Sei sulla pagina 1di 9

Aerospace Science and Technology 17 (2012) 6573

Contents lists available at ScienceDirect

Aerospace Science and Technology


www.elsevier.com/locate/aescte

Relationship between fuel consumption and altitude for commercial aircraft during descent: Preliminary assessment with a genetic algorithm
Enis T. Turgut a, , Marc A. Rosen b
a b

Anadolu University, School of Civil Aviation, Department of Aircraft Airframe and Engine Maintenance, TR26470 Eskisehir, Turkey Faculty of Engineering and Applied Science, University of Ontario Institute of Technology, 2000 Simcoe Street North, Oshawa, Ontario, L1H 7K4, Canada

a r t i c l e

i n f o

a b s t r a c t
Despite fuel ow rates being low during descent compared to climb and cruise, the potential exists for signicant fuel savings during that phase of ight. Increased fuel use, costs and related environmental impacts are associated with stepped descents leading to low level ights and holding in the air due to delays or slot conicts. Except for the Base of Aircraft Data (BADA) model, little research has been reported on this relationship for aircraft fuel consumption. In this study, the negative impact of low level ight is examined and a relationship between fuel consumption and altitude established using real ight data and the genetic algorithm (GA) technique. Exponential models are established between fuel consumption and altitude, and corresponding results exhibit similar model coecients and low residuals for three of four ights. Moreover, it is found that the relationship between fuel ow and altitude for similar ight congurations in the descent phase can be explained by a generic exponential model. 2011 Elsevier Masson SAS. All rights reserved.

Article history: Received 13 September 2010 Received in revised form 3 February 2011 Accepted 17 March 2011 Available online 22 March 2011 Keywords: Aircraft Flight data records Fuel consumption Descent Genetic algorithm

1. Introduction Compared to previous decades, considerable focus is now placed on environmental concerns in many areas, including aviation. It is reported in the Emission Trading Scheme, which included aviation in 2009 through legislation introduced by the EU, that the total emission allowance allocated to aircraft operators will be equivalent to 97% of aviation emissions for the period between Jan. 1, 2012 and Dec. 31, 2013, with lower allowances for future periods [1]. This legislation will impact operators ying in the European Union Emission Trading Scheme (ETS) area, creating competitive advantages for operators that invest in recent new and clean aircraft and engine technologies and/or ecient schedules and eet procedures. Fuel cost is a signicant direct operational cost for air transportation operators and variations in fuel prices directly affect the cost of ights. Fuel prices in the U.S. between 2003 and 2008 increased by an annual average of 28.2%, but decreased by 38.1% in 2009 and further to 0.5 U.S.$ per liter by 2010 June [2]. Numerous studies on optimizing fuel consumption of aircraft have been reported, mainly in three key topic areas [4,5,12,23,3]: eliminating unnecessary aircraft weight, using intelligent air trafc management procedures and providing best performance ight procedures for aircraft in all phases (e.g., optimum cruise altitude,

Corresponding author. E-mail addresses: etturgut@anadolu.edu.tr (E.T. Turgut), marc.rosen@uoit.ca (M.A. Rosen). 1270-9638/$ see front matter doi:10.1016/j.ast.2011.03.007

most economic climb). Other methods exist to improve fuel consumption and reduce associated emissions, which can be applied mainly by the operator, such as correct scheduling of eets, optimization of routes to obtain a maximum load factor, reducing the duration of ground operation of engines and auxiliary power units (APUs), use of single engine taxiing, and using blends of alternative fuels. However, the unavoidable occurrence of delays or slot conicts, due to numerous factors, increases fuel consumption, particularly for airports having relatively high air trac. When such situations occur, aircraft on the ground need to wait before taking off, while those in the air are put on hold. For example, the extreme case of the closure of an airport due to a severe air transport system disruption has been assessed by Pejovic et al. in terms of fuel consumption, CO2 emissions, and fuel and environmental costs [20]. In that work, two scenarios are investigated for a one hour closure at Londons Heathrow airport, which is not an unlikely occurrence. In the rst scenario, 43 aircraft planning to land at Heathrow airport are diverted to alternative airports. In the second, some of these 43 aircraft are diverted to alternate airports as in the rst scenario, while some of them are put on hold for 20 minutes. The fuel consumption and CO2 emissions are found to be 33 tonnes and 103 tonnes for the rst scenario, and 22 tonnes and 70 tonnes for the second. The ight levels of the holding aircraft are not specied. Little research has been reported in the literature concerning the modeling of fuel consumption and its relation to engine power setting or aircraft ight phase (rather than its correlation with aircraft or eet selection or route determination, etc.). Neverthe-

2011

Elsevier Masson SAS. All rights reserved.

66

E.T. Turgut, M.A. Rosen / Aerospace Science and Technology 17 (2012) 6573

less, one widely-known resource exists, called Base of Aircraft Data (BADA), which was developed by EUROCONTROL and covers almost 300 aircraft types [22]. In the fuel consumption section of the BADA, the fuel ow function is expressed by three different functions which incorporate the thrust specic fuel consumption as a function of true air speed (TAS) and equation correction coecients for all ight phases except cruise and descent idle. During the descent idle and cruise phases, other expressions for fuel consumption apply which are functions of altitude and equation correction coecients. A genetic algorithm (GA) can be used to optimize aspects of ight operations, and is mainly preferred for effective air trafc management and aircraft sequencing and to resolve problems with aircraft scheduling for either ight or maintenance. The potential of genetic search algorithms in scheduling time of arrival and runway assignments of aircraft is investigated by Hansen [9]. The objective is to minimize the difference between the scheduled time of arrival and the estimated time of arrival considering constraints such as separation rules based on aircraft categories, using the same runway for landing and takeoff, management of multiple runway operation, etc. Hu and Paolo examined the aircraft arrival sequencing and scheduling problem for multiple runways using a rst-come-rst-served approach [11]. The approach considers the landing time interval of aircraft, which differs for various types of aircraft due to safety concerns, performance (landing speed) and other factors relevant to runway and operational concerns. The objective, minimization of the total difference between allocated landing times and planned landing times, is similar to that of Ref. [9]. Hu and Paolo found that a new GA developed with uniform crossover exhibited better results compared to GAs developed in previous studies. A more recent study [21] focuses on the arrival problem for a single runway, and uses an objective function related to ight time and its cost. Soomer claims that the developed GA identies signicant potential cost savings compared to a schedule resembling the current scheme. The GA is also utilized to improve the impact of excessive wind turbulence on the automatic landing system [13] and in optimization of ight prole of aircraft [19]. Apart from air trac management, other utilization areas for the GA include design of axial ow compressor blades [8,14], testing of engine dynamics [6], aircraft structural design [10], unmanned air vehicle navigation [15], design of micro air vehicles [16], design and optimization of ight control systems [7] and design of algorithms for aircraft stall recovery mechanisms using a squared cost function [17,18]. In this study, the relationship between fuel ow and altitude during descent is assessed for a commercial aircraft using genetic algorithms. Actual data on fuel ow, aircraft speed, engine speed (N1, N2), altitude and so on are obtained from aircraft ight data records (FDRs), which were provided in the form of raw data records from the 2009 ights of Pegasus Airlines operating in Turkey. Four different descents (denoted F1 through F4) are analyzed performed by the same type of aircraft (but not aircraft of the same tail number) and engines. The aircraft type considered is the B737-800, which is a twin-engine medium-haul aircraft. The engine type used in these aircraft is the CFM56-7B, a high by-pass turbofan engine with a by-pass ratio of 5.5, an overall pressure ratio of 32, and a thrust rating of 115 kN. The relationship between fuel ow and altitude can be used in state-of-the-art delay management systems. When a delay condition becomes necessary at an arrival airport and leads to aircraft being held in the air, a delay management system could advise pilots of options, or could provide information related to default scenarios. Taking into consideration the performance and weight of aircraft with wind direction and speed, and utilizing fuel ow and altitude models, aircraft could be directed during holding to

higher altitudes if possible, yielding substantial fuel savings compared to holding aircraft at lower altitudes. Therefore, this study and the models developed are believed to contribute to the decision process for determining proper altitudes. 2. Genetic algorithm program setup The GA is a powerful tool which can be utilized to optimize linear or nonlinear models. A typical GA starts with the selection of a random population of a proper size. A population is composed of a dened number of inputs, and its size is important as it affects the output and the duration of solution procedure. The corresponding binary code of each input is called a genome (or chromosome) and includes a different number of bits, i.e., 0 and 1 values. The objective is to nd the best output through the binary codes by using some reproduction processes, which are referred to as crossover and mutation. Accordingly, in each iteration, some of the genomes are applied by these reproduction processes. In crossover, a specic number of bits of a specic number of genomes are converted, as shown in the following example: Before crossover: Genome A = 101 010 Genome B = 011 111 After crossover: Genome A = 101 111 Genome B = 011 010 In mutation, on the other hand, a specic number of bits are modied reciprocally, as illustrated in the following example: Before mutation: Genome A = 10 1010 After mutation: Genome A = 10 0101 The number of genomes and the bits of each genome which are to be reproduced should be determined carefully, since inappropriate selections can lead to a signicant increase in the duration of solution or losing valuable outputs found before reproduction. Once the reproduction process is completed, all the new outputs are ranked from best to worse, and some of the worse outputs can be replaced by best ones. This completes the single iteration. After that, the GA tries to approach to the best result with subsequent iterations utilizing the reproduction processes. The approach can be gradual or sudden. The base GA program developed for this study is composed of ve modules. In the rst module the features of the program are determined. The features relating to the GA program include population, iteration, crossover and mutation ratios and accuracy degree. Other features such as ight ID, range of observation data and minimum and maximum limits for variables are related to ight parameters. The new accuracy degree, according to the variable limits, is also determined in this module. In the second module, the genomes are ranked to the values of the actual count of each genome. At the end of this task, the genomes that have a zero value of actual counts are replaced by genomes which have a maximum actual count. Modules three and four involve codes that perform the tasks of crossover and mutation. In the fth module,

E.T. Turgut, M.A. Rosen / Aerospace Science and Technology 17 (2012) 6573

67

Fig. 1. Demonstration of variation of observed fuel ow rate with altitude for F1.

the outputs of the rst iteration are recorded on an excel spreadsheet. The output having the best solution in terms of objective function is also recorded on another spreadsheet with the values for variables along with the results from the iterations. Within the modules, auxiliary tasks are also conducted, such as routines for checking the proper performance of crossover and mutation and to prevent the loss of the best output from the population after the tasks of crossover and mutation. A comparison of fuel ow variation with altitude during descent indicates an exponential trend of decrease of fuel ow (see Fig. 1). In order to model this trend, therefore, an exponential function is used. GA is run to nd the coecients of the function which are given in detailed at the subsequent section. In the GA program, the assumed parameter values are given as 0.4 for the crossover ratio and 0.02 for the mutation ratio, with the accuracy degree given for the rst coecient ( ) as 106 and for the second coefcient ( ) as 108 . The reason for utilizing such low values is to be able to provide a suciently sensitive bit length for each genome so as to obtain a sensitive variation of both coecients. The regular bit length of each genome is 30 which allows a wide range of values to be sought. The limits are selected to be 1 to 1000 for and 103 to 107 for , respectively. The population is xed at 100. 3. Analysis In the present section, the GA is performed for four ights. Two approaches and methods are used, depending on the objective function. This section is correspondingly divided into two subsections. 3.1. Approach 1 In this method a GA is programmed for all fuel ow and altitude pairs (hereafter pairs of observations) separately to maximize the base objective function y:

for nding the optimum outputs. The counts of the observation pairs are 28, 20, 19 and 18 for ights F1 through F4, respectively. The maximum value for y is obtained for the minimum residual fuel ow at a given altitude. For each pair of observations, the GA program is run at 500 cycles and each cycle involves 100 iterations. The total calculation count of the program is 4.25 106 . We now have 500 different y functions resulting from the various and coecients found for each observation pair for each ight. Each y function is the best value of the 100 iterations. At this point a question can arise as to why we do not eliminate the y functions rather than the maximum one. If the objective is to nd the minimum residual for of the each pairs, then the maximum y function for each pair is the solution. However, since the aim here is to nd the optimum y function (and the coecients of and ) that is suitable for all pairs of observations for each of the ights (and then for all ights) we do not know which y function provides the best solution. Further steps are therefore needed. In this context, initially, all 500 and values calculated for the rst pair of observations are input to the objective function for all the pairs of observations (since there are 28 observations in Flight 1, there are total of 14,000 (500 28) pairs of and ). Hence, each pair of and is used for each of the pairs of observations. Second, all the y outputs are brought to desired values by 1/ y to observe the value of the residuals since the goal is to nd the minimum value of denominator of Eq. (1). Third, the minimum residuals for each pair of observations are identied by calculating the sum of the squares of the residuals. Hence, the coecients of and offer the minimum residuals and are selected as the best solution. 3.2. Approach 2 From the results of the rst method it can be seen that there is no relationship between the maximum y function of a single pair of observations and that the function provides the optimum solution for all pairs of observations. Therefore, in this method, a different objective function is considered:
n

y ( , ) =

1 f f kj

e h

(1)

y = ( , ) = 1
i =1

f f ki e hki

(2)

where k and j denote the ight and the observation (altitude and corresponding fuel ow rate) item numbers, while f f and h denote fuel ow (in kg h1 ) and altitude (in ft). Since the altitude measurements (plus other measurements, such as speed and distance) are recorded by the FDR in English units, h values are utilized in the models in feet. Also, and are the coecients used

where i identies pairs of observations, k denotes ight number, and n is the count of the pairs of observation of each ight. The objective is to nd the coecients of and that yield a maximum y function, which can also be dened as the lowest sum of the squares of the residuals. However, at this time all of the observation pairs are considered together. It is noted that this method

68

E.T. Turgut, M.A. Rosen / Aerospace Science and Technology 17 (2012) 6573

Table 1 Objective functions and coecients of each observation pair resulting from GA program of the rst approach (Eq. (1)) for F4. Fuel ow rate (kg h1 ) 210 223 228 236 239 246 252 259 267 275 283 295 301 311 320 331 339 351 358
a

Altitude m (ft) 7919 7616 7315 7008 6702 6396 6097 5789 5490 5185 4879 4575 4269 3964 3657 3353 3044 2743 2438 (25,982) (24,986) (23,999) (22,992) (21,988) (20,985) (20,004) (18,992) (18,013) (17,010) (16,006) (15,011) (14,006) (13,005) (11,997) (11,002) (9988) (8999) (7999)

y (min.) 0.09 3.43 12.66 1.72 2.43 2.45 15.73 0.74 6.28 4.09 5.75 29.74 2.15 1.71 2.44 5.26 1.64 3.26 23.71

y (max.) 1,684,443 3,416,528 13,821,300 1,288,244 1,975,252 5,777,948 3,527,785 2,535,966 614,761 2,032,063 170,377 1,859,227 14,519,173 26,437,360 2,769,475 666,714 755,088 596,936 2,547,768

y (yielding min. SSRa ) 301 777 21 868,520 62 88 50 4679 263 28 90 49 27 438 89 291 467 29 208

Min. SSRa 276.69 148.06 113.03 145.18 73.83 65.45 100.55 162.25 95.24 113.97 111.41 153.34 69.18 65.70 66.30 84.65 65.91 154.85 72.66

449.90 445.80 446.11 445.68 455.36 452.70 453.34 456.46 451.50 451.05 452.27 453.31 450.62 451.79 451.84 455.50 451.62 460.38 449.73

2.93 105 2.77 105 2.79 105 2.77 105 2.94 105 2.90 105 2.93 105 2.99 105 2.91 105 2.91 105 2.92 105 2.86 105 2.88 105 2.88 105 2.88 105 2.92 105 2.89 105 3.03 105 2.87 105

Sum of the squares of the residuals.

Fig. 2. Progress of the rst approach of GA program for F4 (fuel ow rate is 246 kg h1 , altitude is 6396 m (20,985 ft)).

yields the best solution in the least time and the solutions are found to be superior than determined previously for all ights. The results are discussed in the next section. 4. Results and discussion In the analysis, for the rst method, the GA program is run 500 times for each pair of observation for four ights. In this method, the objective is to nd the maximum y function which yields the minimum residual at the corresponding pairs of observations. Therefore, for each pair of observations the maximum y functions and the relevant coecients of and are selected. The selected data are then used to nd the optimum results among all the pairs of observations. The results for F4 are shown in Table 1. As can be seen from Table 1, for some pairs of observations the GA program is able to nd very large outputs for the y function. For instance, at an altitude of 3964 m (13,005 ft) it can be seen that the y function exceeds 26 million which leads to very small residuals for the same pairs. As pointed out earlier, the t of all the coecients to all the pairs of observations is investigated. This investigation is performed in terms of minimum sum of the squares of the residuals (SSR). As the SSR decreases, the tness of model

increases. Accordingly, in the column y (yielding min. SSR), the y functions can be seen yielding the minimum sum of the squares of the residuals. Except for one (the fourth), all of the ndings obtained are relatively low y functions compared to the maximum relevant y functions. In the sixth column, the minimum sum of the squares of the residuals can be seen. The minimum is found to be 65.45 which is a good output, as it provides a standard error of almost 2.3 kg h1 . Considering the fuel ow rates change between 210 and 358 kg h1 , the standard error found represents an error of only 0.61%. The progress of the GA program and the residuals related to this solution are shown as a function of iteration number in Figs. 2 and 3. It can be seen in Fig. 2 that the output of the rst iteration ( y function) is 3.28 (yielding a residual of 0.3), while after 100 iterations the output is approximately 5.7 million, which yields a residual of 1.73 107 . Considering as an example the actual fuel ow of 246 kg h1 , the output of the rst iteration is suciently signicant as the residual is quite small. However, since we are investigating the best coecients for all fuel ow rates (i.e., those that yield the minimum sum of the squares of the residuals for all observations), at this stage of analysis it is not possible to know

E.T. Turgut, M.A. Rosen / Aerospace Science and Technology 17 (2012) 6573

69

Fig. 3. Residuals for the minimum sum of squares of the rst approach (Eq. (1)) for F4 (fuel ow rate is 246 kg h1 , altitude is 6396 m (20,985 ft)). Table 2 Optimum outputs for all ights resulting from the rst approach. Flight F1 F2 F3 F4

442.69 813.57 459.08 452.70

2.92 105 4.01 105 2.78 105 2.90 105

y (at min. SSR) 5945.77 564.71 95.32 88.80

Min. SSR 4582.1 1813.1 117.3 65.5

which output is better. Therefore, increasing the number of the outputs increases the occurrence of optimum solutions. Considering all observations of F4, using the coecients yielding the minimum SSR ( = 452.70 and = 2.90 105 ), it can be seen in Fig. 3 that the largest residuals are 3.65 kg h1 and 3.00 kg h1 for the fuel ow rate, which are very small compared to the observed fuel ow rate. Performing the analysis for all ights, the minimum residuals of the sum of the squares are calculated and tabulated in Table 2. As seen from the second column, the outputs are not large values compared to the levels of millions. Also, regarding the minimum residuals of the sum of the squares, the last two ights have very low values, while the rst ights have higher values. The increase in the minimum residuals of the sum of the squares may be attributable to deviations in the observations. In order to understand these deviations, we examine the trend of fuel ow observations

with altitude, as shown in Fig. 4. In that gure, the gray and black solid curves indicate a signicant relationship between fuel ow and altitude for F3 and F4. Another relationship between fuel ow and altitude is observed for F2. However, there are two vertical regions (around 7010 m (23,000 ft) and 3658 m (12,000 ft)) where fuel ow and altitude do not appear to be related in F1. The existence of these regions affects the coecients and the outputs of the GA program, and the impact of these regions is shown in Fig. 5. Two residuals in Fig. 5 (48.13 and 34.87) are far from average residuals. Taking into consideration the sum of the squares of these two residuals, it can be seen that they constitute 77% of the total residual sum of the squares. If these regions are omitted from the analysis, the residual sum of squares decreases to 1049.53, which is lower than that for F2. At this point in the analysis, the GA program has been run for each ight 500 times for each pair of observation. The coefcients output from the GA program are then input to the inverse of Eq. (1) with the all pairs of observations for the relevant ight. Considering the inverse of Eq. (1) is necessary since the objective is to determine coecients which offer the minimum deviation between the observation and the prediction for all pairs. After completing this task, there are 500 different outputs (residuals, and coecients of and ) for each pair of observation (e.g., 500 19 for F4). Now, each set of the 500 different and values are used

Fig. 4. Demonstration of variation of fuel mass ow rate with altitude for all ights.

70

E.T. Turgut, M.A. Rosen / Aerospace Science and Technology 17 (2012) 6573

Fig. 5. Residual analysis of the results of the rst approach: the effect of two regions (around 7010 m (23,000 ft) and 3658 m (12,000 ft)) on residuals for F1, where fuel ow rates uctuate unusually.

Fig. 6. The coecients of models obtained from the results of the rst approach for all ights and all observation pairs. The numbers given in the charts indicate the coecients providing the minimum sum of the squares of the residuals.

in the inverse of Eq. (1) for all pairs of observations, to nd the unique coecients of and which offer the minimum residual of the sum of the squares throughout the ight. In Fig. 6, the coefcients of and which yield the minimum residuals of the sum of squares are shown for all four ights. The values are given corresponding to the pairs of observation. The square markers denote

the values of series which satisfy the requirement for the minimum residuals of the sum of the squares. The long black dash shows the average value for the coecient of . The impact of the deviation region in F1 on the coecient of is reected in Fig. 6a. While the average of is around 441 for this ight, its calculation includes the lowest (402) and highest

E.T. Turgut, M.A. Rosen / Aerospace Science and Technology 17 (2012) 6573

71

Fig. 7. Comparison of observations and predictions for fuel mass ow rate, where observations are related to F4 and predictions are obtained from the coecients obtained for the observations of F3 (based on the results of the rst approach).

(518) values in these regions. This signicant variability affects the standard deviation, increasing it to 28.9. In Fig. 6b, better values for pairs of observations are found. The minimum residual of the sum of the squares obtained is around 813 for value, while the values found for other pairs of observations result in a lower standard deviation (10.7). The coecients of for F3 and F4 are found to be consistent for both ights (see Figs. 6c and 6d). The minimum sums of squares for the residuals are found to be 117.3 and 65.5 for the coecients of 459 and 452, for F3 and F4, respectively. Furthermore, the standard deviations are found to be as low as 7.7 and 3.5, respectively. The second coecient, , is also shown for all ights in Figs. 6a6c. Except for F2, all the coecients of are observed to be similar in value, ranging between as 2.77 105 and 2.91 105 . To interpret the signicance of the models for all the ights, it is useful to know how the values of the residuals compare to the actual fuel ow observations. The minimum residual sums of squares for F1 through F4 are given in Table 2. It can be shown with these results that the ratio, on a percentage basis, of the maximum positive and minimum negative residuals to the observed fuel mass ow rate are +3% and 13% for F1, +10% and 7% for F2, +2% and 2% for F3 and +1% and 1% for F4, respectively. The residuals for the last two ights seem to be highly signicant, while the residuals for F1 in particular are high. The high residual in this case is likely attributable to the deviations in the observed fuel mass ow rates noted earlier. Since the results for F3 and F4 are so similar, the focus of the last two ights is to investigate how the low residual of the sum of the squares could be obtained when the coecients of F3 are used for the observations of F4 and vice versa. In this regard, the minimum residual of the sum of the squares is obtained for the rst case (using the coecients of F3 in the observations of F4) as 203.1 and for the second case (using the coecients of F4 in the observations of F3) as 212.8, respectively. Taking into consideration that the minimum residuals of the sum of the squares for F3 and F4 are 65.5 and 117.3, the increase in the residual of the sum of the squares is 225% and 73%, respectively. Moreover, this increase

is reected on the ratio of the residuals to the observed fuel mass ow rates, in that the maximum positive residuals and minimum negative residuals are +3.6% and 1.2% of the actual fuel ow for the rst case and +3.6% and 2.3% of the actual fuel ow for the second case. Therefore, using the rst case (coecients of F3 in observations of F4) yields better results in terms of deviation from the actual fuel mass ow rate. This result also suggests that, for similar ight congurations, a unique model could be signicant for all ights. The relationship between observations and predictions relevant to the method described above is shown in Figs. 7 and 8. In the second analysis method, all pairs of observation are considered together for each ight. Then, the model can be expressed as an extension of Eq. (2):

y ( , ) = 1/

f f k 1 e hk 1
2

+ f f k 2 e hk 2

2 2

+ f f k 3 e hk 3

+ + f f kn e hkn

(3)

where k denotes the ight identication. The current objective is to determine the minimum residuals of the sum of the squares for all pairs of observations for each ight. In this approach, the GA program is run 1000 times. The population and iterations are both set to 200. The mutation and crossover ratios are kept unchanged. The results obtained from the GA program based on Eq. (3) strongly corroborate previous predictions. Also the minimum residuals of the sum of the squares are found to be slightly lower compared to those obtained from previous predictions for all ights. The outputs and coecients are shown in Table 3. 5. Sensitivity analysis In the previous section, indicators that form the models are found. However, these indicators do not provide numerical information regarding the change in fuel ow with changes in altitude. In this section, the models are analyzed according to changes in the amount of fuel ow. Since the model coecients obtained from different approaches are similar (for each ight), a sensitivity analysis is applied to models obtained with only one approach (the rst approach).

72

E.T. Turgut, M.A. Rosen / Aerospace Science and Technology 17 (2012) 6573

Fig. 8. Comparison of observations and predictions for fuel mass ow rate, where observations are related to F3 and predictions are obtained from the coecients obtained for the observations of F4. It is clear that residuals are not a minimum since optimum coecients are used to determine the minimum residuals of the sum of the squares for both F3 and F4 (based on the results of the rst approach).

Table 3 Outputs of the second approach GA giving the minimum SSR for all ights based on Eq. (3). Flight F1 F2 F3 F4

442.60 812.50 457.75 452.30

2.92 105 4.01 105 2.77 105 2.89 105

Min. SSR 4577.7 1812.6 116.6 65.2

imum residual of the sum of the squares (which is calculated as 65.45) is found to be 88.8. To obtain these results for one ight, the GA program was run for a total of 26.4 hours (assuming an average of 10 seconds for each pair, totaling 19 pairs and a 500 runs of the GA program). Since the time needed for each pair is 83 minutes (10 seconds 500 run), the maximum y output is found after 3476 seconds (57.9 minutes), while the y output giving the minimum residual of the sum of the squares is found after 3239 seconds (53.9 minutes). 6. Assessment of advantages of the GA approach To determine the advantage of the GA method over the simpler approach using MS Excel, an additional analysis has been carried out. Considering the coecients of the functions obtained by MS Excel, we can observe that all of the descents have similar values, except those for F2. The rst coecients are between roughly 400 and 500 for F1, F3 and F4, while it is around 800 for F2. From this information, a range between 0 and 1000 is accepted during the process of determining this coecient. To nd the best possible solution the accuracy is kept as low as 106 . This means that 109 numbers are investigated leading to a very slight interval. A similar process is performed for the second (exponential) coecient. The computationally generated rough result is around 3 105 . For this coecient, the limits are determined as 1 103 and 1 107 . The corresponding accuracy utilized is 108 , leading to 99,990 ((103 107 )/108 ) different decimal values between those limits. The reason for using such low numbers is to obtain coecient values that are as precise as possible. This leads to provide better results and, hence, less residuals. Another benet of the GA approach is that it provides the ability to generate a model for a given descent, which can be generalized and then utilized for other descents. In Table 4, the coecient values obtained via the MS Excel and GA approaches are shown. It can be seen that the coecients obtained by the GA approach differ from those found with MS Excel. The effects of these different coecients on the SSR of the fuel

Firstly, the impact is assessed of increasing altitude by increments of 30 m (100 ft), 152 m (500 ft) and 305 m (1000 ft) from a 3050 m (10,000 ft) altitude. Secondly, the same assessments are performed from an altitude of 6100 m (20,000 ft). Similar results are observed for these two groups. The results fall into two parts, as expected, one involving F1, F3 and F4 and the other involving only F2. In the rst part, the altitude increases of 30 m (100 ft), 152 m (500 ft) and 305 m (1000 ft) yield average decreases in fuel ow of 0.3%, 1.2% and 1.9%, respectively, while in the second part, the same altitude increases yield average decreases in fuel ow of 0.3%, 1.2% and 1.5%, respectively. These results suggest that, during descent, performing low level ight at an altitude of 305 m (1000 ft) higher could lead to a decrease in fuel ow of 1.9%. Since the potential time in this mode of low level ight is on the order of up to 10 or 15 minutes, the amount of fuel saving appears to be small for a single aircraft ight. However, this fuel saving could be signicant when all descent operations of an airline annually are considered. Lastly, results are reported on the time duration required to achieve solutions with the GAs. Since the investigated differentials for variables are small, the time required for the GA program to nd solutions can be long (for a standard PC having a Core Duo E4500 2.2 GHz processor and 3 GB of RAM). For instance, nding a solution for only one observation pair with 100 iterations may take 620 seconds, due to the population increasing from 100 to 200. Also, the time required to determine a maximum y output and a y output giving a minimum residual of the sum of the squares can be demonstrated through an example. For F4, the maximum y output is found to be 5,777,948, while the y output for the min-

E.T. Turgut, M.A. Rosen / Aerospace Science and Technology 17 (2012) 6573

73

Table 4 Summary of coecients determined by different methods. Flight ID MS Excel R2 F1 F2 F3 F4 h denotes altitude (ft). 0.959 0.990 0.995 0.998 Function 439.44e0.000029h 825.73e0.000041h 456.73e0.000028h 451.73e0.000029h Function obtained by GA First approach (Table 2) 442.69e0.000029h 813.57e0.00004h 459.08e0.000028h 452.70e0.000029h Second approach (Table 3) 442.60e0.000029h 812.50e0.00004h 457.75e0.000028h 452.30e0.000029h

Table 5 Summary of the SSR values determined by different methods. Flight ID Fuel ow SSR Obtained by MS Excel F1 F2 F3 F4 4601 1867 118 66 Obtained by the rst approach (Table 2) 4590 1815 117 66 Obtained by the second approach (Table 3) 4578 1813 117 65

References
[1] Anonymous, Directive 2008/101/EC of the European Parliament and of the Council, Ocial Journal of the European Union 52 (2009), L 8:3-21. [2] Bureau of Transportation Statistics (BTS), Airline fuel cost and consumption (U.S. Carriers Scheduled): January 2000May 2010, Research and Innovative Technology Administration, U.S. Department of Transportation, Washington, DC, 2010. Retrieved July 4, 2010, from http://www.transtats.bts.gov/ fuel.asp?pn=0&display=data4. [3] J.P. Clarke, M. Lowther, L. Ren, W. Singhose, S. Solak, A. Vela, L. Wong, En route trac optimization to reduce environmental impact, PARTNER Project 5 Report, Partnership for Air Transportation Noise and Emissions Reduction, 2008. [4] EUROCONTROL, Continuous descent approaches, Implementation guidance information, 2008. Retrieved July 4, 2010, from http://www.eurocontrol.int/ environment/public/subsite_homepage/homepage.html. [5] EUROCONTROL, Airport environmental partnership, 2008. Retrieved July 4, 2010, from http://www.eurocontrol.int/environment/public/subsite_homepage/ homepage.html. [6] C. Evans, P.J. Fleming, D.C. Hill, J.P. Norton, I. Pratt, D. Rees, K. RodrguezVzquez, Application of system identication techniques to aircraft gas turbine engines, Control Engineering Practice 9 (2) (2001) 135148. [7] R. Fantinutto, G. Guglieri, F.B. Quagliotti, Flight control system design and optimisation with a genetic algorithm, Aerospace Science and Technology 9 (2005) 7380. [8] L. Gallar, M. Arias, V. Pachidis, R. Singh, Stochastic axial compressor variable geometry schedule optimisation, Aerospace Science and Technology 15 (5) (2011) 366374. [9] J.V. Hansen, Genetic search methods in air trac control, Computers & Operations Research 31 (3) (2004) 445459. [10] L.U. Hansen, P. Horst, Multilevel optimization in aircraft structural design evaluation, Computers & Structures 86 (12) (2008) 104118. [11] X.-B. Hu, E. Di Paolo, An ecient genetic algorithm with uniform crossover for air trac control, Computers & Operations Research 36 (2009) 245259. [12] IATA Report, Flight eciency plan: Fuel and emission savings, IATA, EUROCONTROL, CANSO, 2008. Retrieved July 4, 2010, from http://www.iata.org/ pressroom/Documents/Flight%20Eciency%20Plan.pdf. [13] J.-G. Juang, H.-K. Chiou, L.-H. Chien, Analysis and comparison of aircraft landing control using recurrent neural networks and genetic algorithms approaches, Neurocomputing 71 (1618) (2008) 32243238. [14] A. Keskin, D. Bestle, Application of multi-objective optimization to axial compressor preliminary design, Aerospace Science and Technology 10 (2006) 581 589. [15] Y. Kuroki, G.S. Young, S.E. Haupt, UAV navigation by an expert system for contaminant mapping with a genetic algorithm, Expert Systems with Applications 37 (6) (2010) 46874697. [16] T.T.H. Ng, G.S.B. Leng, Application of genetic algorithms to conceptual design of a micro-air vehicle, Engineering Applications of Articial Intelligence 15 (5) (2002) 439445. [17] A. Omran, A. Kassem, Optimal task space control design of a Stewart manipulator for aircraft stall recovery, Aerospace Science and Technology 15 (5) (2011) 353365. [18] A. Omran, A. Kassem, G. El-Bayoumi, M. Bayoumi, Mission-based optimal control of Stewart manipulator, Aircraft Engineering and Aerospace Technology 81 (3) (2009) 226233. [19] R. Pant, J.P. Fielding, Aircraft conguration and ight prole optimization using simulated annealing, Aircraft Design 2 (4) (1999) 239255. [20] T. Pejovic, N.B. Noland, V. Williams, R. Toumi, A tentative analysis of the impacts of an airport closure, Journal of Air Transport Management 15 (2009) 241248. [21] M.J. Soomer, G.J. Franx, Scheduling aircraft landings using airlines preferences, European Journal of Operational Research 190 (1) (2008) 277291. [22] User Manual for the Base of Aircraft Data (BADA), Revision 3.6, 2004. Retrieved May 21, 2010, from http://www.eurocontrol.int/eec/gallery/content/ public/document/eec/report/2004/022_BADA_User_Manual.pdf. [23] F.J.M. Wubben, J.J. Busink, Environmental benets of continuous descent approaches at Schiphol Airport compared with conventional approach procedures, Report NLR-TP-2000-275, National Aerospace Laboratory, 2000.

ow are summarized in Table 5. In almost all descents, the coecients found by the GA approach yield a better SSR. In the case of F4, the SSRs from using MS Excel and the rst GA approach were equal, but the second GA approach nonetheless provided a slightly better SSR than that for MS Excel. It can also be observed that the GA model response to noise data is better than that for the MS Excel approach. Considering the relationship between fuel ow and altitude for all descents, it can be seen that the relationship is better for F3 and F4, compared to that for F1 and F2 (see Table 5). Also, it can be seen from the SSRs of these descents that the SSR difference between the MS Excel and GA models is not large for F3 and F4 (where there is good relationship between fuel ow and altitude), while relatively large differences (in favor of GA models) are exhibited for F1 and F2. This result indicates that, when data have noise or the R 2 of data is relatively low, a higher signicance can be obtained via the GA approach. 7. Conclusions The present analysis and results indicate that there is an inverse relationship between fuel mass ow rate and altitude during descent. In this study this relationship is investigated using a genetic algorithm which is an effective technique to solve minimization and maximization problems in terms of duration, sensibility and accuracy. Small intervals of variables are investigated for a wide range of data, which are obtained for real ight conditions from the ight data records. Models have been developed to identify the relationship between fuel ow and altitude. The predictions obtained from the models agree well with actual observations. It is also determined that a best result found for a single pair of fuel ow and altitude is not in general the best result for all data pairs. The present results are consistent with the view that, during aircraft descent, maintaining a higher altitude as long as possible (avoiding low level ight) provides better fuel economy. The relationship developed in this study can also be useful for airline fuel saving programs. An understanding of the relation between vertical descent prole and fuel ow will allow the total fuel saving potential for an airline to be evaluated straightforwardly for various descent and approach methods or holding procedures for all airports and all landings. In addition, it may be possible to assess the environmental impact connected to fuel consumption during landing should actual fuel data be available.

Potrebbero piacerti anche