Professional Documents
Culture Documents
discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/284010053
CITATIONS READS
0 62
6 authors, including:
All in-text references underlined in blue are linked to publications on ResearchGate, Available from: Kin Keung Lai
letting you access and read them immediately. Retrieved on: 21 October 2016
Expert Systems With Applications 45 (2016) 331–340
a r t i c l e i n f o a b s t r a c t
Keywords: Transportation disruption is the direct result of various accidents in supply chains, which have multiple neg-
Transportation disruptions ative impacts on supply chains and member enterprises. After transportation disruption, market demand
GM(1,1) model
becomes highly unpredictable and thus it is necessary for enterprises to better predict market demand and
Neural network
optimize purchase, inventory and production. As such, this article endeavors to design an improved model
Prediction
of grey neural networks to help enterprises better predict market demand after transportation disruption
and then the empirical study tests its feasibility. This improved model of grey neural networks exceeds the
conventional grey model GM(1,1) with respect to the fact that the raw data tend to show exponential growth
and data variation is required to be moderate, demonstrating the good attribute of nonlinear approximation
in terms of neural networks, setting up standards for selecting the number of neurons in the input layer of
BP neural networks, increasing the fitting degree and prediction accuracy and enhancing the stability and
reliability of prediction. It can be applied to sequential data prediction in transportation disruption or muta-
tion, contributing to the prediction of transportation disruption. The forecasting results can provide scientific
evidence for demand prediction, inventory management and production of supply chain enterprises.
Crown Copyright © 2015 Published by Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.eswa.2015.09.052
0957-4174/Crown Copyright © 2015 Published by Elsevier Ltd. All rights reserved.
332 C. Liu et al. / Expert Systems With Applications 45 (2016) 331–340
found that transportation disruption is one of the factors generat- 2. An improved prediction model of grey neural networks
ing supply chain risks. He stated that risk can be mitigated by in-
creasing the visualization of the transportation process and optimiz- Deng (1982) proposed the grey system theory which addresses
ing the inventory decisions and partnership. Wilson (2007) examined samples of some known and some unknown information and is an
the effects of transportation disruption on the attributes of supply uncertain system of inadequate information. The advantages are a
chains by systematic dynamics and found that transportation dis- simple modelling process and concise expressions, which are exten-
ruption affects the first suppliers and their inventory most. Zhang sively employed in multiple disciplines. By contrast, this prediction
and Lam (2015) investigate the transportation disruption caused by method is suitable for sequences with stable rates of variation. In par-
typhoons at ports and the loss they cause and finds that loss con- ticular, when the sequence of data changes dramatically, where devi-
sists of 4 parts: reputation loss, shippers’ loss, transporters’ loss ation varies in terms of prediction results, many scholars have made
and port operators’ loss. Taking China’s railways and airline com- improvements accordingly (Chang, Li, Huang, & Chen, 2015; Cristóbal
panies as examples, Ouyang, Pan, Liu, and He (2015) discuss mea- et al., 2015; Zeng & Liu, 2011).
sures to mitigate the fragility of various transportation systems in Neural networks refer to concurrent nonlinear dynamic systems
terms of the complementary relationship of the infrastructure sys- on a large scale, with strength to handle nonlinear problems. Partic-
tem. Hishamuddin, Sarker, and Essam, (2013), (2015) consider the ularly, back propagation is most widely employed, in e-commerce,
two-stage recovery model of production and inventory in transporta- industry, environmental science and many other fields (Chi & Zhao,
tion disruption, and find that the optimal order quantity and pro- 2012; Ray, Mukhopadhyay, Datta, & Pal, 2013; Wang, Zeng, Feng, &
duction can be determined and the total cost can be minimized Xia, 2013). In contrast, this method is likely to show flaws such as
during the recovery period. Khademi et al. (2015) take Tehran with local minimizers in longer operations.
frequent earthquakes as an example, analyze the fragility of trans- There are merits and demerits of each algorithm and thus it is
portation networks, and point out that after disasters the demand common to integrate two or more methods and build a combined
for tourism becomes highly uncertain and the traditional methods model (Bates & Granger, 1969). Some scholars have endeavored to in-
cannot be applied to predictions. They recommend some methods corporate the grey model with other methods to build a combined
to evaluate the loss caused by earthquakes and recover the trans- model, including particle swarm optimization (Hodzic & Tai, 2016),
portation. Bravo and Vidal (2013) develop an optimal model for sup- DEMATEL method (Rajesh & Ravi, 2015), uncertainty theory (Memon,
ply chain goods by mathematical planning and provide an exam- Lee, & Mari, 2015), multivariate model (Intharathirat, Salam, Kumar,
ple and build a model in terms of abnormal transportation. The & Untong, 2015), fuzzy mathematics method (Dewangan, Gangopad-
literature discussed above has shown that many researchers of- hyay, & Biswas, 2015) and Markov method (Xie, Yuan, & Yang, 2015),
fer strategies for transportation disruption from multiple perspec- among many others. A number of researchers have attempted to inte-
tives and note that the market demand is highly unpredictable af- grate grey prediction model and neural network to handle nonlinear
ter transportation disruption, affecting the production and inventory problems, tackle unpredictable demand and increase prediction pre-
significantly. cision (Alvisi & Franchini, 2012; Hsu, 2011; Pai et al., 2015; Zhang, Xu,
The uncertainty of market demand after transportation disrup- Cui, He, & Tian, 2012).
tion can be optimized from many aspects such as production, inven- It can be observed that in the combination of grey model and neu-
tory and distribution in supply chain enterprises. Chiu, Lee, Chiu, and ral network model, the number of neurons in the input layers of BP
Cheng (2013) explore the supply chain in multinational companies to neural networks can be either 1 or random, which can control the in-
ensure that products can be sold in multiple places; quality is guar- trinsic errors of the prediction model of neural networks, whereas
anteed; the optimal quantity of products and means of transportation the deviation of grey model and neural network model cannot be
can be determined, thus minimizing the cost of production, inventory controlled at the same time. This leads to failure to optimally con-
and transportation of supply chain members. Based on Generic Bill of trol the total deviation. As such, based on the grey model and the
Materials (GBOM), Shu, Chen, Wang, and Lai (2014) investigate the neural network model, this article designs an improved prediction
risk control of supply chain enterprises and tackle the unpredictabil- model of grey neural networks, setting up the standards for choos-
ity of production after disruption by combining random simulation ing the number of neurons in the input layer of BP neural networks,
and neural networks. Costantino, Gravio, Shaban, and Tronci (2015) namely, the optimal number of dimensions of the improved grey pre-
design the real time statistics process control (SPC) system to coun- diction model is the figure for neurons in the input layer of BP neural
teract the whip effect through control charts, hence achieving the sta- networks. This improved prediction model for determining the num-
bility of competitive inventory. The results show that SPC system can ber of neurons in the input layer of BP neural networks exceeds the
satisfy the principle of smooth complements. Pasandideh, Niaki, and conventional grey GM(1,1) which has strict requirements for the se-
Asadi (2015) consider a three-tier supply chain consisting of manu- quence of raw data, demonstrating the good attribute of nonlinear
facturers, distribution centers with uncertain service and customers approximation in terms of neural network, increasing the fitting de-
to achieve optimization of multiple products and multiple stages in gree and prediction accuracy, and enhancing the stability and relia-
supply chains. They deploy the model of mixed integral linear plan- bility of prediction. It can be applied to sequential data prediction in
ning and the method of multiple-target decisions, to achieve mini- transportation disruption or mutation.
mization of total cost and maximization of average quantity of goods
for customers. 2.1. A grey prediction model
Market demand is unpredictable and highly uncertain particularly
after transportation disruption and thus enterprises have to manage 2.1.1. Traditional GM(1,1) model
their inventory and production accordingly. Now, if the market de- GM(1,1) is a fundamental prediction model based on the grey the-
mand is well predicted, the cost of inventory and production as well ory. GM is formed by the initial letters of the grey model; regard-
as risk can be reduced and sales can be optimized. Little research so ing (1.1), the first 1 means that the prediction model only consists
far has involved prediction of market demand for supply chain en- of a 1-step equation and the second 1 means that there is only one
terprises after transportation disruption. For this reason, this article dependent variable in the model. As such, GM(1.1) is based on the
designs an improved model of grey neural networks to help supply premise that 1-step equation and 1 dependent variable are satisfied.
chain enterprises better forecast market demand after transportation In the Grey prediction model GM(1,1), the original sequence is cumu-
disruption. Also, an empirical study is conducted to test the feasibility latively produced as the time sequence increases progressively. The
of the model. corresponding approximate differential equations are constructed to
C. Liu et al. / Expert Systems With Applications 45 (2016) 331–340 333
display the development of data sequence. What follows are the pro- n − 1 equations are insolvable in terms of two unknown co-
cesses of model construction and solutions. efficients (a and b), and then the method of minimal squares
is used to achieve  = (BT B)−1 BT Y , and the approximate solu-
(1) If the original time sequence is given as x(0) = (x(0) (1),
tions â and b̂ can be achieved.
x(0) (2), . . . , x(0) (n)), in which n is the figure for the origi-
(5) Put â and b̂ into the grey differential equation, and the discrete
nal data and x(1) is cumulatively generated data in relation
corresponding functions are achieved:
to x(0) , i.e., 1-AGO, then x(1) = (x(1) (1), x(1) (2), . . . , x(1) (n)) is
achieved, in which x(1) (i) = ni=1 x(0) (i), i = 1, 2, . . . , n. If there x
(1) (1) = x(0) (1)
i≥1 (5)
x
is no special notification, value i will remain the same in the (1) (i + 1) = b̂
â
+ e−âi x(1) (0) − b̂
â
following. Here, GM(1.1) model is shown as follows:
x(0) (i) + ax(1) (i) = b(a, b areunderminedcoefficients) (1) (6) x
(1) (i + 1) is recovered by cumulative reduction and the pre-
diction value of the original data is achieved.
(2) Test the quasi-smoothness of x(0) and the quasi-index pat-
tern of x(1) . The sequence smoothness can be achieved by x
(0) (1) = x(0) (1)
i≥1 (6)
continuous functions while if the time sequence is com- x (i + 1) = x
( 0) (1) (i + 1) − x
(1) (i)
posed of discrete individual points, it is difficult to obtain
smoothness by this derivation. Here, the time sequence is (7) Precision test
supposed to show the attribute of being smooth, and the
x(0) (i+1) p(i+1)
• Pointwise test
quasi-smooth coefficient is p(i + 1) = . As p(i)
< 1 Absolute error: ε (0) (i) = x(0) (i) − x
(0) (i)
x(1) (i)
ε(t )
and p(i) < 0.5 are satisfied, the original time sequence x(0) is Relative error: e(i) = (0) × 100%
x (i)
considered a quasi-smooth sequence (Liu & Guo, 1991). The • Posterior difference ratio
normal non-negative quasi-smooth is derived by accumulat-
1 (0) 2 1 (0)
n n
ing it once and the randomness is reduced, showing a gen-
S1 = x (i) − x̄(0) , in which x̄(0) = x (i)
eral trend of exponential growth. p(i + 1) < 0.5 together with n n
(1) (0) (i+1)+x(1) (i) i=1 i=1
σ (1) (i + 1)= x x(1()i+1
(i)
) x
=
x(1) (i)
= 1+ p(i + 1) < 1.5 are
given; here x(1) can be derived by accumulating the original 1
n 2 1
n
S2 = ε(i) − ε̄ (0) in which ε̄ (0) = ε (0) (i)
time sequence x(0) once, which satisfies the quasi-index pat- n n
i=1 i=1
x(1) (i+1)
tern. The quasi-index pattern target is σ (1) (i + 1) = . If
x(1) (i)
S2
σ (1) (i) ∈ 1, 1.5 is given, then the quasi-index pattern is gen- C=
S1
erally satisfied (Liu & Guo, 1991).
(3) If p(i + 1) < 0.5, σ (1) (i + 1) ∈ 1, 1.5 is given, then the albe- Accuracy range is acceptable if C ∈ [0.35, 0.65] (Wa, 2010). When
faction differential equation of GM(1,1) is: C becomes smaller, the range of prediction errors decreases and
therefore the accuracy increases.
dx(1)
+ ax(1) = b (2) • Frequency of minor errors
dt
P = {P |ε(i) − ε̄| < 0.6745S1 }
where, a is the coefficient of x(1) and dx(1) is the differential of x(1) ,
and thus a can show the tendency of x(0) and x(1) . If a is grey develop- Accuracy range is acceptable if P ∈ [0.7, 0.95] (Wa, 2010). When
ment scale and b is the data derived from the background value, in- P becomes larger, the probability of minor errors increases and
dicating the changing patterns of the data, this can be viewed as the therefore the accuracy increases.
significant evidence to differentiate the grey model and the model
based on the general input and output. If b is the grey action quan- (8) Test of fitting degrees
tity; t indicates time; and x(1) (i) − x(1) (i − 1) = x(0) (i), x(0) (i) is the
(1) Grey correlation analysis involves judging degrees of closeness by
grey derivative. (dxdt ) + ax(1) = b is the simple transformation from determining approximation degrees of geometrical shapes related to
x(0) (i) + ax(1) (i) = b in Eq. (1). The background value x(1) (i) is the grey the original and forecast sequences. Then, the tendency of dynamic
number; x(1) (i) and grey derivatives do not satisfy the relationship development of the original sequence is quantified and analyzed by
of flat projection. As the time sequence consists of discreet individ- comparison. The procedure is to make the prediction and original se-
ual points, there are boundary points rather than inner points and quence non-dimensional, and to calculate the correlation coefficients
thus it is vacant between points. For this reason, the inner points are and degrees. Next, the closeness degree is determined by the correla-
derived from the mean value of the right and left boundary points. tion degree (Table 1).
Here, x(1) (i) and the grey derivative x(0) (i) fail to satisfy the rela- In the first step, variables become non-dimensional. Then calcu-
tionship of flat projection. The traditional method is that the se- late the initial image of the original and forecast sequence. Here, the
quence x(1) uses the adjacent mean value to derive the sequence w(1) , initial image is selected.
that is, w(1) = (w(1) (1), w(1) (2), . . . , w(1) (n)), in which w(1) (i + 1) =
0.5x(1) (i) + 0.5x(1) (i + 1). The albefaction differential equation in (2) x(0) x
(0)
x(0) = (0) , x
(0) =
can be transformed into the following grey differential equation: x ( 1)
x(0) (1)
x(0) (i) + aw(1) (i) = b (3) In the second step, calculate the related number ξ i :
(4) Construct min min |y(k) − xi (k)| + ρ max max |y(k) − xi (k)|
⎡ (1matrix B, ⎤constant matrix Y and matrix A, in which
⎡ ⎤
−w ) (2) 1
ξi (k) = i k i k
x(0) (2) |y(k) − xi (k)| + ρ max max |y(k) − xi (k)|
⎢ . .⎥
.. ⎦, Y = ⎣ . . . ⎦, A =
a i k
B=⎣ .. , and therefore
b
( ) x (0) (n ) where ρ is the resolution ratio. If ρ becomes smaller, the resolution
−w (n)
1 1
ratio increases. If ρ ≤ 0.5463, the resolution ratio is the best. Gener-
Eq. (2) can be equally changed into:
ally, if ρ = 0.5 (Liu, Dang, & Fang, 2010), the calculation remains the
Y = BA (4) same here.
334 C. Liu et al. / Expert Systems With Applications 45 (2016) 331–340
Table 1
Accuracy test.
Levels Results Average relative error (ARE) Posterior difference ratio C Frequency of minor errors P
In the third step, calculate the correlation degree ς i : (3) With fixed dimensional numbers, such as 5-dimension, the
model set with n − 4 sub-sequences can be built.
1 n
ςi = ξi (k), k = 1, 2, . . . , n (0)
n k=1 x (1), x(0) (2), x(0) (3), x(0) (4), x(0) (5)
(0)
In the fourth step, consider the fitting degree. If ς i ≥ 0.6, the fitting x (2), x(0) (3), x(0) (4), x(0) (5), x(0) (6)
degree is good and satisfies the fitting requirements. If ς i < 0.6, the ···
(0)
fitting degree is worse and cannot satisfy the fitting requirement (Li x (n − 4), x(0) (n − 3), x(0) (n − 2), x(0) (n − 1), x(0) (n)
& Hu, 2005).
The treatment methods of other dimension numbers are ex-
2.1.2. An improved GM(1,1) model actly the same.
In building of the actual prediction model, the forecasting data are (4) Make predictions by sub-models of the fixed dimensional
more likely to be affected by short-term data rather than the long- numbers in the traditional models GM(1,1) and test the fit-
term data. As a result, if all data in the original sequence are used ting degree and accuracy, determine the corresponding mean
to build the model, forecasting results are excessively affected and value and then compare each mean value, to select the optimal
interfered by long-term data, particularly when the long-term data model to make predictions from the model set.
are mutated and disrupted, and thus the recent changeable pattern
does not show. In order to improve the traditional Model GM(1,1), 2.2. The improved neural network prediction model
the index pattern should be satisfied or approximately satisfied; the
speed of change of data sequence should not be too fast; the fitting 2.2.1. Back-propagation neural network
degree and forecasting accuracy should be improved; stability and At the moment, in actual use, most neural network models adopt
reliability should be enhanced; and the isochronal sequence within either feed-forward back-propagation network (BP network) or oth-
x(0) (n) should be included in data for building the models. As such, ers, including the multi-layer feed-forward error back-propagation
two methods are used to build the models. Firstly, change of number network, containing input, output and other implicit layers. Each
of dimensions shows that x(0) (n) is the termination of the value se- layer is made up of a number of neurons and the super-stratum and
lected for building the models and then the values selected forward substratum neurons are connected via weight; complete connection
can form a set of forecasting models. The optimal model can be se- is used to combine the neurons between layers; and there is no con-
lected from the prediction models for future prediction, which can nection between neurons within layers. When the learning samples
be called the changeable starting point model. Another point to note are provided for the neural network and are then spread forward
is the fixed number of dimensions. That is to say, the oldest informa- through advanced line in the network, if the error between the out-
tion is deleted and the latest information is embedded for continuous put results and target output is beyond expectations, this forward
metabolism, which is called the metabolism model. spreading process reverses to the backward process. The mistaken
The model building and solutions of the changeable models are signals return through the original route and the mistaken signals are
shown as follows. decreased by modifying the weight of neurons in each layer. As the
modification of the error of backward spreading continues, the corre-
(1) The original time sequence is x(0) = (x(0) (1), x(0) (2), . . . , sponding accuracy of the network to the input mode increases until
x(0) (n)), in which n is the number of the original data; it meets the requirements.
(2) Model GM(1,1) built on x(0) = (x(0) (1), x(0) (2), . . . , x(0) (n)) is BP network consists of the input, output and several explicit lay-
the complete data model; and Model GM(1,1) is built on x(0) = ers. Cybenko (1989) and Ito (1991) showed that the three-layer BP
(x(0) (1), x(0) (2), . . . , x(0) (n)), i ≥ 2 is the partial data model; neural networks can satisfy the omnipotent approximation theorem,
(3) The number of the least time sequence is 4 (Deng, i.e., in the three-layer neural networks containing one explicit layer,
2005), and thus a four-dimensional sequence is built: if the number of the explicit points is sufficient, any continuous func-
(x(0) (n − 3), x(0) (n − 2), x(0) (n − 1), x(0) (n)). This continues tion can be approximated within the boundaries, which, however,
one by one, and finally n-dimensional sequence can be built: does not mean that the three-layer neural networks are the most rea-
(x(0) (1), x(0) (2), . . . , x(0) (n)), the original time sequence, and sonable. In terms of the same target function, Tamura and Tateishi
n − 3 sub-sequences can be built in total; (1997) showed that less neurons are used in the four-tier structure
(4) Make predictions by each sub-sequence in the traditional than in the three-layer neural structure. The implicit layer is more
Model GM(1,1) and then test the fitting degree and accuracy, than one layer and can improve the network training in terms of
and select the optimal models to make predictions from the speed. However, it takes longer for training in actual applications. The
set of models. training speed can be achieved by increasing the number of implicit
The model building and solutions of the metabolism models are layers which simulates nonlinear continuous projections. Therefore,
shown as follows. three-layer BP neural network with one implicit layer is selected for
prediction here. The input layer of the BP neural networks is the sam-
(1) The original time sequence is x(0) = (x(0) (1), x(0) (2), . . . , ple data, while the output layer is the prediction results.
x(0) (n)), in which n is the figure for the original data; (1) The design of the model structure and functions selected
(2) n − 3 dimensions in the n-dimensional original time sequence What BP model actually completes is the projection of input sam-
are fixed, such as four dimensions, five dimensions, till n- ple space on the output expectation space: the output sample of
dimension; ai → bj is i-dimension; the output expectation is j-dimension; and
C. Liu et al. / Expert Systems With Applications 45 (2016) 331–340 335
the dimensional number is the figure for the corresponding neu- 2.2.2. The improved back-propagation network
rons. The dimensional numbers of the input, output space show the • Attached momentum method
characteristics of vectors, and the number of neurons in each layer Rumelhart et al. (1986) put forward an approach to improve the
is determined by features of the historic data and prediction in the BP training time to ensure process stability. This method adds the
training. previous weighting variation of a positive proportion to each weight
If there are fewer neurons in the implicit layer, it is very difficult modulating value. The specific representation is as follows.
for the network to identify the sample and the training is hard to com-
∂ ET
plete. At the same time, the fault tolerance declines. If the number is ω(t + 1) = ω(t ) − (1 − α)η + α ω(t )
bigger, then the iterations increase and it takes longer to train the
∂ω(t )
network. Meanwhile, the universality of the network decreases, low- where ω(t ) = ω(t ) − ω(t − 1); α is the momentum factor with
ering the prediction capacity. In the specific design, the first step is 0.95 as the normal value. Here, the modification of weight value is
to determine the number of neurons in the implicit layer in terms of added to the memory of the modification direction for the previous
the experiential formula. Next, training and contrast are followed for moment of weight value.
different numbers of neurons. Finally, the number of neurons is de- When the attached momentum method is used to modify the
termined. The number of neurons in the implicit layer is determined weight of the network, the function of error on slopes is also consid-
by Kolmogorov Theory (Qian & Gong, 2000). If i is the number of neu- ered and the influence of the changing tendency of the error curved
rons in the implicit layer and n is the number of neurons in the input surface is also taken into account. The function is similar to a low
layer, the equation is as follows. filtering device which ignores minor changes. Without the attached
momentum, the network might fall into local minima. The function
i = 2n + 1 of the attached momentum is to skip the minimal values. The nature
The function of the neuron transmission in the network implicit of the attached momentum is to pass the effect of the final weight
layer is Sigmoid (Rumelhart, Hintont, & Williams, 1986); the neu- variation through the momentum factor. In the training process, the
ron transmission in the input layer adopts the linear function Pure- conditions to determine the attached momentum method are as
lin (Tian, 2009); and the training function uses the Trainbr approach follows.
based on the Bayes algorithm (Tian, 2009), which can increase net- 0, SSE (t ) > 1.04SSE (t − 1)
work universality. α= 0.95, SSE (t ) < SSE (t − 1)
(2) The pre-treatment of sample data α , SSE (t − 1) ≤ SSE (t ) ≤ 1.04SSE (t − 1)
In reality, most data are incomplete, noisy and inconsistent and
need to be packed by accumulating, deleting redundant features or where SSE is the square sum of the network output errors.
• Self-adaption learning ratio
clustering; it is necessary to normalize the input and output sample
data in the network. That is to say, the quantitative value is limited in Generally, the principles of modulating learning rate involve
the range of [0.1, 0.9]; the normalizing equation is shown as follows. checking whether the corrected weight reduces the error function ac-
tually. If so, the learning rate is small and some amount can be added.
w(0) (k) − w(0) min If not, then the result is redundancy, which means the learning rate
(0) (k) = 0.1 + (0.9 − 0.1)
w(0) max − w(0) min must be reduced. This method can ensure that the network can use
the maximal learning rate.
(3) The selection of initial weighting and learning rate The modulation equation of the self-adaptation ratio is as follows
In BP neural networks, the initial weight of the network is distinct (Tian, 2009).
and thus the training results differ, which is caused by numerous local
minima related to errors in the curved surface. Generally speaking, 0.75ηt, SSE (t ) > 1.04SSE (t − 1)
the initial weight of the network is a small random value, ensuring η(t + 1) = 1.05ηt, SSE (t ) < SSE (t − 1)
each neuron works in the maximal range of the motivated function ηt, SSE (t − 1) ≤ SSE (t ) ≤ 1.04SSE (t − 1)
slope variations. Also, it can prevent infinite changes of some unrea- The initial learning rate is highly random, implying that the num-
sonable weight after adequate continuous learning. Thus, the random ber of networks trained by the self-adaptive learning rate is only a
number (−1, +1) is commonly selected. fraction of networks trained at the fixed learning rate. Therefore, net-
The learning rate decides the weight variation generated in each work training at self-adaptive learning rate is extremely valid.
cycle of training. A larger learning rate contributes to system instabil-
ity whereas a smaller learning rate costs longer training time, which
2.3. An improved prediction model of grey neural networks in
means very slow convergence. Thus, the smaller learning rate is se-
transportation disruption
lected to ensure system stability and the range (0.01, 1) is generally
selected.
The number of neurons in the input layer of BP neural networks
(4) Model evaluation index
can be either 1 or random, which is not vigorous or scientific. An im-
The performance assessment indices of the model prediction in-
proved prediction model of grey neural networks provides a scientific
clude absolute errors of the predicted and actual values and relative
method to select the number of neurons in the input layer of the BP
errors of the predicted and actual values and r coefficient.
neural networks. In other words, high precision and fitting degrees in
Absolute error: ε (0) i = x(0) i − x
(0) i
the optimal prediction model is selected from grey prediction model,
Relative error: e(i) = |ε(0t)| × 100% determining the optimal number of dimensions, which is the figure
x i
for neurons in the input layer of BP neural networks.
x(0) i − x(0) 1 x
(0) i − x
(0) 1
r=
2
2 3. An improved prediction model of grey neural networks in
x(0) i − x(0) 1 x
(0) i − x
(0) 1 transportation disruption and its empirical study
Table 3
Calculation processes and results of traditional model GM(1,1).
Time x(0) x(1) p(i) σ (1) (i) w(1) (i) ɛ(0) (i) e(i) x
(0)
Table 4 Table 6
Calculation processes and results of testing fitting degrees. Calculation results of the 5-dimension metabolism model GM(1,1).
Time x(0) x
(0) x(0) x
(0) ξi Location of original e(i) C P ς
sequential data selected
1 5285 5285 1.0000 1.0000 1.0000
2 5313 4536.7058 1.0053 0.8584 0.3578 1–5 0.0340 0.4906 0.8 0.5806
3 4492 4578.4529 0.8455 1.0092 0.3333 2–6 0.3956 2.6850 0 0.2540
4 4480 4620.5842 0.9973 1.0092 0.8733 3–7 0.0021 0.5935 0.8 0.4686
5 4486 4663.1032 1.0013 1.0092 0.9124 4–8 0.0033 0.4481 1 0.7136
6 4489 4706.0134 1.0007 1.0092 0.9056 5–9 0.0033 0.2381 1 0.6724
7 4512 4749.3186 1.0051 1.0092 0.9525 6–10 0.0017 0.0805 1 0.5431
8 4583 4793.0222 1.0157 1.0092 0.9261 7–11 0.0005 0.0207 1 0.6232
9 4680 4837.128 1.0212 1.0092 0.8725 8–12 0.0076 0.2175 1 0.6036
10 4795 4881.6396 1.0246 1.0092 0.8419 9–13 0.0063 0.1597 1 0.5942
11 4900 4926.5608 1.0219 1.0092 0.8657 Mean 0.0505 0.5482 0.8444 0.5615
12 5215 4971.8955 1.0643 1.0092 0.5977
13 5340 5017.6472 1.0240 1.0092 0.8472
Table 5
Calculation results of transferable starting point model in disruptions.
13 493008.6660e 0.0092i
− 487723.6660 0.0421 0.8165 0.3077 0.7912
12 23568.5028e0.0181i − 23555.5028 0.0215 0.3567 1 0.5246
11 207259.2003e 0.0206i
− 202767.2003 0.0186 0.3453 1 0.5145
10 183916.9932e0.0233i − 179436.9932 0.0160 0.3081 1 0.4865
9 163827.0572e0.0264i − 159341.0572 0.0122 0.2329 1 0.5323
8 148746.9916e0.0294i − 144257.9916 0.0097 0.1871 1 0.5609
7 138779.0351e0.0321i − 134267.0351 0.0075 0.1662 1 0.6056
6 254940.5447e0.0187i − 250357.5447 0.0205 0.4137 1 0.4907
5 121492.6285e0.0395i − 116812.6285 0.0063 0.1597 1 0.5942
4 113815.0285e0.0424i − 109020.0285 0.0062 0.1831 1 0.6162
338 C. Liu et al. / Expert Systems With Applications 45 (2016) 331–340
Table 7 and the corresponding 7 neurons in the second set are (x(0) (2), x(0) (3),
Calculation results of metabolism model GM(1,1).
x(0) (4), x(0) (5), x(0) (6), x(0) (7), x(0) (8)). In turn, the corresponding 7
Numbers of sequential dimensions e(i) C P ς neurons in the final set are (x(0) (7), x(0) (8), x(0) (9), x(0) (10), x(0) (11),
x(0) (12), x(0) (13)). The corresponding neuron in the output layer is
13 0.0421 0.8165 0.3077 0.7912
12 0.0319 0.5990 0.9167 0.5800
the prediction value, and x(0) (8) can be predicted according to the
11 0.0327 0.3634 0.7273 0.5619 corresponding 7 neurons in the first group. In turn, x(0) (9), x(0) (10),
10 0.0211 0.4261 0.9000 0.5692 x(0) (11), x(0) (12) and x(0) (13) could be predicted. The neuron trans-
9 0.0169 0.3673 0.9556 0.6059 mission of the network explicit layer uses the Sigmoid function; the
8 0.0131 0.3167 0.9583 0.6067
neuron transmission of the network output layer uses the linear Pure-
7 0.0107 0.2927 0.9796 0.6166
6 0.0108 0.3115 0.8958 0.5705 lin function; and the training uses Trainbr on the basis of the Bayes
5 0.0505 0.5482 0.8444 0.5615 algorithm. The precision of the training target is 0.005; the terminal
4 0.0085 0.3342 0.9400 0.5652 training steps are 6000; and by 2705 times of training, the prediction
results are (x (0) ); the normalization results are ( (0) (k)); and the
index of model evaluation can be found in Table 8.
Table 8
Results and processes of the improved grey neural network
prediction model. 3.3. A comparison of prediction results
Time x (0)
x
(0) ϖ (0)
(k) ɛ (i)
(0)
e(i)
Comparison of the improved grey neural network model, the
1 5285 – 0.8488 – – traditional GM(1,1) and the improved GM(1,1) model (the model
2 5313 – 0.8749 – – of changing starting points is used instead for comparative pur-
3 4492 – 0.1112 – – poses), prediction values of the traditional GM(1,1), improved GM(1,1)
and the improved grey neural network model are x (0) (1), x
4 4480 – 0.1000 – –
(0) (2)
5 4486 – 0.1056 – –
6 4489 – 0.1084 – –
( )
dx (3), respectively and the corresponding relative errors are e(1),
0
7 4512 – 0.1298 – – e(2) and e(3), respectively. Each prediction value and the correspond-
8 4583 4584.8 0.1958 −1.8 0.0004
ing relative error can be found in Table 9.
9 4680 4680.2 0.2860 −0.2 0.0000
10 4795 4825.5 0.3930 −30.5 0.0064 The comparisons show that prediction accuracy of the improved
11 4900 4955.7 0.4907 −55.7 0.0114 model of grey neural networks is far greater than that of the tradi-
12 5215 5211.4 0.7837 102.6 0.0007 tional GM(1,1) model and that it is also superior to prediction accu-
13 5340 5354.4 0.9000 −14.4 0.0027 racy of the improved GM(1,1) model. As such, the improved model of
grey neural networks is a feasible prediction method for transporta-
tion disruption.
average relative error is 1.07%. Hence, there is a difference between
the average relative error in the model of changing starting points 4. Conclusions
and the model of metabolism.
Transportation disruption is the direct result of varied accidents
3.2.3. An improved prediction model of grey neural networks in such as natural disasters (e.g., ice damage, ocean disasters and earth-
transportation disruptions quakes, etc.) and social accidents (e.g., strikes and terrorist attacks),
The BP neural network consists of input, output and several ex- exerting multiple negative effects on supply chains and their member
plicit layers. Comparing the traditional GM(1,1) model, the model of enterprises, even leading to paralysis of supply chains. Market de-
changing starting points and the model of metabolism, it is optimal mand after transportation disruption is highly unpredictable while
to choose the seven-dimension sequence for the neural network pre- production and inventories of enterprises are affected to a large
diction. Therefore, the input layer consists of seven neurons: 4512, extent.
4583, 4680, 4795, 4900, 5215 and 5340. The implicit layer contains This article proposes an improved model of grey neural networks
15 neurons; the output layer contains 1 neuron, and the output is the for prediction of demand after transportation disruption. Firstly, the
ultimate prediction results. So the network typology constructed is conventional GM(1,1) model is deployed to predict the sequence of
7-15-1. It is important to note that 7 neurons are input into the out- transportation disruption; the inaccurate prediction is due to the mu-
put layer once, and there are 6 sets. The corresponding 7 neurons in tation and instability of the data. Secondly, based on this, GM(1,1)
the first set are (x(0) (1), x(0) (2), x(0) (3), x(0) (4), x(0) (5), x(0) (6), x(0) (7)) model of varying starting points and GM(1,1) model of metabolism
Table 9
A comparison of prediction results.
Time x(0) x
(0) (1) e(1) x
(0) (2) e(2) x
(0) (3) e(3)
are used for improvement of the prediction. The optimal models with Bravo, J. J., & Vidal, C. J. (2013). Freight transportation function in supply chain opti-
high precision and high fitting degrees are selected to determine the mization models: a critical review of recent trends. Expert Systems with Applica-
tions, 40(17), 6742–6757.
optimal number of dimensions. The optimal number of dimensions is Chang, C. J., Li, D. C., Huang, Y. H., & Chen, C. C. (2015). A novel gray forecasting model
the number of neurons in the input layer of neural networks. Finally, based on the box plot for small manufacturing data sets. Applied Mathematics and
the improved neural networks are employed for optimal prediction. Computation, 265(8), 400–408.
Chi, M., & Zhao, J. (2012). Application of BP neural network in evaluating e-business
A good example is Hunan HX Office and School Furniture Fac- performance for service industry. International Journal of Networking and Virtual
tory, whose sales were partially disrupted by the snow storm in Organisations, 10(3/4), 374–385.
South China in 2008. First, the traditional model is built. The cal- Chiu, S. W., Lee, C. H., Chiu, Y. S. P., & Cheng, F. T. (2013). Intra-supply chain system with
multiple sales locations and quality assurance. Expert Systems with Applications,
culation results show that the requirement of prediction accuracy
40(7), 2669–2676.
cannot be satisfied due to the influence and interference of the long- Costantino, F., Gravio, G. D., Shaban, A., & Tronci, M. (2015). A real-time SPC inventory
term data mutation and disruption, although the test of fitting de- replenishment system to improve supply chain performances. Expert Systems with
Applications, 42(3), 1665–1683.
gree is passed. As such, the changeable starting points model and
Cristóbal, J. R. S., Correa, F., González, M. A., Navamuel, E. D. R. D., Madariaga, E.,
metabolism model improve the traditional model GM(1,1). The cal- Ortega, A., López, S., & Trueba, M. (2015). A residual grey prediction model for pre-
culation results of the changeable starting points model show 7- dicting s-curves in projects. Procedia Computer Science, 64(11), 586–593.
dimension and 4-dimension can be used as the prediction models. Cybenko, G. (1989). Approximation by superpositions of a sigmoidal function. Mathe-
matics of Control, Signals and Systems, 2(4), 303–314.
Results of the metabolism model suggest that it is optimal when the Deng, J. L. (2005). Basic methods of grey system. Wuhan: Central-China Science and
dimension number is fixed at 7. By comparing and integrating these Technology Press.
two models, the 7-dimension is selected as the prediction model, and Deng, J. L. (1982). Control problems of grey systems. Systems & Control Letters, 5(1),
288–294.
the optimal number of dimensions is 7. Ultimately, the BP neural net- Dewangan, S., Gangopadhyay, S., & Biswas, C. K. (2015). Multi-response optimization
work is applied here and the network typology 7-15-1 is built. The of surface integrity characteristics of EDM process using grey-fuzzy logic-based
neuron transmission of the network explicit layer uses the Sigmoid hybrid approach. Engineering Science and Technology, 18(3), 361–368.
Hishamuddin, H., Sarker, R., & Essam, D. (2013). A recovery model for a two-echelon
function; the neuron transmission of the network output layer uses serial supply chain with consideration of transportation disruption. Computers &
the linear Purelin function; and the training function uses Trainbr Industrial Engineering, 64(2), 552–561.
based on the Bayes algorithm. The accuracy of the training target Hishamuddin, H., Sarker, R., & Essam, D. (2015). A simulation model of a three echelon
supply chain system with multiple suppliers subject to supply and transportation
is 0.005; the terminal training steps are 6000. By MATLAB 2011, in- disruptions. IFAC-PapersOnLine, 48(3), 2036–2040.
put 6 sets of data, and the accuracy requirement of 0.005 is satis- Hodzic, M., & Tai, L. C. (2016). Grey predictor reference model for assisting particle
fied through 2705 times of training. This shows the feasibility of the swarm optimization for wind turbine control. Renewable Energy, 86(2), 251–256.
Hsu, K. T. (2011). Using a back propagation network combined with grey clustering
method.
to forecast policyholder decision to purchase investment-inked insurance. Expert
This approach to determining the number of neurons in the input Systems with Applications, 38(6), 6736–6747.
layer of BP neural network through the optimal number of dimen- Intharathirat, R., Salam, P. A., Kumar, S., & Untong, A. (2015). Forecasting of municipal
sions in the improved grey prediction model has exceeded the tradi- solid waste quantity in a developing country using multivariate grey models. Waste
Management, 39(5), 3–14.
tional grey model GM(1,1), which limits the sequence of the original Ito, Y. (1991). Representation of functions by superpositions of a step or sigmoid func-
data. It also shows the characteristics of a better nonlinear approxi- tion and their applications to neural network theory. Neural Networks, 4(3), 385–
mation in neural networks so that the fitting degree and forecasting 394.
Khademi, N., Balaei, B., Shahri, M., Mirzaei, M., Sarrafi, B., Zahabiun, M., &
precision are increased and the stability and reliability of prediction Mohaymany, A. S. (2015). Transportation network vulnerability analysis for the
are enhanced. This can contribute to the prediction of the market de- case of a catastrophic earthquake. International Journal of Disaster Risk Reduction,
mand after transportation disruption and the forecasting results may 12(7), 234–254.
Li, H., & Hu, Q.Y. (2005). Forecasting and decision-making. Xi’an: Xi’an Electronic
provide information for production and inventory management deci- Science.
sions, as well as demand prediction in supply chain enterprises. This Liu, S. F., & Guo, T. B. (1991). Grey system theory and its applications. Kaifeng: Henan
method can also be based on big data to construct similar model or University Press.
Liu, S. F., Dang, Y. G., & Fang, Z. G. (2010). Grey system theory and its applications. Beijing:
expert system, predicting the consequence of transportation disrup- Science Press.
tion caused by accidents, and then some measures can be taken in Memon, M. S., Lee, Y. H., & Mari, S. I. (2015). Group multi-criteria supplier selection
advance. using combined grey systems theory and uncertainty theory. Expert Systems with
Applications, 42(21), 7951–7959.
The empirical study in this article is grounded in the single supply
Ouyang, M., Pan, Z. Z., Liu, H., & He, Y. (2015). Vulnerability analysis of complementary
chain enterprise, which has not explored the upstream and down- transportation systems with applications to railway and airline systems in China.
stream supply chain enterprises as well as other enterprises in other Reliability Engineering & System Safety, 142(11), 248–257.
regions affected. However, it is likely that results derived from the Pai, T. Y., Lo, H. M., Wan, T. J., Chen, L., Hung, P. S., Lo, H. H., Lai, W. J., & Lee, H. Y. (2015).
Predicting air pollutant emissions from a medical incinerator using grey model and
data of single enterprises will not reflect the performance of corre- neural network. Applied Mathematical Modelling, 39(5/6), 1513–1525.
sponding regions, markets and other relevant enterprises after dis- Pasandideh, S. H. R., Niaki, S. T. A., & Asadi, K. (2015). Optimizing a bi-objective multi-
ruption. In the future research, it is important to analyze the big data product multi-period three echelon supply chain network with warehouse relia-
bility. Expert Systems with Applications, 42(5), 2615–2623.
concerning the upstream and downstream enterprises and the cor- Qian, M. P., & Gong, G. L. (2000). Random process theory. Beijing: Beijing University
responding regions and markets after transportation disruption. As Press.
such, it is possible that the findings might be more universal. Rajesh, R., & Ravi, V. (2015). Modeling enablers of supply chain risk mitigation in elec-
tronic supply chains: a Grey-DEMATEL approach. Computers & Industrial Engineer-
ing, 87(9), 126–139.
Ray, A., Mukhopadhyay, S., Datta, B., & Pal, S. (2013). Prediction of atmospheric pressure
Acknowledgments at ground level using artificial neural network. International Journal of Research in
Computer Science, 3(1), 11–18.
This article is financially supported by the Natural Science Foun- Rumelhart, D. E., Hintont, G. E., & Williams, R. J. (1986). Learning representations by
back-propagating errors. Nature, 323(9), 533–536.
dation of China [Grant numbers 71172194, 71390330, 71390331 and
Sheffi, Y. (2001). Supply chain management under the threat of international terrorism.
71221001]. International Journal of Logistics Management, 12(2), 1–11.
Shu, T., Chen, S., Wang, S. Y., & Lai, K. K. (2014). GBOM-oriented management of produc-
tion disruption risk and optimization of supply chain construction. Expert Systems
References with Applications, 41(1), 59–68.
Tamura, S., & Tateishi, M. (1997). Capabilities of a four-layered feed forward neural net-
Alvisi, S., & Franchini, M. (2012). Grey neural networks for river stage forecasting with work: four layers versus three. IEEE Transactions on Neural Networks, 8(2), 251–255.
uncertainty. Physics and Chemistry of the Earth, Parts A/B/C, 42-44, 108–118. Tian, Y. B. (2009). Mixed neural network technology. Beijing: Science Press.
Bates, J. M., & Granger, C. W. J. (1969). Combination of forecasts. Operations Research Wa, D. (2010). The GM(1,1) forecast model of the Tibet rural consumption expenditure
Quarterly, 20(4), 451–468. trends and effect factors analysis. Statistical Research, 27(4), 79–82.
340 C. Liu et al. / Expert Systems With Applications 45 (2016) 331–340
Wang, R. J., Zeng, J. L., Feng, X. X., & Xia, Y. J. (2013). Evaluation of effect of plastic in- Zeng, B., & Liu, S. F. (2011). Direct modeling approach of DGM(1,1) with approximate
jection molding process parameters on shrinkage based on neural network simu- non-homogeneous exponential sequence. Systems Engineering Theory and Practice,
lation. Journal of Macromolecular Science, Part B: Physics, 52(1), 206–221. 31(2), 297–301.
Wilson, M. C. (2007). The impact of transportation disruptions on supply chain per- Zhang, H. N., Xu, A. J., Cui, J., He, D. F., & Tian, N. Y. (2012). Establishment of neural
formance. Transportation Research Part E: Logistics and Transportation Review, 43(4), network prediction model for terminative temperature based on grey theory in
295–320. hot metal pretreatment. Journal of Iron and Steel Research, 19(6), 25–29.
Xie, N. M., Yuan, C. Q., & Yang, Y. J. (2015). Forecasting China’s energy demand and self- Zhang, Y., & Lam, J. S. L. (2015). Estimating the economic losses of port disruption due
sufficiency rate by grey forecasting model and Markov model. International Journal to extreme wind events. Ocean & Coastal Management, 116(11), 300–310.
of Electrical Power & Energy Systems, 66(3), 1–8.