PREDICTINGAIRPLANEWEIGHTANDBALANCE

Presented by: Tommaso Ferrari (xbl574)

Alicia Mosquera Rodríguez (dlx861)Andreas Malthe Faber (qzj517)

Data provided by: A-ICEAviation Information and

Communication Engineering

TheDatabase:Inputs

The database contains data referring 64’000 flights to and from the MPXairport in Milan.Most of the features are various parameters that indicate weights andcenters of gravity in commercial and passenger flights. To cite a few of themost important examples:

● DOW_W/I: Dry Operating Weight, weight/index (float that expressesCOG)

● TOW/I: TakeOff weight/index● PAX_W/I: Passenger weight/index● Dist: Distance of the flight

2

TheDatabase:Outputs

There are four features that we are trying to predict with our machinelearning algorithms:

● ZFW/I: Zero Fuel, weight/index These are the ideal values ofweight and center of gravity to maintain the ideal flight attitude, whichreduces wear on the aircraft.

● Blockfuel/Burneoff: Weight of fuel at takeoff/Weight of consumed fuel

Fuel is one of the main expenses for an airline3

Pre-ProcessingCategoricalInputs:ACtype,Config,Version.

● ACtypeandConfigturnedintointegers OrdinalEncoder● Versiondividedinto3integers(NaNsaswell):F(firstclass),C(club),Y(economy)

Normalizationofnon-categoricaldata(inputandprediction):

● Multi-peakdataQuantilerange10/90%● Scaled*200(tomatchcategorical)

4

NeuralNetworkStructureInputLayer:17inputsNumberofDenselayers:1-3LayerSize:4-30Dropout:0-0.5OutputLayer:4regressions

Activation:relu,softplus,seluOptimizer:Adam,AdaMaxLearningrate:0.0001-0.1(logsampling)loss=MSE

5

ResultsofFirstHyperparameterSearch

● Arandomsearchof250trials,trainingfor20epochsmaxbutwithearlystoppingandapatienceof4epochs

6

ExampleModel● AfterfindinggoodHPcandidatesinthesearch,modelsweretrainedwithcrossvalidation

● EvenwhentheMSEwasn’tterrible,manyofthenetworksstruggledseverelywiththeoutliers

7

FurtherSearchesandTraining

● 3totalfurtherHPsearcheswereperformed● Arandomsearchof500trialswithreducedsearchspacefromthelasttime

● Ahyperbandsearchof270totaltrialsusingthesamespaceaspreviously

● Ahyperbandsearchof270totaltrialsexploringdeeperandbroadernetworks,increasingthenumberoflayersfrom1-3to2-5andlayersizefrom4-30to4-60

8

TrainingtheTopContendersofEachSearch

9

TheFinalNeuralNetModel● Anensemblemodelwasconstructedusingthetopnmodelsfromatotalof150trainedmodels

● Theensemblesize(n)andaggregationmethodwaschosentominimizethelossonthetrainingset

● n=4,aggregation=median 10

XGBoostRegressor(1)

Implementation of gradient boosting that produces a prediction modelin the form of an ensemble of decision trees.1. HPO through random search of 50 trials

1. Train the model2. Perform cross-validation3. Score on test set:MSE = 0.032Score on individual decision tree: MSE = 0.442

Hyperparameter Range OptimizedHyperparameter

Learningrate 0.1-0.3 0.3

Numberofestimators 100-300 200

Maximumdepth 5-15 10

11

XGBoostRegressor(2)

Precisepredictionsandbettertreatmentofoutliers

12

FeaturePermutationImportance 13

Conclusionandoutlook

● LeastMSEfoundinthetree(0.032)followedverycloselybytheNN

ensemble(0.054)

● Alwaystryatreefirst:D

● Finalsuper-ensemble:tree+NN

● Thedecisionsinthetreearenotdifferentiable,sotheycannotbe

optimizedfinalmodelwouldbetheNN

● Optimizationproblemcanbedonewiththefinalmodel14

THANKYOUFORYOURATTENTION

15