Aerospace Science and Technology 49 (2016) 52–62

Contents lists available at ScienceDirect

Aerospace Science and Technology

www.elsevier.com/locate/aescte

Modeling the fuel flow-rate of transport aircraft during flight phases

using genetic algorithm-optimized neural networks
Tolga Baklacioglu ∗
Department of Aircraft Airframe and Engine Maintenance, Faculty of Aeronautics and Astronautics, Anadolu University, 26470 Eskisehir, Turkey

a r t i c l e i n f o a b s t r a c t

Article history: Predicting the fuel consumption of transport aircraft is vital for minimizing the detrimental effects of fuel
Received 30 September 2014 emissions on the environment, saving fuel energy sources, reducing flight costs, achieving more accurate
Received in revised form 13 January 2015 aircraft trajectory prediction, and providing effective and seamless management of air traffic. In this study,
Accepted 23 November 2015
a genetic algorithm-optimized neural network topology is designed to predict the fuel flow-rate of a
Available online 30 November 2015
transport aircraft using real flight data. This model incorporates the cruise flight phase and the fuel
Keywords: consumption dependency with respect to both the variation of true airspeed and altitude. Feed-forward
Aircraft backpropagation and Levenberg–Marquardt algorithms are applied, and a genetic algorithm is utilized to
Fuel flow-rate prediction design the optimum network architecture regarding time and effort. The predicted fuel flow-rates closely
Neural networks match the real data for both neural network training algorithms. Backpropagation gives the best accuracy
Genetic algorithms for the climb and cruise phases, whereas the Levenberg–Marquardt algorithm is optimal for the descent
Learning algorithms phase.
© 2015 Elsevier Masson SAS. All rights reserved.

1. Introduction results; it requires improvement in terms of the input parameters

Because fuel rates constitute the largest portion of operational
costs, airlines, aircraft manufacturers, and air traffic authorities at-
tempt to implement procedures to reduce fuel consumption, such
as continuous descent and tailored arrivals, and employ modeling
approaches to predict the fuel flow-rate [1]. Conservation of fuel
energy sources, as well as the reduction of detrimental effects of
fuel emissions on the environment, can only be fulfilled in this
way. Furthermore, the accomplishment of more accurate aircraft
trajectory predictions, which will provide more effective manage-
ment of air traffic, demands the development of better propulsive
and fuel flow-rate prediction models [2].
Early attempts to estimate fuel consumption use Collins’ model
[3], which was also used in the FAA’s Airport and Airspace Sim-
ulation Model, SIMMOD. Later, Trani et al. [4] developed a neural
network (NN) model for the prediction of fuel consumption us-
ing the Fokker F-100 aircraft performance flight manual data. As
noted by Senzig et al. [1], this model requires detailed aerody-
namic information or a large database of airplane operations and
associated airplane state data, meaning its acceptance was lim-
ited. While Trani et al.’s [4] NN model provided some accurate
* Tel.: +902222261457; fax: +902223221619.
E-mail address: tbaklacioglu@anadolu.edu.tr.
and optimization of the modeling architecture. Trani et al. intro-
duced the effect of a variable Mach number as an input parameter,
and considered the initial and final altitudes, but not the effect
of a variation in altitude. Additionally, the initial aircraft weight
and temperature were also defined as input parameters. It should
be noted that the specific air range, but not the fuel flow-rate,
was given as an output parameter in their modeling of the cruise
flight phase. They tested eight candidate topologies via a sensitiv-
ity analysis, and selected a three-layer model with eight neurons
in the first two layers and one neuron in the output (third) layer.
A trial-and-error method was used to select the number of lay-
ers and neurons, rather than an optimization for the NN model
architecture [4]. Bartel and Young [5] developed a thrust-specific
fuel consumption model for the cruise flight phase. Their model
considered fuel consumption with respect to the Mach number,
whereas the effect of the temperature ratio was excluded. The tem-
perature term was replaced by the ratio of any cruise thrust to
the cruise thrust in some reference condition. Senzig et al. [1] in-
vestigated a fuel consumption model with regards to the terminal
area. Their proposed arrival thrust-specific fuel consumption algo-
rithm was based on that of Hill and Peterson with modifications
by Yoder. The temperature ratio effect, variation of Mach number,
and net corrected thrust with respect to the pressure ratio were
included in their model for the terminal area. Turgut and Rosen
[6] recently proposed a fuel consumption model based on a ge-

http://dx.doi.org/10.1016/j.ast.2015.11.031
1270-9638/© 2015 Elsevier Masson SAS. All rights reserved.
 T. Baklacioglu / Aerospace Science and Technology 49 (2016) 52–62
Fig. 1. Topology of a three-layer fuel flow rate FNN.
netic algorithm (GA) for the descent flight phase by considering
the fuel flow-rate with respect to altitude. According to its user-
manual, Eurocontrol’s BADA (current version 3.12) [7] calculates
the nominal fuel flow using the thrust and the thrust-specific fuel
consumption, which is specified as a function of the true airspeed
for all flight phases except the idle descent and cruise. The cruise
fuel flow expression differs from the nominal fuel flow by an ad-
ditional correction coefficient, whereas for the idle thrust descent
condition, the fuel flow is expressed as a function of pressure, al-
titude, and descent fuel flow coefficients.
The GA-NN model derived in this study considers the variation
of both altitude and true airspeed as input parameters, and pro-
vides the fuel flow-rate as an output parameter for climb, cruise,
and descent flight phases. This model is developed using real raw
flight data records of the medium weight transport aircraft type,
Boeing 737-800 operating in Pegasus Airlines of Turkey in 2009. As
a modeling approach, flight altitude and true airspeed variations
are inserted as inputs to the model to prevent the need for ex-
tensive detailed information of airplane operations and data. This
enables accurate fuel flow-rate modeling in a simplified manner.
With regard to the optimization of the NN modeling architec-
ture, the proposed model utilizes a GA method to determine the
network input parameters and number of neurons in the hidden
layer(s), thus achieving an optimal or nearly optimal model in a
shorter time with less effort.
2. Neural networks
An NN is an information or signal processing system composed
of a large number of simple processing elements, called neurons.
These are interconnected by weighted links (or weighted connec-
tions) that cooperate to perform parallel distributed processing and
solve computational tasks. NNs attempt to (at least partially) simu-
late the structure and functions of the brain and nervous system of
living creatures [8]. Because the inputs of the first layer of neurons
are connected to external data, this is called the input layer. Sim-
ilarly, as the outputs of the last layer of neurons are the result of
the total NN, this is called the output layer. All neuron layers be-
tween the input and output layers are called hidden layers, whose
action cannot be seen directly from the outside. If the outputs of
one layer are all connected to the input of the next layer, and there
are no connections within the same layer or from a later layer back
to an earlier layer, this is regarded as a feed-forward network [9].
Collecting the values from all of its input connections, each neuron
produces a single output passing through an activation function
[10]. In this paper, a multilayer feed-forward NN (FNN) structure
is used for the proposed fuel flow-rate prediction model. A num-
ber of parameters, such as the number of layers and the type and
number of units per layer, must be defined for any FNN. Adjusting
the weights of the network to create the desired output constitutes
the last design step. This process is called training the NN. A fuel
flow rate FNN model with two inputs (flight altitude and true air-
speed), three neurons, as an example, in a single hidden layer and
one output (fuel flow rate) is shown in Fig. 1. Coefficients asso-
53
ciated with the hidden layer (weights and biases) are grouped in
matrices W 1 and B 1 , whereas coefficients associated with the out-
put layer are grouped in matrices W 2 and B 2 . The output of the
network can be expressed as
 
Y = f 2 W 2 f 1 (W 1 X + B 1 ) + B 2 (1)
where X is the matrix of the input variables, f 1 and f 2 are the
activation functions in the hidden and output layers, respectively,
and Y is the matrix of the output variables [11].
2.1. Backpropagation algorithm with momentum
Backpropagation (BP) is a well-known training algorithm. After
measuring the output error, the BP algorithm then calculates the
gradient of this error and adjusts the NN weights in the direction
of descending gradient. As a gradient-descent local search tech-
nique, BP is highly accurate, but may also fall into local optima in
complex search landscapes.
The squared error of the NN for a set of patterns can be defined
as

P 
N
 p p 2
E= xi − o i (2)
p =1 i =1
p p
where E is the squared error of the NN, whereas xi and o i are
th
the i components of the expected vector and the current output
vector for the pattern p, respectively, N is the number of output
neurons, and P is the number of patterns. The weights of the net-
work define the actual value of the error function. Calculating the
gradient of E, the BP algorithm updates the weights by moving
them in the gradient descent direction, which can be expressed as
∂E
w i j (t + 1) = w i j (t ) − η (3)
∂ wij
where w i j (t ) and w i j (t + 1) correspond to the network weights
at the steps t and t + 1, respectively. The parameter η > 0 is the
learning rate controlling the learning speed. A momentum term is
added to Eq. (3) to increase the stability of the search process:
∂E
w i j (t + 1) = w i j (t ) + m w i j (t ) − η (4)
∂ wij
where  w i j (t ) shows the change in the weight w i j at step t, and
m is the momentum rate, which takes a value in the interval [0, 1).
When the error function does not change, this term accelerates the
error minimization [12].
2.2. Levenberg–Marquardt algorithm
Being a variation of the Gauss–Newton (GN) method, the
Levenberg–Marquardt (LM) algorithm is designed to minimize
functions that are sums of squares of other non-linear expres-
sions [13]. The LM algorithm acts as an intermediate optimization
54 T. Baklacioglu / Aerospace Science and Technology 49 (2016) 52–62

algorithm between the GN method and a gradient descent algo-

rithm. In the vicinity of a minimum, the LM algorithm exhibits
quadratic convergence, approximating the GN method. LM utilizes
gradient descent to select better initial parameters, and converts
to the GN method when approaching the minimum value of the
cost function. To improve accuracy, LM converts back to the gradi-
ent descent algorithm once it approaches the minimum. According
to the GN method, the weights of the neural network are updated
by
  −1
w = − ∇ 2 E (w) .∇ E (w) (5)

where ∇ 2 E (w) denotes the Hessian matrix, and can be expressed

as

∇ 2 E (w) = J T (w). J (w) + S (w)


N
S (w) = e i (w).∇ 2 e i (w) (6)
i =1

Here, J (w) is the Jacobian matrix and e i (w) represents the error
vector of the NN for pattern i. Because the number of calculations
increases exponentially with the network size, calculating S (w) is
costly; therefore, it is neglected and assumed to be equal to zero.
As a result, the NN weights can be updated by the following
expression for the GN method:
  −1
w = − J T (w). J (w) . J T (w)e (w)
The LM modification to the GN method can be stated as
 −1
w = − J T (w). J (w) + λI . J T (w)e (w)
For large values of the parameter λ, the LM algorithm approx-
imates the gradient descent, whereas for small values of λ, it ap-
proximates the GN method. Adaptively adjusting λ allows the LM
algorithm to combine the advantages of the gradient descent and
GN methods, and bypass the limitations of both approaches [14].
When the algorithm commences, λ is set to a small value (i.e.,
λ = 0.01). If a step yields an increased value for E (w), λ is multi-
plied by a factor ϑ > 1 to drive the algorithm towards the gradient
descent. Conversely, if a step yields a decreased value for E (w), λ
is divided by ϑ to drive the algorithm towards GN [13,14].
3. Genetic algorithm-optimized neural networks
GAs are stochastic search techniques that utilize the mecha-
nisms of natural selection and natural genetics [15]. These algo-
rithms have been successful in solving hard optimization prob-
lems, such as training NNs, i.e., problems in which the steepest
descent techniques fall into local minima or fail because of the
complexity and/or non-differentiability of the search space [12].
Starting with an initial set of random solutions, known as a popula-
tion, GAs modify these multiple individuals, each of which encodes
a tentative solution, toward an optimal solution to the problem
being considered [12,15]. To create this optimal solution, some
measure of the suitability, or fitness, of each possible solution is
required [16]. To create a new population, the individuals in the
current population are manipulated by several variation operators,
a step known as generation [12]. In GAs, individuals are encoded
as strings, so that the chromosome values are uniquely mapped
onto the decision variable domain. Although other representations
can be used, the binary alphabet is most commonly applied [17].
Individuals are then selected according to their fitness values, and
two basic operations are used to obtain the next generation from
the selected individuals. The features of parent individuals are re-
combined to one or more individuals using the crossover operation.
Some features of an individual are also modified by mutation. The
(7)
Fig. 2. Flowchart of BP algorithm hybridized with GA.
(8) GA continues to generate successive populations until a prespeci-
fied termination criterion [18].
GAs can be used to train NNs and optimize their architecture.
This utilizes the ability of GAs to escape from local optima. The
search features of GAs are population-driven, and differ from those
of classical algorithms such as BP, which have trajectory-driven
search methods. In trajectory-driven methods, the points in the
search space are visited individually, depicting a trajectory in the
search space. However, population-driven methods promote the
exploration of the entire search space by modifying a set of so-
lutions in each step. This brings the advantage of being able to
avoid local optima [12].
In this study, a GA was hybridized with BP and LM algorithms
to optimize the structure and connection weights. For LM, this
meant obtaining the initial weights (mean, variance) and num-
ber of neurons, whereas for BP, the learning and momentum rates
were adjusted in addition to the previously mentioned parameters.
The chromosomes in the GA were encoded as binary strings, which
were defined by the number of neurons in the hidden layer(s) (in
the case of training with the LM or BP algorithm), and the learning
and momentum rates (in the case of training with the BP algo-
rithm) of the NN. The substring length was chosen as 16-bit and
normalized to represent a real value in the interval [−1, +1]. The
weights related to input links of a neuron were placed contigu-
ously in the chromosome. The fitness function for the developed
GA was determined as the mean squared error (MSE), which pro-
vided the NN output error with respect to the desired output.
A population size of 100 was defined and a maximum generation
of 50 was chosen as the termination criterion for the GA. The num-
ber of epochs used in the NN was set to 10,000. Roulette-wheel
selection, two point crossover, uniform mutation, and elitism were
applied. The crossover and mutation rates were chosen to be 0.85
and 0.01, respectively. The choice of these parameters and values
were based on the computational results in terms of accuracy dur-
ing the model implementation. In the GA-BP algorithm, the learn-
ing and momentum rate optimizations were carried out between
 T. Baklacioglu / Aerospace Science and Technology 49 (2016) 52–62 55

Fig. 3. Training and cross-validation minimum MSE results obtained for different NN topologies: a) One HL, climb, b) One HL, cruise, c) One HL, descent, d) Two HL, climb,
e) Two HL, cruise, f) Two HL, descent (T: Training, CV: Cross-validation; HL: Hidden layer).

lower and upper bounds of 0 and 1, respectively, whereas in the
GA-BP and GA-LM algorithms, the number of neurons were opti-
mized between lower and upper bounds of 4 and 15, respectively.
A steady-state strategy in which one new solution in every step is
computed was utilized in the GA coding mechanism. 50 indepen-
dent runs were performed in order to accomplish the statistical
significance of the results. A flowchart of the GA-optimized BP-NN
model is shown in Fig. 2.
4. Application of fuel flow-rate prediction model
Randomly selected 2009 flight data records of the medium-
weight Boeing 737–800 transport aircraft operating in Pegasus
Airlines, Turkey were utilized in this study. These raw data in-
clude flight altitude, Mach number, engine speeds, flight time, and
fuel flow rate. The flight altitude, and true airspeed data of the
aircraft were used as inputs to the NN, and the fuel flow-rate
56 T. Baklacioglu / Aerospace Science and Technology 49 (2016) 52–62

Fig. 3. (continued)

value associated to the same time instant for these inputs in the
flight data records was taken as the output of the NN. The true
airspeed, altitude, and fuel flow-rate values were normalized by
multiplying by factors of 10−3 , 10−6 , and 10−5 , respectively. The
fuel flow-rate values range between 4113–6621 kg/h for the climb,
1465–2884 kg/h for the cruise, and 146–1465 kg/h for the descent
phases, whereas the variation for the flight altitude is between
approximately 10,000–30,000 ft for the climb, 32,000 ft for the
cruise (constant flight altitude-level flight), 7,000–29,000 ft for the
descent phases. There are 347, 404, and 483 data points for the
climb, cruise, and descent phases respectively. The data were ran-
domized and split into training (60%), cross-validation (15%), and
testing (25%) sets, allowing a supervised learning process to be
performed.
As the assignment of initial weight values, the number of neu-
rons in the hidden layers, and the activation functions all affect the
accuracy of the NN model; several configurations were tested using
the usual training processes. The BP and LM algorithms were ex-
ecuted for 10 runs of 10,000 epochs, using random initial weights
with different numbers of neurons for one- and two-hidden-layer
network architectures. Different momentum rates were also ap-
plied in the BP algorithm. The hyperbolic tangent activation func-
 T. Baklacioglu / Aerospace Science and Technology 49 (2016) 52–62 57

Fig. 3. (continued)

tion was used for the one-hidden-layer test, whereas sigmoid and
hyperbolic tangent activation functions were used for the two-
hidden-layer networks. A linear activation function was used for
the output layer in both cases. These choices of activation func-
tions were seen to provide the minimum MSEs for the training
and cross-validation.
Fig. 3 shows the training and cross-validation results for the
determination of the optimum number of neurons in one- and
two-hidden-layer network structures for the climb, cruise, and de-
scent phases. The network topology with the minimum MSE for
the cross-validation data was considered to be the best NN model
candidate. Note that the MSE values correspond to normalized fuel
flow-rate values. As it can be seen from Fig. 3, for the climb phase,
the best candidate was found to be the LM-trained NN archi-
tecture of 2-14-1, which indicates the number of neurons in the
input, hidden, and output layers, respectively. Likewise, as shown
in Fig. 3, for the cruise and descent phases, the best NN model
candidates were found to be 2-11-1 and 2-13-1 LM-NN architec-
tures, respectively. For two-hidden-layer networks, the number of
neurons in hidden layer 2, which used the hyperbolic tangent acti-
58 T. Baklacioglu / Aerospace Science and Technology 49 (2016) 52–62
Fig. 4. Training results of GA optimization in terms of best fitness (MSE) for one- and two-hidden-layer networks: a) climb, b) cruise, c) descent.
vation function, was fixed to the optimal value obtained from the
one-hidden-layer case, and then the number of neurons in hidden
layer 1, which utilized the sigmoid activation function, was var-
ied.
For example, the number of neurons in hidden layer 2 was
fixed at 14, and the number of neurons in hidden layer 1 was
determined experimentally for the previous best NN architecture
candidate of the climb phase. As well as the best candidate ar-
chitectures found for the one-hidden-layer case, all of the other
architectures were tested for the two-hidden-layer case, so as not
to miss any better NNs. The best NN model candidate architectures
were found to be 2-11-14-1 (LM), 2-13-9-1 (BP with momentum
rate m = 0.6), and 2-8-13-1 (LM) for the climb, cruise, and de-
scent phases, respectively. For the GA optimization of the BP and
LM algorithms, a population size of 100 chromosomes was used.
Roulette selection, two-point crossover (with probability 0.85), and
uniform mutation (with probability 0.01) were applied. The termi-
nation criterion was set to 50 generations. The training algorithms
were again executed for 10,000 epochs. For BP hybridized with GA,
the lower and upper bounds for the learning and momentum rates
were selected as 0 and 1, respectively. The initial mean and vari-
ance of the weight, number of neurons (for BP and LM hybridized
with GA), and momentum and learning rates (for BP hybridized
with GA) were optimized using the GA. Because the success of the
training algorithms is critically dependent on choosing these pa-
rameters, and as there are no guidelines for this procedure, using
a stochastic algorithm (such as a GA) provides a fine-tuned search
[19]. The best fitness values achieved by the GA are shown in Fig. 4
for one- and two-hidden-layer networks. As it can be seen from
Fig. 4(a), for the climb phase, a one-hidden layer network topology
trained by LM hybridized with GA provided the minimum MSE,
which was equal to 2.414 × 10−5 . Meanwhile, applying the trial-
error approach on the climb phase, the minimum cross-validation
MSE was obtained as 2.877 × 10−5 by 2-11-14-1 two-hidden-layer
network structure trained by LM as seen from Fig. 3(d). For the
cruise phase, a two-hidden-layer network trained by LM hybridized
with GA gave the minimum MSE, which was equal to 0.006, as
shown in Fig. 4(b). While, trial-and-error case provided the mini-
mum cross-validation MSE as 0.0013 by 2-11-1 one-hidden layer
network trained with LM for the cruise flight condition as seen in
Fig. 3(b). For the descent phase, the minimum MSE was obtained
as 0.0008 by a two-hidden-layer network architecture trained by
 T. Baklacioglu / Aerospace Science and Technology 49 (2016) 52–62 59

Fig. 4. (continued)

Table 1
Test performance of the best candidate NNs found by trial-and-error and GA-optimized NNs in terms of linear
correlation coefficients.

Climb
HL 1 PEs 14 11 6 6 14 13
HL 2 PEs – 14 – – 7 8
Training LM LM GA-BP GA-LM GA-BP GA-LM
R 0.99976 0.99984 0.99988 0.99984 0.99979 0.99986

Cruise
HL 1 PEs 11 13 11 14 7 6
HL 2 PEs – 9 – – 14 12
Training LM BP (m = 0.6) GA-BP GA-LM GA-BP GA-LM
R 0.92178 0.80666 0.95492 0.94484 0.91708 0.90952

Descent
HL 1 PEs 13 8 7 15 8 11
HL 2 PEs – 13 – – 14 13
Training LM LM GA-BP GA-LM GA-BP GA-LM
R 0.99017 0.98687 0.95426 0.99097 0.94087 0.98976

LM hybridized with GA as shown in Fig. 4(c). On the other hand,
in the trial-error-case for the descent flight, the minimum cross-
validation MSE was found as 0.0018 by 2-13-1 one-hidden-layer
network trained with LM in Fig. 3(c). Consequently, it is clear from
the above results that the training of the networks with GA im-
proved the obtained MSE values compared to the trial-and-error
approach for each flight case.
In the testing of the NN model topologies, it was also concluded
that this hybrid learning approach, whereby GAs are utilized to
initialize the parameters of the BP algorithm, drastically improved
the convergence of the fuel flow-rate prediction NN model. One-
hidden-layer networks trained with GA-BP provided the most ac-
curate results for the climb and cruise phases, whereas LM hy-
bridized with GA in a one-hidden-layer network structure gave the
best performance for the descent phase. As summarized in Table 1,
the test performance of GA-optimized NNs and the best trial-and-
error NN model architecture candidates, as chosen by considering
the cross-validation MSE, were determined using the linear cor-
relation coefficient R between the actual fuel flow-rate data and
the NN output. The best testing performance in the trial-and-error
method was achieved by 2-11-14-1 (LM) as R = 0.99984 for the
climb, 2-11-1 (LM) as R = 0.92178 for the cruise, and 2-13-1 (LM)
as R = 0.99017 for the descent. The usage of GA optimization for
the NN parameters even improved these results. Consequently, the
best testing performance for the GA-optimized NN architectures
was accomplished by 2-6-1 (GA-BP) as R = 0.99988 for the climb,
2-11-1 (GA-BP) as R = 0.95492 for the cruise, and 2-15-1 (GA-LM)
as R = 0.99097 for the descent. The learning and momentum rates
achieved for GA-optimized BP-NNs are shown in Table 2. As can be
seen in Fig. 5, there is a good agreement between the actual and
predicted fuel flow-rate values for the test dataset at each flight
phase.
5. Comparison with previous models
In this section, the accuracy of the proposed fuel flow-rate pre-
diction model is compared with that of previous models. The input
variables and mean errors in fuel flow rate for the new model and
previous models are summarized in Table 3. In terms of the mean
percentage error, the proposed NN approach clearly provides su-
perior results to that of Trani et al. [4] for the climb and descent
phases. Note that no comparison of the cruise phase can be made,
because Trani et al.’s model does not predict the fuel flow-rate for
this part of the flight. In this sense, our model accomplishes the
first accurate fuel flow-rate prediction for the cruise phase. The
percentage error of the GA model of Turgut and Rosen [6] is found
60 T. Baklacioglu / Aerospace Science and Technology 49 (2016) 52–62

Fig. 5. Test results of the best NN model architectures by the comparison of estimated and actual fuel flow rates values for a) climb, b) cruise, c) descent.

to range between −13% and −3%, −7% and 10%, −2% and 2%,
−1% and 1% for different cases [6]. Therefore, the mean percent-
age error of their model is greater than 1%, which is less accurate
than the new GA-NN model presented in this paper. Furthermore,
Turgut and Rosen’s GA model covers only the descent flight phase.
Comparing the mean errors of the derived model and the previous
models as shown in Table 3, it is also clearly seen that incorpo-
rating both flight altitude and airspeed as input variables in the
modeling approach improves the accuracy of the model to esti-
mate fuel flow rate values.
6. Conclusions
This paper constitutes the first attempt to optimize the initial
parameters and design of the NN architecture by using GAs in fuel
flow rate prediction modeling of transport aircraft for emissions
reduction. The originality of this study also lies in the usage of
real random raw flight data in the development of the NN model.
This study used real flight data to develop the NN model, rather
than the aircraft flight manual data used to develop the existing
NN model of Trani et al. Unlike previous approaches, the variation
of both flight altitude and true airspeed was simultaneously in-
corporated into the proposed model. Moreover, for the first time,
the fuel flow-rate of the cruise phase was predicted, represent-
ing an improvement on current models. The accuracies of the new
GA-NN model were achieved in terms of percentage mean error
as 0.020% and 0.745% for the climb and descent phases, respec-
tively, which are superior compared to the previous models. BP
and LM algorithms hybridized with GA provided more accurate
results than a simple trial-and-error approach. GA-NNs with one
hidden layer showed better test performance than two-hidden-
layer GA-NNs. When real flight data or flight manual data of any
 T. Baklacioglu / Aerospace Science and Technology 49 (2016) 52–62 61

Fig. 5. (continued)

Table 2 Table 3
Learning and momentum rates obtained by GA-optimization for networks trained Comparison of the input variables and accuracies in fuel flow rate for the new GA-
by BP. NN model and previous models [4,6].

Climb Climb
One-hidden-layer GA-BP Two-hidden-layer GA-BP Model Input variables Mean error in fuel
flow rate (%)
HL 1 PEs 6 14
HL 2 PEs – 7 GA-NN model flight altitude, airspeed 0.020
η (HL 1) 0.4375 0.3929 Trani et al.’s NN model airspeed 1.026
η (HL 2) – 0.8123
Descent
m (HL 1) 0.7768 0.3740
m (HL 2) – 0.9071 Model Input variables Mean error in fuel
flow rate (%)
Cruise
GA-NN model flight altitude, airspeed 0.745
One-hidden-layer GA-BP Two-hidden-layer GA-BP
Trani et al.’s NN model airspeed 1.423
HL 1 PEs 11 7 Turgut and Rosen’s GA model flight altitude >1
HL 2 PEs – 14
η (HL 1) 0.5844 0.3228
η (HL 2) – 0.6157 tionship between thrust and required fuel energy is accomplished,
m (HL 1) 0.6203 0.7440 FADEC will be able to send correct and minimum required fuel
m (HL 2) – 0.5276
flow rate to engine combustor. Furthermore, providing correct fuel
Descent flow rate means minimization of emission, detrimental environ-
One-hidden-layer GA-BP Two-hidden-layer GA-BP mental effects, wasted energy as well as achievement of maximum
HL 1 PEs 7 8 sustainability of an aircraft. Moreover, robust fuel flow rate mod-
HL 2 PEs – 14 eling approach derived in this paper can be applied to energetic
η (HL 1) 0.7841 0.8772
and exergetic models of aircraft engines in analyzing environmen-
η (HL 2) – 0.3014
m (HL 1) 0.2984 0.9274 tal impact and sustainability of modern transport aircraft.
m (HL 2) – 0.4200
7. Glossary
aircraft is available to the user, this newly proposed GA-optimized
NN model will enable a more efficient NN modeling methodology Abbreviation
for fuel flow-rate prediction in real applications, such as decreas-
BADA Base-of-aircraft data
ing fuel emissions in environmental studies, fuel-saving strategies,
BP Backpropagation
air traffic management, and aircraft trajectory prediction in air
CV Cross-validation
traffic control simulations and decision support tools. In order to FAA Federal Aviation Administration
achieve accurate fuel flow rate predictions to calculate the re- FADEC Full authority engine control unit
quired thrust and energy of an aircraft engine, the model derived FNN Feed-forward neural network
in this study must be embedded into the main engine control unit, GA Genetic algorithm
namely FADEC. While an accurate formulation to describe the rela- GN Gauss–Newton
62 T. Baklacioglu / Aerospace Science and Technology 49 (2016) 52–62

HL Hidden layer [3] B. Collins, Estimation of aircraft fuel consumption, J. Aircr. 19 (11) (1982)
LM Levenberg–Marquardt 969–975, http://dx.doi.org/10.2514/3.44799, also AIAA Paper 81-0789R.
[4] A. Trani, F. Wing-Ho, G. Schilling, H. Baik, A. Seshadri, A neural network model
MSE Mean squared error
to estimate aircraft fuel consumption, in: 4th AIAA Aviation Technology, Inte-
NN Neural network gration and Operations (ATIO) Forum, Chicago, 20–22 Sept. 2004, AIAA Paper
PE Processing elements (neurons) 2004-6401.
T Training [5] M. Bartel, T.M. Young, Simplified thrust and fuel consumption models for
modern two-shaft turbofan engines, J. Aircr. 45 (4) (2008) 1450–1456,
http://dx.doi.org/10.2514/1.35589.
Symbols
[6] E.T. Turgut, M.A. Rosen, Relationship between fuel consumption and altitude
for commercial aircraft during descent: preliminary assessment with a genetic
B Matrix of the biases algorithm, Aerosp. Sci. Technol. 17 (2012) 65–73.
e Error vector of the neural network for a pattern [7] User Manual for the Base of Aircraft Data (BADA), Revision 3.12, https://
E Squared error of the neural network www.eurocontrol.int/sites/default/files/field_tabs/content/documents/sesar/user
-manual-bada-3-12.pdf, 2014 (30.09.2014).
f Activation function
[8] A. Cichocki, R. Unbehauen, Neural Networks for Optimization and Signal Pro-
F Fuel flow-rate cessing, John Wiley & Sons Ltd., West Sussex, 1993.
J Jacobian matrix [9] T. Bräunl, Embedded Robotics: Mobile Robot Design and Application with Em-
m Momentum rate bedded Systems, third ed., Springer-Verlag, Berlin, Heidelberg, Germany, 2008.
N Number of output neurons [10] A. Savran, R. Tasaltin, Y. Becerikli, Intelligent adaptive nonlinear flight control
for a high performance aircraft with neural networks, ISA Trans. 45 (2) (2006)
o Output vector of the NN for a pattern
225–247.
P Number of patterns [11] C.I. Ochoa-Martinez, A.A. Ayala-Aponte, Prediction of mass transfer kinetics
R Linear correlation coefficient during osmotic dehydration of apples using neural networks, Lebensm.-Wiss.
w Network weight Technol. 40 (2007) 638–645.
W Matrix of the weights [12] E. Alba, F. Chicano, Genetic algorithms, in: E. Alba, R. Marti (Eds.), Metaheuristic
Procedures for Training Neural Networks, Springer Science + Business Media,
x Expected vector of the NN for a pattern
LLC, New York, 2006, pp. 109–137.
X Matrix of the input variables of the NN [13] V. Singh, I. Gupta, H.O. Gupta, ANN-based estimator for distillation using
Y Matrix of the output variables of the NN Levenberg–Marquardt approach, Eng. Appl. Artif. Intell. 20 (2007) 249–259.
η Learning rate [14] B.G. Kermani, S.S. Schiffman, H.T. Nagle, Performance of the Levenberg–
Marquardt neural network training method in electronic nose applications,
Sens. Actuators B 110 (2005) 13–22.
Conflict of interest statement
[15] M. Gen, R. Cheng, Genetic Algorithms and Engineering Design, John Wiley &
Sons Inc., New York, 1997.
There is no conflict of interest. [16] J. Love, Process Automation Handbook: A Guide to Theory and Practice,
Springer-Verlag London Limited, London, 2007.
References [17] A. Chipperfield, Introduction to genetic algorithms, in: A.M.S. Zalzala, P.J. Flem-
ing (Eds.), Genetic Algorithms in Engineering Systems, in: IEE Control Engineer-
[1] D.A. Senzig, G.G. Fleming, R.J. Iovinelli, Modeling of terminal-area airplane ing Series, vol. 55, IET, London, 1997, pp. 1–45.
fuel consumption, J. Aircr. 46 (4) (2009) 1089–1093, http://dx.doi.org/10.2514/ [18] R. Klein, Genetic algorithms, in: H. Stadtler, C. Kilger (Eds.), Supply Chain Man-
1.42025. agement and Advanced Planning, third ed., Springer, Berlin, Heidelberg, Ger-
[2] M. Cavcar, A. Cavcar, Aero-propulsive modeling of transport aircraft for air many, 2005, pp. 485–492.
traffic management applications, in: AIAA Guidance, Navigation, and Control [19] D. Montana, Neural network weight selection using genetic algorithms, in: S.
Conference and Exhibit, Providence, Rhode Island, 16–19 Aug. 2004, AIAA Pa- Goonatilake, S. Khebbal (Eds.), Intelligent Hybrid Systems, John Wiley & Sons
per 2004-4792. Ltd., West Sussex, 1995, pp. 85–101.