1、PDF外文：http:/ 附录 3 Electric load forecasting methods: Tools for decision making Heiko Hahn, Silja Meyer-Nieberg *, Stefan Pickl Fakult fr Informatik, Universitat der Bundeswehr, 85577 Neubiberg, Germany Abstract For decision makers in the electricity sector, the decision process is complex wit
2、h several different levelsthat have to be taken into consideration. These comprise for instance the planning of facilities and anoptimal day-to-day operation of the power plant. These decisions address widely different time-horizonsand aspects of the system. For accomplishing these tasks load foreca
3、sts are very important. Therefore,finding an appropriate approach and model is at core of the decision process. Due to the deregulationof energy markets, load forecasting has gained even more importance. In this article, we give an overviewover the various models and methods used to predict future l
4、oad demands. 2009 Elsevier B.V. All rights reserved. 1. Load forecasts in deregulated markets Decision making in the energy sector has to be based on accurateforecasts of the load demand. Therefore, load forecasts areimportant tools in the energy sector. Forecasts of different timehorizonsand differ
5、ent accuracy are needed for the operation ofplants and of the complex power system itself: The system responsefollows closely the load requirement” (Kyriakides and Polycarpou, 2007, p. 392). The decision maker is faced with a multitudeof decision problems on different time-scales as well as on diffe
6、renthierarchies of the power system: These problems comprisefor instance the determination of an optimal secure scheduling ofunit commitment and energy allocation. But decisions do not havemade only with respect to the day-to-day operation of the powersystem but also with respect to investment decis
7、ions on new facilitiesbased on the anticipation of future energy demands. For bothends, reliable forecasts are needed. The deregulation of energymarkets has increased the need for accurate forecasts even more(see e.g. Feinberg and Genethliou, 2005； Kyriakides and Polycarpou, 2007). To participate in
8、 the market, a player needs an accurateestimate howmuch energy is needed at a certain time. On theone hand, an underestimation of the energy demand by a suppliermay lead to high operational costs because the additional demandhas to be met by procuring energy in the market. An overestimationon the ot
9、her hand wastes scarce resources (see e.g. Tzafestasand Tzafestas, 2001； Feinberg and Genethliou, 2005； Kyriakidesand Polycarpou, 2007). Furthermore, demand is one of the mainfactors for pricing. Load forecasting is therefore at the core ofnearly all decisions made in energy markets. 章及标题 Due
10、 to the highimportance of accurate load forecasting, the history of this fieldis quite long: A 1987 survey paper (Gross and Galiana, 1987) listsan impressive number of publications devoted to load analysisand forecasting reaching back as far as 1966 (Heinemann et al.,1966). Up to now, various approa
11、ches have been introduced. Theycan be grouped into two main classes: Models and methods whichfollow a more classical approach, i.e., which apply concepts stemmingfrom time series and regression analysis and methods whichbelong to the fields of Artificial and Computational Intelligence. This paper gi
12、ves a short survey over models and methods forload forecasting. Further survey and review papers are for exampleKyriakides and Polycarpou (2007), Feinberg and Genethliou (2005),Tzafestas and Tzafestas (2001) and Hippert et al. (2001). 2. Short-term, medium-term and long-term forecasts As we have see
13、n, forecasts are made for various purposes: theday-to-day operation of the power system (e.g. Kyriakides andPolycarpou, 2007) requires the prediction of the load for a dayahead whereas the decision whether to undertake major structuralinvestments requires a far longer prediction horizon. Forecasts c
14、anbe distinguished therefore firstly by the time-horizon or the leadtime: short-term load forecasts (STLF) usually aim to predict theload up to one-week ahead (Kyriakides and Polycarpou, 2007).Frequently, the term very short-term load forecast is used forforecasts with a time-horizon of less than 24
15、 hours (see Yang,2006, p. 7). Up to now, the main focus in load forecasting has beenon STLF since it is an important tool in the day-to-day operation ofutility systems (see e.g. Gonzalez-Romera et al., 2006). Morerecently with the deregulation of energy markets, more and moreattention is also paid t
16、o load forecasts with a greater time-horizon,i.e., medium-term load forecasts. As stated in (Gonzalez-Romeraet al., 2006), medium-term load forecasts enables companies toestimate the load demand for a longer time interval which helpsthem for example in the negotiation of contracts with other compani
17、es.Medium-term load forecasts (MTLF) are from one weekto one year. Forecasts aiming at load prediction for more than ayear ahead are usually termed long-term load forecasts (LTLF)(see e.g. Feinberg and Genethliou, 2005). As stated in Kyriakidesand Polycarpou (2007) the time-horizon in LTLF is usuall
18、y 20 yearsalthough longer lead times of 2530 years can be found. The differencesin lead times have consequences for the models and methodsapplied and for the input data available and selected. The load demandis influenced by numerous factors ranging from weatherconditions over seasonal effects to so
19、cio-economic factors. Whichinput data has to be selected usually depends on the task and dataat hand. The decision maker, therefore, is not only faced with thetask of selecting an appropriate model type but also with determiningimportant external factors. Both tasks usually depend oneach other. Some
20、 general observations can be made, however. 章及标题 As stated in Fidalgo et al. (2007), it depends on the region andthe climatic conditions whether weather-dependent factors havea significant influence on the prediction. There is a common agreementthat the air temperature is the most important w
21、eather influence(see e.g. Hippert et al., 2001； Feinberg and Genethliou, 2005).This was already recognized in the 1930s (Hippert et al., 2001).Generally, the demand is high on cold days which can be attributedto electric heating. Similarly on hot days, the increased usage ofair-conditioning generate
22、s a higher demand of energy. In manycountries, this results in a U-shaped and clearly non-linear responsefunction of the load towards the temperature (Hippertet al., 2001). However, the exact shape of the curve depends onthe region, the climatic conditions and of course on the consumersbehavior. Add
23、itionally, the designated time-horizon and the availabilityof the datadetermine the input variables. As mentioned in (Tayloret al., 2006) univariate models are standard for very short-termload forecasts for up to 6 hours ahead. Furthermore, it should benoted that sometimes obtaining accurate weather
24、 forecasts maybe difficult. Therefore, univariate models are also applied for longerlead times (Taylor et al., 2006； Soares and Souza, 2006). In Kyriakides and Polycarpou (2007) three main groups of inputdata for short-term load forecasts are identified: seasonal inputvariables, weather forecast var
25、iables, and historical load data(Kyriakides and Polycarpou, 2007). Short-term load forecasts usuallyaim at providing the daily, hourly, or half-hourly load and thepeak load (day, week) (see e.g. Tzafestas and Tzafestas, 2001)although even smaller time intervals occur. Forecasting the loadprofile, i.
26、e., the load of the next 24 hours, is also a main target(Tzafestas and Tzafestas, 2001； Hippert et al., 2001). Medium-term load forecasts usually incorporate several additionalinfluences - especially demographic and economic factors.These forecasts often provide the daily peak and average load,altho
27、ugh hourly loads are also sometimes given, e.g. Bruhns et al.(2005). In the case of long-term load forecasts, even more indicatorsfor the demographic and economic development have to be takeninto account (Kyriakides and Polycarpou, 2007). These are forinstance thepopulation growth and the gross dome
28、stic product.Long-term load forecasting usually aims at predicting the annualload and the peak load (Kyriakides and Polycarpou, 2007). The time series of the loads itself has generally three seasonalcycles: an intra-daily cycle (the daily load curve or the load profile),a weekly cycle, and a yearly
29、seasonal cycle. The weekly cycle usuallyshows two main groups: week-days and weekends. Due toindustrial demand, the load tends to be higher during week-days.The weekend tends to influence the neighboring days so that Mondaysand Fridays are often treated separately. Saturday is also often found to show a different load profile than Sunday. However, theexact weekly pattern depends on the particular region under considerationand furthermore on the
