Article Preview
TopIntroduction
Loadforecasting (LF) is the prediction of future electrical load by utilizing present or past data. The efficiency of electric power system is strongly linked to accurate LF (Fan & Hyndman, 2012; Raza, Baharudin, & Nallagownden, 2013), and therefore it is being applied in management, daily grid operation and load scheduling (De Felice & Yao, 2011). Its precise predication may improve power system generation, planning reliability and allocation of financial resources (Singh, Singh, & Tripathy, 2012), e.g., scheduling of fuel purchase, revising electricity tariffs and organizing events like switching off/on load response appliances, etc. It has been reported that a 1% decrease in the prediction error for a 10000 MW utility can save upto about $1.6 million annually (Hobbs et al., 1999). Load forecasting is a complicated technique as the electrical energy consumption is affected by several factors, including weather, vacations and economic status of the consumer (Fan & Hyndman, 2012).
Load forecasting can be classified into Very Short Term (VST), Short Term (ST), Medium Term (MT) and Long Term (LT) load forecasting. Short term load forecasting is the most common load forecasting category and is performed generally on daily basis for the day ahead with hourly or lesser granularity down to 5 minutes. A perfect short term load forecasting system requires automatic data access, bad data detection, portability, friendly interface, reduced response time and accuracy.
In the literature several methods have been proposed for LF including, statistical and artificial Intelligence methods. A statistical method has two major types, the time series that includes Moving Average (MA) method (Khashei & Bijari, 2010), Auto Regressive (AR) method and Auto Regressive Moving Average (ARMA) method (Lee & Ko, 2011) and the regression method (Ruzic, Vuckovic, & Nikolic, 2003). Statistical techniques (Chakhchoukh, Panciatici, & Mili, 2011) involve the exploitation of data gathered from earlier records to estimate the future loads and possible energy utilization with the help of the total load (kilowatts) and energy consumed (kilowatt-hours) against years. The statistical method in (Khashei & Bijari, 2010) describes the energy pattern as a time series signal with known daily, weekly as well as seasonal periodicity, providing the forecasting energy at the specified day, hour of the day and season. These types of techniques work well unless there is a sudden variation in the sociological and environmental variables.
Regression strategies make an effort to find out functional relationships among current load demand and weather variables (Ruzic et al., 2003). The conventional techniques use linear representations for predicting functions. By the linear arrangement of these representations, regression method observes the functional relationships among the selected load demand and the weather related variables. Both the time-series and regression schemes require considerable computational duration, as they employ several complex relationships.
Artificial Intelligence methods are categorized into three groups, Expert system and Fuzzy logic approach, Neural Network approach (NN) and Hybrid neural network approach (HNN). The forecasting results of an expert system are superior to that of any theoretical methods. For example, time-series model (Khashei & Bijari, 2010), which is an effective forecasting method for weekdays, shows undesirable results, for load prediction of holidays.