolute percentage error (MAPE), whose denitions are given in Eqs. (10) and (11), respectively. r 1 XY XW XD XH RMSE Lh; d; w; y Lm h; d; w; y2 y1 w1 d1 h1 HDWY 10

Fig. 17. 2-D analytical model obtained by Lm of years from 2002 to 2005. 3;1

2.5

x 10

4
Actual load values Forecasted load values

MW

x 10

Actual load values Forecasted load values

MW

1.5

1.5

1 6100

6150

6200 Hours

6250

6300

0.5

1.5

2.5

3.5
x 10
4

Hours

Fig. 18. Actual load (L) and model output model obtained by Lm of years from 2002 to 2005 (MAPE = 5.74%, RMSE = 1373.3, normalized RMSE = 57.220) a portion is 3;1 zoomed for visual purposes.

Fig. 19. Error surface of the Lm of years from 2002 to 2005. 3;1

.Bas. Filik et al. / Energy Conversion and Management 52 (2011) 199211 Table 3 MAPE, RMSE, and normalized RMSE for weekly total loads of years between 2002 and 2005 for different models. Lm i;j i 1 2 3 4 1 2 3 4 j 1 1 1 1 2 2 2 2 2.69 3.35 1.87 1.98 3.76 4.59 2.65 2.74 90,233 109,560 70,250 75,630 120,400 140,380 86,590 90,270 537.101 652.148 418.154 450.178 716.666 835.595 515.416 537.321 MAPE (%) RMSE Normalized RMSE

209

MAPE

! Y W D H X X X X jLh; d; w; y Lm h; d; w; yj 1 HDWY y1 w1 d1 h1 Lh; d; w; y 100% 11

Table 4 MAPE, RMSE, and normalized RMSE for monthly total loads of years between 2002 and 2005 for different models. Lm i;j i 1 2 3 4 1 2 3 4 j 1 1 1 1 2 2 2 2 2.48 3.11 1.50 1.59 3.67 4.52 2.40 2.50 345,400 434,540 237,800 264,300 489,800 593,440 333,100 348,250 479.722 603.527 330.277 367.083 680.277 824.222 462.638 483.680 MAPE (%) RMSE Normalized RMSE

Table 5 MAPE, RMSE, and normalized RMSE for yearly total loads of years between 2002 and 2005 for different models. Lm i;j i 1 2 3 4 1 2 3 4 j 1 1 1 1 2 2 2 2 1.84 2.74 0.94 0.73 3.50 4.39 1.95 1.84 2,962,600 4,026,300 1,394,900 1,082,900 5,143,000 6,339,000 3,056,000 2,815,700 339.125 460.885 159.672 123.958 58.713 725.618 349.816 322.309 MAPE (%) RMSE Normalized RMSE

where L(h, d, w, y) and Lm(h, d, w, y) are actual load and forecasted load values respectively. Since the numerical values of the load entity increases every year, in order to make a fair comparison in terms of the squared error, we also present a normalized RMSE, which corresponds to the RMSE value normalized by the average year load value for each year. The error measurements are made within a time window of interest that starts from 1, and ends at the last recording hour, say, HDWY. In our case, we take 4 year long comparisons, therefore HDWY = 24 7 52 4 = 34,944. Table 1 summarizes the MAPE, RMSE, and normalized RMSE values for hourly 4 years (20022005) and different model candidates mentioned in Section 3. As seen from the Table 1, the minimum forecasting error is obtained for the Lm . Therefore, a more detailed analysis is per3;1 formed over this method. In this new set of experiments. For example, the hourly actual load of 2002 are used as a base data (for training), and then forecasting error values are found of the other years 2002, 2003, 2004 and 2005, respectively, and result are presented in Table 2. As can be noticed from Table 2, as test year gets farther away from the training year, the accuracy slightly decreases. Nevertheless considering the large time span (of several years) with an impressive hourly resolution, the amount of decrease is still reasonable. The 2-D analytical models obtained from Lm for years from 3;1 2002 to 2005 are given in Fig. 17. Actual data and model output data for years 20022005 are again rendered in 1-D and given Fig. 18 in an overlaid plot. The difference (error) data is presented in 2-D surface form in Fig. 19. It must be noted that the error values (less than 6000) are signicantly small as compared to the actual values (about 20,000) of the load demand. In Tables 35 the MAPE, RMSE, and normalized RMSE values of the different models of years between 2002 and 2005 for weekly, monthly, and yearly total load values are given respectively. Actual and model output data for years between 2002 and 2005 are rendered in 1-D and given in Figs. 2022 for weekly, monthly, and yearly in an overlaid plot, respectively.

3.6

x 10

3.4

3.2

MW
2.8 2.6 2.4
Forecasted load values Actual load values

2.2 20 40 60 80 100 120 140

160

180

200

Weeks
Fig. 20. Weekly 1-D actual load (L) and model output obtained by Lm of years from 2002 to 2005 (MAPE = 1.87%, RMSE = 70,250, normalized RMSE = 418.154). 3;1

210
7

.Bas. Filik et al. / Energy Conversion and Management 52 (2011) 199211

1.45 1.4 1.35 1.3 1.25

x 10

MW

1.2 1.15 1.1 1.05 1 0.95 5 10 15 20 25 30 35


Forecasted load values Actual load values

40

45

Months
Fig. 21. Monthly 1-D actual load (L) and model output obtained by Lm of years from 2002 to 2005 (MAPE = 1.50%, RMSE = 237,800, normalized RMSE = 330.277). 3;1

1.65

x 10

1.6

1.55

1.5

MW
1.45 1.4 1.35
Forecasted load values Actual load values

1.3 1 1.5 2 2.5 3 3.5 4

Years
Fig. 22. Yearly 1-D actual load (L) and model output obtained by Lm of years from 2002 to 2005 (MAPE = 0.73%, RMSE = 1,082,900, normalized RMSE = 123.958). 4;1

5. Conclusions Load forecasting is important for energy suppliers, nancial institutions, and other participants in electric energy generation, transmission, distribution, and markets. The three load forecasting types, which are short-, medium-, and long-term, are very important for power planning and operation. Until now, the motivations behind these three forecast ranges were kept different due to their inherently different resolutions. However, this work provides a unied approach that enables the same hourly resolution property for all of the mentioned forecast ranges. Consequently, unique and versatile applications and utilizations may emerge from the availability of such a versatile tool. It must always be kept in mind that the results of a shorter term time windowed system constantly monitors the load values and incorporates them into the prediction of a near future, say, several hours or days, whereas this system is capable of making predictions for fairly far future hours, say, within several years without need of constantly recording newly available data.

The proposed method consists of a nested combination of three sub-sections for modeling. Several mathematical functions are applied at each level of the nested system for achieving the minimal forecasting error. Unlike other hourly prediction methods, the system is capable of making predictions for several years ahead, making it unique in the eld of load demand forecasting. The three sub-sections govern different variation resolutions corresponding to yearly, weekly, and hourly load demand. The reason for splitting the model into three parts is to keep each part fairly simple in terms of mathematical complexity. Despite the simplicity of each resolution sub-section, plausible prediction results are obtained. Clearly, other model types with increased complexities could be used and tested in the work. It is also possible that better forecasting results can be achieved with such (probably more complex) models. On the other hand, the aim of this work is only to illustrate the availability of a multi resolution framework for the forecasting problem. The plausible results indicate that the proposed approach is a promising tool for the problem. Testing several other function candidates at each sub-section would complicate the

.Bas. Filik et al. / Energy Conversion and Management 52 (2011) 199211

211

manuscript, and probably occlude the point of the work. Therefore, an extensive suit of models is left beyond the scope of this paper. References
[1] Bridger MM. A short guide to electric utility load forecasting; 1986. [2] Papalexopoulos AD, Hesterberg TC. A regression based approach to short term system load forecasting. In: Proceedings of PICA conference, vol. 3; 1989. p. 41423. [3] Haida T, Muto S. Regression based peak load forecasting using a transformation technique. IEEE Trans Power Syst 1994;9:178894. [4] Charytoniuk W, Chen MS, Van Olinda P. Nonparametric regression based shortterm load forecasting. IEEE Trans Power Syst 1998;13:72530. [5] Hill T, OConnor M, Remus W. Neural networks models for time series forecasts. Manage Sci 1996:108292. [6] Hagan MT, Behr SM. The time series approach to short term load forecasting. IEEE Trans Power Syst 1987;PWRS-2:78591. [7] Cho MY, Hwang JC, Chen CS. Customer short-term load forecasting by using ARIMA transfer function model. Proc Int Conf Energy Manage Power Deliv 1995;1:31722. [8] Fan JY, McDonald JD. A real-time implementation of short-term load forecasting for distribution power systems. IEEE Trans Power Syst 1994;9:98894. [9] Amjady N. Short-term hourly load forecasting using time series modeling with peak load estimation capability. IEEE Trans Power Syst 2001;16:798805. [10] Irisarri GD, Widergren SE, Yehsakul PD. On-line load forecasting for energy control center application. IEEE Trans Power Appar Syst 1982;101:718. [11] Ho KL, Hsu YY, Chen FF, et al. Short-term load forecasting of Taiwan power system using a knowledge based expert system. IEEE Trans Power Syst 1990;5:121421. [12] Rahman S, Hazim O. Load forecasting for multiple sites: development of an expert system-based technique. Electr Power Syst Res 1996;39:1619. [13] Park DC, El-Sharkawi MA, Marks II RJ, et al. Electric load forecasting using an articial neural network. IEEE Trans Power Syst 1991;6:4429. [14] Papalexopoulos AD, Hao S, Peng TM. An implementation of a neural network based load forecasting model for the EMS. IEEE Trans Power Syst 1994;9:195662. [15] Bakirtzis AG, Petridis V, Kiartzis SJ, et al. A neural network short-term load forecasting model for the Greek power system. IEEE Trans Power Syst 1996;11:85863. [16] Khotanzad A, Rohani RA, Maratukulam D. Articial neural network short-term load forecastergeneration three. IEEE Trans Neural Networks 1998;13:141322. [17] Kandil N, Wamkeue R, Saad M, et al. An efcient approach for short-term load forecasting using articial neural networks. Electr Power Energy Syst 2006;28:52530. [18] Mandal P, Senjyu T, Funabashi T. Neural networks approach to forecast several hour ahead electricity prices and loads in a deregulated market. Energy Convers Manage 2003;47:212842. [19] Topalli AK, Erkmen I, Topalli I. Intelligent short-term load forecasting in Turkey. Electr Power Energy Syst 2006;28:43747. [20] Santos PJ, Martins AG, Pires AJ. Designing the input vector to ANN-based models for short-term load forecast in electricity distribution systems. Electr Power Energy Syst 2007;29:33847. [21] Kurban M, Basaran Filik U. Next day load forecasting using articial neural network models with autoregression and weighted frequency bin blocks. Int J Innovative Comput, Inform Control 2009;5(4):88998.

[22] Kiartzis SJ, Bakirtzis AG. A fuzzy expert system for peak load forecasting: application to the Greek power system. In: Proceedings of the 10th mediterranean electrotechnical conference, vol. 3; 2000. p. 1097100. [23] Miranda V, Monteiro C. Fuzzy inference in spatial load forecasting. In: Proceedings of IEEE power engineering winter meeting, vol. 2; 2000. p. 10638. [24] Bakirtzis AG, Theocharis JB, Kiartzis SJ, Satsois KJ. Short term load forecasting using fuzzy neural networks. IEEE Trans Power Syst 1995;10(3):151824. [25] Srinivasan D, Chang DS, Liew AC. Demand forecasting using fuzzy neural computation, with special emphasis on weekend and public holiday forecasting. IEEE Trans Power Syst 1995;10(4):1897903. [26] Saini LM. Peak load forecasting using Bayesian regularization, resilient and adaptive backpropagation learning based articial neural networks. Electr Power Energy Syst 2008;78(7):130210. [27] Lauret P, Fock E, Randrianarivony RN, et al. Bayesian neural network approach to short time load forecasting. Energy Convers Manage 2008;49(5):115666. [28] Amin-Naseri MR, Soroush AR. Combined use of unsupervised and supervised learning for daily peak load forecasting. Energy Convers Manage 2008;49(6):13028. [29] Yao SJ, Song YH, Zhang LZ, et al. Wavelet transform and neural networks for short-term electrical load forecasting. Energy Convers Manage 2000;41(18):197588. [30] Tai N, Stenzel J, Wu H. Techniques of applying wavelet transform into combined model for short-term load forecasting. Electr Power Syst Res 2006;76:52533. [31] El-Telbany M, El-Karmi F. Short-term forecasting of Jordanian electricity demand using particle swarm optimization. Electr Power Syst Res 2008;78(3):42533. [32] Barakat EH. Modeling of nonstationary time-series data. Part II. Dynamic periodic trends. Electr Power Energy Syst 2001;23:638. [33] Abdel-Aal RE, Al-Garni AZ. Forecasting monthly electric energy consumption in Eastern Saudi Arabia using univariate time-series analysis. Energy 1997;22(11):105969. [34] Ghiassi M, Zimbra DK, Saidane H. Medium term system load forecasting with a dynamic articial neural network model. Electr Power Syst Res 2006;76(5):30216. [35] Islam SM, Al-Alawi SM, Ellithy KA. Forecasting monthly electric load and energy for a fast growing utility using an articial neural network. Electr Power Syst Res 1995;34:19. [36] Gonzlez-Romera E, Jaramillo-Morn MA, Carmona-Fernndez D. Monthly electric energy demand forecasting with neural networks and Fourier series. Energy Convers Manage 2008;49:313542. [37] Wills HL, Tram HN. Load forecasting for transmission planning. IEEE Trans Power Syst 1984;103:5618. [38] Alexander GP, Esmaeil O, Muthusami J, et al. Development of an intelligent long-term electric load forecasting systems. IEEE Trans Power Syst 1996;11(2):85863. [39] Padmakumari K, Mohandas KP, Thiruvengadam S. Long term distribution demand forecasting using neuro fuzzy computations. Electr Power Energy Syst 1999;21:31522. [40] Jia NX, Yokoyama R, Zhou YC, Gao ZY. A exible long-term load forecasting approach based on new dynamic simulation theory GSIM. Electr Power Energy Syst 2001;23:54956. [41] Carpinteiro OAS, Leme RC, Souza ACZ, et al. Long-term load forecasting via a hierarchical neural model with time integrators. Electr Power Syst Res 2007;77:3718. [42] Hong WC. Electric load forecasting by support vector model. Appl Math Model 2009;33:244454.

Anda mungkin juga menyukai