Indonesian J our nal of Electrical Engineering and Computer Science V ol. 20, No. 2, No v ember 2020, pp. 1000 1006 ISSN: 2502-4752, DOI: 10.11591/ijeecs.v20i2.pp1000-1006 r 1000 F or ecasting f or smart ener gy: An accurate and efficient negati v e binomial additi v e model Y ousef-A wwad Daraghmi 1 , Eman TT aser Daraghmi 2 , Motaz Daadoo 3 , Samer Alsaadi 4 1,3,4 Colle ge of Engineering and T echnology , P alestine T echnical Uni v ersity - Kadoorie, P alestine 2 Departent of Applied Computing, P alestine T echnical Uni v ersity - Kadoorie, P alestine Article Inf o Article history: Recei v ed Feb 15, 2020 Re vised Apr 17, 2020 Accepted Apr 31, 2020 K eyw ords: Ne g ati v e binomial additi v e models Nonlinearity Ov erdispersion Seasonal patterns Short-term load forecasting Smart ener gy T emporal autocorrelation ABSTRA CT Smart ener gy requires accurate and ef ficient short-term electric load forecasting to enable ef ficient ener gy management and acti v e real-time po wer control. F orecasting accurac y is influenced by the characteristics of electrical load particularly o v erdisper - sion, nonlinearity , autocorrelation and seasonal patterns. Although se v eral fundamen- tal forecasting methods ha v e been proposed, accura te and ef ficient forecasting meth- ods that can consider all electric load characteris tics are still needed. Therefore, we propose a no v el model for short-term electric load forecasting. The model adopts the ne g ati v e binomial additi v e models (NB AM) for handling o v erdispersion and capturing the nonlinearity of electric load. T o address the seasonality , the daily load pattern is classified into high, moderate, and lo w seasons, and the autocorrelation of load is mod- eled separa tely in each season. W e also consider the ef ficienc y of forecasting since the NB AM captures the beha vior of predictors by smooth functions that are estimated via a scoring algorithm which has lo w computational demand. The proposed NB AM is applied t o real-w orld data set from Jericho city , and its accurac y and ef ficienc y outper - form those of the other models used in this conte xt. Copyright c 2020 Insitute of Advanced Engineeering and Science . All rights r eserved. Corresponding A uthor: Y ousef-A ww ad Daraghmi, Colle ge of Engineering and T echnology , P alestine T echnical Uni v ersity-Kadoorie, P alestine. Email: y .a ww ad@ptuk.edu.ps 1. INTR ODUCTION Smart ener gy requires short-term forecasting for predicting the future load se v eral hours ahead and for e v aluating control strate gies before putting them in use [1]. Thus, accurate and ef ficient f o r ecasting is required to enable ef fecti v e and timely decision making process. The accurac y of the load forecasting models is af fected by electric load characteristics such as nonlinearity [2-5] high fluctuations [5, 6], autocorrelation and seasonal patterns [5, 7, 8]. The high fluctuation in electric load causes o v erdispersi o n e xpressed by a lar ge v ariance v alue that is clearly lar ger than the mean [9]. Research states that electric load fluctuation depends on the life style, day time and location [10, 11]. Ov erdispersion reduces the accurac y if it is not handled properly , particularly in the short-term forecasting, because forecasting methods are sensiti v e to the fluctuation [12]. In the meantime, the ef ficienc y of the forecasting methods is af fected by the data size, and model’ s time comple xity that deter - mines the time needed for an algorithm to produce accurate results [9]. Although se v eral short term forecasting models ha v e been proposed, o v erdispersion has not been appropriately handled i n these models. The statistical forecasting models such as re gression based models and smoothing based models assume that the v ariance is equal to the mean [4, 7, 13], or the y utilize distrib utions that ignore the high v ariations [14, 15]. In other models, the high v ariance in time series is treated to become homogeneous and stationary by Box-Cox and dif ferentiating transformations which increase the computational demand [7]. In the artificial int elligence based models (e.g. [11, 16, 17]), considering the high v ariation costs J ournal homepage: http://ijeecs.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 r 1001 much computational time [18]. Therefore, this paper focuses on the problem of o v erdispersion, nonlinearity and temporal autocorrelation to increase the accurac y of forecasting models. Simultaneously , the paper aims to reduce the computational demand to increase the ef ficienc y of the forecasting models. This paper proposes a temporal Ne g ati v e Binomial Additi v e Model (NB AM) to handle the o v erdis- persion precisely . The temporal NB AM is nonparametric and capture nonlinearity via smoothing functions. T o address the autocorrelation and seasonality , the dai ly load pattern is classified into lo w , moderate and high seasons in which the autocorrelation can be modeled separately . The proposed model is ef ficient because the method utilizes a lo w-comple xity optimization technique to estimate the smoothing function. The proposed model is tested on r eal-w orld data set collected from Jericho city - P alestine, and its results are compared with other classical load forecasting models including ARIMA, ARMA, Holt-W inters (HW) and ne g ati v e binomial linear model. The results sho w that the proposed temporal NB AM is more accurate than other models because its mean absolute percentage error (MAPE) is lo wer than the others. Also, the NB AM is more ef ficient be- cause the computational time needed for training and forecasting is lo wer than the time of the other models. The ne g ati v e bi nomial based models were successfully applied to forecasting traf fic data that is autocorrelated, o v erdispersed and ha v e seasonal patterns [19-22]. 2. RESEARCH METHOD This section proposes a method for forecasting electrical load which is nonlinear , o v erdispersed, and temporally autocorrelated. The proposed method adopts the Generalized Additi v e Models for handling non- linearity and Ne g ati v e Binomial for handling o v erdispersion, and then de v elops a temporal model for handling the temporal autocorrelation. 2.1. Generalized additi v e models (GAMs) The GAMs focus on capturing the nonlinearity by permitting the correlating v ariables to ha v e non- linear relationship [23]. In GAMs, a response v ariable Y has a mean v alue ( ) that is assumed dependent on predictors X via a function which is nonlinear . The importance of the GAMs is generalizing the con v entional additi v e models so that the response v ariable can fit into an y distrib uti on from the e xponential f amily [23, 24]. Consequently , an y GAM allo ws for more fle xibility than parametric-based models. The mean of Y , which is = E ( Y j X = x ) can be link ed to some predictors by G ( ) = s 0 + s ( x i ) + " i (1) where G indicates that the correlation is controlled by a link function, and s 0 is the o v erall mean or the intercept of Y , x i is the i th records in a data set, s ( x i ) is a smooth function for the i th record of the predict or X , " i is the error which is i ndependent of the predictors, and i N I D (0 ; 2 ) , and 2 is the v ariance [23, 24]. The importance of the smooth function is that it can beha v e as the original data and it can catch nonlinearity . Smoothers est imate the smooth function by fitting the data of a predictor through de v eloping a continuous curv e that consists of multipl e sections combined by knots [23]. Each curv e contains a total number of sections z , and each section can ha v e an equation such as the linear re gression equation or the cubic polynomial equation. Each equation contains a base function and coef ficients. Th e base function is used to generate the model matrix X n z , and the coef ficients form the parameter matrix M z 1 . Details on estimating the base functions and the model matrix X are in [24]. The smoot h function has a de gree of smoothness parameter that decides the number of sections ( z ) . is estimated in cross v alidation process such that the mean square error is minimized [23]. In GAMs, a local scori ng algorithm is used to computationally esti mate the smooth functions by maximizes a lik elihood function as stated in [23, 24]. The scoring algorithm is chosen carefully so that the computational time is v ery lo w . Also, the generalized cross-sectional v alidation is used for a v oiding o v erfitting of data [24]. 2.2. Negati v e binomial additi v e model (NB AM) The NB AM is a special case of the GAMs to o v ercome its main limitation which is only modeling data of e xponential distrib ution [25]. Besides ha ving nonlinear dependenc y , the response v ariable and the predictors may follo w other distrib utions. F or e xample, if o v erdispersion is found in the data, i.e. the v ariance 2 > , the best distrib ution for describing that data will be the Ne g ati v e Binomial [12]. In f act, a NB AM e xtends the GAM to treat the o v erdispersion by allo wing Y to ha v e Ne g ati v e Binomial distrib ution [25]. The Ne g ati v e binomial distrib ution can accounts for the high fluctuations in the data by permitting the v ariance 2 of Y to be F or ecasting for smart ener gy: An accur ate and ef ficient ne gative ... (Y ousef-A wwad Dar a ghmi) Evaluation Warning : The document was created with Spire.PDF for Python.
1002 r ISSN: 2502-4752 greater than the mean as 2 = + ' 2 where ' is the o v erdispersion parameter [12]. The ne g ati v e binomial model for Y gi v en predictors X is additi v ely fit to data by choosing a natural log link function. The NB AM is written as log E ( ) = s 0 + s ( x i ) + " i (2) where a local scoring algorithm is used to train the smooth function iterati v ely by estimating ' and that maximize a log-lik elihood function e xplained in [25]. 2.3. De v elopment of temporal NB AM The NB AM can be applied to electric load by considering the temporal autocorrelation of the load. First, the electric load daily pattern is classified into three seasons, lo w , moderate and high seasons. In the data set sho wn in Figure 1, the seasons are: (1) a lo w load season from 02:00AM to 10:00AM (8 hours) and e xists in the early morning when the load is less than the daily mean; (2) a high load season from 06:00PM to 01:00AM when the load is bigger than the daily mean (8 hours); (3) a moderate load season between 01:00AM-02:00AM and 10:00AM-06:00PM(9 hours) when the load is around the mean. Each season will ha v e its o wn forecasting model. The benefit of this classification is allo wing training each season indi vidually , remo ving the outliers and decreasing the v ariability as the electric load is not fluctuating between the maximum load v alue and the minimum v alue during a single season. Second, the temporal correlation is modelled for each season. W e let y i;t be the response v ariable corresponding to the load measured from station i at time frame t . The temporal autocorrelation is considered by incorporating a one-step lag load of the dependent station as a predictor , i.e. y i;t 1 . Thus, the proposed temporal NB AM is e xpressed by ln y i;t = i + s i ( y i;t 1 ) + " t : (3) The lag v alues y i;t 1 accounts for the electric load persistence, and it re gularly updates the model with an y change of the data trend. Equation 3 is for accommodating the smooth function to data based on the ne g ati v e binomial distrib ution, and for computing the model matrix M and the parameter matrix P that are used in the forecasting process. The estimation be gins by selecting a dependent station, selecting the season based on the time, select- ing the data belonging to the same season from historical data, and then starting the estimation of the smooth function of load for that station. The R mg cv package introduced in [24] is used for performing the analyses. This R package contains all libraries needed for smooth function estimation and forecast results generation. F orecasting by GAM models requires a data set (data for the season of interest), a model matrix M and a model parameter matrix P which are computed in the training stage [24]. Figure 1. Single-day electric load and an e xample of the smooth function for the load in Aqabat Jaber station 2.4. Data set A data set consisting of electrical current measured e v ery hour from 15 stations w as collected in Jericho city in the period between January to December in 2015. The total number of records collected from Indonesian J Elec Eng & Comp Sci, V ol. 20, No. 2, No v ember 2020 : 1000 1006 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 r 1003 each st ation is 4380 records. Figure 2 illustrates the electric load pattern for 17 consequent days. Before the analyses start, the in v alid records such as missing v alues, ne g ati v e v alues and zero v alues in the data set were treated. The number of the these records is less than (2%) per day which is v ery small and does not reduce the forecasting accurac y . Figure 2. The load o v er 17 days from 13/No v/2015 to 29/No v/2015 in Aqabat Jaber station The electric loads of the selected st attions during 17 days are represented by time series. Figure 2 illustrates an e xample of the load time serie s during 17 days, while Figure 1 illustrates the load time series during a single day . The figures demonstrate that the load data ha v e c yclic seasonal patterns that are repeated e v ery day . Also,tThe electric load illustrated in Figure 2 and 1 is temporally autocorrelated, that is, the v alue of the load at an y moment depends on the pre vious one. The electric load has o v erdispersion because the v ariance of the load at each station is lar ger than the mean as sho wn in T able 1. The table sho ws the mean, the v ariance and the dispersion v alues for three stations during t he lo w , moderate and high seasons. Ov erdispersion e xists in the data if the dispersion v alue which is ”the Pearson statis tic ( 2 ) di vided by the de grees of freedom is lar ger than one” [12]. In the data set, all loads are o v erdispersed as the corresponding dispersion v alues sho wn in T able 1 are lar ger than one. The aforementioned characteristics are common among electrical netw orks loca lly and globally . The o v erdispersion of the load refers to the v ariations in the consumption of electric ener gy within the seasonal daily patterns. The v ariations depend on the place in the city (urban, industry , mark et), changing conditions of weather , time of the day , and life style. T able 1. The statistical measures of four stations during the four load seasons Area Aqabat Jaber Sea le v el 1 Sea le v el 2 Sea le v el 3 for high season 153.0 149.2 87.3 205.9 2 for high season 168.1 181.4 127.7 327.5 dispersion for high season 2.4 2.3 2.5 2.8 for moderate season 125.7 127.6 68.3 189.3 2 for moderate season 143.1 155.3 98.0 242.6 dispersion for moderate season 2.7 2.8 1.9 1.9 for lo w season 101.8 97.8 46.1 151.4 2 for lo w season 117.2 152.6 77.3 197.0 dispersion for lo w season 1.9 2.5 1.7 2.0 3. RESUL T AND DISCUSSION This section sho ws the results of the proposed method whic h is the temporal NB AM. The s ection firstly sho ws ho w the method is trained, and then sho ws its accurac y and ef ficienc y . F or ecasting for smart ener gy: An accur ate and ef ficient ne gative ... (Y ousef-A wwad Dar a ghmi) Evaluation Warning : The document was created with Spire.PDF for Python.
1004 r ISSN: 2502-4752 3.1. T raining of the pr oposed temporal NB AM The proposed NB AM is trained for the four stations in the data set, and ten-months data collected from January-2015 to October -2015 are used. The proposed model is trained separately for each load season such that the complete data set is di vided to three data sets including a set for daily lo w season of eight hours length, a set for daily moderate season of se v en hours length, and a set for daily high season of nine hours length. Therefore, each load season will ha v e a dif ferent forecasting model. The result of the training process is described in T able 2 and Figure 1. As sho wn in T able 2, the three load seasons ha v e significant load autocorrelation because the P- v alue is smaller than the significance le v el. In our analyses, the statistical confidence interv al is 95% and the significance le v el is 5%, similar to [10]. T able 2 sho ws the main statistical output of the NB AM when Aqabat Jaber is the response v ariable. The use of the Ne g ati v e Binomial is justified because the v alues of ' are greater than zero as sho wn in T able 2. Also, the smooth function for the data in Aqabat Jaber station is plotted in Figure 1. The NB AM describes the real data with tin y dif ferences which moti v ates us to use this model for forecasting. T able 2. The main outputs of the NB AMs for the Aqabat Jaber station lo w season Moderate season high season autocorrelation P-v alue 7 10 14 2 10 14 5 10 14 Intercept ( ) 3.45 4.57 2.39 o v erdipsersion ' 3.63 2.33 3.18 data size n 2432 2128 2736 z 41 38 33 size of M 2432 41 2128 38 2736 33 3.2. F or ecasting accuracy The NB AM utilizes the R mg cv package mainly the pr edict function.Tthe proposed NB AM is e v alu- ated during each load season using the models generated in the training stage. The smooth functions obtained from the training stage based on the coef ficients sho wn in T able 2 is used to produce the Aqabat Jaber forecasts for multiple steps ahead, ranging from one hour to 24 hours. The models used for comparison are trained as in [9]. These models are classical forecasting models and the y are widely used in in this field. The models are: the temporal NBLM [9], the Holt-W int ers (HW) [26], Auto Re gressi v e Inte grated Mo ving A v erage (ARIMA) [14], and the Auto Re gressi v e Mo ving A v erage (ARMA) method as in [7]. A single season for the HW , the ARMA and the ARIMA models is used because the y are c yclic-based models and need a repeated pattern. The Mean Absolute Percentage Error (MAPE) is calculated to compare the accurac y of the models forecasting results, similar to [3, 10]. T able 3 sho ws the MAPE v alues of the v e models for the selected stations during the three load seasons for a ten-hours horizon. T o sho w the accurac y of the NB AM, the results of the models during the high load season are also presented in Figure 3. T able 3. The MAPE v alues of dif ferent models for three electric stations during the three load seasons load MAPE % MAPE % MAPE % season Method Aqabat Jaber sea le v el 1 seas le v el 2 NB AM 0.88 1.09 0.83 Lo w T emporal NBLM 1.51 1.72 1.39 Load HW 3.17 3.62 3.49 Season ARMA 2.36 3.12 2.97 ARIMA 6.27 5.09 4.24 NB AM 2.38 1.97 2.11 Moderate T emporal NBLM 3.57 1.95 2.77 Load HW 5.24 3.29 4.33 Season ARMA 4.77 2.83 3.91 ARIMA 6.79 4.65 5.18 NB AM 3.35 1.61 1.24 High T emporal NBLM 4.62 2.12 2.61 Load HW 5.87 3.68 4.36 Season ARMA 5.29 3.63 3.61 ARIMA 7.19 5.95 5.98 Indonesian J Elec Eng & Comp Sci, V ol. 20, No. 2, No v ember 2020 : 1000 1006 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 r 1005 Figure 3. The forecasting of load for 10-hours horizon during the high load season in Aqabat Jaber station The MAPE v alues of the NB AM are less than those of the temporal NBLM, the HW , the ARMA and the ARIMA which concludes that the NB AM is more accurate than the others. Also, Figure 3 emphasizes that the NB AM has better accurac y than the other models. The temporal NBLM handles the o v erdispersion b ut its structure contains a re gression which incorporates fe w correlating hours from the r ecent past. On the contrary , the NB AM captures the entir e trend and pattern because its structure contains a smooth function that beha v es e xactly the same as the real data. W e find that the smooth function of the NB AM performs better than the temporal correlation with the pre vious v e hours as in NBLM. Classifying the load into three dif ferent seasons, the v ariance is ensured to be lo wer than the v ariance of the entire daily season, as in T able 1. Thus, the forecast accurac y is increased. 3.3. F or ecasting efficiency T o measure the ef ficienc y of the proposed NB AM, the computational time during the training and forecasting stages is recorded. More focus is gi v en to the high load season since this season has the lar gest data size which is 2736. As in [9], the model is tested on Intel CPU of 2.8 GHz, 64-bit operating system and 16GB RAM. The NB AM is found f aster than the ot her models during training for producing the smooth functions, and during the forecasting, as presented in T able 4. The NB AM ef ficienc y also outperforms the other models during the forecasting and the training stages. The reason behind the lo w computational time is that the smooth function is optimized using re gression spli ne where each section coef ficie nts are comput ed by minimi zing the mean square error . T able 4. The computation time of the models during the training and forecasting stages Model T raining (sec) F orecasting (sec) NB AM 1.69 0.08 NBLM 2.124 0.26 ARMA 5.94 3.19 ARIMA 6.82 4.78 HW 6.38 0.86 4. CONCLUSION Nonlinearity , o v erdispersion, temporal autocorrelation and seasonality are important characteristics of El ectrical load. These characteristics ha v e dramatic ef fect on the forecasti ng accurac y and ef ficienc y , and the y requires careful handling during de v eloping a forecasting model. Therefore, this paper adopted the NB AM because of its ability to handle nonlinearity and o v erdispersion. A temporal NB AM w as deri v ed by allo wing the current elec trical load to autocorrelate with pre vious loads. Also, In the proposed model, the seasonal patterns of the electric load is classified into lo w , moderate and high load patterns. The future w ork will include applying the de v eloped NB AM to a real time electrical load monitoring system in P alestine. F or ecasting for smart ener gy: An accur ate and ef ficient ne gative ... (Y ousef-A wwad Dar a ghmi) Evaluation Warning : The document was created with Spire.PDF for Python.
1006 r ISSN: 2502-4752 REFERENCES [1] Z. Y asin, et al., An accurate medium-term load fore- casting based on h ybrid technique, Indonesian Journal of Electrical Engineering and Computer Science, v ol. 12, no. 1, pp. 161–167, 2018. [2] N. T . T ran, “Gri d search of e xponential smoothing method: a case study of ho chi minh city load demand, Indonesian Journal of Electrical Engineering and Computer Science, v ol. 19, no. 3, pp. 161–167, 2020. [3] R. E. Caraka, et al., “The step construction of penalized spline in electrical po wer load data, TELK OM- NIKA T elecommunication, Computing, Electronics and Control, v ol. 17, no. 2, pp. 1023–1031, 2019. [4] S. F an and R. J. Hyndman, “Short-term load forecasting based on a semi-parametric additi v e model, IEEE T ransactions on Po wer Systems, v ol. 27, no. 1, pp. 134–141, 2012. [5] F . L. Quilumba, et al., “Using smart meter data to impro v e the accurac y of intraday load forecasting considering customer beha vior similarities, IEEE T rans. on Smart Grid, v ol. 6, no. 2, pp. 911–918, 2015. [6] Y . W ang, Q. Chen, M. Sun, C. Kang, and Q. Xia, An ensemble forecasting method for the aggre g ated load with subprofiles, IEEE T ransactions on Smart Grid, v ol. 9, no. 4, pp. 3906– 3908, 2018. [7] K. G. Borooj eni, et al., A no v el multi-time-scale modeling for electric po wer demand forecasting: From short-term to medium-term horizon, Electric Po wer Systems Research, v ol. 142, pp. 58 73, 2017. [8] P . G. D. Silv a, D. Ili’c, and S. Karnousk os, “The impact of smart grid prosumer grouping on forecasting accurac y and its benefits for local electricity mark et trading, IEEE T ransactions on Smart Grid, v ol. 5, no. 1, pp. 402–410, 2014. [9] Y .-A. Daraghmi, et al., Accurate and time-ef ficient ne g ati v e binomial linear model for electric load fore- casting in ioe, T ransactions on Emer ging T elecommunications T echnologies, 2019. [10] B. Dha v al and A. Deshpande, “Short-term load forecasting with using multiple linear re gression, Inter - national Journal of Electrical and Computer Engineering, v ol. 10, no. 4, pp. 3911–3017, 2020. [11] A. Khw aja, X. Zhang, A. Anpalag an, and B. V enkatesh, “Boosted neural netw orks for impro v ed short- term electric load forecasting, Electric Po wer Systems Research, v ol. 143, pp. 431–437, 2017. [12] J. M. Hilbe, ”Ne g ati v e Binomial Re gression, 2nd ed. Cambridge: Cambridge Uni v ersity Press, 2011. [13] J. W . T aylor and P . E. McSharry , “Short-term load forecasting methods: An e v aluation based on european data, IEEE T ransactions on Po wer Systems, v ol. 22, no. 4, pp. 2213–2219, 2007. [14] I. Mado1, A. Soeprijanto, and Suhartono, Applying of double seasonal arima model for electrical po wer demand forecasting at pt. pln gresik indonesia, International Journal of Electrical and Computer Engi- neering, v ol. 8, no. 6, pp. 4892–4901, 2018. [15] K. B. Sahay and M. M. T ripathi, “Day ahead hourly load forecast of pjm electricity mark et and ISO ne w england mark et by using artificial neural netw ork, IEEE Inno v ati v e Smart Grid T echnologies-Asia (ISGT Asia), pp. 1–5, 2013. [16] L. Is mael and A. I. Ismael, “Prediction the data consumption f o r po wer de mandsby elman neural net- w ork, International Journal of Electrical and Computer Engineering, v ol. 9, no. 5, pp. 4003–4009, 2019. [17] R. K. V and P . Dixit, “Daily peak load forecast using artificial neural netw ork, International Journal of Electrical and Computer Engineering, v ol. 9, no. 4, pp. 2256–2263, 2019. [18] A. Bertoni and B. P alano, “Structural comple xity and neural netw orks, Neural Nets, ser . Lecture Notes in Computer Science, v ol. 2486, pp. 190–216, 2002. [19] Y .-A. Daraghmi, C.-W . Y i, and T .-C. Chiang, “Space-time multi v ariate ne g ati v e binomial re gression for urban short-term traf fic v olume prediction, 12th Int. Conference on ITS T elecommunications, 2012. [20] Y .-A. Daraghmi, C.-W . Y i, and T .-C. Chiang, “Ne g ati v e binomial additi v e models for short- term traf fic flo w forecasting in urban areas, IEEE T ra n s . on Intel. T ransport. Sys., v ol. 15, no. 2, pp. 784–793, 2014. [21] Y .-A. Daraghmi, C.-W . Y i, and T .-C. Chiang, “Mining o v erdispersed and autocorrelated v e- hicular traf fic v olume, 5th International Conference on Computer Science and Information T echnology , 2013. [22] Y .-A. Daraghmi and M. Daadoo, “Impro v ed dynamic route guidance based on holt-winters-taylor method for traf fic flo w prediction, Int. J. of T raf fic and T ransport. Enginee., v ol. 4, no. 3, pp. 83–83, 2015. [23] T . Hastie and R. T ibshirani, “Generalized additi v e models (with discussion), Statistical Science, v ol. 1, no. 3, p. 297–318, 1986. [24] S. N. W ood, ”Generalized Additi v e Models An Introduction with R. Chapman and Hall/CRC, 2006. [25] S. W . Thurston, M. P . W and, and J. K. W ienck e, “Ne g ati v e binomial additi v e models, Biomet- rics, v ol. 56, pp. 139–144, 2000. [26] J. W . T aylor , “T riple seasonal methods for short-term electricity demand forecasting, European Journal of Operational Research, v ol. 204, no. 1, pp. 139–152, 2010. Indonesian J Elec Eng & Comp Sci, V ol. 20, No. 2, No v ember 2020 : 1000 1006 Evaluation Warning : The document was created with Spire.PDF for Python.