Energies 2011, 4, 1495-1507; doi:10.

3390/en4101495
OPEN ACCESS

energies
ISSN 1996-1073
www.mdpi.com/journal/energies

Article

Forecasting Monthly Electric Energy Consumption Using
Feature Extraction
Ming Meng *, Dongxiao Niu and Wei Sun

School of Economics and Management, North China Electric Power University, Baoding 071003,
Hebei, China; E-Mails: niudx@126.com (D.N.); sunweichina@sohu.com (W.S.)

* Author to whom correspondence should be addressed; E-Mail: ncepumm@126.com;
Tel.: +86-13731661516; Fax: +86-0312-7525110.

Received: 22 July 2011; in revised form: 14 September 2011 / Accepted: 20 September 2011 /
Published: 28 September 2011

Abstract: Monthly forecasting of electric energy consumption is important for planning the
generation and distribution of power utilities. However, the features of this time series are so
complex that directly modeling is difficult. Three kinds of relatively simple series can be
derived when a discrete wavelet transform is used to extract the raw features, namely, the
rising trend, periodic waves, and stochastic series. After the elimination of the stochastic
series, the rising trend and periodic waves were modeled separately by a grey model and
radio basis function neural networks. Adding the forecasting values of each model can yield
the forecasting results for monthly electricity consumption. The grey model has a good
capability for simulating any smoothing convex trend. In addition, this model can mitigate
minor stochastic effects on the rising trend. The extracted periodic wave series, which
contain relatively less information and comprise simple regular waves, can improve the
generalization capability of neural networks. The case study on electric energy consumption
in China shows that the proposed method is better than those traditionally used in terms of
both forecasting precision and expected risk.

Keywords: monthly electric energy consumption; forecasting; feature extraction; discrete
wavelet transform; neural network; grey model

sample simplification is also a feasible method for STLF [6]. the steady monthly consumption trends of electric energy often change every few years as a result of macroeconomic conditions and social development. weather conditions. The periodic wave is forecasted using the Fourier series. and other unexpected factors. The width of the data window in the moving average is selected by measuring the fitting accuracy and the smoothness of the obtained rising figure. were proven to be capable of yielding better forecasting results compared with the time series and single NN methods. González–Romera et al.. However. is difficult. the data used for modeling can only be derived from a continuous number of years. First. 4 1496 1. however. In addition. whereas the rising trend is predicted by NNs. and the stochastic series. Classical techniques. On the other hand. during which macroeconomic conditions may have varied slightly. namely. Introduction Monthly forecasting of electric energy consumption plays an important role in the operation of thermal power plants and is one of the most important basis for coal dispatch. electricity trading.g. time series methods [3–5] use a moving average to simplify the raw trends. compared with short-term load data (e. useful information may be mistakenly eliminated and the forecasting results of these methods are often unsatisfactory because not all the useful information in the raw data is considered. neural networks (NNs) have been widely applied in short-term load forecasting (STLF) [6–10] and short-term price forecasting [11. Furthermore. In the past several years. such as regression [1] and expert systems [2]. To date.12]. monthly data often contain more information.. based on the aforementioned reasons. although NN has been widely used for monthly data forecasting. a number of difficulties are associated with forecasting. numerous periodical waves. As a result. Zhao and Wei [18] have summarized a number of methods for extracting the series features. are incapable of generating precise forecasting results because of low adaptability. [19–21] adopted a moving average algorithm to extract the rising trend from a monthly electric energy demand series. 96-point daily ones). For monthly forecasting. monthly forecasting of electric energy consumption is vital for the planning and maintenance of the grid. a long-term rising trend. the monthly consumption trends often comprise at least three kinds of sub-trends.14]. and so on. Thus. The primary advantage of a NN is its capability for modeling non-linear relations without the supervision of human experts [13. even for the short term. Therefore. time series methods have been proven to be better than classical techniques. Second. .Energies 2011. combined with other methods. collecting adequate steady monthly data. The extracting techniques. while requiring relatively less information from each sample (the relationship between input and output vectors is relatively simple). NNs require significant volumes of data for training [15].g. peak load prediction) [17]. These methods remove tiny waves to smoothen the data figures. When used for load forecasting. The key to further improvement of the forecasting precision of NN in monthly forecasting is the reduction of the amount of information in each sample. the monthly consumption trend data is extremely complex because of the effects of people’s living habits. In fact. However. feature extraction is an effective method for reducing the amount of information in each sample. the primary limitation of these methods is the inability to distinguish stochastic waves from a number of other useful waves. These factors are important for forecasting precision. As a result. its applications are limited to either special trends [16] or special points (e.

DWT is used in the present work with the aim of decomposing the monthly time series.k is derived from: +∞ C j . A schematic diagram of the DWT process is shown in Figure 1. Third. two other transitional methods are designed to test the contribution of each innovation. but it can also distinguish stochastic behavior.1. but is incapable of eliminating the stochastic effect. which can simulate their increasing amplitude. 2. . the amplitude of periodic waves extracted from the raw trends will inevitably increase because of rapid increases in total monthly electric energy consumption. Second.k (3) j =−∞ k =−∞ where C is a constant. Periodic waves are forecasted by NNs. First. the moving average method is only capable of separating waves from the rising trend. As a feasible method for simulating convex function. the continuous occurrence of freak weather). both of which are derived from a continuous wavelet transform (CWT) by discreteness. the previously designed models were still observed to have certain limitations. The discrete wavelet coefficient can be used to reconstruct the function f(t) as follows: ∞ ∞ f (t ) = C ∑ ∑C ψ j .k (t ) = a ψ⎜ − j/2 j ⎟ = a ψ (a t − kb) − j/2 −j (1) ⎝ a ⎠ where aj is a scale function and kajb is a translation function.k(t) can be written as: ⎛ t − ka j b ⎞ ψ j . DWT is not only capable of extracting the rising trend and periodic waves. the Fourier series can only simulate waves with invariable amplitude.. Aside from the forecasting method discussed in Section 2. especially in developing countries. 4 1497 However. However. The DWT function ψj.k (t ) j . Section 4 concludes the present paper. The paper is structured as follows: Section 2 briefly introduces the principles and primary algorithms used in the present work. the use of NNs to simulate the extracted rising trend cannot easily mitigate minor unexpected factors (e. Feature Extraction Based on DWT As a time-frequency decomposition method for a discrete signal. and both methods are compared with the traditional algorithm through forecasting error analysis. thereby affecting generalizability. the grey model (GM) is adopted to forecast the rising trend.Energies 2011. For any discrete function f(t). Section 3 presents the case study on the monthly electric energy consumption in China as an example to test the model.k = ∫ f (t )ψ *j .g. Feature Extraction and Forecasting Algorithms 2. The present paper adopts a discrete wavelet transform (DWT) to decompose the raw forecasting figure. the discrete wavelet coefficient Cj. k (t )dt (2) −∞ where * denotes the complex conjugation of ψ.

GM for Exponential Trend Forecasting If the socio-economy runs smoothly. These wavelets are an orthogonal family defining a discrete wavelet transform and are characterized by a maximal number of vanishing moments for a given support [22].2. Equation (4) can simulate any smoothing convex trend. and b3.2. the reconstructed result of low frequency extracted from the monthly electric energy consumption series using DTW has a steadily increasing trend. by adjusting the parameters of b1. and b3 are the shape parameters. As a result. Daubechies wavelets were selected for extracting the features of the monthly electric energy consumption series.2.24] proposed a method for solving this problem. The DWT decomposes a signal into low-frequency (zL) and high-frequency coefficients (y0 − yL). and x(1)(k) (k = 1. which can be simulated using the following equation: x = b1 + b2 exp(b3t ) (4) where b1. The optimum parameters of Equation (4) cannot be obtained by minimizing the residual sum of square (RSS) because of the location of b3.…) refers to the raw series. The following differential form of Equation (4) can be processed to obtain the optimum parameters of a and u by minimizing the RSS: dx + ax = u (5) dt To process the discrete series. The relationship between these values is defined by f(t) = zL + y0 + y1 + … + yL. b2.…) stands for the accumulated k generating series and is written as x (k ) = ∑ x (0) (i ) . In fact.Energies 2011. (1) i =1 . Deng [23. both the economic growth rate and electric elasticity coefficient tend to be constant. 2. respectively. 4 1498 Figure 1. Equation (5) should be discretized as follows: 1 x (0) (k + 1) + a[ x (1) (k + 1) + x (1) (k )] = u (6) 2 where x(0)(k) (k = 1. The low-frequency and high-frequency coefficients comprise the approximate features and the details of f(t). For a wavelet function. b2. Multi-resolution analysis using DWT.

Considering the characteristics of Equation (11) k should equal i − 1 when forecasting the ith value. whereas the output layer yields the output vector y. f can be simulated using a three-layer feedforward NN. The input layer gathers the input vector x. given any n-dimension continuous mapping f. RBF NNs for Periodical Waves In 1988. and B = ⎢ 2 (0) (0) (0) T T ⎥ ⎢ # # ⎥⎥ ⎢ ⎢ 1 (1) ⎥ ⎢ − [ x (n − 1) + x (n)] (1) 1⎥ ⎣ 2 ⎦ The RSS can then be minimized to obtain the optimum parameters of Equation (5): ˆ ||= min(N − B ⋅ A min || N − B ⋅ A ˆ )T (N − B ⋅ A ˆ) (8) Using a matrix derivative will yield: ˆ = (B T B) −1 B T N = [aˆ uˆ ]T A (9) Using the optimum parameters obtained in Equation (9) and adopting x(1)(1) as a particular solution will yield the following solution for Equation (5): uˆ uˆ x (1) (k + 1) = [ x (0) (1) − ]exp(−ak ˆ )+ (10) aˆ aˆ Through inverse accumulated generating operation.…. the forecasting model of x(0) is derived as follows: uˆ x (0) (k + 1) = (1 − exp(aˆ ))( x (0) (1) − ) exp(− ak ˆ ) (11) aˆ where k = 0. In Equation (11). the hidden.2. A = [a u] . 2. For any nonlinear fluctuant. As a classic multilayer feedforward NN. The forecasting value is obtained only by entering an appropriate k. . the RBF NN often consists of three layers. 4 1499 Integrating the raw and the accumulated generating series with the iterative Equation (6) will yield the following: N = B⋅A (7) ⎡ 1 (1) ⎤ ⎢ − 2 [ x (1) + x (2)] (1) 1⎥ ⎢ ⎥ ⎢ − 1 [ x (1) (2) + x (1) (3)] 1⎥ where N = [x (2) x (3) … x (n)] .3. According to the Kolmogorov theorem. a three-layer feedforward NN can be used for theoretical simulation at any precision. k is the only variable.1. 1].Energies 2011. Broomhead and Lowe introduced the radial basis function (RBF) to the NN field. the input. namely. U × U…… × U → R × R…… × R. and the output layers. where f(x) = y and U ∈ [0.

That is.Energies 2011. which adopts an S-like kernel function that has reflections in infinite space.…. 2. Figure 2 shows the output distribution of the RBF NN with two Gauss function nodes. 4 1500 Unlike other feedforward NNs. Based on Equation (12). Forecasting Model Design The operation of the monthly electric energy consumption forecasting model involves the following steps: first. j = 1. and other periodical waves (the remaining reconstructed results from different high . the historical data are decomposed by DWT and then reconstructed by each frequency. Reconstructed results with a small amplitude and an indistinct wave period (considered as stochastic waves) are eliminated. Figure 2. The output of the RBF NN is clearly seen to rapidly approach 0 when the inputs deviate from the centers. each Gauss function has a reflection (significantly unequal 0) only when the input vector is very close to the space center (wj). wj is the center of the Gauss function of the jth node. σj2 is the spread of the Gauss function. the RBF NN has better learning efficiency compared with the traditional error back propagation NN.2. the output of each Gauss function will be between 0 and 1. and N1 is the number of nodes in the hidden layer. the kernel function of an RBF NN is a Gauss function that is usually written as: ⎡ (x − w j )T (x − w j ) ⎤ u j =exp ⎢ − ⎥ . the rising trend (the reconstructed result of low-frequency) is modeled using GM (1.4. 1). Furthermore. second. which are very important because NNs will be frequently utilized in the proposed forecasting model.N1 (12) ⎢⎣ 2σ 2j ⎥⎦ where uj is the output value of jth node in the hidden layer. Outputs of RBF NN with two nodes in the hidden layer. For high-dimension input vectors. the RBF NN often has a rapid convergence rate and enhanced training speed. x is the input vector of the hidden layer.

In fact. Data were obtained from the website of the Chinese Economic and Financial Database of the China Center for Economic Research (CCER) [25].Energies 2011.4% similar to the actual data. The monthly electric energy consumption curve (Figure 3) obviously exhibits a basic exponential rising trend and is characterized by numerous waves. 4 1501 frequencies) are determined using the RBF NNs. Monthly electricity consumption in China from January 1990 to December 2006. Results (5) and (6) have no distinct wave periods. 2600 2400 2200 2000 Electric energy consumption 1800 1600 1400 1200 1000 800 600 400 200 1 13 25 37 49 61 73 85 97 109 121 133 145 157 169 181 193 204 Month (1=Jan. third. The next section will demonstrate the forecasting steps in detail. which exhibits an exponential rise. Amplitudes were observed to gradually increase with the development of the social economy. 2006) Figure 3 was decomposed using the db(5) wavelet and was further reconstructed by each frequency. Figure 3. 1990. 204=Dec. the summation of Results (1)–(4) is 98. and their amplitudes were generally very small. Experimental Setup and Forecasting Results The actual values of the monthly electric energy consumption (108 kWh) in China from January 1990 to December 2006 were selected to validate the aforementioned methods. The reconstructed results of each frequency are shown in Figure 4. Thus. . the forecasting results of monthly electric energy consumption are obtained by adding the forecasting values of the previous models. Results (2)–(4) reflect periodical waves of different frequencies. The reconstructed low-frequency Result (1) reflects the primary monthly electric energy consumption trend. these results were classified as stochastic waves and consequently eliminated. 3.

RBF NN has been adopted to simulate the relationship between the current and a number of its previous values. and the 12 previous values of each of these outputs were selected as the inputs. the values from January to December of 2004 were used as the inputs for the network. For the reserved reconstructed high-frequency results (Results (2)–(4) in Figure 4).. when forecasting the jth point. i. if the value of certain remaining reconstructed results of January 2005 were to be forecasted. the neuron number of the RBF layer refers to the dimension of the input vector and the training goal defaults to 0.Energies 2011. Here it is set at 50. . For each forecasting point. 3000 1500 ① 0 1 13 25 37 49 61 73 85 97 109 121 133 145 157 169 181 193 204 200 0 ② -200 1 13 25 37 49 61 73 85 97 109 121 133 145 157 169 181 193 204 200 0 ③ -200 1 13 25 37 49 61 73 85 97 109 121 133 145 157 169 181 193 204 200 0 ④ -200 1 13 25 37 49 61 73 85 97 109 121 133 145 157 169 181 193 204 200 0 ⑤ -200 1 13 25 37 49 61 73 85 97 109 121 133 145 157 169 181 193 204 200 0 ⑥ -200 1 13 25 37 49 61 73 85 97 109 121 133 145 157 169 181 193 204 ① : low-frequency. When constructing training samples. ② -⑥ : high-frequency For the extracted rising trend. For each output value. the values from January 2001 to December 2004 were successively selected as outputs. the value of spread will greatly affect the feature of RBF. After training. all RBF NNs were implemented using the Matlab Neural Network Toolbox and were designed to have similar parameters. 12 consecutive previously obtained values have been adopted to model the exponential forecasting equation. In the current paper. After the NN structures were established. The 13th to 60th values were taken as outputs. Based on the same rule. the reconstructed low-frequency result has been modeled using GM (1. For example. 1). The larger that spread is the smoother the function approximation will be.e. Furthermore. The consequent output was the forecasting result of January 2005. 60 of its consecutive previous values were selected for RBF NN training. the 12 consecutive previous values are entered and the output of NN is considered the forecasting result. The trend value of forecasting was derived assuming k = 12 in Equation (11). 12 previous values were selected as the inputs. 4 1502 Figure 4. the NN is used for forecasting. For each forecasting point. Results of db(5) wavelet transformation. the values from January 2000 to December 2004 from the same time series were selected to establish the NN structure.

75 2426.91 −2.16 1. The forecasting results and errors from January 2005 to December 2006 are listed in Table 1.73 2158. 2301.09 −4.1 Sep. 2573.89 1957.17 −2.25 2035. 2364. 1940.66 1912.27 2250.Energies 2011.73 8.48 −1 2164.74 −1.88 2083.79 −4.3 −5.98 5.14 2422.35 1770.24 2422.47 1879.89 1970.18 2215.01 2266.67 2104.18 −1.7 1957.74 1878. 2005 1914.6 1741. To demonstrate the forecasting results intuitively.84 −5.4 2439.62 −1. The forecasting results of Methods 2.13 2153.9 −4.9 2148.39 2107. 2570. 2363.33 7.67 3. 2178.75 2465.82 2.61 5.16 0. 2517. 2324.43 −3.15 2092.1 −1.85 Sep. 1601.68 −0.77 −10. 2041.76 2287.59 2195.1 2057.19 −3.77 −3. Table 1.49 2191.76 3.28 2049.17 0.08 2462.28 2165.56 2.36 2052 0. 4 1503 The forecasting values of GM and RBF NNs were combined to yield the forecasted results of monthly electric energy consumption in China (hereinafter Method 1).7 2093.79 May 2175.49 2468.73 −3.6 2059. 1892.79 1684.05 Oct. .04 −2.2 0.97 1916.68 −2. 2006 2063. The NN used in the traditional method was replaced with GM (hereinafter Method 3).89 2419.93 2289.54 −4.53 −2.16 1912. Actual Error Error Error Error Method 1 Method 2 Method 3 Method 4 Data (%) (%) (%) (%) Jan.22 Nov.7 −1. 2010.05 2124. They are the applications of DWT and GM. 2116.85 4.05 −0.88 0.48 2308.16 Dec.47 1954. Analysis of forecasting results. 2162. 2246.42 1893.43 Jan.88 2.38 Jul.53 1.05 −1.78 5.01 1811.75 −6.07 2137.79 2146. Furthermore.71 Jun. the curves of actual data and the forecasting results of Methods 1 and 2 were drawn (see Figure 5).78 1988 1.92 −0.17 −3.66 −1.8 −3.97 2114.43 −0. 2006.48 Aug.33 1907.15 2057.3 2434. and 4 are also listed in Table 1.06 1686.44 Oct.93 −4. 2053.32 Mar.61 0.62 2258. 3.61 −0. [21] (hereinafter referred to as Method 2).69 2137.46 2018.25 −1.72 1843.36 1927.64 −4.38 −0.48 2288.29 −1.87 Aug.51 2130.98 0.2 2391.17 −0. the traditional method was also used as a basis for modeling the monthly electric energy consumption data for China.13 2371. To verify the contribution of each innovation.61 2 Feb.2 2122. and the moving average algorithm in the traditional method was replaced with DWT (hereinafter Method 4).15 −1.71 2194.43 2172.18 −0.58 May 1924.46 5.23 Apr.51 −0.9 −1.75 2370.88 Nov.76 −1.6 −2.35 −7.24 −6.12 1931.42 2006.5 1966.4 2.2 −3.33 3.06 1843.96 2361.78 −3.53 1.22 Mar.84 2096.56 1906.04 −1.81 2.52 8.3 −1.37 1797.68 2296.91 −0.62 2128.41 2160. 1962. two transitional methods were designed.94 2298.02 1962.55 2160.71 2031.78 −3.09 −0.37 −3. two primary innovations are notable in Method 1.1 2024.52 2228.39 0.75 2075.73 2346.86 1.02 −2.83 2368.61 0.16 2145.71 1950.1 2125. 2161.55 −2.43 −0.36 2199.3 2300.19 −0.64 −0.01 1990.51 −4.89 1934.59 −5.35 −3.66 Apr.23 2074.7 Dec.85 1.42 0.41 Compared with the traditional monthly electric energy consumption forecasting method proposed by González–Romera et al.85 Feb.51 Jun.62 Jul.44 1877.15 −7.

.. 2005. and N is the number of data used for the MAPE calculation: 1 N x(i ) − xˆ (i ) MAPE = N ∑ i =1 x(i ) ×100 (14) ⎛ x(i ) − xˆ (i ) ⎞ MdAPE = median ⎜ ⎟ × 100. median absolute percentage error (MdAPE). 2600 Actural Data Method 1 2400 Method 2 Electric energy consumption 2200 2000 1800 1600 1400 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 Month(1=Jan. 2006) The maximal absolute percentage error (MaxAPE). and geometric mean relative absolute error (GMARE) were used as indicators of forecasting precision: ⎛ x(i ) − xˆ (i ) ⎞ MaxAPE = max ⎜ ⎟ × 100. 24=Dec.. mean absolute percentage error (MAPE).... . 2. i = 1. 4 1504 Figure 5.. The results of the above indicators are listed in Table 2. xˆ (i ) represents its forecasting result.Energies 2011. N (13) i ⎝ x(i ) ⎠ where x(i) is the electric energy consumption value in the ith month. N (15) i ⎝ x(i ) ⎠ 1 ⎡ N x(i ) − xˆ (i ) ⎤ N GMARE = ⎢∏ ⎥ (16) ⎣ i =1 x(i ) − xˆ (i ) ⎦ * where xˆ * (i ) is the forecasting result obtained from the benchmark method.. Curves of actual data based on the forecasting results of Methods 1 and 2. i = 1. 2.

65. The application of DWT (Method 4) improves both the forecasting precision and forecasting risk.33 0. In the current paper. 3.42 1. . Method 1 Method 2 Method 3 Method 4 MaxAPE 6.18. Compared with the traditional method. Acknowledgements This work was supported by “the Fundamental Research Funds for the Central Universities (11MR57)” and “the National Natural Science Foundation of China (NSFC) (71071052)”.44 2. the primary contribution of GM is the reduction of the maximum forecasting error.76 2. and third.61 0. 8. and 4 were 2. The variance of absolute percentage error of Methods 1. Results of forecasting precision indicators. thereby resulting in poor forecasting performance.01 2. the following conclusions are confirmed: first.51 0.6 8. This finding proves that Method 3 performs better than Method 2 in terms of mitigating the stochastic effect on the primary trend.79 7. To mitigate the stochastic effect on the primary trend. the rising trend and other periodic waves are retained. the application of DWT in forecasting may result in a simultaneous reduction of the MAPE and maximum error. the method proposed in the present study performs better in terms of forecasting precision and expected risk.Energies 2011. The wave amplitude of Method 2 tends to be invariable. as shown in Figures 3–5. the application of GM (Method 3) has similar forecasting precision (MAPE) with less forecasting risk (MaxAPE).73 MdAPE 2. 4. The values of for 24 consecutive months of electric energy consumption in China are forecasted to test the effect of the proposed method.63 Compared with the Method 2. the GM is selected to model the rising trend of the retained components. the RBF NN is selected for use in simulation.82. and 2. 2. 5. second.76 2. 4 1505 Table 2. the MAPE and maximum error are both further reduced. In summary. the wave amplitude of the actual data increases with time. Furthermore. when GM and DWT are used simultaneously. Conclusions The difficulties in forecasting monthly electric energy consumption are attributable to the excessive information in the values relative to the limited number of samples.72. the MAPE and maximum error are both further reduced. When Method 1 is used. This finding proves that Method 1 performs better than Method 2 in terms of both forecasting precision and risk.67 2. This result proves the positive effects of DWT in simplifying the periodic waves and eliminating the stochastic effects. The biggest variance value of Methods 2 was the root cause of smallest GMARE and made it possible to obtain the smallest MAPE. Considering the increasing amplitude of retained periodic waves. Summing up the forecasting values of the GM and RBF NNs will yield the forecasting result of monthly electric energy consumption. The performance of Method 1 is acceptable for simulating this feature.16 MAPE 2. After the elimination of the stochastic series. the DWT is adopted to extract the features of monthly data into several relatively simple series.44 GMARE 0.94 10.

: Upper Saddle River.C. 2008. Saab..-Y.A. Randrianarivony.. 1059–1069. Demuth. 219–241. R. Hong..-F.. L. Ye. Nasr. 1535–1550.A. 24. Appl. G. R. Abdel-Aal. 5. NJ. Power Energy Syst.-J. Fan.-J. A regression-based approach to short-term load forecasting. Zhao. Kandil.. C. 2006. R. 2008. S. E. M. Xiao. N. 21. 2011.Z. Dong. 10. 1–14. 1996.-X. Power Syst. M. 2008. 4.R. 1441–1448. Saad. . 36. Hazim. SVR with Hybrid Chaotic Immune Algorithm for Seasonal Load Demand Forecasting. 2005. C. Keynia. Appl. Harrison. Electr. IEEE Trans. Al-Garni. 16.P. 5. Stat. S. Day-ahead price forecasting of electricity markets by mutual information technique and cascaded neuro-evolutionary algorithm. 1156–1166. Jump process for the trend estimation of time series. Neural Network Design... Univariate modeling and forecasting of monthly energy demand time series using abductive and neural networks. S. S. Soft Comput. Sun. Energy 1997. Energies 2011.D. 2011.W. F..P. C.. 4 1506 References 1. Neural Networks: A Comprehensive Foundation.. M. 19. 4246–4256. Parkpoom. A. 4. Badr. F. Amjady. Wei. Chen. Wei. 525–530. T. Monthly electric energy demand forecasting based on trend extraction. Papalexopoulos. N.. 17.. P.. Power Syst. 17–27. USA. M.-C. Keynia.. 960–977. Univariate modeling and forecasting of energy consumption: The case of electricity in Lebanon. Expert Syst... 508–514. 28. 273–279. E. A. 18. PWS Publishing Company: Boston. 13. B.. Hagan.J. A new neural network approach to short term load forecasting of electrical power systems.D.. 4. P. 12. Pearson Education Inc. Manicom-Ramsamy. Chang. 11. Georges. J. Eng. Power Syst. 8. 26.. 6. A generalized knowledge-based short-term loadforecasting technique. 42. J.. Lauret.N. Power Syst. Comput. A new prediction strategy for price spike forecasting of day-ahead electricity markets. Wamkeue. Lin. N. 3. S. Int. 2009. Feature extraction via multiresolution analysis for short-term load forecasting. S.E. 20. Energy 2001. 8. D. A. Power Energy Syst. W. IEEE Trans. N.. 33. Rahman. Jaramillo-Morán. 189–198. 488–503. Zhong. Amjady. IEEE Trans. Keynia. Haykin. Z.J. Y. Comput. Silva. Electr. 903–917. Int. 1990.-Y. Carmona-Fernández. González-Romera. Forecasting monthly electric energy consumption in Eastern Saudi Arabia using univariate time-series analysis.. USA. BP neural network with rough set for short term load forecasting.. Fock. 22. Beale. J. Amjady. Hesterberg. Bayesian neural network approach to short time load forecasting. Power Syst.T. 2006. Manag. IEEE Trans. 49. Analyzing the impact of climate change on future electricity demand in Thailand.. An efficient approach for short term load forecasting using artificial neural networks. 54.. J. 2009.. 2009. Power Syst. Monthly electricity demand forecasting based on a weighted evolving fuzzy neural network approach.Energies 2011. MA. G.. Data Anal. 1993. 23. Ind. A. IEEE Trans. 15. E.E. G.C. S-Y. 9. F. 2.H. Energies 2011.. Energy Convers. Reis.-Y. R. 11. 14.. 2003. O. IEEE Trans. 7. 306–318. Lai.. 1946–1953. Abdel-Aal.B. S. H.

Energy Convers. 23. Ramachandran. Eng.0/). 49. M. Ind. 288–294. D. 4 1507 20. licensee MDPI. Control problems of grey systems. E.Energies 2011. E. 22. Appl.I. Expert Syst.-L. © 2011 by the authors. González-Romera. J. Carmona-Fernández. Jaramillo-Morán. 1. 1982. 3135–3142. 24. Control Lett. 25. Incipient gear box fault diagnosis using discrete wavelet transform (DWT) for feature extraction and classification using artificial neural network (ANN). 37.. Comput. 52. Syst. 336–343. Manag.-L. M.org/licenses/by/3.. Saravanan. China Center for Economic Research (CCER). Forecasting of the electric energy demand trend and monthly fluctuation with neural networks.. . K.com (accessed on 30 April 2011).. Jaramillo-Morán. Grey Forecasting and Decision.ccerdata. China. Deng. Switzerland.: Hangzhou. Carmona-Fernández. Deng. 21. 2007. Available online: http://www.A. D. N. J. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.. Chinese Economic and Financial Database 2011. 2008. 2002. 4168–4181. Monthly electric energy demand forecasting with neural networks and Fourier series.A. Basel. Huazhong University of Science & Technology Press Co. 2010. González-Romera.