Escolar Documentos
Profissional Documentos
Cultura Documentos
a r t i c l e
i n f o
Article history:
Received 5 August 2013
Received in revised form 26 June 2014
Accepted 7 July 2014
Keywords:
Holt-Winters
Weighted nearest neighbor
Haar wavelet
Load forecasting
Hybrid model
a b s t r a c t
Short term load forecasting (STLF) is an integral part of power system operations as it is essential for
ensuring supply of electrical energy with minimum expenses. This paper proposes a hybrid method based
on wavelet transform, Triple Exponential Smoothing (TES) model and weighted nearest neighbor (WNN)
model for STLF. The original demand series is decomposed, thresholded and reconstructed into deterministic and uctuation series using Haar wavelet lters. The deterministic series that reects the slow
dynamics of load data is modeled using TES model while the uctuation series that reects the faster
dynamics is tted by WNN model. The forecasts of two subseries are composed to obtain the 24 h ahead
load forecast. The performance of the proposed model is evaluated by applying it to forecast the day
ahead load in the electricity markets of California and Spain. The results obtained demonstrate the forecast accuracy of the proposed technique.
2014 Elsevier Ltd. All rights reserved.
Introduction
The accurate forecast of electricity demand (load) for a range of
lead times is important for the effective management of power systems [1]. Short term load forecast (STLF) of power system loads
with lead times ranging from an hour to several days ahead has
a signicant impact on the operational efciency of electrical utilities [2]. The forecasts are needed for a variety of utility attributes
such as generation scheduling, the scheduling of fuel purchase,
maintenance scheduling, security analysis and energy transactions
[3]. This importance of load forecasts on decision making in energy
sector, calls for the development of accurate, fast and simple prediction algorithms. Unfortunately, the load demand is a nonstationary process inuenced by a multitude of factors ranging
from weather conditions over seasonal effects to socio-economic
factors and random effects, which makes load forecasts difcult
[4,5].
The existing models for load forecasting can be broadly
classied into conventional (statistical) approaches, articial intelligence based models and hybrid/combination models. The survey
papers [4,69] give an account of different techniques devoted to
electrical load analysis and forecasting. The conventional models
Corresponding author. Tel.: +91 0891 2739144, +91 08885043344.
E-mail addresses: g.sudheer@gvpcew.ac.in (G. Sudheer), suseelatha.a@gmail.
com (A. Suseelatha).
1
Tel.: +91 09866000118.
http://dx.doi.org/10.1016/j.ijepes.2014.07.043
0142-0615/ 2014 Elsevier Ltd. All rights reserved.
are linear and make certain assumptions regarding the characteristics of the load series. Among these models, the autoregressive
integrated moving average (ARIMA) model has excellent natural
statistical characteristics and is the most popular [10]. However,
these models are known to show some weakness in the presence
of special events and nonlinearities [11,12]. The low adaptability
of the statistical techniques to deal with nonlinear series has generated an increasing interest in using Articial intelligence based
techniques [13,14]. Among the AI based techniques, articial neural networks (ANN) have been successful applied to STLF [8]. However, a certain regularity of the data is an important precondition
for the successful application of Neural Networks (NN) [15] and
in addition they suffer from a number of weaknesses such as network construction problem, over tting issue, connection weight
estimation and the need for a large number of data for training
[16]. Moreover, these methods cannot capture well the rapid
changes in load [17]. Currently the emphasis is on hybrid models
that blend different technique to enhance the performance and
eliminate the limitations of existing individual models [18,19].
Both theoretical and empirical ndings in the literature show that
combining different methods can be an effective and efcient way
to capture the different patterns in the data and improve the forecasting performance [3,20].
Recently, hybrid models that integrate wavelet transforms with
other techniques [2,15,16,2125] have been extensively used for
short term load forecasting. In these studies, wavelet transforms
have been effectively employed to extract/analyze the characteristic
341
where N is the length of the lter. These lters are constructed from
the wavelet kernel w(t) and the companion scaling function /(t)
through the relations
/t
pX
2 hk/2t k
wt
pX
2 gk/2t k
Using the wavelet lters, the data can be decomposed into a set
of low and high frequency components named approximations
aj[k] and details dj[k] respectively. They are given by
dj1 k
X
aj ng2k n
aj1 k
X
aj nh2k n
then xk aJ;k
J
X
dj;k k 1; 2; . . . ; l
j1
342
constructed from an exponentially weighted average of past observation [34]. The method gains its popularity due to its simple
model formulation and good forecasting results [35]. An excellent
survey of the state of art in exponential smoothing models is
brought out in [36]. The application of this method to load forecasting problem can be found in [1,3739]. In particular, the triple
exponential smoothing (Holt-Winters) method, allows us to deal
with univariate time series that contain both trend and seasonal
factors. The HoltWinters exponential smoothing approach [40]
includes methods for both additive and multiplicative seasonal
patterns. The additive form of the TES method works with the following recursive updating equations:
F nh Ln hbn Snhc
10
11
Ld1 P
1
i2N ai
X
i2N
ai Li1 where ai
distqk ; d disti; d
distqk ; d distq1 ; d
12
Proposed methodology
The discrete Haar wavelet transform is applied to the historical load data to decompose it into several other scales with different levels of resolution. Applying the wavelet lters on the
data at each scale transfers the noisy characteristics of the data
into a set of coefcients whose absolute values are smaller than
the rest of the coefcients [44]. When the Haar wavelet lter is
applied to the load signal, the components under different time
scale levels show different characteristics. Both the values and
variation characteristics of wavelet coefcients in the deterministic series and uctuation series are obviously different and an
accurate wavelet coefcient threshold will separate the two series [45]. Using wavelet thresholding, the uctuation coefcients
can be separated to recover the deterministic component of
the data. Thus utilizing the Haar DWT together with wavelet
de-noising technique the demand series is decomposed into
deterministic and uctuation components. The deterministic
component contains the trend and seasonality factors and is
therefore modeled using TES method. The uctuation component
that contains the rapid changing dynamics of data is modeled
using WNN method. The proposed forecast strategy is summarized in the following steps:
1. Decompose the historical load data (X) via the DWT into two,
three, four levels of decomposition. In the rst level decomposition, the Haar lters splits up the load series into a low
frequency component (approximation) and high frequency
component (details). The approximation represents a general pattern of the load signal where as the details can be
viewed as a difference between two successive approximations. In the second level of decomposition, these approximation coefcients are further decomposed into
approximation and detail coefcients. In this manner, the
Haar wavelet splits up the load signal into one approximation series and some detail series in the wavelet domain.
The best resolution level is chosen based on the approximation signals ability to describe the general pattern of the
load series.
2. The detail/wavelet coefcients are subjected to thresholding
which separates the variation below a certain value. Here
the wavelet coefcients are thresholded adaptively at each
level using Steins Unbiased Risk Estimation (SURE) method
[46,47].
3. After thresholding the coefcients, wavelet reconstruction of
the decomposed levels is carried out using the inverse wavelet transform to obtain the deterministic component (Xd) of
the data. The uctuation series (Xf) is then obtained using
the relation Xf = X Xd.
4. The deterministic series Xd is now modeled using the TES
method through Eqs. (7)(10). The strategy for obtaining
the seed values to be used is the following:
i. Compute L24
24
1 X
xi
24 i1
13
where (x1, x2,. . ., x24) is the deterministic part in the hourly load of
the rst day of the test week. This value is taken to be the initial/
seed value L0 of the level term (Lt) in Eq. (7).
ii. The deterministic part (x1, x2,. . ., x48) in the rst two days of
the hourly load is used to initialize the trend term b0 in Eq.
(8). It is computed as
b0
1 hx25 x1 x26 x2
x48 x24 i
...
24
24
24
24
14
343
MAPE
i 1; 2; . . . ; 24
15
5. With the initial values L0, b0, S1, S2, . . .S24 determined, each of
the parameters a, b, c are varied in steps of 0.01 from 0 to 1
to determine the forecast values of the deterministic
component.
6. The uctuation series (Xf) is modeled using WNN. The
parameters window length (m) and number of neighbors
(k) are determined by sliding a window of length (m) along
the training data and minimizing the forecast error. For a
training data of two weeks, m is varied from 1 to 12 and k
from 2 to 14 m. For each pair of values m, k the forecast
values of the uctuation component is obtained.
7. For every possible combination of the parameters
a, b, c, m, k, the 24 h ahead load is forecast by composing
the forecast values of the deterministic and uctuation components. The best parameters are those resulting in the
smallest 24 h ahead forecasting errors. The minimum value
of the mean squared forecast error determines the optimum
values of the parameters.
8. With these optimum values and Eqs. (10) and (12), the 24 h
^ d and uctuation compoahead deterministic component X
^ f are determined. The next day load is obtained using
nent X
the relation
^X
^d X
^f :
X
16
N
1X
jli ^li j
100
N i1 li
17
where li and ^li are the actual load and the forecast load at hour i
respectively and N is the number of predictions. The forecasts for
the test weeks are further compared on the two metrics MAPEWeek
and weekly error variance EVWeek [50,51]. The weekly MAPE is calculated as
MAPEWeek
168
1 X
jli ^li j
168 i1 lWeek
18
P168
1
where lWeek 168
i1 li . The weekly error variance which is an index
of the uncertainty of a model is given by
EVWeek
168
1 X
jli ^li j
MAPEWeek
168 i1 lWeek
!2
19
344
Table 1
The daily forecast errors (MAPE) for the analyzed weeks in California energy market.
Winter (January 17th23rd)
TES
WNN
HWNN
Proposed model
TES
WNN
HWNN
Proposed model
TES
WNN
HWNN
Proposed model
1.7188
1.4312
0.6060
0.6510
1.0042
2.5742
2.5820
2.7163
1.3947
0.8391
0.7904
1.3361
2.0349
1.0618
2.601
1.4332
0.8179
0.5964
1.0731
2.0349
1.0607
1.2093
1.1833
0.6040
0.4594
0.6793
1.8400
0.9827
1.2485
1.2225
0.6704
0.7271
1.3304
2.2315
2.4814
2.0983
2.0669
1.4034
1.7177
1.617
2.2248
1.786
1.8632
1.5373
0.9655
1.3536
1.6183
1.7234
1.5315
1.0672
0.9696
0.5146
0.5154
1.0819
1.6620
1.6416
2.0506
1.2747
0.7819
0.5084
0.8176
1.7130
2.3591
1.3655
2.1205
3.8194
1.2548
1.5807
1.5671
3.5665
1.4303
2.2422
3.1618
1.2346
1.5309
1.4204
2.866
1.2755
0.8408
0.7226
0.4915
0.6879
1.5682
1.4231
Table 2
The daily forecast errors (MAPE) for the analyzed weeks in Spain energy market.
Winter (January 21st27th)
TES
WNN
HWNN
Proposed model
TES
WNN
HWNN
Proposed model
1.4830
0.6998
0.4017
0.3483
0.5443
3.1206
2.6959
1.5742
1.5738
0.5027
0.6051
1.2139
1.9353
3.465
1.2489
1.0239
0.3716
0.5801
1.0659
1.5939
4.1628
1.3374
0.6452
0.3040
0.4111
0.5485
1.1651
2.3995
1.6730
1.2779
0.3520
0.4050
0.7121
4.0101
3.1665
2.4037
1.4595
0.6188
1.0022
0.8962
2.583
3.0863
2.254
1.3018
0.5341
0.9311
0.7708
2.583
3.012
1.6107
0.7196
0.3040
0.3565
0.5883
2.4004
2.7681
TES
WNN
HWNN
Proposed model
TES
WNN
HWNN
Proposed model
2.2078
0.6878
0.6282
0.2806
0.3588
2.3799
2.2180
2.6488
0.7636
1.0742
0.3758
0.9567
1.8789
2.087
2.4542
0.5656
0.9139
0.4559
0.9152
1.8789
1.7056
1.8085
0.5504
0.6084
0.2723
0.3049
1.3946
1.8313
2.6176
1.0408
0.6183
0.4264
0.5404
3.5552
3.9553
2.1434
0.9244
0.9499
1.092
0.6428
1.4396
2.0888
2.1434
0.9123
0.8437
1.0095
0.6528
1.4396
2.0396
2.2177
0.7314
0.5656
0.4066
0.5225
1.4267
2.4188
Table 3
Weekly forecast errors for the analyzed weeks of different energy markets.
Energy market
Typical week in
MAPEWeek
TES
WNN
HWNN
Proposed model
Winter
Spring
Summer
1.5097
1.4160
1.3579
1.4533
1.8449
2.1821
1.3739
1.5132
1.9837
0.9940
1.0646
1.0967
Winter
Spring
Summer
Fall
1.3276
1.6567
1.2516
1.8220
1.5529
1.7214
1.3978
1.3273
1.4353
1.6267
1.2699
1.2915
0.9730
1.2497
0.9672
1.1842
California
Spain
Table 4
Weekly error variances for the analyzed weeks of different energy markets.
Energy market
Typical week in
EVWeek
TES
WNN
HWNN
Proposed model
Winter
Spring
Summer
0.000157
0.000123
0.000130
0.000160
0.000165
0.000328
0.000155
0.000161
0.000246
0.000085
0.000075
0.000086
Winter
Spring
Summer
Fall
0.000234
0.000316
0.000816
0.000330
0.000182
0.000211
0.000136
0.000120
0.000181
0.000199
0.000120
0.000122
0.000102
0.000155
0.000134
0.000156
California
Spain
345
Conclusion
A new approach for forecasting the 24 h ahead electricity load is
presented. The Haar wavelet transform is used as a preprocessor in
the approach to decompose the original load series into deterministic and uctuation series. The seasonal oscillations and trend factors contained in the deterministic component is modeled using
TES method. The faster dynamics of data contained in the uctuation component is modeled using the WNN technique. The proposed approach is illustrated using the load data of the energy
markets of California and Spain for the years 2000 and 2002
respectively. The accuracy of the technique which is based on using
the Haar wavelet combined with HoltWinters additive seasonal
model and WNN model outperforms the direct use of HW model,
WNN model and Haar wavelet combined with WNN model. The
performance of the other wavelet transforms such as the Daubechies wavelets in improving the forecast accuracy will be investigated in the future.
References
Fig. 1. Actual load and forecast load using proposed model for typical weeks in (a)
winter, (b) spring and (c) summer seasons of California energy market.
Fig. 2. Actual load and forecast load using proposed model for typical weeks in (a)
winter, (b) spring, (c) summer and (d) fall seasons of Spain energy market.
[1] Taylor JW. Triple seasonal methods for short-term electricity demand
forecasting. Eur J Oper Res 2010;204:13952.
[2] Bashir ZA, El-Hawary ME. Applying wavelets to short-term load forecasting
using PSO-based neural networks. IEEE Trans Power Syst 2009;24(1):207.
[3] Wang J, Wang J, Li Y, Zhu S, Zhao J. Techniques of applying wavelet denoising
into a combined model for short-term load forecasting. Int J Elect Power
Energy syst 2014;62:81624.
[4] Hahn H, Nieberh MS, Stefan Pickl. Electrical load forecasting methods: tools for
decision making. Eur J Oper Res 2009;199:9027.
[5] Ceperic E, Ceperic V, Baric A. A strategy for short term load forecasting by
support
vector
regression
machines.
IEEE
Trans
Power
Syst
2013;28(4):435664.
[6] Gross G, Galiana FD. Short term load forecasting. Proc IEEE 1987:155873.
[7] Tzafestas S, Tzafestas E. Computational intelligence techniques for short-term
electric load forecasting. J Intell Robot Syst 2001;31:768.
[8] Hippert HS, Pedreira CE, Souza RC. Neural networks for short term load
forecasting: a review and evaluation. IEEE Trans Power Syst 2001;16(1):4455.
[9] Alfares HK, Nazeeruddin M. Electric load forecasting: literature survey and
classication of methods. Int J Syst Sci 2002;33:2334.
[10] Fard AK, Fard FK. A new hybrid correction method for short-term load
forecasting based on ARIMA, SVR and CSA. J Exp Theo Artif Intell
2013;25(4):55974.
[11] Moghram I, Rehman S. Analysis and evaluation of ve short term load
forecasting techniques. IEEE Trans Power Syst 1989;4(4):148491.
[12] Amina M, Kodogiannis VS, Petrounias I, Tomtsis D. A hybrid intelligent
approach for the prediction of electricity consumption. Int J Elect Power
Energy Syst 2012;43:99108.
[13] Romere EG, Jaramillo-Moran MA, Diego CF. Monthly electric energy demand
forecasting based on trend extraction. IEEE Trans Power Syst
2006;21(4):194553.
[14] Jaramillo-Maran MA, Romere EG, Diego CF. Monthly electric demand
forecasting with neural lters. Int J Elect Power Energy Syst 2013;49:25363.
[15] Reis AJR, da Silva APA. Feature extraction via multiresolution analysis for
short-term load forecasting. IEEE Trans Power Syst 2005;20(1):18998.
[16] Fard AK, Akbari-Zadeh MR. A hybrid method based on wavelet, ANN and
ARIMA model for short-term load forecasting. J Exp Theo Artif Intell
2014;26(2):16782.
[17] Chen Y, Luh PB, Guan C, Zhao Y, Michel LD, Coolbeth MA, et al. Short-term load
forecasting: similar day-based wavelet neural networks. IEEE Trans Power Syst
2010;25(1):32230.
[18] Zhu Suling, Wang Jianzhou, Zhao Weigang, Wang Juie. A seasonal hybrid
procedure for electricity demand forecasting in China. Appl Energy
2011;88:380715.
[19] Hanmandlu M, Chauhan BK. Load forecasting using hybrid models. IEEE Trans
Power Syst 2011;26(1):209.
[20] Cagdas Hakan Aladag, Egrioglu Erol, Kadilar Cem. Forecasting nonlinear time
series with a hybrid methodology. Appl Math Lett 2009;22:146770.
[21] Lee Cheng-Ming, Ko Chia-Nan. Short term load forecasting using lifting
scheme and ARIMA models. Expert Syst Appl 2011;38(5):590211.
[22] Amjady N, Keynia F. Short term load forecasting of power systems by
combination of wavelet transform and Neuro-evolutionary algorithm. Energy
2009;34(1):4657.
[23] Pandey AK, Singh Devender, Sinha SK. Intelligent hybrid wavelet models for
short-term load forecasting. IEEE Trans Power Syst 2010;25(3):126673.
[24] Suseelatha A, Sudheer G, Bharathi D. A Simple Hybrid Model for Short-Term
Load Forecasting, J Engg, Hindawi 2013; Article ID 760860. doi:10.1155/2013/
760860.
346
[39] Gould PG, Koehler AB, Ord JK, Snyder RD, Hyndman RJ, Vahid-Araghi F.
Forecasting time-series with multiple seasonal patterns. Eur J Oper Res
2008;191:20722.
[40] Winters P. Forecasting sales by exponentially weighted moving averages.
Manag Sci 1960;6:32442.
[41] Fix E, Hodges J.L. Discriminatory analysis, nonparametric discrimination:
consistency properties. USAF School of Aviation Medicine Randolgh eld, Tx,
Project No. 21-49-004 1951;4, Contract No. AF41(128)-31.
[42] Dudani SA. The distance-weighted nearest-neighbor rule. IEEE Trans Syst Man
Cyber 1976:3257.
[43] Lora AT, Santos R, Exposito G, Ramos M. Electricity market price forecasting
based on weighted nearest neighbors techniques. IEEE Trans Power Syst
2007;22(3):1294301.
[44] Akbaryan F, Bishnoi PR. Smooth representation of trends by a wavelet-based
technique. Comput Chem Eng 2000;24:191343.
[45] Sang YF, Wang D, Wu JC, Zhu QP, Wang L. Entropy based wavelet denoising
method for time series analysis. Entropy 2009;11(4):112347.
[46] Gencay Ramzan, Seluk Faruk, Whitcher Brandon. An introduction to
wavelets and other ltering methods in nance and economics. Academic
press; 2002.
[47] Donoho DL, Johnstone IM. Ideal spatial adaptation via wavelet shrinkage.
Biometrika 1994;81:42555.
[48] http://www.ucei.berkely.edu/ucei/datamine/ucei-data.
[49] http://www.omel.es/.
[50] Conejo AJ, Plazas MA, Espnola R, Molina AB. Day-ahead electricity price
forecasting using the wavelet transform and ARIMA models. IEEE Trans Power
Syst 2005;20(2):103542.
[51] Tan Zhongfu, Zhang J, Wang J, Jun Xu. Day-ahead electricity price forecasting
using wavelet transform combined with ARIMA and GARCH models. Appl
Energy 2010;87(11):360610.