Application of Seasonal and Multivariable Grey Prediction Models for Short-Term Load Forecasting

Short-term electricity load forecasting is one of the most important operations in electricity markets. The success in the operations of electricity market participants partially depends on the accuracy of load forecasts. In this paper, three grey prediction models, which are seasonal grey model (SGM), multivariable grey model (GM (1,N)) and genetic algorithm based multivariable grey model (GAGM (1,N)), are proposed for short-term load forecasting problem in day-ahead market. The effectiveness of these models is illustrated with two real-world data sets. Numerical results show that the genetic algorithm based multivariable grey model (GAGM (1,N)) is the most efficient grey forecasting model through its better forecast accuracy.


Introduction
Electricity is a crucial resource, which must be planned and managed carefully.Electricity market involves generation, transmission, distribution, wholesaler, retailer companies and governmental authorities.Electricity load forecasts are one type of information that these market participants need in their decision-making processes.Short and long-term success in the operations of these electricity market participants partially depends on the accuracy of load forecasts.
Electricity load forecasting can be divided into three categories.The first category includes long term load forecasting.The time span in consideration can be six months, one year or longer.It is particularly important for growth strategies at the governmental level and it also has importance on the strategic decision making process of the electricity market operators.The second category deals with medium term load forecasting.The time span is weeks or months.It is vital for electricity generation companies because the stock levels and resource management decisions need this information (Hoffman & Wood, 1976).The last category is short term, hourly load forecasts.Hourly load forecasting of the next day is an essential operation in electricity markets.
This study aims to predict short-term loads of one-day ahead electricity market using grey prediction models.At this point, a seasonal grey model (SGM), a multivariable grey model (GM(1,N)) and a multivariable grey model optimized with genetic algorithm (GAGM(1,N)) are proposed to predict the electricity loads of one-day-ahead hours.
Many studies can be found in the literature concerning the application of grey forecasting methods for short-term load forecasting problem.These studies are summarized below: Yao et al. (2003) presented an improved GM(1,1) based prediction algorithm to forecast a very short electric power demand.In that study, the forecast error is reduced using adaptive model parameters.Li et al. (2006) proposed an improved GM(2,1) model to perform short term load forecasting.Zhou et al. (2006) suggested GM(1, 1) model with the trigonometric residual modification technique for forecasting electricity demand.Niu et al. (2008) developed genetic algorithm based GM(1,1) model to solve the problem of short-term load forecasting.In the proposed model, the value of parameter in the GM(1,1) is optimized using genetic algorithm.Bianco et al. (2010) used a trigonometric grey model with rolling mechanism for nonresidential electricity consumption in Romania.Li et al. (2012) proposed the Alphanumeric Journal Volume 5, Issue 2, 2017 adaptive grey model for forecasting short term load.In this study, the performance of the proposed grey model is compared with other methods based on back propagation neural networks and support vector regression.Jin et al. (2012) developed a hybrid optimization grey model using grey correlation contest for shortterm power load forecasting.The efficiency of developed model is illustrated by comparing with the results of basic grey models.Bahrami et al. (2014) presented a new model which is based on combination of the wavelet transform and GM(1,N) model.In this paper, to improve the forecast accuracy, the parameters of GM(1,N) model are determined using particle swarm optimization.
In this paper, performance analysis is conducted of grey prediction models for shortterm load forecasting problem.Also, in order to improve the forecast performance of the original GM (1,N) model, genetic algorithms is used to estimate the parameters of this model.
The organization of the paper is as follows.Section 2 describes the time series representation of load forecasting and defines basic performance criteria used in load forecasting.Section 3 gives brief introduction about seasonal grey model (SGM), multivariable grey model (GM(1,N)) and GM(1,N) model optimized with genetic algorithm.In section 4, an application of the proposed models with the experimentations on real data for the short-term load forecasting is given.The results are discussed in Section 5.

Short Term Load Forecasting Problem
Short-term load forecasting of hourly electricity load, usually handled by time series approach.If t is the index of the hour to be forecasted and Dt is the load of the hour t; training dataset S T with an autoregressive time series structure at each row is generated by Equation (1).
Apart from load values itself, temperature and other environmental or economic parameters can also be added to the data.
Mean Absolute Percentage Error (MAPE) is a commonly used performance criteria in load forecasting.Its formula can be seen in Equation (2).Here, Ai is the actual load value and Fi is the forecasted value for the same hour.
Seasonality is an important aspect in load forecasting.Several types of seasonality should be handled accordingly.The main types are day of the week seasonality and monthly seasonality.Generally, weekdays and weekends should be modeled separately.But even in this case, day of the week seasonality occurs, i.e. the behavior of the load pattern may be different on Mondays and Fridays.

Seasonal Grey Model
As mentioned earlier, grey prediction is one of the most important parts in the grey theory.The GM (1,1) model, which provides good prediction using limited data, is the basic grey forecasting model.GM (1,1) indicates one variable and one order grey forecasting model (Wang & Hsu, 2008).Short-term electricity consumption data has multiple seasonal patterns such as monthly, weekly, daily and hourly periodicity.Therefore, the use of the GM (1,1) model without considering of seasonal factors leads to inaccuracy load demand prediction.At this point, Xia and Wong (2014) proposed a seasonal discrete grey forecasting model.The steps of this model can be summarized as follows (Xia & Wong, 2014): Step 1. Assume X ( 0) ={x (0) (1), x (0) (2),…, x (0) (i),…, x (0) (n)} be an original time series data.A new sequence X ( 1) is obtained by cycle truncation accumulated generating operation (CTAGO).
(1) (0 where q is the periodicity of the time series.
Step 2. The seasonal grey forecasting model is established as follows. ( Step 3. In the Equation ( 4), the values of parameter d1, d2 and  can be calculated as follows using the least square method.

Multivariable Grey Model -GM(1,N)
Short-term electric load is affected by many factors such as temperature, humidity and time.These factors should be taken into account for a better-forecast accuracy.At this point, GM (1,1) and SGM (1,1) models are inadequate.
Alphanumeric Journal Volume 5, Issue 2, 2017 The calculation steps of GM(1,N) model which is multi variables first-order grey model, are described as follows (Tseng, Yu, & Tzeng, 2001;Hsu L. C., 2009): x n  be an original sequence of variable i.For i=1, this sequence includes output values of the data set, the others contain input data sets.
Step 2. For each input data set and output data set, new sequences are generated using accumulated generating operation (AGO).
Step 3. A sequence Z ( 1) Step 4. The first order -multi variables grey differential equation of GM(1,N) is established as follows (Pai, Chiou, & Wen, 2008): In this equation, parameters a,b2, b3, …, bN can be calculated as follows using the least square method.
  1 23 , , ,..., ( ) where Step 5.The predicted values of the accumulated sequence are obtained using following equation. ( Step 6.The predicted values of the original sequence are calculated by using the inverse accumulated generating operation (IAGO).
Traditional GM(1,N) forecasting model often sets the parameter α to 0.5.This value is not optimal for all data sets.Therefore, in this study, in order to improve the forecast performance of the original GM (1,N) model, genetic algorithms can be used to estimate the parameters of this model.To this end, the following model is established.
where 0 () xkis predicted value and n is the number of sequence data.

Short-Term Load Forecasting with Proposed Algorithms
In this section, application and adaptation of the previously explained algorithms to hourly load forecasting is introduced.In day-ahead market, the problem of load forecasting concerns making forecasts for each hour of the next day.The historical data and temperature data are used to make load prediction for each hour of the next day.In other words, short-term load forecasting for day-ahead market makes 24 different predictions for the next day.

Data Sets
In this study, two different load datasets used for benchmarking.BSE dataset (Hong, Wang, & Willis, 2011)

Methodology
In this study, two different approaches are proposed for each dataset.First, all data is used to build single training datasets, which include all hourly loads for all hours.In the second approach, each hour is modeled separately; consequently 24 different models is developed in this approach.Predictions for each hour are done with its own trained model.
Moving cross-validation procedure is used in this study.In this procedure, a part of the data is employed as training data, and the model generated by this training data is used in forecasting (test) phase.Then, for each new day, training data is shifted, thus, the actual values of the previously forecasted day become the last line of the training dataset.
In the first approach, t-1 (previous day), t-2 (the day before), t-5 (the same day in the previous week) are.In the second approach, when whole data is used to make predictions, then the structure is t-24, t-25, t-120.Here, the load of the same hour is used in the previous day and so on.For the selected datasets, temperature data is also available.Same lag structure is used as temperature input in the time series models.In the seasonal grey model, the value of the periodicity of the data sets (q) is taken as 24.
In the GAGM(1,N) model, the parameters of GA are as follows: the population size is 20, the crossover rate and mutation rate is set to be 0.8 and 0.1 respectively.The default values of the MATLAB optimization toolbox are used for the other parameters such as stopping criteria, elite count and selection function.
Forecast range for each dataset is reported in

Conclusions
Result on BSE dataset can be seen in Table 2. Also results from EUNITE dataset is given in Table 3  Further studies may include modeling impacts of holidays and other special days in load data by applying outlier analysis.Also for the distribution companies, unread meters are a source of problem.Some meters are read rarely, therefore there are big gaps in the load data for some customers.Future models may handle this randomness and imperfectness.

Figure 1 .
Figure 1.(a) Average hourly load of BSE dataset; (b) Average hourly load of EUNITE dataset

Forecasting
results are also represented in Figure 2 for EUNITE dataset.Forecasting results show that GAGM (1,N) model is promising techniques in hourly load forecasting.For the BSE dataset, Hong et al. (2011) has reached a minimum MAPE of 4.98% with updating their forecast model every day.In this study, GAGM (1,N) model has reached a minimum MAPE of 3.86% by using separate models for each hour.Similarly, in EUNITE dataset, GAGM (1,N) model has reached a MAPE value of 3.13%.

Figure 2 .
Figure 2. Forecasting results of EUNITE dataset in case of a single trained model contains hourly loads of an electricity distribution company between 01/01/2003 and 12/31/2008 and accompanying hourly temperature values are included in.In this study, each weekday of 2008 is forecasted by using this trained model.EUNITE dataset (Chen & Chang, 2004) originally contains half hourly loads.But they are combined into hourly values in this study.The data range is between 01/01/1997 and 12/31/1998 and accompanying daily temperature values are also presented.The first month of 1999 is predicted with the proposed methods.Hourly average load values of these datasets can be seen in Figure 1 and 2.

Table 1 .
Proposed algorithms are programmed on MATLAB 2016b.

Table 1 .
Datasets used in experimental analysis

Table 2 .
. In the approach lines of the tables, single model represents training a single forecasting model with the whole data, including all hours.Besides separate models represent building 24 different models for each hour.Performances of the proposed grey prediction algorithms on BSE dataset

Table 3 .
Performances of the proposed grey prediction algorithms on EUNITE dataset