T”ixti0ns on Power Systems, Vol. zyxwvutsrqp 6. No. zyxwvutsrqpo 4, November 1991 zyxwvutsrqp A PRIORITY VECTOR BASED TECHNIQUE FOR LOAD FORECASTING Saifur Rahman Govinda Shrestha Senior Member Member Energy Systems Research Laboratory Electrical Engineering Department Virginia Polytechnic Institute and State University Blacksburg, VA 24061. U.S.A. ABSTRACT A new short term load forecasting technique is presented which utilizes the attractive features of both the statistical and expert system based methods, but avoids their drawbacks. The priority vector based load forecasting technique uses pairwise comparisons to extract relationships from pre-sorted historical hourly load and weather records for up to two years. The pre-sorting is done to identify seasonal boundaries and to categorize the day types (weekdays, weekends, holidays, etc.). The technique is adaptive in the sense that it internally generates the coefficients of relationships among the governing variables (i.e., weather parameters) and the load. As these relationships change over lime, such coefficients are automatically updated. The resulting linear method is robust and fairly accurate. This technique has been applied to forecast the hourly loads for a week, using 168-hour lead time, in different seasons. The only forecast variable used in this study is the dry-bulb temperature. When tested for the historical data in the service area of a Virginia electric utility for four weeks (in different seasons of the year), the average forecast error remained mostly under 4%. Further. only 23 individual errors. out of a total of 672 cases, exceeded zyxwvutsrqpon 6%. Keywords: Priority vector. Eigenvector, Pairwise comparison, Load-temperalure relationships, Load forecast. Expert systems, Self-learning. INTRODUCTION Short term load forecast is a necessary element for the eleclric utility operation. One-hour to 168-hour load forecast is needed for economic dispatch, unit commitment, energy purchase-sale decisions, load management, and hydro-thermal scheduling including pumped-storage operations. Elaborate reviews of short term load 91 WM 1 9 8 - 2 PWRS by the IEEE Power System Engineering Committee o f the IEEE Power Engineering Society for presentation at t h e IEEE/PES 1991 Winter Meeting, New York, New York, February 3-7, 1991. Manuscript submitted January 31, 1990; made available for printing January 3, 1991. A paper recommended and approved 1459 forecasting is reported by GI-oss and Galiana 111 and Abu-El-Magd and Sinha[P]. In addition to seasonal and diurnal variations, the load shape is generally governed by the ambient temperature lhis load-temperature relationship. Iiowever. is not stationary The day-type. humidity and temperature ineilia elfects are soine of the reasons for lhis non-stationary characteristic Many statistical techniques liave been proposed. and used. to identify the short term variations in the load. Moghrani and Rahnian 131 liave discussed some of these in a recent paper. These techniques are: 1 Multiple Linear Regression; 2. Stochastic Time Series; 3. General Exponential Smoothing; and 4. State Space Method. Most of these techniques woi-k generally well for the applications they were developed for. Errors. however. begin to creep in as parameters start to drift from the modeled conditions. Thus a periodic updating of these models is necessary This point is emphasized by Hubele and Cheng [4] They have presented seasonal short term load forecasting models using decision funclions The average errors for 24-hnur forecasts range from zyxwvuts 39’” to 5% Rahman et al. zyxw [5,fi,7] and Jabbour et al, [O] have put forward expert system based techniqiies using which the effect of weather and ollicr governing variables (that affect the load) can be modeled. lhesc techniques are generally more convenient for upd3ting the changing relationqhips ovrr time. Expert system based (or knowledge based) techniques depend on human expertise, and/or (historical) statistical relationships to generate rules. These rules are then applied to utilize independent vai-iables (like weather conditions) to produce load forecasts. An inherent difficulty with expert opinions, however. is that they may not always be consistent, or the reliability of such opinions may be in question There are inconsistent dala analysis techniques that can be applied lo resolve this situation. In a recent paper Rahman and Shrestha 191 have shown how an inconsistent dala analysis technique can be used to estimate various attribute values subject to uncertainty when applied to a power planning pi-oblem. As mentioned earlier, there are statistical lerhriiques which can be used to extract relationships belween load and weather variables. The basic methods and algorithms can be found ill many statistical texts [Kill], while some recent papers [12,13,14] preserit the current developments in statistical approach Most of these techniques are computationally very burdensome as well as labor intensive because statistical models have to be ripdated with changing conditions. Expert systems, on the other hand are more adaptable to changing conditions, but can fail because of inconsistent rules. OS85-8950/!31$01.01991 EEE