REVSTAT – Statistical Journal

Volume 0, Number 0, Month 0000, 000-000

A NOVEL FRACTIONAL FORECASTING MODEL

FOR TIME DEPENDENT REAL WORLD CASES

Authors: Ummugulsum Erdinc

– Department of Mathematics, Aksaray University, Aksaray,
TURKEY ([email protected])

Halis Bilgil *
– Department of Mathematics, Aksaray University, Aksaray,
TURKEY ([email protected])
Zafer Ozturk
– Institute of Science, Nevsehir Haci Bektas Veli University,
TURKEY ([email protected])

Received: Month 0000 Revised: Month 0000 Accepted: Month 0000

Abstract:
ˆ The grey modelling in the prediction of time series has been one of the interesting
study fields recently due to its efficiency and convenience. Fractional grey models,
on the other hand, have become preferable, despite the difficulty in calculations,
since they give more effective results than standard models. Difficulties in fractional
accumulation and difference calculations have begun to be overcome thanks to new
definitions and theorems made in recent years. The new trend in grey modelling is
to compose models that are more useful than the previous ones and give results with
less error. In this paper, a new grey model derived from the conformable fractional
order is defined and it is shown that more effective estimates are made compared to
the models created in recent years. The verification of the method is shown with real
data set. The results show that the proposed conformable fractional grey model is
more effective than the existing models.

Key-Words:
ˆ Grey systems; least square method; nonresponse; conformable fractional calculus; frac-
tional grey model; ECFGM model.

AMS Subject Classification:

ˆ 26A33, 60G25, 81T80.

*
Corresponding author
2 U. Erdinc, H. Bilgil, and Z. Ozturk

1. INTRODUCTION

The grey system is a discipline that studies the problem of uncertainty,

which was first proposed by Deng in 1982, plays an important role in the grey
system theory [4]. It has been a useful tool in processing uncertain or excursive
systems with small samples and limited data set as distinct from the machine
learning models, hybrid models, the empirical models ..., etc. The grey prediction
models have been widely and successfully applied to various fields, such as science
and technology, energy, environmental problems, economy, health and other fields
[24, 35, 36, 7, 1, 33, 5, 18].

Recent studies on grey modelling focus on two main purpose: practicality

and prediction accuracy. For these purposes, important studies have been car-
ried out in recent times. Ma and Lui [15] proposed a time-delayed polynomial
grey prediction model called TDPGM (1,1) model, the grey polynomial model
with a tuned background coefficient was proposed by Wei et al. [20], Cui et al.
[3] developed a parameter optimization method to improve the ONGM (1,1,k)
model, Bilgil [1] proposed an exponential grey model named EXGM(1,1). Fur-
thermore, Ma et al. [14] developed a novel nonlinear multivariate forecasting
grey model based on the Bernoulli equation named NGBMC (1, n), Wang et al.
[19] introduced a seasonal grey model called SGM (1,1), Wu et al. [25] proposed
a new grey model called BNGM (1,1,t2 ) model, Liu and Wu [12] proposed the
ANDGM model, Ma [13] proposed kernel-based KARGM(1,1) model, Li et al.
[10] developed structure-adaptive intelligent grey forecasting model, Wu et al.
[26] developed a novel grey Riccati model (GRM), the modified grey prediction
model with damping trend factor was proposed by Liu et al. [11], the nonlinear
grey Bernoulli model with improved parameters, INGBM (1, 1), was proposed
by Jiang et al. [8].

It is clear that most of the existing grey models are defined with a first-order
whitening differential equation. If the original data is a disordered sequence, the
characteristic features of the sequence may not be exactly found out by first-
order accumulative generation operation (1-AGO) [32]. Moreover, first-order
derivative models are ideal memory models, which are not suitable for describing
irregular phenomena. As a result of this, the parameters of the model may
not be compatible according to the data characteristics of the actual problem
for a sequence with large data fluctuation. Therefore, fractional accumulation
generating operation and fractional derivative should be introduced into the grey
model to overcome this problem [32].

Wu et al. developed traditional GM(1,1) with fractional order accumulated

operator named FAGM(1,1) [23]. Some researchers optimized the FAGM(1,1)
model and reached better prediction accuracy in recent years. Wu et al. sug-
gested a fractional FAGMO(1,1,k) model with linear grey input of time in lieu
of constant grey input in the initial FAGM(1,1) model and optimized it with
optimal order and optimal parameters [27]. Besides, Mao et al. introduced the
A novel fractional forecasting model 3

fractional grey model FGM(q,1) [17], a power-driven fractional accumulated grey

model named PFAGM is introduced by Zhang et al. [34], Yuxiao et al. proposed
the multivariable Caputo fractional derivative grey model with convolution in-
tegral named CFGMC(q,N), Xie et al. developed a conformable fractional grey
model in opposite direction CFGOM(1,1) [30].

However, the definition is given by Wu [23] only show us a single situation of

fractional order calculus and differencing. From the perspective of computational
complexity, frequently operated descriptions of fractional accumulation and its
suitable fractional differencing in the present grey models are not easy to apply,
and it causes serious rigours to the more deeply theoretical analysis. Yang and
Xue [31] submitted the fractional order calculus, but the exact solution of this
kind of model include infinite series, and it is clearly difficult to use and analyse.
Furthermore, this kind of complications would also impede the improvement of
the fractional order grey models thanks to some new operators [22, 21].

Lately, Khalil et al. defined a limit-based fractional derivative in 2014 [9],

which is named the conformable fractional derivative. The structure of this new
definitions of fractional derivative is simpler than that of other popular frac-
tional derivatives, such as the Caputo derivative and Riemann-Liouville deriva-
tive. Then, Ma et al. [16] introduced the new useful definitions of the fractional
order difference and accumulation based on the conformable fractional deriva-
tive in which the computational complexity of accumulation is lower than that
of the traditional fractional accumulated operator and they firstly proposed an
improved fractional order grey model named CFGM(1,1). Recently, a continuous
grey model named CCFGM based on the conformable fractional derivative was
described by Xie et al. [29].

In this paper, we introduced the novel exponential conformable fractional

grey model (denoted as ECFGM(1,1) for short) by using the new definitions
of conformable fractional difference and conformable fractional accumulation by
Ma et al. [16]. The structural order of ECFGM(1,1) is sought out by using the
Brute Force algorithm. The effectiveness of ECFGM(1,1) is validated by real
data sets in comparison with other used new grey models and it is seen that the
performance of the ECFGM(1,1) model is very successful.

The rest of this paper is organized as follows: Section 1 includes relevant

literature. Some useful properties and definitions of the conformable fractional
calculus are given in section 2. The presentations and modelling mechanism of
the ECFGM(1,1) are introduced in section 3. In Sections 4 and 5, we present a
series of samples to validate ECFGM(1,1). Finally, the conclusions of this study
are given in Sections 6.
4 U. Erdinc, H. Bilgil, and Z. Ozturk

2. SOME DEFINITIONS AND PROPERTIES ON CONFORMABLE

FRACTIONAL CALCULUS

In this section, some useful definions and properties of the conformable

fractional derivative are summerized.

2.1. The Conformable fractional derivative

Definition 2.1. (See [16]). If f : [0, ∞) → R is a differentiable function,

the conformable fractional derivative of f with α ∈ (n, n + 1] order is defined as

f (t + εt⌈α⌉−α ) − f (t) df (t)

(2.1) Tα (f ) (t) = lim = t⌈α⌉−α
ε→0 ε dt
where ⌈.⌉ is the ceil function, i. e. the ⌈α⌉ is the smallest integer no larger than
α. It is clear that ⌈α⌉ = 1 for α ∈ (0, 1]. Hence, Eq. (2.1) can be written as

f (t + εt1−α ) − f (t) df (t)

Tα (f ) (t) = lim = t1−α
ε→0 ε dt
for all t > 0.

The following theorem gives the properties of the definition (Khalil et al.[9]).

Theorem 2.1. (See [9]). If the function f and g are differentiable, α ∈

(0, 1], then we have

1. Tα (f ) (t) = t1−α dfdt(t) .

2. Tα (mf + ng) = mTα (f ) + nTα (g) for all m, n ∈ R

3. Tα (f.g) = f Tα (g) + gTα (f )

 
4. Tα fg = gTα (f )−f
g2
Tα (g)

5. Tα (c) = 0 for all constant c

6. Tα (tp ) = ptp−α for all p ∈ R.

7. Tα (ecx ) = cx1−α ecx for all c ∈ R.

Proof: The proof is omitted.

A novel fractional forecasting model 5

2.2. The conformable fractional accumulation and difference

New definitions to calculate the conformable fractional accumulation (CFA)

and the conformable fractional difference (CFD) are given by Ma et al. [16] as
follows:

Definition 2.2. (see [16, 28]) The conformable fractional difference (CFD)
of f with α order is defined as
(2.2) ∆α f (k) = k 1−α ∆f (k) = k 1−α [f (k) − f (k − 1)]
for all k ∈ N + , α ∈ (0, 1] and

k  
α ⌈α⌉−α n+1 ⌈α⌉−α
X k−j ⌈α⌉
(2.3) ∆ f (k) = k ∆ f (k) = k (−1) f (j)
k−j
j=k−⌈α⌉

for all k ∈ N + , α ∈ (n, n + 1] .

Definition 2.3. (see [16]) The conformable fractional accumulation (CFA)

of f with α order is defined as
  Xk
α f (k) f (j)
(2.4) ∇ f (k) = ∇ 1−α
=
k j 1−α
j=1

for all k ∈ N + , α ∈ (0, 1] and

 
α n+1 f (k)
(2.5) ∇ f (k) = ∇
k ⌈α⌉−α
for all k ∈ N + , α ∈ (n, n + 1] .

3. Presentation of exponential conformable fractional grey model

In this section, a novel exponential conformable fractional grey model,

named ECFGM(1,1), is introduced, which optimizes the classical CFGM(1,1)
model with an exponential grey action quantity.

3.1. Formulation of proposed fractional grey model

The original series X (0) = x(0) (1), x(0) (2), ..., x(0) (n) is given. CFA with

α order is calculated as follow,

6 U. Erdinc, H. Bilgil, and Z. Ozturk

 
(3.1) X (α) = x(α) (1), x(α) (2), ..., x(α) (n)

where
k 
⌈α⌉ x(0) (i)
X 
(α) α (0)
(3.2) x (k) = ∇ x (k) = , α ∈ R+
k − i i⌈α⌉−α
i=1
 ⌈α⌉  Γ(k−i+⌈α⌉) (k−i+⌈α⌉−1)!
where k−i = Γ(k−i+1)Γ(⌈α⌉) = (k−i)!(⌈α⌉−1)! (see [28]).

Definition 3.1. The first-order whitening differential equation of the

ECFGM(1,1) is defined as,
dx(α) (t)
(3.3) + ax(α) (t) = b + ce−t .
dt
Where a is a development coefficient, b is called driving coefficient and ce−t is
an exponential grey action quantity. So that, the monotone decreasing term ce−t
will suppress the growth of the prediction error.

When α = 1, the ECFGM(1,1) model yields the EXGM(1, 1) [9].In ad-

dition, the proposed model can be translated to the conventional CFGM(1,1)
model for c = 0 [23].

3.2. Parameters estimation

Theorem 3.1. For the computed CFA and the value of fractional order,
the system parameters a, b and c of the ECFGM(1, 1) satisfy the following
equation,

−1 T
(3.4) [a, b, c]T = B T B B Y,
where the matrix B and Y are
−0.5(x(α) (2) + x(α) (1)) 1 (e − 1)e−2 x(α) (2) − x(α) (1)
   
 −0.5(x(α) (3) + x(α) (2)) −3
1 (e − 1)e  (α) (α)
 x (3) − x (2) 
   
(3.5) B = 
 . . . ,Y = 
  . .

 . . .   . 
−0.5(x(α) (n) + x(α) (n − 1)) 1 (e − 1)e−n x(α) (n) − x(α) (n − 1)

Proof: Integrating both side of the whitening equation (3.3) within the
interval [k − 1, k] the discrete form of ECFGM(1, 1) model is obtained as follow:

Zk Zk Zk
dx(α) (t)
(3.6) dt + a x (α)
(t)dt = (b + ce−t )dt.
dt
k−1 k−1 k−1
A novel fractional forecasting model 7

According to the Newton-Leibniz formula, the first integral of Eq. 3.6 can
be expressed as

Zk
dx(α) (t)
(3.7) dt = x(α) (k) − x(α) (k − 1).
dt
k−1

Rk
It is clear that the integration term x(α) (t)dt denotes the area between t-axis
k−1
and the curve x(α) (t) in the interval [k − 1, k] . Then, using the generalized
trapezoid formula as in many recent studies [23, 25, 16, 29, 26, 17, 10, 21], the
second integral of Eq. (3.6) can be obtained as

Zk
a  (α) 
(3.8) a x(α) (t)dt = x (k) + x(α) (k − 1) .
2
k−1

and the right side of Eq. (3.6) is equal to

Zk
(3.9) (b + ce−t )dt = b + c(e1−k − e−k ).
k−1

Substituting Eqs. (3.7) - (3.9) into Eq. (3.6), it can be written as

  a 
(3.10) x(α) (k) − x(α) (k − 1) + x(α) (k) + x(α) (k − 1) = b + c(e1−k − e−k )
2
where k = 2, 3, ..., n.

The linear equations system (3.10) can be written as follows

x(α) (2) − x(α) (1) = −0.5a(x(α) (2) + x(α) (1)) + b + c(e−1 − e−2 )
x(α) (3) − x(α) (2) = −0.5a(x(α) (3) + x(α) (2)) + b + c(e−2 − e−3 )
(3.11) .. .. ..
. . .
x(α) (n) − x(α) (n − 1) = −0.5a(x(α) (n) + x(α) (n − 1)) + b + c(e1−n − e−n )

and system (3.11) can be written as

(3.12) Y = Bρ,

where
−0.5(x(α) (2) + x(α) (1)) 1 (e − 1)e−2 x(α) (2) − x(α) (1)
   
 −0.5(x(α) (3) + x(α) (2))
 1 (e − 1)e−3 

 x(α) (3) − x(α) (2) 
 
(3.13) B = 
 . . . ,Y = 
  . .

 . . .   . 
−0.5(x(α) (n) + x(α) (n − 1)) 1 (e − 1)e−n x(α) (n) − x(α) (n − 1)
8 U. Erdinc, H. Bilgil, and Z. Ozturk

and ρ = [a, b, c]T in which n is the number of samples used to construct the
model.

The parameter estimation of the ECFGM(1,1) model using the least squares
method can be obtained. For the estimated value of the parameter sequence ρ,
the x(α) (k) − x(α) (k − 1) on the left side of the equation (3.11) is replaced with
−0.5a(x(α) (k) + x(α) (k − 1)) + b + c(e1−k − e−k ) , the error sequence ϵ = Y − Bρ
is obtained. Here,

(3.14) ϵ = [ϵ2 , ϵ3 , ..., ϵn ]T

and ϵk are represent the error for each equation in the system (3.11) for k =
2, 3, ..., n.

Notice, S(ρ) is defined as the sum of squares of errors, which yields

n
X
S(ρ) = ϵ2k
k=2
=ϵT ϵ
(3.15) =(Y − Bρ)T (Y − Bρ)
=(Y T − ρT B T )(Y − Bρ)
=Y T Y − Y T Bρ − ρT B T Y + ρT B T Bρ
=Y T Y − 2ρT B T Y + ρT B T Bρ.

The parameter vector ρ = [a, b, c]T that minimize S(ρ) satisfy

∂S
(3.16) = −2B T Y + 2B T Bρ = 0,
∂ρ
so

(3.17) B T Y = B T Bρ.

Thus

(3.18) ρ = (B T B)−1 B T Y,

or

−1
(3.19) [a, b, c]T = B T B B T Y,

Thence the proof is completed by using the least square estimation method.
A novel fractional forecasting model 9

3.3. Response function and restored values

Theorem 3.2. The discrete form of the response function of ECFGM(1,

1) model is given as
 
(α) (0) b c −1 a(1−k) b c −k
(3.20) x̂ (k) = x (1) − − e e + + e
a a−1 a a−1
where k = 2, 3, ..., n.

Proof: It is clear that, the solution of the first order linear whitening
differential equation (3.3) can be obtained as
b c −t
(3.21) x(α) (t) = + e + de−at
a a−1
where d is integral constant. By using the initial condition x(α) (1) = x(0) (1), the
constant d can be found as,
 
(0) b c −1 a
d = x (1) − − e e .
a a−1

Therefore the grey prediction model Eq. (3.21) can be obtained as following
 
(α) (0) b c −1 a(1−t) b c −t
x̂ (t) = x (1) − − e e + + e
a a−1 a a−1
and the discrete form of the response function can be write as
 
(α) (0) b c −1 a(1−k) b c −k
x̂ (k) = x (1) − − e e + + e .
a a−1 a a−1

The proof is completed.

Theorem 3.3. Then restored values can be given as

(3.22) x̂(0) (k) = ∆α x̂(α) (k) = k ⌈α⌉−α ∆n+1 x̂(α) (k), α ∈ (n, n + 1]
where k = 2, 3, ..., n.

Proof: From Eq. (3.2) it can be seen that x̂(α) (k) = ∇α x̂(0) (k). If we
apply the inverse operator ∆α , it is obtained as
x̂(0) (k) = ∆α x̂(α) (k)
and from Definition 2.2 it can be written as
x̂(0) (k) = ∆α x̂(α) (k) = k ⌈α⌉−α ∆n+1 x̂(α) (k), α ∈ (n, n + 1].
This completes the proof.

It is clear that, the restored values for α ∈ (0, 1] can be written as,
(3.23) x̂(0) (k) = k 1−α (x̂(α) (k) − x̂(α) (k − 1)).
10 U. Erdinc, H. Bilgil, and Z. Ozturk

3.4. Evaluative accuracy of the forecasting model

The relative percentage error (RPE) and the mean absolute percentage
error (MAPE) are used to evaluate the fitting and predicting performance of
ECFGM(1,1). The lowest MAPE value indicates the best prediction model. They
are defined as follows:

x̂(0) (k) − x(0) (k)

(3.24) RP E(k) = × 100%,
x(0) (k)

n
1X
(3.25) M AP E = RP E(k).
n
k=1

where x(0) (k) is the original series, and x̂(0) (k) is the predicted series.

For a raw sequence with n samples, the metric M AP Ef it is defined as the

fitting performance metric while M AP Epre is defined as a prediction performance
metric. Mathematically, they can be formulated as:
p
1X
(3.26) M AP Ef it = RP E(k),
p
k=1

n
1 X
(3.27) M AP Epre = RP E(k)
n−p
k=p+1

where p represents the number of samples used for fitting a model while the rest
of the raw sequence is used to examine the prediction accuracy of the model. The
total MAPE is given Eq. (3.25) and it is used to evaluate the whole performance
of a model.

3.5. Computation steps

The computation steps of ECFGM(1,1) model with given sample and α

can be summarized as follows:

Step 1: Create a raw data set x(0) (1), x(0) (2), ..., x(0) (n) ;

Step 2: Take as α = 0.01 to the initial value and designate an initial value
of MAPEmin ;

Step 3: Compute the CFA series with α order of the given raw data set
by using Eq. (3.2);
A novel fractional forecasting model 11

Step 4: Build the matrix B and Y using Eq. (3.5);

Step 5: Calculate the parameters a , b and c using Eq. (3.4);

Step 6: Calculate the predicted values and the response function using
Eq. (3.20) and Eq. (3.22) ;

Step 7: Calculate the mean absolute percentage error (MAPE) using Eq.
(3.25);

Step 8: If MAPE is greater than MAPEmin , set α as α = α + h (where h

is the step size), otherwise take as MAPEmin =MAPE and go to Step 3; where
step 8 is continued until α reaches the predetermined value.

Brute Force is a straightforward approach, which is also known as the

Naive algorithm, for solving optimization problems that rely on sheer computing
power and trying every possibility rather than advanced techniques to improve
efficiency. Unlike some of the other popular swarm intelligence algorithms, Brute
Force is applicable to a very wide variety of problems and it is an effective and
easy method to find the optimum value in the solution interval.

Despite the convergence speed advantages of other algorithms, it is a dis-

advantage that they may focus on the local extremum point rather than the
global extremum. However, the Brute Force algorithm scans the whole domain,
evaluates each point, then calculates the MAPE’s on these points and reaches
optimum parameters without any delusion.

In this paper, our purpose is to find the optimum parameter α that min-
imizes the model’s mean absolute percentage error (MAPE). Therefore we enu-
merate all the values in an interval with step 0.01. For suitability, in the next
section, α will be generated in (0, 1]. In this way, all the α points in the whole
interval and the MAPE’s at these points are calculated. Hence α, which give the
minimum of the calculated MAPE’s, is determined as the optimum parameters.
The computational steps mentioned above are employed for all the α points.

From Eq. (2.5), as the α value approaches 1, the CFA series approaches
the 1-AGO series, and as the α value approaches 2, the CFA series approaches
the 2-AGO series. Obviously, it can be seen that for each point the CFA value
x(α) (k) becomes larger with larger α, and the growing speed also increases with
larger α [16].

According to the theorem 3.3, if α ∈ (0, 1] the following first-order fractional

difference must be calculated to evaluate the x̂(0) (k) values,

(3.28) x̂(0) (k) = k 1−α (x̂(α) (k) − x̂(α) (k − 1)),

where k = 2, 3, ..., n. This shows that the error in x̂(0) (k) values is due to two
points of the CFA series. If α ∈ (1, 2], the following second-order fractional
12 U. Erdinc, H. Bilgil, and Z. Ozturk

difference must be calculated to evaluate the x̂(0) (k) values,

(3.29) x̂(0) (k) = k 2−α ∆2 x̂(α) (k) = k 2−α (x̂(α) (k) − 2x̂(α) (k − 1) + x̂(α) (k − 2)),

where k = 3, ..., n. If k = 2, x̂(α) (0), x̂(α) (1), and x̂(α) (2) values are needed
to determine the value of x̂(0) (2). However, according to equation (3.20), since
there is no x̂(α) (0) value in the CFA series, the second-order difference can only
be calculated for k = 3, 4, ..., n. In this case, although x̂(α) (1) is known as x(0) (1)
according to the initial value, the value of x̂(α) (2) can only be calculated with the
help of a first-order fractional difference. This will increase the total error rate.
According to equation (3.29), the error of restored value is effected by the errors
of the CFA series with three points, which will be another factor that increases
the error rate. Moreover, 99% optimal α values in the CFGM model are in the
range of [0.1). In the FGM model, 72% optimal α values are in the range of (0, 1)
(for more detailed information, see [16]). In the ECFGM model, 89% optimal α
values are in the range of (0.1) and 10% optimal α are obtained at 1.

All these processes are completed in about five seconds by writing a simple
FORTRAN code.

4. Validation of the ECFGM(1,1)

In this section, two instructive examples are given to demonstrate the effi-
cacy of the proposed model.

4.1. Example A

In this subsection, a numerical example is presented to show the computa-

tional steps of the ECFGM(1,1) model with the raw data
X (0) (k) = (13.21, 18.82, 26.45, 36.04, 42.34, 51.00, 59.12) . In this example, we
select the raw data as a monotone increasing series.

4.1.1. Selecting the optimal α

By using the Brute Force strategy, calculated MAPEs with α in the interval
(0, 1] with step 0.01 are given in Fig. 1. It can be seen that the values of MAPE
increase when the α moves away from its optimum value. For this example, the
optimal α is obtained at α = 0.14, and the MAPE is calculated for the optimal
α as MAPE=0.9843. It is seen that the optimal α is easily obtained by using
the Brute Force strategy. Furthermore, it is clear that the performance of the
proposed ECFGM(1,1) model is respectable.
A novel fractional forecasting model 13

2 .0

1 .8

1 .6
M A P E

1 .4

1 .2

1 .0

α
0 .0 0 .2 0 .4 0 .6 0 .8 1 .0
A L P H A

Figure 1: MAPEs of ECFGM model with α in (0, 1] for example A.

4.1.2. Computing the CFA of the original series and modelling the ECFGM(1,1)

Computation of the CFA of the original series is the first step to build
the ECFGM(1,1) model. For the optimum α = 0.14, the conformable frac-
tional accumulation series can be obtained using Eq. (3.2) as X (0.14) (k) =
(13.2100, 23.5789, 33.8615, 44.8014, 55.4095, 66.3329, 77.4234) .

The matrices B and Y can be constructed as

−18.3945
   
1 0.2325 10.3689
 −28.7202 1 0.0855   10.2826 
   
 −39.3314 1 0.0315   10.9399 
B=
 −50.1054
, Y =
 
 1 0.0116 
  10.6081 

 −60.8712 1 0.0043   10.9235 
−71.8781 1 0.0016 11.0905

Then we obtain the parameters â, b̂ and ĉ using the least squares solution
as
−1
[a, b, c]T = B T B BT Y
14 U. Erdinc, H. Bilgil, and Z. Ozturk

−•− R a w d a t a
−•− E C F G M
6 0

5 0

4 0

3 0

2 0

1 0

k
1 2 3 4 5 6 7

Figure 2: Actual values and forecasting values of example A.

and
a = −0.0136199261
b = 10.0950218700
c = −0.0670244100
By substituting the parameters into the response function Eq. (3.20) we have
(4.1) x̂(α) (k) = 754.3807 e−0.0136(1−k) − 741.1951 + 0.06612 e−k .
Then the restored can be obtained using (4.1) by k from 1 to 7 as
X̂ (0.14) (k) = (13.2100, 23.5396, 34.0206, 44.6491, 55.4247, 66.3485, 77.4223) .
Then the restored values can be obtained using the CFD in (3.23) as
X̂ (0) (k) = (13.2100, 18.7485, 26.9607, 35.0141, 43.0085, 51.0018, 59.0312) .
The original raw series X (0) (k) and predicted values X̂ (0) (k) are plotted in Fig.
2.

4.2. Example B

Different from example A, the raw data are not a monotone increasing series
here. The computational mechanism is similar to example A. In this example,
the raw data select as
X (0) (k) = (120.21, 131.83, 143.45, 150.02, 134.34, 121.04, 110.15) .
A novel fractional forecasting model 15

1 .5 0

1 .4 5

1 .4 0

1 .3 5
M A P E

1 .3 0

1 .2 5

1 .2 0

1 .1 5
0 .0 0 .2 0 .4 0 .6 0 .8 1 .0
α
A L P H A

Figure 3: MAPEs of ECFGM model with α in (0, 1] for example B.

4.2.1. Selecting the optimal α

The calculated MAPEs with α in the interval (0, 1] with step 0.01 is given
in Fig. 3. For this example, the minimum MAPE is calculated at optimal α as
α = 0.89. The MAPE is calculated for the optimal α as M AP E = 1.1836. It
is seen that the prediction performance of the proposed ECFGM(1,1) model is
successful again.

4.2.2. Computing the CFA of the original series and modelling the ECFGM

Computation of the conformable fractional accumulation series of the orig-

inal series is the first step to build the ECFGM model. For α = 0.89, the CFA
series can be obtained using Eq. (3.2) as,
X (0.89) (k) = (120.2100, 242.3621, 369.4831, 498.2851, 610.8281, 710.2157, 799.1407) .
16 U. Erdinc, H. Bilgil, and Z. Ozturk

The matrices B and Y can be constructed as

−181.2861 1 0.2325
   
122.1521
 −305.9223 1 0.0855   127.1210 
   
 −433.8841 1 0.0315   127.1210 
B= 
,
 Y = 
 −554.5566 1 0.0116   112.5430 

 −660.5219 1 0.0043   99.3876 
−754.6782 1 0.0016 88.9250

Then we obtain the parameters â, b̂ and ĉ using the least squares solution as

−1
[a, b, c]T = B T B BT Y

and
a = 0.1266697621
b = 184.88810740
c = −174.9875634
By substituting the parameters into the response function Eq. (3.20) we have

(4.2) x̂(α) (k) = −1413.1086 e0.12667(1−k) + 1459.6073 + 200.3681 e−k .

Then the restored can be obtained using (4.2) by k from 1 to 7 as

X̂ (0.89) (k) = (120.2100, 241.7407, 372.7221, 496.9157, 609.5692, 710.0100, 798.9388).

Then the restored values can be obtained using the CFD in (3.23) as

X̂ (0) (k) = (120.2100, 131.1594, 147.8062, 144.6525, 134.4718, 122.3227, 110.1547) .

The original raw series X (0) (k) and predicted values X̂ (0) (k) are plotted
in Fig. 4.

5. Validation of ECFGM(1,1) with a real case

One of the renewable clean energy sources is wind. Generally, comparing

with the many other energy sources, producing energy using wind has fewer
effects on the environment. There is no released emission that can pollute the air
or water and they do not require water for cooling. Wind turbines may also have
the benefit that reduce the amount of power generation from fossil fuels, which
outcomes in lower total air pollution and carbon dioxide emissions that could
help solve the shortage problem of energy. From prehistoric to today, human
beings have used wind energy for sailing, windmills, and wind turbines. Electric
generators convert wind energy to electrical energy [16].

One of the biggest countries with large land mass and coastline is China
has rich wind resources. With regard to the evaluations by China Meteorological
A novel fractional forecasting model 17

1 6 0 −•− R a w d a t a
−•− E C F G M

1 4 0

1 2 0

1 0 0 k
0 1 2 3 4 5 6 7

Figure 4: Actual values and forecasting values of example B.

Administration, based on the relatively low height of 10 m above ground, the

total theoretical wind power reserves in China 4350 GW, while the technically
exploitable wind resources estimated at 297 GW [9].

In this section, we use the novel ECFGM(1,1) model to predict wind en-
ergy consumption in China. The data set from [34, 2, 6] is used to test for
the efficacy and applicability of the proposed grey model. Furthermore, the
ECFGM model compared with the six effective models, including the GM(1,1),
EXGM(1,1), FAGM(1,1), FAGMO(1,1,k), PFAGM(1,1) and CFGM(1,1). The
response function of the grey models can be given as follows.

The response function of standart GM(1,1) model is obtained as

(5.1) x(1) (k) = 56.12166e0.22483(k−1) − 49.87166.

The response function of EXGM(1,1) is

(5.2) x(1) (k) = 67.78567e−0.20588(1−k) − 70.97226.

For FAGM(1,1) model, the response function is

(5.3) x(0.36872) (k) = 37.55924e0.17072(k−1) − 31.35924.

The response function of FAGMO(1,1,k) is

(5.4) x(1.13366) (k) = 283.1282e0.13851(k−1) − 30.4471k − 246.4811.
18 U. Erdinc, H. Bilgil, and Z. Ozturk

The response function of PFAGM(1,1) is,

(5.5) x(0.17874) (k) = 19.69290e0.17874k − 14.47094e−0.08124(k−1) − 2.87604.

The response function of CFGM(1,1) model is obtained as

(5.6) x(0.07) (k) = 81.63469e0.06355(k−1) − 75.38469.
Here, the optimal value of α is obtained as 0.07 by using the Brute Force strategy
for the CFGM (1,1).

The response function of ECFGM(1,1) model is obtained as

(5.7) x(0.3319) (k) = 66.64788e0.10331(k−1) + 3.98468e−k − 61.86377.

In addition, the mean absolute percentage error (MAPE) is used to assess

the prediction accuracy of these grey models. Firstly, we split the raw sequence
into two groups to build a model and test the model. The first group, including
the consumption from 2009 to 2017, is used to build models for the seven grey
models separately. The second group, including wind energy consumption from
2018 to 2020, is used to verify the prediction accuracy of these grey models. In this
section we enumerate all the values in the interval [0, 2] with step 0.0001, then use
the computational steps presented in Section 3.4 and select the α corresponding to
the minimum M AP Ef it as the optimal value. Optimum α is found as α = 0.3319
by using the Brute Force strategy and values of α and calculated MAPEs are
shown in Fig. 5. From figure 5 it is clear that the values of MAPE increase
when the α moves away from its optimum value. The optimal parameters are
calculated as α = 0.3319, a = −0.10331, b = 6.39111 and c = −4.39633. The

2 0
3 .7

3 .6
1 9

3 .5 1 8

3 .4 1 7

3 .3
1 6
M A P E
M A P E

3 .2
1 5
3 .1
1 4
3 .0
1 3
2 .9

2 .8 1 2
α α
0 .0 0 .2 0 .4 0 .6 0 .8 1 .0 1 .0 1 .2 1 .4 1 .6 1 .8 2 .0

A L P H A A L P H A

(a) (b)

Figure 5: MAPEs of ECFGM model for the real data set. (a) α ∈ (0, 1], (b)
α ∈ (1, 2]

prediction results and the mean absolute percentage errors of the recent models
are shown in Table 1.

According to the Table 1, the essential conclusions can be drawn as follows:

A novel fractional forecasting model 19

Table 1: The results generated by the proposed model and other comparative
grey models for forecasting values of China’s wind energy consumption (million
tonnes oil equivalent).

Year Actual Value GM (1,1) EXGM (1,1) FAGM(1,1) FAGMO(1,1,k) PFAGM CFGM ECFGM
2009 6.25 6.2500 6.2500 6.2500 6.2500 6.2500 6.2500 6.2500
2010 10.10 14.1489 9.5314 10.9059 10.7861 10.8294 10.2050 10.0534
2011 15.91 17.7160 16.8448 15.9000 15.9521 15.9001 15.8554 16.0464
2012 21.72 22.1824 22.5845 21.4733 21.7572 21.5803 22.0785 22.2010
2013 31.95 27.7749 28.4425 27.8495 28.3244 28.0628 28.9531 28.8472
2014 35.32 34.7772 35.2004 35.2395 35.7864 35.5571 36.5538 36.2435
2015 42.03 43.5450 43.3416 43.8718 44.2900 44.3005 44.9565 44.5937
2016 53.64 54.5231 53.2845 54.0032 53.9997 54.5677 54.2405 54.0796
2017 66.75 68.2690 65.4787 65.9310 65.1024 66.6817 64.4903 64.8821
MAPEfit 8.4111 3.6099 3.1595 3.1901 3.1104 3.0416 2.8418
2018 82.82 85.4804 80.4526 80.0030 77.8110 81.0239 75.7962 77.1931
2019 93.31 107.0310 98.8466 96.6289 92.3686 98.0478 88.2555 91.2232
2020 107.30 134.0148 121.4447 116.2921 109.0538 118.2925 101.9728 107.2063
MAPEpre 14.2714 7.3248 5.1128 2.8971 5.8303 6.2875 3.0392
MAPE 9.8762 4.5386 3.6478 3.1169 3.7904 3.8531 2.8912

G M (1 ,1 )
1 0

6
M A P E

5 E X G M

4 P F A G M C F G M
F A G M
F A G M O
3 E C F G M

Figure 6: Total MAPE values of the models

i. Table 1 reveals that, the seven grey models’ M AP Ef it values are 8.4111%,
3.6099%, 3.1595%, 3.1901%, 3.1104%, 3.0416% and 2.8418%, respectively.
So that, in the fitting period, the fitting performance of the ECFGM(1,1)
model is best.

ii. The M AP Epre values of seven models are calculated as 14.2714%, 7.3248%,
5.1128%, 2.8971%, 5.8303%, 6.2875% and 3.0392%, respectively. At this
point, the FAGMO(1,1,k) model has the smallest M AP Epre value while
20 U. Erdinc, H. Bilgil, and Z. Ozturk

Table 2: Forecasted wind energy consumption of China

Years Forecasting values of wind energy consumption

2021 125.4042
2022 146.1105
2023 169.6550
2024 196.4082
2025 226.7862

−•− R a w d a t a
2 7 5 −•− E C F G M
2 5 0
W in d e n e r g y c o n s u m p tio n o f C h in a

2 2 5
s
u e
2 0 0 g v a l
in
a s t
u e
s o re c
1 7 5
d v a l F
d fitte
1 5 0 l a n
c tu a
A
1 2 5

1 0 0

7 5
v ie w
rg e d
E n la
5 0

2 5 E n d o f d a ta f o r b u ild in g th e m o d e l
0
2 0 0 9 2 0 1 2 2 0 1 5 2 0 1 8 2 0 2 1 2 0 2 4
Y e a rs

Figure 7: Actual, fitting and forecasting values of wind energy consumption of

China.

the ECFGM(1,1) model has the second smallest M AP Epre .

iii. It is observed from Table 1 that for the whole period, the total MAPEs of
seven models are 9.8762%, 4.5386%, 3.6478%, 3.1169%, 3.7904%, 3.8531%
and 2.8912%, respectively. Thence, in the whole period, the performance
of the ECFGM(1,1) model is the best (see Fig. 6).

iv. The wind energy consumption of China can be estimated properly with
the proposed ECFGM (1,1) model. Hence, the forecasting values of wind
energy consumption of China are given in Table 2.

It is seen from Figure 7 that the actual (red line) and forecasted (black
line) values of wind energy consumption of China are matched to each other.
A novel fractional forecasting model 21

6. Conclusion

Since the fractional calculations are most important to the grey prediction
model, there are many scholars proposing new methods on the fractional grey
models. Hence, a novel optimization for the CFGM (1,1) and EXGM(1,1) mod-
els have been developed in this study. The results of the numerical examples
indicated that the proposed grey prediction model aims to achieve very effective
performance. The structural parameters (a, b and c) of the model can be dy-
namically adjusted according to the real world systems. The optimal value of
fractional order, α, is calculated by using the brute-force approach. The pro-
posed ECFGM(1,1) model is suitable for predicting the data sequence with the
characteristics of non-homogeneous exponential law. Comparison results indi-
cate, ECFGM(1,1) performs better than those achieved by the other grey models
such as GM(1,1), EXGM(1,1), FAGM(1,1), FAGMO(1,1,k), PFAGM(1,1) and
CFGM(1,1). However, they can all be employed for estimations.

Wind energy with cleanliness and pollution-free will have a positive attitude
on global energy transformation. Because of this, research on more accurate pre-
diction of wind energy consumption is quite important for wind power generation.
Therefore, the wind energy consumption of China is predicted successfully by us-
ing a novel proposed fractional grey model based on the conformable fractional
difference and conformable fractional accumulation in the paper. The forecast-
ing results show that wind energy consumption of China will develop rapidly in
recent years, and will reach approximately 200 million tones oil equivalent by
2023.

Using two examples and a case study in Sections 4 and 5, we show that the
MAPE of the ECFGM(1,1) model is very low.

The proposed ECFGM (1,1) model may play an important role in enriching
the theoretical system of grey forecasting theory and it can be used for other real
cases of small sample forecasting in the future. Besides, the combination of other
fractional forecasting models, especially for the time series with highly effective,
is also an interesting direction for next studies.

Acknowledgements

The authors would like to thank the reviewers for their valuable comments
and suggestions which have improved many aspects of this article.
22 U. Erdinc, H. Bilgil, and Z. Ozturk

Conflict of interest

The author declares no conflict of interest in this paper.

REFERENCES

[1] Bilgil, H. (2021). New grey forecasting model with its application and computer
code, AIMS Mathematics, 6, 1497–1514.
[2] Cao, J.; Liu, L.; Yang, L. and Xie, S. (2020). Application of a novel fractional
order grey support vector regression model to forecast wind energy consumption
in china, Journal of Advances in Mathematics and Computer Science, 35 (2),
58–69.
[3] Cui, J.; Liu, S.; Zeng, B. and Xie, N. (2013). A novel grey forecasting model and
its optimization, Appl. Math. Model., 37, 4399–4406.
[4] Deng, J. L. (1982). Control problems of grey systems, Syst. Control Lett., 1,
288–294.
[5] Ene, S. and Ozturk, N. (2018). Grey modelling based forecasting system for
return flow of end-of-life vehicles, Technol. Forecast. Soc. Change, 117, 155–166.
[6] https://chinaenergyportal.org/2020-electricity-other-energy-statistics prelimi-
nary/. (2021). 2020 electricity and other energy statistics (preliminary). China
Energy Portal.
[7] Javed, S.A. and Liu, S. (2018). Predicting the research output/growth of selected
countries: application of even gm (1,1) and ndgm models, Scientometrics, 115,
395–413.
[8] Jiang, J.; Feng, T. and Liu, C. (2021). An improved nonlinear grey bernoulli
model based on the whale optimization algorithm and its application, mathe-
matical problems in engineering, Mathematical Problems in Engineering, 2021,
6691724.
[9] Khalil, R.; Al Horani, M.; Abdelrahman, Y. and Mohammad, S. (2014). A new
definition of fractional derivative, J Comput Appl Math, 264, 65–70.
[10] Li, S.; Ma, X. and Yang, C. (2018). A novel structure-adaptive intelligent grey
forecasting model with full-order time power terms and its application, Com-
puters & Industrial Engineering, 120, 53–67.
[11] Liu, L.; Chen, Y. and Wu, L. (2021). The damping accumulated grey model and
its application, Communications in Nonlinear Science and Numerical Simulation,
95, 105665.
[12] Liu, L. and Wu, L. (2021). Forecasting the renewable energy consumption of the
european countries by an adjacent non-homogeneous grey model, Appl. Math.
Modell., 89, 1932–1948.
[13] Ma, X. (2016). Research on a novel kernel based grey prediction model and its
applications, Mathematical Problems in Engineering, 2016:1–9.
A novel fractional forecasting model 23

[14] Ma, X.; Liu, Z. and Wang, Y. (2019). Application of a novel nonlinear multivari-
ate grey bernoulli model to predict the tourist income of china, J. Comput. Appl.
Math., 347, 84–94.
[15] Ma, X. and Liu, Z.B. (2017). Application of a novel time-delayed polynomial
grey model to predict the natural gas consumption in china, J. Comput. Appl.
Math., 324, 17–24.
[16] Ma, X.; Wu, W.; Zeng, B.; Wang, Y. and Wu, X. (2020). The conformable
fractional grey system model, ISA Transactions, 96, 255–271.
[17] Mao, S.; Gao, M.; Xiao, X. and Zhu, M. (2016). A novel fractional grey system
model and its application, Appl Math Model, 40, 5063–5076.
[18] Sahin, U. and Sahin, T. (2020). Forecasting the cumulative number of confirmed
cases of covid-19 in italy, uk and usa using fractional nonlinear grey bernoulli
model, Chaos Solitons Fract., 138, 109948.
[19] Wang, Z.X.; Li, Q. and Pei, L.L. (2018). A seasonal gm (1,1) model for forecasting
the electricity consumption of the primary economic sectors, Energy, 154, 522–
534.
[20] Wei, B.; Xie, N. and Hu, A. (2018). Optimal solution for novel grey polynomial
prediction model, Applied Mathematical Modelling, 62, 717–727.
[21] Wei, M.; Li, Q. and Zeng, B. (2016). Study on fractional order grey reducing
generation operator, Grey Syst Theory Application, 6, 80–95.
[22] Wu, L.; Liu, S.; Yang,Y.; Ma, L. and Liu, H. (2016). Multivariable weakening
buffer operator and its application, Inform. Sci., 339, 98–107.
[23] Wu, L.; Liu, S.; Yao, L.; Yan, S. and Liu, D. (2013). Grey system model with
the fractional order accumulation, Commun. Nonlinear Sci. Numer. Simul., 18,
1775–1785.
[24] Wu, L. F.; Liu, S. F.; Chen, D.; Yao, L. and Cui, W. (2014). Using gray model
with fractional order accumulation to predict gas emission, Nat. Hazards, 71,
2231–2236.
[25] Wu, L.Z.; Li, S.H.; Huang, R.Q. and Xi, Q. (2020). A new grey prediction model
and its application to predicting landslide displacement, Appl. Soft Comput., 95,
106543.
[26] Wu, W.; Ma, X.; Wang, Y.; Cai, W. and Zeng, B. (2020). Predicting China’s
energy consumption using a novel grey riccati model, Appl. Soft Comput., 95,
106555.
[27] Wu, W.; Ma, X.; Zeng, B.; Wang, Y. and Cai, W. (2018). Application of the novel
fractional grey model fagmo(1,1,k) to predict china’s nuclear energy consumption,
Energy, 165, 223–234.
[28] Wu, W.; Ma, X.; Zhang, Y.; Li, W. and Wang, Y. (2020). A novel conformable
fractional non-homogeneous grey model for forecasting carbon dioxide emissions
of brics countries, Sci Total Environ, 707, 135447.
[29] Xie, W.; Caixia, L.; Wu, W.; Weidong, L. and Chong, L. (2020). Continuous
grey model with conformable fractional derivative, Chaos, Solitons Fractals, 139,
110285.
[30] Xie, W.; Wu, W. Z.; Liu, C. and Zhao, J. (2020). Forecasting annual electric-
ity consumption in china by employing a conformable fractional grey model in
opposite direction, Energy, 202, 117682.
24 U. Erdinc, H. Bilgil, and Z. Ozturk

[31] Yang, Y. and Xue, D. (2016). Continuous fractional-order grey model and elec-
tricity prediction research based on the observation error feedback, Energy, 115,
722–733.
[32] Yuxiao, K.; Shuhua, M.; Yonghong, Z. and Huimin, Z. (2020). Fractional deriva-
tive multivariable grey model for nonstationary sequence and its application,
Journal of Systems Engineering and Electronics, 31, 1009–1018.
[33] Zeng, B.; Tan, Y.; Xu, H.; Quan, J.; Wang, L. and Zhou, X. (2018). Forecasting
the electricity consumption of commercial sector in hong kong using a novel grey
dynamic prediction model, J. Grey Syst., 30, 157–172.
[34] Zhang, P.; Ma, X. and She, K. (2019). A novel power-driven fractional accumu-
lated grey model and its application in forecasting wind energy consumption of
china, Plos One, 14, e0225362.
[35] Zhang, Y. G.; Xu, Y. and Wang, Z. P. (2009). Gm(1,1) grey prediction of lorenz
chaotic system. Chaos Solitons Fract., 42, 1003–1009.
[36] Zhou, W. and He, J. M. (2013). Generalized gm (1,1) model and its application
in forecasting of fuel production. Appl. Math. Modell., 37, 6234–6243.