Load Forecasting

26-Jan-22 1
 Objectives
The main objectives of this lecture are:
♦ Defining the different classes of load forecasting
for distribution systems; short, medium and long
load forecast.
♦ Presenting the different types of classification
for short and medium load forecast.
♦ Identifying the load forecasting methods suitable
for distribution systems.
♦ Presenting the mathematical formulation for
different load forecasting methods.
♦ Presenting the state of the arts load forecasting
methods.
26-Jan-22 2
 Outlines
♦ Introduction.
♦ Spatial Load Forecast.
♦ Classifications of load Forecasting Methods.
♦ Classification of the Methods Used for Short and
Medium Load Forecasts.
♦ Data Collection and Preprocessing for Short and
Medium Load Forecast.
♦ Naïve Approach.
♦ Regression-Based Load Forecast.
♦ State Space and Kalman Filter.
♦ Soft Computing and Expert Systems.

26-Jan-22 3
 Introduction
♦ Forecasting is a statement of what will happen if certain
conditions or trends continue and assumes that the causes
of the events are under human control.
– Objective
– Scientific
– Reproducible
– Error analysis Possible

♦ Prediction is generally a statement or a guess of

something which will happen regardless of whether the
people want it or not and is something beyond one's
control
– Subjective
– Intuitive
– Non- Reproducible
– Error Analysis Limited
26-Jan-22 4
 Introduction, cont’d
♦ Forecasting is essential for a number of planning decisions

♦ LONG TERM DECISIONS

– New Distribution System Infrastructure
– Distribution System Expansions

♦ MEDIUM TERM DECISIONS

– Purchasing or Sell Electricity from other Neighboring Networks
– Fuel Required by the Utility in The near Future.

♦ SHORT TERM DECISIONS

– Operational Planning
– Unit Commitment.

26-Jan-22 5
 Introduction, cont’d
♦ Load forecasting has always been important for planning
and operational decision conducted by utility companies

♦ Load forecasting at the distribution level is essential input

for planning and proper operation of the distribution
network.
System
Objectives

System to be Plan of
Load Forecast
Managed Action

Constraints
-Budget/ Land
Resources

26-Jan-22 6
 Introduction, cont’d
Factors Affecting Load Forecasting

♦ Historical load data

♦ Load density
♦ Land use
♦ City planning
♦ Alternative energy source available
♦ Population growth

26-Jan-22 7
 Introduction, cont’d
Horizon Based Classification

Electric Load Forecast

Short Term Medium Term Long Term

1 hr – 1 week 1 week- 12 weeks 1 year-20 years

♦ Short term: For power system operation studies, losses

reduction, voltage regulations, unit commitment and maximizing
the utility revenues in the deregulated environment.

♦ Medium term: For estimating the necessary power to purchase

or sell from other neighboring networks and also the fuel required
by the utility in the near future.

♦ Long term: For planning and expansion studies.

26-Jan-22 8
 Introduction, cont’d
Forecasting Methods

Long and Medium Term Load Short and Medium Term Load
Forecast Forecast

Trending

Naïve Regression State Soft

Approach Space Computing
Land-Use
Simulation
ANN
Kalman
Multi Variant Time Series Filter
Econometric
Fuzy Logic
Modeling
MA

ES
AR

ARMA

26-Jan-22 9
 Outlines
♦ Introduction.
♦ Spatial Load Forecast.
♦ Long and Medium Methods of Forecasting.
♦ Classification of the Methods Used for Short and
Medium Load Forecasts.
♦ Data Collection and Preprocessing for Short and
Medium Load Forecast.
♦ Naïve Approach.
♦ Regression-Based Load Forecast.
♦ State Space and Kalman Filter.
♦ Soft computing and Expert Systems.

26-Jan-22 10
 Spatial Load Forecast
♦ It provides the load magnitude as well as where
it will be located (the load geographical
distribution) and where it will grow.
♦ A large service area is divided into smaller areas.
Load forecasting is then carried out in each small
area.
♦ The traditional approach was to forecast load
growth on a substation-by-substation or feeder-by-
feeder basis, allowing equipment service areas to
implicitly define the small areas.

26-Jan-22 11
 Spatial Load Forecast, cont’d
♦ Spatial load forecast depends on the
following principles:
– Small area forecast
– Small area load forecast.
– Spatial load forecast methods.
• Trending methods.
• Land-use based methods.

26-Jan-22 12
 Outlines
♦ Introduction.
♦ Spatial Load Forecast.
♦ Long & Medium Methods of Forecasting
♦ Classification of the Methods Used for Short and
Medium Load Forecasts.
♦ Data Collection and Preprocessing for Short and
Medium Load Forecast.
♦ Naïve Approach.
♦ Regression-Based Load Forecast.
♦ State Space and Kalman Filter.
♦ Soft computing and Expert Systems.

26-Jan-22 13
Long & Medium Methods of Forecasting

Long and Medium Term Load Forecast

Extrapolation Land-Use Econometric

(Trending) Simulation Modeling

26-Jan-22 14
 Extrapolation (Trending) Methods
♦ Trending or Regression analysis is the study of the behavior
of a time series or a process in the past and its mathematical
modeling so that future behavior can be extrapolated from it.
♦ These methods are suitable for estimating loads for short
periods ahead (from one to three years).
♦ Two general approaches:
– The fitting of continuous mathematical functions through
actual data to achieve the least overall error, known as
regression method.
– The fitting of a sequence on a discontinuous lines or
curves to the data, known as decomposition method.
♦ These methods have two main advantages:
– They employ simple, direct algorithm, and
– They require a minimum data.

26-Jan-22 15
Extrapolation (Trending) Methods, cont’d
♦Trending methods use the historical data to
project the future demand for the utility
service area under consideration.
♦Trending methods may be used to forecast
the distribution system load of:
– Single area.
– Multiple areas .

26-Jan-22 16
 Single Area Load forecasting
Regression-based method

♦ The oldest and most widely used technique for

developing forecast.
♦ Once a functional relationship is assumed to exist
between the dependent and the independent
variables, two steps must be taken in order to
forecast:
– The form of the relationship between the variables
must be determined
– The parameters of the relationship must be estimated.

26-Jan-22 17
 Single Area Load Forecasting
Regression-based method, cont’d
♦ The principle of regression theory is that any function y = f(x)
can be fitted to a set of points (x1,y1), (x2, y2) so as to
minimize the sum of errors squared to each point, that is
n
ε = ∑ [ yi − f ( xi )] = minimum
2 2

i =1

♦ Sum of squared errors is used as it gives a significant

indication of goodness of fit.

26-Jan-22 18
 Single Area Load Forecasting
Regression-based method, cont’d
♦ Typical regression functions used in power
system forecasting are:
Linear : y = a + bx
Exponential : y = a( 1 + b)x
Power : y = ax b
Polynomial : y = a + bx + cx 2
−be -cx
Gompertz : y = ae
The coefficients used in these equations are called regression coefficients.

26-Jan-22 19
 Single Area Load Forecasting
Regression-based method, cont’d
♦ Linear Regression: Using the method of least squares.
n
ε = ∑ [ yi − (a + bxi )] 2 = minimum
2

i =1

the regression coefficients can be determined by finding the partial

differentiation with respect to a and b and equate it to zero.

20
 Single Area Load Forecasting
Regression-based method, cont’d

♦ Regression analysis can deal only with linear

relationships. The non-linear functions have to be
transformed to linear functions.
– If y = a( 1 + b) then this function can be
x

transferred to ln y = ln a + x ln ( 1 + b)
– If y = a + bx + cx 2 + dx 3 ,
it can be transferred to y = a + bx1 + cx2 + dx3
where x1 = x , x 2 = x 2
and x 3 = x 3

21
 Single Area Load forecasting
Regression-based method, cont’d
♦ Once the regression coefficients are determined, a correlation
test must be applied to the model.
♦ Correlation is a measure of how well a line or curve fits a
given set of data. In general, it is a measure of how one
variable is related to another.
♦ Typical correlations are shown below

22
 Single Area Load forecasting
Regression-based method, cont’d
♦ The standard error of estimate is defined by

where yest is the estimated value of the dependent variable y for each
value of the independent variable(s), here x. This is analogous to standard
deviation. The coefficient of correlation, often called "goodness of fit", is
given by:

This varies from −1 to +1 and is ±1 for perfect +ve/ −ve correlation (or fit)
and 0 for no correlation.

23
Several sets of (x, y) points, with the Pearson correlation coefficient of x and y
for each set. Note that the correlation reflects the noisiness and direction of a
linear relationship (top row), but not the slope of that relationship (middle), nor
many aspects of nonlinear relationships (bottom). N.B.: the figure in the center
has a slope of 0 but in that case the correlation coefficient is undefined because
the variance of Y is zero

24
 Single Area Load forecasting
Decomposition method
♦ Decomposition method: The fitting of a sequence on discontinuous lines or
curves to the data.
♦ A time varying event such as power system load can be broken down into the
following four major components.
– The basic trend represents the long run behaviour of the load data and can
be increasing, decreasing or constant. The selected trend function is usually
the analytical function for the electric power demand, which is either
exponential, straight line, parabolic, or modified exponential.
– The seasonal variation relates to the periodic fluctuations of constant length
that are caused by such things as temperature, rainfall, or month of the year.
– The cyclic variation represents ups and downs of the economy of a specific
utility, such as Gross National Product (GNP), the index of industrial
production, demands for housing and interest rates.
– Random variations which occur on account of the day-to-day changes and
in the case of power systems, are usually dependent on the time of the week,
e.g. weekend, week day, weather, etc.
♦ Such decomposition can frequently facilitate improved accuracy in the forecast,
and aid in better understanding of the behaviour of the load data.

25
 Single Area Load forecasting
Decomposition method, cont’d

Decomposition of typical load growth curve: (a) Total process; (b) Decomposition

26
 Single Area Load forecasting
Decomposition method, cont’d
♦ The decomposition method assumes that the data are
made up of a certain pattern plus error, i.e.
– Data = pattern + error
– Data = f (trend, cycle, seasonality) + error
– The error is assumed to be the difference between the combined effect
of the sub-patterns of the data series and the actual data.

♦ There are several decomposition approaches, all of which

aim at separating each component of the data series as
accurately as possible.
♦ The basic concept in this separation is empirical and
consists of removing first seasonality, then trend, and
finally cyclic factors.
♦ Any residual is assumed to be random which while it can
not be predicted, can be identified.
27
 Single Area Load forecasting
Decomposition method, cont’d
♦ Typical power system load curves in the classical
decomposition method can be represented by the equation
(multiplicative model):
Xt=Tt×Ct×St×Rt
where,
Tt = long-term trend
Ct= cyclical trend (often over several years)
St= seasonal trend (1 year cycle)
Rt = Randomness ("noise")

Tt
St
Identify Obtain Xt
Ct
Time “t” Rt

28
 Single Area Load forecasting
Decomposition method, cont’d

De-seasonalizing the Time Series:

♦ If the time series represents a seasonal pattern of L periods,
then by taking a moving average Mt of L periods, we would
get the main value for the year.
♦ This would be free of seasonality and contain little random-
ness (owing to averaging)
♦ Thus Mt=Tt×Ct

29
 Single Area Load forecasting
Decomposition method, cont’d

Determination of the Trend:

♦ To the de-seasonalized series, a suitable trend line could be
fitted using Regression.
♦ The choice could be
– Linear
– Quadratic
– Exponential
– Other

30
 Single Area Load forecasting
Decomposition method, cont’d

Estimation of Cycle Component:

♦ After the trend Tt has been estimated , one can use
Ct =Mt/Tt
to estimate the Cycle Component, Ct

31
 Single Area Load forecasting
Decomposition method, cont’d

Determination of Seasonal Indices:

♦ To isolate seasonality, one could simply divide the original
series by the moving average
Xt/Mt=Tt×Ct×St×Rt/ Tt×Ct
= St×Rt
♦ Averaging over same months eliminates randomness and
yield seasonality indices.

32
 Single Area Load forecasting
Decomposition method, cont’d

Procedure for Decomposition:

1. Decompose the time series into its components
– Find seasonal component
– Deseasonalize the electricity demand (load)
– Find trend component

2. Forecast future values of each component

– Project trend component into the future
– Multiply trend component by seasonal component

33
 Single Area Load forecasting
Decomposition method, cont’d

Example 1:

Past Load Average Load Seasonal Factors

MW (1000/4)
Spring 200 250 200/250=0.8
Summer 350 250 350/250=1.4
Fall 300 250 300/250=1.2
Winter 150 250 150/250=0.6
Total = 1000

34
 Single Area Load forecasting
Decomposition method, cont’d

Example 1 (cont’d):

Expected Average Load Next year

Next year load (1100/4) Forecast
MW
Spring 275 *0.8=220
Summer 275 *1.4=385
Fall 275 *1.2=330
Winter 275 *0.6=165
Total = 1100

Using any regression function Seasonal Factors

35
 Single Area Load forecasting
Decomposition method, cont’d
Example 2:
Computing trend and Seasonal factor on a 2 year demand history

Quarter Load (MW) Quarter Load (MW)

I-2000 300 I-2001 520

II-2000 200 II-2001 420
III-2000 220 III-2001 400
IV-2000 530 IV-2001 700

36
 Single Area Load forecasting
Decomposition method, cont’d
Example 2 (cont’d): From Trend equation Tt=170+55t

Quarter Load (MW) Ratio of Seasonal

actual/Trend factor

2000
I 300 225 1.33 I 1.25
II 200 280 0.71 II 0.78
III 220 335 0.66 III 0.69
IV 530 390 1.36 IV 1.25
2001
I 520 445 1.17
II 420 500 0.84
III 400 555 0.72
IV 700 610 1.15
Cyclic factor included 37
 Single Area Load forecasting
Decomposition method, cont’d

Example 2 (cont’d):
Forecast for 2002 using Trend and Seasonal factors
I-2002 [170+55*09] 1.25 =831
II-2002 [170+55*10] 0.78 =562
III-2002 [170+55*11] 0.69 =535
IV-2002 [170+55*12] 1.25 =1038

Trend * Seasonal Factors = Forecast

38
 Single Area Load Forecasting
Cluster-based method
♦ The cluster-based method is proposed to overcome
the actual S load shape forecasting problem.

♦ Determines a set of clusters, with each cluster

containing a group of small areas with similar S-
curve, and then it combines these curves to obtain the
projected load in the future.

♦ The cluster-based technique is divided into two parts:

– Determination of the group S curve
– Forecasting of the loads.

39
 S-Curve
The S-curve shape reflects the load growth

Early Mass Slow

Take off
adopters adoption down
 Transforming to a linear model

The S-curve has a complicated formula but a change of

variables produces a simple linear equivalent.

Pt  1 
=
S 1 + e −(α + β .t ) 
The non-linear S-curve

Can be changed by ‘logistic’ transformation to

 Pt 
ln  = α + β .t Its linear transformation
 S − Pt 

S: load saturation level (i.e. max possible load)

Pt: load estimation at time t
 Single Area Load Forecasting
Cluster-based method, cont’d
♦ Determination of the set of curve S shapes: This is done
using the following steps:

– Shifting of load data for every small area that shows a load growth that
begins during the historical period
– All small areas are divided into clusters, each cluster sharing
approximately the same load growth curve.
– Extension of growth histories is carried out for small areas that were
not previously assigned to a cluster.

♦ Determination of each small area load forecast: To forecast

each small area load, one uses the cluster curve to which the
area belongs, and shifts the curve until it matches the small
area load history. This curve will usually extend far into the
future and represents the load forecast.

26-Jan-22 42
 Multiple Area Load forecasting
Vacant Area Method
♦ Reasons for the vacant area interference method:
– Load growth usually occurs in small areas which have no load in
them.
– These areas have zero historical load data and it is impossible to
trend such zero data into the future.
♦ Vacant area interference (VAI) method:
– This method is based on the assumption that the region’s smooth
continuous load growth can occur and continue in the future if
vacant areas within the region begin to grow.
– The VAI method uses this assumption in a repetitive and
hierarchical manner to forecast the small areas load.

43
 Multiple Area Load forecasting
Load Transfer Coupling Method
♦ Reasons for the load transfer coupling method:
– For different reasons, loads are transferred from one equipment
service area to another, some transfers are permanent and others are
temporary.
– These shifts of loads have no relation to load growth and produce
considerable error in load forecast results, especially when most of
these shifts are not measured neither their directions are known.
♦ The load transfer coupling method:
– This method introduces simple modification to the regression curve
fit, which does not require information on the magnitude nor on the
direction of the shift of the load transfer.

44
 Land User Simulation Methods
♦ Use the data available from different utilities billing files to
construct annual consumption values for each type of customers.

♦ Then modify these values for the local area under study.

♦ By knowing the customer numbers and the types of the loads for
this specific service area, load estimation can be obtained.

♦ Willis defined the simulation method as “Simulation-based

method attempt to produce, or model, the process of load growth
itself in order to forecast where, when and how load will develop,
as well as to identify some of the causes driving that growth”.

♦ In general, simulation methods are iterative methods in which the

data of consumption and load maps of one year are transferred
into another year.

45
Land User Simulation Methods, cont’d

46
 Econometric Modeling
♦ Relationships between power consumption and different
economical parameters are considered in the forecast.

♦ These econometric parameters are; gross national product,

index of industrial production, rate of inflation, rate of
increase of national population, electricity prices, etc.

♦ Large-scale econometric models are very complicated and

require many data. These type of models are suitable for
long rang forecast for transmission/generation systems, they
are rarely used for small area load forecast used in
distribution systems.

47
Econometric Modeling, cont’d

48
Econometric Modeling, cont’d

49
 Another Load Forecast Classification
Non-Analytical Forecast Methods
♦ Use computers to classify and store the data.
♦ Human intuition is then employed with this classified data
to forecast loads.
♦ The most common method is the coloring book method.
– The utility service area is divided by a series of grid lines into
small area.
– The existing load classes (industrial, commercial and residential)
are then placed on this map by simply coloring each class with
different color.
– Future land use is then placed on the map using the same set of
coloring code.
– Average load density for every class of loads is developed from
experience or previous billing data of the utility.
– These densities are used to convert the colored map into load map
by simply multiplying each colored area by its corresponding load
density.
– 25% of the US utilities use this method for their load forecast.
50
 Analytical Forecast Methods
♦ Analyze the past and present data and then extract the
trending in the data to project the load in the future.

♦ These methods are broadly divided into two main

categories:
– Trending Load Forecasting Methods
– Multivariate Load Forecasting Methods

♦ There are some conditions that should be satisfied before

applying the analytical load forecasting methods. These
conditions are:
– There is enough information about the past load.
– This information can be quantified as data.
– The load pattern of the past will continue in the future (this is
known as the assumption of constancy).

51
Factor Influencing Load Forecasting
Method Selections

52
Different Load Forecasting Methods

53
 Outlines
♦ Introduction.
♦ Spatial Load Forecast.
♦ Long & Medium Methods of Forecasting
♦ Classification of the Methods Used for Short and
Medium Load Forecasts.
♦ Data Collection and Preprocessing for Short and
Medium Load Forecast.
♦ Naïve Approach.
♦ Regression-Based Load Forecast.
♦ State Space and Kalman Filter.
♦ Soft computing and Expert Systems.

54
Short & Medium Load Forecasting
Methods Classification
Short and Medium Term Load Forecast

Naïve Regression State Soft

Approach Space Computing

ANN
Kalman
Multi Variant Time Series Filter
Fuzy Logic
MA

ES
AR

ARMA

55
 Outlines
♦ Introduction.
♦ Spatial Load Forecast.
♦ Long & Medium Methods of Forecasting
♦ Classification of the Methods Used for Short and
Medium Load Forecasts.
♦ Data Collection and Preprocessing for Short and
Medium Load Forecast.
♦ Naïve Approach.
♦ Regression-Based Load Forecast.
♦ State Space and Kalman Filter.
♦ Soft computing and Expert Systems.
56
 Data Collection and Preprocessing for
Short and Medium Load Forecast
♦ The following data are important for proper load
forecast studies:
– Load in MW and MVAR.
– Weather conditions.
– Economic indicators.
– Customer counts and types as a function of system,
substation, or buss.

♦ Important factor in short term electric load

forecast
– Time factor.
– Weather conditions.
– Historical load demand.

57
 Data Collection and Preprocessing for
Short and Medium Load Forecast, cont’d

♦ Chronicle steps are generally required for pre-

processing the collected raw data:
– Remove all negative and zero values.
– Remove outlying data that have values that are much
greater than, or less than, the average values.
– Remove patterns that have time gradients that are too
large.

58
 Outlines
♦ Introduction.
♦ Spatial Load Forecast.
♦ Long & Medium Methods of Forecasting
♦ Classification of the Methods Used for Short and
Medium Load Forecasts.
♦ Data Collection and Preprocessing for Short and
Medium Load Forecast.
♦ Naïve Approach.
♦ Regression-Based Load Forecast.
♦ State Space and Kalman Filter.
♦ Soft computing and Expert Systems.
59
 Naïve Approach
♦ The simplest approach for electric load forecast.

♦ Utilized extensively by the utility engineers for short

term electric load forecast.

♦ Based on searching historical data for days within

one, two, or three years with similar characteristics to
the forecast day.

♦ The load of a similar day is considered as a forecast.

♦ A more advanced is based on regression procedure

that can include several similar days with the trend
coefficients be extracted using similar days in the
previous years.
60
 Outlines
♦ Introduction.
♦ Spatial Load Forecast.
♦ Long & Medium Methods of Forecasting
♦ Classification of the Methods Used for Short and
Medium Load Forecasts.
♦ Data Collection and Preprocessing for Short and
Medium Load Forecast.
♦ Naïve Approach.
♦ Regression-Based Load Forecast.
♦ State Space and Kalman Filter.
♦ Soft computing and Expert Systems.
61
 Regression-Based Load Forecast
Time-Series-Based Methods

♦A time series is a set of observations on the values

that a variable take at different times.
♦Univariate time series
– it is a time series that consists of single
observations recorded over regular time interval
such as daily wind speed
 Regression-Based Load Forecast
♦ Several time series models have been developed for
Load forecasting. These developed models include
♦ Autoregressive model (AR)
♦ Moving average model (MA)
♦ Autoregressive moving average model ARMA
♦ Autoregressive integrated moving average ARIMA

63
 Regression-Based Load Forecast

– Auto Regression (AR) model:

An AR model is one in which 𝑌𝑌𝑡𝑡 depends only on its own past values 𝑌𝑌𝑡𝑡−1 , 𝑌𝑌𝑡𝑡−2 , 𝑌𝑌𝑡𝑡−3 , etc.
𝑌𝑌𝑡𝑡 = 𝑓𝑓(𝑌𝑌𝑡𝑡−1 , 𝑌𝑌𝑡𝑡−2 , . . … , ℇ𝑡𝑡)
– A common representation of AR model is
𝑌𝑌𝑡𝑡 = 𝛽𝛽0 + 𝛽𝛽1 𝑌𝑌𝑡𝑡−1 , 𝛽𝛽2 𝑌𝑌𝑡𝑡−2 +. . … + 𝛽𝛽𝑝𝑝 𝑌𝑌𝑡𝑡−𝑝𝑝 + ℇ𝑡𝑡
– Moving Average (MA) model:
A moving average model is one when 𝑌𝑌𝑡𝑡 depends only on random error terms which follow a white noise process, i.e.,
𝑌𝑌𝑡𝑡 = 𝑓𝑓(𝜀𝜀𝑡𝑡 , 𝜀𝜀𝑡𝑡−1 , 𝜀𝜀𝑡𝑡−3 , … … . )

𝑌𝑌𝑡𝑡 = 𝛽𝛽0 + 𝜀𝜀𝑡𝑡 , ϕ1 𝜀𝜀𝑡𝑡−1 , ϕ2 𝜀𝜀𝑡𝑡−2 +. . … + ϕ𝑝𝑝 𝜀𝜀𝑡𝑡−𝑝𝑝

– The error terms 𝜺𝜺𝒕𝒕 are assumed to be white noise process with mean zero and variance 𝜎𝜎 2

– Auto Regression Moving Average (ARMA) model:

There are situations where the time series may be represented as a mix of both AR and MA models refereed as ARMA
𝑌𝑌𝑡𝑡 = 𝛽𝛽0 + 𝛽𝛽1 𝑌𝑌𝑡𝑡−1 , 𝛽𝛽2 𝑌𝑌𝑡𝑡−2 +. . … + 𝛽𝛽𝑝𝑝 𝑌𝑌𝑡𝑡−𝑝𝑝 + ℇ𝑡𝑡 + ϕ1 𝜀𝜀𝑡𝑡−1 , ϕ2 𝜀𝜀𝑡𝑡−2 +. . … + ϕ𝑝𝑝 𝜀𝜀𝑡𝑡−𝑝𝑝
 Regression-Based Load Forecast

– Auto Regression Integrating Moving Average (ARIMA) model:

ARIMA is applicable to non-stationary time series. A non stationary series can be made stationary by taking
a difference between two consecutive points. A series which is stationary after being differentiated is said to
be integrated.

• AI-based time series models:

− A time series models based on using ANN have been also used for wind speed
forecasting. Among these ANN based methods; Elman Recurrent Network (ERN),
Adaptive Network based Fuzzy Inference System (ANFIS), Radial Basis Function
network (RBF), and Neural Logic Networks (NLN).
− These models require large sets of historical data for their parameters and model
training (up to weeks of recorded data). Moreover, they have been only proved to be
effective for short term prediction (few hours ahead).
 Load Forecasting Using AR Model
♦ Example: Wind Speed Prediction Using an AR Time Series Model
- A simple second order AR(2) model can be constructed as

y(t) = a + b1 y(t − 1) + b2 y(t − 2)

♦ where y is the wind speed at time t, and a, b1, b2 are the model parameters
- For a vector of 𝑛𝑛 components, the model parameters can be obtained by the least-square
method (LS) as follows:

𝑎𝑎
𝐴𝐴 = 𝑏𝑏1 = 𝐵𝐵𝑇𝑇 𝐵𝐵 −1 𝐵𝐵 𝑇𝑇 𝑌𝑌
𝑏𝑏2
1 𝑦𝑦 1 𝑦𝑦 0 𝑦𝑦 2
. . . .
where: 𝐵𝐵 = . . . , 𝑌𝑌 = .
. . . .
1 𝑦𝑦 𝑛𝑛 − 1 𝑦𝑦 𝑛𝑛 − 2 𝑦𝑦 𝑛𝑛
 Regression-Based Load Forecast
Application of Stochastic Time Series Analysis in Short and
Medium Term Load Forecast

67
 Outlines
♦ Introduction.
♦ Spatial Load Forecast.
♦ Long & Medium Methods of Forecasting
♦ Classification of the Methods Used for Short and
Medium Load Forecasts.
♦ Data Collection and Preprocessing for Short and
Medium Load Forecast.
♦ Naïve Approach.
♦ Regression-Based Load Forecast.
♦ State Space and Kalman Filter.
♦ Soft computing and Expert Systems.
68
 State Space and Kalman Filter
♦ Kalman Filter (KF) is a dynamic optimal estimator
which can be utilized for electric load forecast.
♦ It can follow the change of loads dynamic and
provide a prediction which is optimal in the statistical
sense.
♦ KF is used to estimate time varying parameters of a
system.
♦ It is a recursive estimator, which can be used to
process noisy measurements.
♦ KF gives a least square optimal estimate.

69
 Outlines
♦ Introduction.
♦ Spatial Load Forecast.
♦ Long & Medium Methods of Forecasting
♦ Classification of the Methods Used for Short and
Medium Load Forecasts.
♦ Data Collection and Preprocessing for Short and
Medium Load Forecast.
♦ Naïve Approach.
♦ Regression-Based Load Forecast.
♦ State Space and Kalman Filter.
♦ Soft computing and Expert Systems.
70
Soft Computing and Expert Systems
Neural Networks
− A time series models based on using ANN have been also
used for Load forecasting. Among these ANN based
methods; Elman Recurrent Network (ERN), Adaptive
Network based Fuzzy Inference System (ANFIS),
Radial Basis Function network (RBF), and Neural Logic
Networks (NLN).
− These models require large sets of historical data for their
parameters and model training (up to weeks of recorded
data). Moreover, they have been only proved to be
effective for short term prediction (few hours ahead).

71
 Soft Computing and Expert Systems
Neural Networks
♦ Artificial Neural Networks, ANNs, are mathematical models
designed to imitate brain works.
♦ They are parallel processing configurations.
♦ Neurons are the main building blocks of ANN.
♦ Any ANN architecture is created from the interconnection of
these building blocks. Computation neuron sums N weighted
inputs and then outputs the result through a non-linear
function. Data flows through this series of non-linear
transformations (e.g., sigmoid), in one or more stages in an
assigned topology of ANN.
♦ Neural networks must be trained prior to use.

72
 Soft Computing and Expert Systems
Neural Networks, cont’d
♦ Categorization of ANNs is based on the method of
training, network topology, algorithm used in
training and so on.
♦ According to the method of training neural network
paradigms, ANN can be classified into two basic
categories: Supervised learning and Unsupervised
learning structures.
♦ According to the topology, neural network may be
categorized into feed forward neural network and
recurrent neural network.
♦ According to the training algorithm, training of
neural network could be done off-line or online.
73
Soft Computing and Expert Systems
Neural Networks, cont’d

74
 Soft Computing and Expert Systems
Application of ANN to load forecast
♦ Basic variables required to use ANN in load
forecasting include:
– Input variable to the ANN (load, temperature,…etc)
– Number of classes ( weekday, weekend, season,…etc)
– What to forecast: hourly loads, next day peak load, next
day total load, …etc
– Neural network structure ( Feedforward, number of hidden
layer, number of neuron in the hidden layer,…etc)
– Training method and stopping criterion
– Activation functions
– Size of the training data
– Size of the test data

75
Soft Computing and Expert Systems
Application of ANN to load forecast

76
 Soft Computing and Expert Systems
Fuzzy Logic
♦ Fuzzy Logic (FL) can be considered as an extension
to crisp set theory.
♦ FL allows partial set membership rather than crisp set
membership or non-membership.
♦ FL model is experimentally based, relying on
operator's experience.
♦ In 1973, Professor Lotfi Zadeh proposed the concept
of linguistic or "fuzzy" variables as linguistic objects
or words, rather than numbers.
♦ FL does not require precise inputs, inherently robust,
and can process any reasonable number of inputs.

77
 Soft Computing and Expert Systems
Fuzzy Logic, cont’d
♦ Fuzzy systems rely on a set of if-then rules. Systems,
which use these rules for a specific task, are called
fuzzy inference systems, fuzzy –based system, fuzzy
models or fuzzy controller.
♦ Fuzzy If-Then Rules: Fuzzy logic employs simple,
plain-language of the form ‘IF A... THEN B…’rules,
that illustrates the system response in terms of
linguistic variables rather than mathematical formula,
where ‘A’ is called the premise part and ‘B’ is called
antecedent .
♦ Fuzzy Inference Systems: The process of producing
the mapping from a given input to an output using
fuzzy logic.

78
Soft Computing and Expert Systems
Fuzzy Logic, cont’d

79
 Soft Computing and Expert Systems
Expert Systems

♦ Expert systems are systems which encode the experience of

the electric utility expert in the form of IF …THEN rules.

♦ Load forecast value is the result of applying suitable inference

approach on the existing expert rule.

♦ Expert systems commonly consists of three main building

blocks; the knowledge base, the rule base and the inference
engine.

♦ Load forecasting done using these expert systems has a good

success for short term forecast but it has limited success for
long term load forecast.

80