2004 IEEElPES Transmission & Distribution Conference 4% Exposition: Latin America Impact of Exogenous Variables on Estimated Values of Demand and Energy Jorge Yasuoka, Reynaldo A. Souza Jr, JosC A. Jardini, Roberto Castro, Fernando A. A. Prado, Liliana M. V. Brittes, Andre L. P. Cruz, Hemin P. Schmidt Abstrocr- Traditional methods for estimating future values of demand and energy do not normally take into account the effect of the so-cailed exogenous variables, which include Ioad geographical location, seasonal variations, availability restrictions of energy and summer time schedules. This work proposes a methodology for assessing the impact of these external variables on the estimation of future values of demand and energy, with a view to improving current practices which dictate demand and energy purchases in both the short term and the medium term. The load curve is broken into components associated with each one of the exogenous variables. The future behavior of each component is estimated through Artificial Neural Networks and the estimated global curve is obtained by aggregating the various components back together. Index Terms- Neural networks, System Load Forecasting.. INTRODUCTION HE exactness of the demand and of the energy forecast is T fundamental for the business success of an Electric Utility. The new organizational structure of the utilities and its entrance into a merchandizing scenario that integrates the new rules yet to be defined, the competition and the entrance of Distributed Generation are some of the new items that make the already complex task of load forecast an immense challenge. The huge amount of the total purchase and sale operations of energy for the utilities is in the order of some billions of Dollars per annum. To develop methodologies that imply in the reduction of the risk of these operations and in the improvement of its accuracy is extremely desirable and necessary, optimizing costs in the planning processes of the expansion, of the operation, and in the bid-offer operations of energy. Iri general, the short and medium term forecast problem has been approached according to two traditional methodologies: temporary series and regression analysis. These conventional computational techniques may not present results sufficiently precise, In a similar way, computational methods based on complex algorithms demand a heavy processing task, and thus, they can converge stowly and in certain cases even diverge. The approach through temporary series normally does J. A. Jardini, F. A. A, Prado, H. P. Schmidt and J. Yasuoka are with Escola PolitCcnica da Universidade de SHo Paulo, Brazit. (e-mail: j.jardini@ieee.org). R. Castro, L. M. V. Brittes and A. L. P. Cnn are with CPFL - Companhia Piratininga de Forp e tuz, Brazil. (e-mail: castro@cpfl.com.br). R. A. Sow Jr. is with Expertise Engenharia, Brazil. (e-mail: repaldo@expertise-eng.com.br) not take into account the climatic information (ambient temperature, humidity, insolation, etc.) that constitutes an important parameter, especially when referring to residential loads. Consequently, problems resulting from little precision and numeric instability of the results are some serious problems observed in this technique. The objective of this model is to recognize a pattern inside the values of historical series and to extrapolate this in the future. On the other hand, the regression analysis tries to associate the climatic variables with the system load. In this case, a fixed hctional relationship among these variables has to be established beforehand, and usually a linear model is adopted. The objective of this model is to discover the form of this relationship and to use it in the forecast of the dependent variables utilizing the future evolution of the independent variables (e.g.: simple and multiple regression). However, as the functional relationship itself varies through time, this approach frequently produces unsatisfactory results. To bypass such obstacles, this article presents a work whose objective is to study the influence of the exogenous variables in the aspect of the demand and energy curve, and the use of Artificial Neural Networks (ANNs) technique to improve the forecast results. The forecast outlook considered in the study are: The daiiy demand curve forecast represented by288 values (5 minutes demand values) and 96 values (5 minutes demand values) for week ahead; The daily demand curve forecast represented by 96 values ( 5 minutes demand values) for month ahead;. The energy forecast of week to 2 months ahead.. ANALYSIS OF DEMAND CURVES For the studies, demand data extracted from CPFL historical data base were used, including a period from 997 to mid 2003. Special days, such as holidays, were not considered. Some of the studies made and the obtained results are described as follows. A. Clusterizution To improve the results of the demand and energy forecast, analyses were made grouping the substations in families according to the shape of the demand curve. Because the system of this utility is comprised of approximately 300 substations, typical demand curves from each substation were compared with the intention of obtaining some typical curves that would 0-7803-8775804/$20.00 02004 leee 344
I 2 be used as a basis for the load forecast for all other substations. By seeking a leveling in the amplitude of the demand curve, curves in 'put were used in the analyses in relation to the daily average energy; in other words, the ratio of demand data (kw) to the average daily Energy (kwh). Firstly, 75 substations were tested, whose had data were inserted in a clusterization software (SAS), that groups the load curves in different families according to statistical models. The results demonstrate that it is possible to identify typical curves for each family (that resemble to the average curve of each group). Ten different clusters were obtained and one of them is shown in figure. In the analysis of the clusters, the discrimination of substations with profiles predominantly residential, commercial and industrial was verified. It was noticed that the substations of the same cluster were not necessarily located geographically close to each other. parison, it was verified that it is necessary to make the forecast for the days of DSTperiod separately. Hence it was verified that there is great similarity in the aspect of the demand curves of the substations out of the peak time, and that the differences lie basically in the peak time, mainly in substations with predominance of industriallcommercial profile. Figures 4 and 5 show the cases of two substations located in the state of Siio Paulo. I 2 I.6 2 a 0.0 CLVSTERI-SAS 7 l0... Z I l Fig.. One of the clusters calculated by the software. The enhanced curve is the cluster curves average and the dotted curve consists of the mean deviation of the same. B. Pre andposf-rationirtgperiod Comparison (200) As it is known, the rationing of energy which occurred in Brazil, in 200, promoted several changes in the energy sector, mainly in the value and in the form of the demand curve. The inherent alterations of the rationing were studied, comparing previous and subsequent data to 200, based on the load data of the main substations of the CPFL electrical system. The results show that the tendency was that the load shape maintained in most of the analyzed places, signaling that the data previous to 200 can be used, provided that only the shape of the curve is considered; in other words, analyzing the curves in pu, it is noticed that the shape of the curve was not altered even after the rationing period. The results of two substations are in the figures 2 and 3. For each year (Apnlll997 to July/2003) the Mondays average of each month was made. C. Daylight Saving Time (DSr) The demand and energy data in periods in which the Daylight Saving Time (DST) is in force are different in relation to the other periods of the year, To verify this, analyses of the curves of main CPFL's substations were made, aiming at this com- 0.4 I 0.2.i 0 rrrrrrrrr~" Fig. 3. Monthly average demand curve @U) of a substation located in the northeast region of the state of SBo Paula....... I.6 I I.4 h I 0.4 I - l Fig. 4. Demand curve in pu of a substation from central region of the state of Si0 Paulo. The curve in continuous line is the average of the demands of the Mondays for the months of December and January, and the dotted curve is relative to the months of April and May. 345
.4.2 - Fig, 5. Average demand curve @U) of a substation in the city of Campinas in the years of 2002 and 2003, highlighting the differentiation in the consumption in the Daylight Saving Time @SZJ (dotted curve). A predominant midenrial profile is noticed.. MAPPING THE EXOGENOUS VARIABLES The relationship between the electric charge and its exogenous factors is complex and non-linear, hindering its modeling through linear mappings. Not only does it lack the necessary precision, but also the traditional statistical techniques are not sufficiently robust. ANNs can represent patterns that are not detectable by traditional statistical methods that utilize auto-correlation coeficients or crossed correlation, or even the power spectrum. The ability of the ANN in mapping complex non-linear relations has been responsible for the growing number of applications in load forecasting. Besides the data of the load themselves, the inclusion of ambient temperature, relative air humidity and insolation (incident radiation) was analyzed as causal variables for the load forecasts. It is noticed intuitively that these variables influence the consumption of energy, although the mechanisms of such impact are unknown. A pre-analysis was made only to detect possible correlations between the load (kwh) and those exogenous factors. If, for a variable, this qualitative analysis had revealed some level of correlation (linear or not) with the load, it would mean that this variable could be considered to aid the forecast. The climatic data, available every 0 minutes, were supplied by an automatic collection station of climatic data located in the city of Campinas, Sgo Paulo. The first variable analyzed was the average temperature TMm defined as: 0 5 20 25 3( Fig. 6. Adjusted straight line for energy in relation to the temperature. By the determination coefficient calculated in the linear modeling, it is noticed that approximately 60% of the energy variance is explained by the variation of the average temperature. The other studied variable was the relative air humidity. Only this item presents very low correlation with the consumed load, as revealed by several types of regression curves adjustments. However, when associated to the temperature, there is the effect of thermal comfort, which occurs because of the air saturation through humidity, that associated to thermal effect, hinders the exchange of heat between the human body and the atmosphere. This effect occurs through the Index of Humidity Temperature Discomfort (ITU)[, that relates ambient temperature to the relative air humidity according to the equation: where T is the ambient temperature in Celsius ("C) and UR is the relative air humidity in percentage ("h). Therefore, it is possible to check the effect of the temperature and relative humidity to, if proved its influence on the consumption of energy, consider it as input in the load forecast. The same data that served for analysis of the temperature were also used for analysis of thr influence of ITU on the load behavior, as shown in figure 7. 6 5 ENERGIAx TTU m m being T- and Tm the maximum and the minimum daily temperature, respectively. Through an analysis of simple linear regression a high correlation level was verified between this variable and the daily energy of a certain place, especially in the working days of the week. Thus, it was defined that the average temperature would be taken into account in the load forecast. Figure 6 shows the dispersion diagram and the calculated regression line, along with the determination coefficient for a Friday. F? =0.587 0 2 la I8 in 20 a 2A ab 28 Fig. 7. Line adjusted for energy in function of daily ITU for Fridays. The determination coefficient calculated far this case was smaller. Note: The ITU was calculated in function of the minimum daily relative humidity and TAvvo. In general, there was a good linear correlation between ITU and the daily energy, although smaller in the case of TAVG. For this variable, the average of the determination coefficients obtained for the working days was 0,5554, and in the weekends it was 0.2053. Now in the case of ITU, the average of the J 346
~ 4 working days was 0.5093 and in the weekends it was 0.936, From these results, it was concluded that the relative air humidity does not significantly influence in the consumption of energy. Therefore, the relative air humidity will not be taken into account in the load forecast. Finally, an evaluation of the influence of direct insolation in the consumption of energy was made. The variable RADI, which consists of the daily average of the incident radiation measured in a certain place, was defined. Preliminary tests showed that this variable has irrelevant impact on the consumption of energy. Statistical analyses based on multiple regressions were carried out, considering not only the RAD variable, but also the humidity temperature index. One of the dispersion diagrams obtained can be seen in figure 8. 3 I 65 I ENERCUX~,RADI R2=02502! j 2 4 r6 ra m zz 2 2a SI Fig. 8. Line adjusted for Saturdays of the daily Energy in relation to the daily TAVG and daily average of incident radiation. This variable has slightly higher influence on the weekends. Comparing with the previous case, it was verified that there is practically no improvement with the insertion of the RAD variable, either on the weekends or the working days. Therefore, it was decided that this last item will not be considered as a causal variable in the load forecast. Consequently, the only climatic variable to be considered for forecast was the ambient temperature. However, other analyses showed that in hotter and not rainy days the insolation and relative humidity, added to the temperature, presented good correlation with the load, even higher than the cases obtained with TAVG only. This shows that a more complex relation can exist between these two variables. IV. NEXTSTEPS Complementing the analyses of exogenous variables described previously, studies are being carried out with econometric variables such as, Gross Domestic Product (GDP), mass income, tendencies for each consumer class, etc. Moreover, the analyses shown were made at substation level, and that will also be extended to the consumer classes, with the purpose of evaluating the behavior of each class in the amount of the supplied energy which will also be made. At the same time, the work of forecast of the curves using ANNs is being carried out, taking into consideration the results obtained until now. Figure 9 shows the result obtained for the forecast of the last Monday of the month of April 2000. The mean absolute percentage error (MAPE) is 3.07%, which is quite satisfactory but still liable for improvement..6 0.4 - ABRlL - zoo0 ERM-3.07% 0..... 6 6 2 26 3 36 4 40 5 56 6 00 7 76 B 06 9 96 Fig. 9. Forecast accomplished for the last Monday of the month of April 2000. The dotted curve is the one calculated and the continuous curve is the measured one. V. CONCLUSION The article presented a work that has the purpose of developing a methodology for demand and energy forecast for more accurate processes than the current models based on typical variables and forecast techniques, with application in the global forecast of the load in the outlook of short and medium terms, directed to the bid-offer operations of energy. The load forecast is of extreme importance in this area, and an improvement in the results implicates in expressive gains for the Electric Utility. The analyses of the exogenous variables prior to the execution of the forecast, contribute greatly in the results of the demand and energy forecasts and the application of more robust techniques (such as ANNs), in substitution to the traditional techniques based on statistical methods allows to treat variables that these last ones do not, such as the climatic variables. [I] [Z] VI. REFERENCES WALLACE, J.M. e P.V. HOBBS, Atmospheric Science: an introductory survey. Academic Press, 977. LUTGENS, F.K. e E.J. TARBUCK, The Atmosphere: an introduction to Metorology. Prentice Hall, 989. VII. BIOGRAPHIES JosC Antonio Jardini (M'66-SM'78-F'90) was born in Siio Paulo, Brazil, on March 27th, 94. He graduated from Escola Politkcnica da Universidade de,520 Paulo, Brazil, in 963 (Electrical Engineering). From the same institution he received the MSc, PhD, Associate Professor and Head Professor degrees in 97, 973, 99 and 999, respectively. For 25 years he worked for Themag Engenharia Ltda., a leading consulting company in Brazil, where he conducted many power systems studies and participated in major power system projects such as the Itaipu hydro plant. He is currently Full Professor at Escola Politicnica da Universidade de S b Paulo, where he teaches power system analysis and digital automation. He represented Brazil at SC-38 of CIGRE and was a Distinguished Lecturer of IASflEEE. Fernando Amaral de Almelda Prado Jr, was bom in Araqatuba, Brasil on 5th June 955. He graduated from Campinas State University, UNI- CAMP, 977 (Civil Engeneering) From the same university he received the MSc (994) and PhD (999) in Energetic System Planning. For 24 years he worked at CESP, Energy Secretariat and Public Utility Comission in S b Paulo State, where used to be Chief Comissioner (999-200). He is currently researcher at Escota Politecnica da Universidade de S5o Paulo and manager of Sinerconsult Consultoria, Treinamento e ParticipaGlo Ltda. j 347
Jorge Yasuoka (5 04) was bom in Guarulhos, Brazil, on July 4th, 974. He received the BSc. and MSc. degrees in Electrical Engineering from Escola Polithica da Universidade de Si0 Paulo, Brazil, in 999 and 2002, respectively. He is currently working towards the PhD. degree at the m e institution. Since 997 he has been working in the field of load forecast and Artificial NeuraI Networks. His research interests include the application of intelligent system in planning and operational problem in electric power systems. Reyddo Ayres de Souza Jr. received the ESc. degree in Electrical Engineering from Universidade Estadual de Campinas QJNICAMP), Brazil, in 200. He is currently researcher at Expertise Engenharia. His research interests include the load forecasting and power quality in planning and operational problems in electric power systems. Roberto Castro, was barn in Presidente Bemardes, Brazil, on December ZZnd, 960. He received the BSc degree from Escola de Engenharia Mad - IMT, Brazil, in 984, and MSc. degree in Electrical Engineering from Universidade Estadual de Campinas (UNICAMP), Brazil, in 994. For 3 years, he worked for CESP - Cia Energdtica de S2o Paulo, Brasil, in the planning of generation expansion and definition of supply contracts. He worked for 3 years for Elektro - Eleiricidade e Servifos S.A., Brazil, where he used to be supervisor of Wholesale Electric Energy Market and Energy Supply Area. He is currently working at the CPFL group, a Brazilian electricity utility, where he rriamges Energy Planning and Energy Supply Contracts. Liliana de Mattos Venegas Britt-, was Bom in Ribeido Preto, Brazil, on July 4th, 966. She received the BSc. degree from Universidade de Sao Paulo - Tnstituto de Citncias MatemDicas e de ComputagBo, Brazil, in 987. Since 997 she has been working for CPFL, in the fieid of Tariff Planning and Energy Planning. She is currently working on energy trade planning for the CPFL group. Andre Lufz Preite Crw, was born in Silo Pado, Brazil, on January 23rd, 978. He graduated in Electrical Engineering from Universihde Paulista, Braril, in 200. Atualmente mestrando em Planejamento Energetic0 pela UNICAMP. For 2 years he worked at ASMAE - Administradota do Mercado Atacadjsta de Energia. He is currently working at the CPFL group, a Brazilian electricity utility, in the field of Energy Trade Planning. Hemin Prieto Schmidt was bom in Montevideo, Uruguay, on March 6th, 960. He received the BSc and MSc degrees in Electrical Engineering from Escola Politknica da Universidade de SBo Paulo in 982 and 989, respectively. In 994 he received the PhD degree in Electrical Engineering from the University of London, UK. Between 98 and 990 he worked for E. J. Robba Consultoria & Cia. Ltda., a leading consulting company in Si% Paulo, where he developed computational models for power system analysis concerning composite generatiodtransmission reliability evaluation, automatic VAr sizing and location, and electrical and thermal analysis of underground power cables, Since 985 he has been with Escola Politknica da Universidade de SZo Paulo, where he teaches power system basics. His current research interests include the application of artificial neural networks in distribution systems and the developmnt of GIs-based tools for the operation of electrical power systems. 348