text
stringlengths
0
120k
Fuel Burn Estimation Using Real Track Data, AIAA-2011-6881, 11th AIAA Aviation Technology, Integration, and Operations (ATIO) Conference
GBChatterji
20-22 Sep. 2011
Virginia Beach, VA
Chatterji, G. B. Fuel Burn Estimation Using Real Track Data, AIAA-2011-6881, 11th AIAA Aviation Technology, Integration, and Operations (ATIO) Conference, Virginia Beach, VA, 20-22 Sep. 2011.
Fuel consumption modeling in support of ATM environmental decision-making
DSenzig
GFleming
RIovinelli
Proceedings of the Eighth Annual FAA/EUROCONTROL Air Traffic Management Research and Development Seminar
the Eighth Annual FAA/EUROCONTROL Air Traffic Management Research and Development SeminarNapa, CA
June 29-July 2, 2009
Senzig, D., Fleming, G., and Iovinelli, R. Fuel consumption modeling in support of ATM envi- ronmental decision-making, Proceedings of the Eighth Annual FAA/EUROCONTROL Air Traffic Management Research and Development Seminar, Napa, CA, June 29-July 2, 2009.
I. IntroductionSince 2015, the National Aeronautics and Space Administration (NASA) in collaboration with the Federal Aviation Administration (FAA) and flight operators, has been deploying some of its technologies developed for Integrated Arrival Departure Surface (IADS) [1][2][3] to a few chosen airports as part of the Airspace Technology Demonstration-2 (ATD-2) project [4].For its phase 1 & 2, the ATD-2 project focused on operations at Charlotte Douglas International Airport (CLT), Charlotte, North Carolina, due to its complex ground surface, its location in one of the busiest air traffic corridors on the East Coast and its abundant air traffic (# 7 in movements worldwide for 2016, [5]).ATD-2 deployment resulted in an improved operational efficiency that saved an estimated 4.3 million pounds of fuel between October 2017 and January 2020 [6].This improved efficiency is due to a better data exchange, coordination of release time for overhead stream insertion and application of surface metering.After this successful deployment, ATD-2 extended its operation for its phase 3 by deploying to a metroplex environment, the Dallas-Ft.Worth (Texas) Terminal Radar Approach Control (TRACON) where multiple airports operate within common TRACON boundaries.For this phase 3, ATD-2 provided coordination between multiple airports and added the functionality to reroute flights expected to be delayed by restrictions (often due to weather conditions).For fiscal year 2021, the ATD-2 project was extended for an additional year during which its focus is to improve the benefits analysis of its phase 3 and implement a version of its scheduler algorithm that could run for most airports in National Airspace System (NAS).This NAS-ready scheduler is to be enabled by the implementation of Machine Learning (ML) services that will substitute the complex physical model of ground movements and airport configurations designed in the prior implementation of the ATD-2 project.These microservices will lend themselves to a quicker and easier adaptation to new airports and easier maintenance.These microservices will need to be sufficiently accurate to enable the scheduler to provide useful schedules to Air Traffic Control personnel.These microservices will use ML algorithms that will be trained on historical data available for most airports in the NAS.To produce predictions, the trained algorithms will be fed FAA System Wide Information Management (SWIM) data from the ATD-2 Fuser [7], a software that ingests data feeds from various sources, formats and deconflicts them.To enable this new implementation of the ATD-2 scheduler, several microservices would be required such as a service that estimates the unimpeded taxi-out time, one that estimates ON time for arrivals, one that predicts departure runways, etc.This paper describes how the model for predicting unimpeded taxi-out was developed and its performance.In section II, we will present an overview of previous works on taxi-out time predictions.In section III, we will describe the frameworks and architecture used to build and deploy the taxi-out ML prediction service.In section IV, we will present some results produced by the microservice and we will conclude in section V.
II. Taxi-out PredictionsTaxi times of aircraft on the airport surface have been a long standing subject of study.Historically, the community started by developing sophisticated physical models of the airport surface and of aircraft movement [8][9][10][11][12][13][14][15].Recently, models based on historical data and machine learning algorithms have been put forward as more sustainable solutions.The literature of taxi time predictions is fairly large, but the wide range of datasets renders a comparison quite difficult.Prior studies focused on various airports, different measured durations, etc., to calculate their models' performance.The largest portion of the literature concentrates on modeling impeded taxi times, which is the actual duration of time it takes an aircraft to travel on the airport surface, including any time waiting for air traffic control (ATC) instruction and surface traffic congestion along the way.Some models used first-come-first-served queueing theory [16] and others used fast-time simulation calibrated on historical data [17].In recent years, using historical data and machine learning techniques, several studies have tried to estimate the impeded taxi time.Early papers studying the estimation of impeded taxi time with ML techniques used either linear or log-linear regression algorithms [18][19][20][21].Later papers often compared linear regression with more complex algorithms such as Singular Vector Machine (SVM), Random Forest or Fully Connected Neural Network [22][23][24][25][26][27].[28] studied a reinforcement learning approach in which the actions are the taxi out time estimate and the states are defined by parameters such as the number of aircraft taxiing, time, previous taxi times, etc. [29] defined a 2-dimensional embedding to capture the airport state using a fully connected Siamese Deep Neural Network, and used a kNN model to compute the taxi time from the 2D-embedding.In a large number of studies previously listed, the unimpeded taxi time is an input to the impeded taxi time calculation.The unimpeded time represents the duration of time the taxiing aircraft would spend if it had no constraints acting to slow it down.One of the most employed estimate of the unimpeded taxi time is to use the 10 th percentile of the taxi time for a given pair of stand and runway.Another technique developed by the EUROCONTROL PRU [30] uses taxi time as a function of the airport congestion for a given stand-runway pair and assigns unimpeded taxi time as the median value of taxi times that have a congestion below a saturation value.Recently [31] estimated the unimpeded taxi time using a node-link model at the Incheon International Airport (RKSI, 1553 nodes and 2034 links) and Airport Surface Detection Equipment (ASDE) data from April 1 st to 20 th , 2015.The unimpeded taxi time is estimated by the sum of the taxi time of each link along a given path which must be provided (for instance by tools such as departure manager DMAN [32]).The model takes into account the plane weight category, whether the flight is an arrival or a departure, as well as whether aircraft are turning before and after each link.The properties of each link are calculated from historical data.In this paper, we will present a method to evaluate unimpeded taxi out time that uses historical data and a ML algorithm instead of the physical models described previously.This method is similar to the techniques employed to estimate impeded taxi time.We will then use the unique capability of ATD-2 to test its performance.
III. Taxi-out Machine Learning Service
A. Frameworks and LibrariesWe decided to use the Python programming language and several common, open-source packages to develop these ML services.The main packages used for this development are Kedro * , MLFlow † and scikit-learn ‡ (see Fig. 1 to visualize where each framework is used).We endeavored to institute best practices across the Python model development.We decided to implement the Kedro framework as the backbone of each training pipeline used for training and storing the models.The Kedro framework enforces software engineering best practices for developing machine learning code.For instance, Kedro eases collaboration among our team given that the structure of all the pipelines is very similar.It also allows abstract data access (avoid hard-coded local directory path, database access, etc.) and organizes each node of the pipeline into direct-acyclic graphs (DAGs), that enforce input/output readiness.These DAGs can be visualized with an additional Python package called kedro-viz.The ML model, which encompasses feature engineering and the proper data science model, is contained in an extended version of a scikit-learn pipeline and is registered and stored on a MLFlow server.MLFlow allows the developer to store each run of a model (see Fig. 2 for an example) with a wealth of information such as the Git hash of code used in the run, the environment in which the model ran, the parameters of the model, performance metrics, etc.The MLFlow-stored models are then deployed as representational state transfer (REST) application programming interface (API) microservices using Flask § framework.These REST APIs are used by the orchestrator to feed the scheduler.Figure 1 shows a summary of the pipeline steps used to train models and the pieces of the training pipeline that are transferred into the live prediction service through its storage on the MLFlow server.
B. MethodologyThe taxi-out pipelines are separated into three phases: data query, data engineering and data science model.The data query is used in the training model phase to access historical data and primarily accesses data through SQL queries and stores them locally into CSV files.This part of the pipelines is common to all our three pipelines, i.e. the pipeline to estimate the ramp taxi time (from stand to spot), the AMA taxi time (from spot to runway) and the total taxi time (from stand to runway).The data engineering does some basic transformations and filtering to feed the data modeling, such as fusing several sources of information, removing data with incomplete information, and marking data for test or train groups.Some of these processes like getting the proper runway will be re-implemented in the predictive ML service and some, like separating train and test groups, won't be necessary (see Fig. 1).One notable filtering applied to our training and testing data is to limit the data to unimpeded taxis.To characterize a taxi as unimpeded, we used Airport Surface Detection Equipment, Model X (ASDE-X) speed data in the AMA and selected flights whose taxiing speed did not drop below 4 knots for 90% of the taxi time.This filter is applied for our AMA and total taxi estimate.For the ramp taxi, we require the number of arriving and departing aircraft in the ramp to be in the lowest percentile (60% and 30%) as well as the number of departing aircraft in the AMA to be low (40 th percentile).These constraints seem to filter a good amount of impeded taxis while keeping about 15% of the flights in our training set.The data science part of the pipeline processes the data into their final form for modeling (feature engineering), fits the data with a model, and computes some performance metrics.The boundary between feature engineering and data engineering is not always well defined but our main criteria for implementation in feature engineering are complexity and specificity.Transformations applied within the data science part will be saved into our pipeline container that will be served through the REST API.A driving goal in this project was to have a process that can easily be adapted to any airport in the NAS.The feature selection, feature engineering, and model algorithms were all carefully developed with this goal in mind.As such, the data science pipeline of the unimpeded taxi-out time takes in a limited number of accessible information: carrier name, departure runway, and departure gate (the aircraft type has also been used but did not seem to improve the models).These features are then encoded with a standard one-hot-encoder for the carrier and departure runways and some dedicated encoder for the departure gates.We looked at different options for encoding the gate information, two looked the most promising: clustering by terminal name, and clustering by similar historical taxi time.The first of these gate encoders uses the departure gate name to identify the terminal and creates one feature column per terminal.It populates the appropriate column with the gate number.The rationale is that gates at the same terminal have a somewhat similar unimpeded taxi time and that gates with similar numbers within a terminal have an even closer taxi time.The latter assumption is unfortunately dependent on how the gate numbers are distributed in the terminal.The second of the gate encoders uses the data to create a number of clusters to group the gates.This encoder uses the taxi time to three runways for each gate to group them in similar buckets.Gates without a sufficient number of taxis to at least 3 runways are added to the clusters using the available runways. Figure 3 shows an example of the resulting clusters for the unimpeded AMA taxi time at KDFW.The left image shows a top-down map of the KDFW airport, each dot represents a gate location along the 5 main terminals and its color indicates the cluster it belongs to.The right image shows the position of each gate (solid blue circle) in a 3 dimensional taxi-time space, each solid blue circle is highlighted by a larger colored circle indicating its cluster.We found similar accuracy between the two methods, most likely because most of the information resides in the terminal.We opted to use the latter one (clustering by historical taxi time) since it provides us with some good insight into the gate behavior.Once the data has been processed by the feature engineering, the model for the unimpeded taxi-out time is defined and fitted on the calculated features with a gradient boosted decision tree from the XGBoost library ¶ .The loss function is defined as the squared error with respect to the actual taxi time.The model hyperparameters and the number of clusters defined in the gate encoding are tuned to improve the squared error.A baseline model can be executed at the same time as the predictive model.For instance, our current baseline model groups flights by pair of runway and terminal and returns the average taxi time.This baseline model tests the proper behavior and improvement in accuracy of the predictive model.All resulting metrics and plots are stored on the MLFlow server along with information about the run (see Fig. 2).The MLFlow UI allows for quick comparison between models and selection of the best predictive model for each airport for each taxi out pipeline (ramp, AMA, total).These models can be retrieved from the MLFlow server and deployed as a microservice that then can be used by an orchestrator to compute schedules for a specific TRACON or airport.
IV. Results and PerformanceThe unimpeded taxi out models were trained on data from June 1 st 2019 to December 31 st , 2019 at the KDFW and KCLT airports.The total number of departing flights for each airports in this date range was about 220k and 170k respectively.When filtering out incomplete data and impeded taxis, the remaining dataset contains about 47k and 20k flights for the AMA and total taxi at KDFW and KCLT respectively and about 20k flights for the ramp taxi at both airports.We split our data further, using 80% to train our models and 20% to test them.To assert the accuracy of our predictions, we calculated the median absolute deviation (multiplied by 1.4826 to match a Gaussian 1 sigma spread, referred to simply as MAD throughout the rest of this paper) and the median value of the residuals (predicted taxi time minus truth value).The median of the residuals informs us about the amount of bias the model introduces, whereas the MAD indicates the precision of our estimate.The following results show accuracy measurements on the test dataset.Calculating accuracy for unimpeded taxi time presents another challenge compared to impeded taxi time: the true value of the taxi time is not clearly defined, since we do not know if a taxi is really unimpeded.Whereas the model loss function minimizes the residuals with respect to the actual taxi time, we decided to use the STBO (Surface Trajectory Based Operations) system estimates as truth values since they are available at KDFW and KCLT and we believe that STBO unimpeded taxi time estimates are closer to the true unimpeded taxi time than the actual taxi time.Table 1 summarizes the accuracy of our models, and demonstrates that the largest source of uncertainty seems to come from the ramp area.The complexity of the ramp operation explains the difficulty in calculating a good estimate of the unimpeded ramp taxi time and could also imply some larger intrinsic spread of its true value.Another issue is that we do not have reliable ASDE-X data in the ramp, preventing us from filtering our flights with the aircraft speed.The training data for the ramp area might not be as free of unimpeded taxi as the training data used for the AMA taxi time model.Overall the total taxi time has an relative uncertainty of 11 to 14%, the AMA taxi time has an relative uncertainty of 9 to 14 % and the ramp taxi time has an relative uncertainty of about 20%.The models also have an overestimating bias of about 1 to 1.5 minutes compared to the STBO estimates.This difference could be due to some impeded taxis making it into our training data or it could be that the STBO system is underestimating unimpeded taxi time.Figure 4 & 5 show some additional insights into the model accuracy.Figure 4 presents the distribution of the ML model residuals (blue area), defined as the model prediction minus the values coming from the STBO, for the ramp, AMA and total taxi at KDFW.In addition to the ML model predictions, the baseline model residuals (orange area) is shown for comparison.On average, ML models perform about 20% better than baseline models.Figure 5 shows that this improvement might be due to a lower discretization of the taxi time estimate, which can be seen as the orange horizontal lines formed by the baseline orange dots.Another improvement is the ability of the ML model to predict more extreme taxi time (especially longer taxi time).Both these effects reflect the inability of the baseline model to distinguish the different taxi time of gates at the same terminal and the fact that the baseline model does not use the carrier information.While Fig. 4 shows that the residual distributions are fairly tight, AMA taxi time residuals at KDFW have some large wings.After digging through our data, we discovered that most of the largest residuals in this distribution might be due to a few unreliable STBO estimates, that for some reason are either very large or very small.For instance, Fig. 5 shows that some STBO unimpeded AMA taxi time at KCLT are less than 100 seconds.More work needs to be done to improve the truth set and understand why the predictions differ from the true taxi time.In order to test if our model could be served to predict the current unimpeded taxi time estimate, we gathered data from August 1 st , 2020 to September 1 st , 2020 as well as from February 5 th , 2021 to April 21 , 2021.These data come from our PostgreSQL database and not from the ML prediction live system, but they allow us to characterize the performance degradation due to systemic change at the airport.One example of these changes is the median number of departures, which for DFW goes from about 1050 in 2019 to about 750 in 2020 back to about 860 in 2021 for the date range listed previously.Figure 6 shows how much the performance of the models degrades.Most taxi time uncertainties increase by at most 10 to 15%, some biases could get as much as 30% larger (e.g.KCLT total taxi time).To reduce the model degradation, models could be trained again on more recent data, however the fast changing traffic volume might make the new models degrade again.Furthermore, a real true test of the model accuracy will come when the unimpeded taxi out model predictions are computed in the live system.
V. ConclusionIn this paper, we showcased a machine learning service to calculate the unimpeded taxi out time using some open-source frameworks that enable a simplified CI/CD (Continuous Integration/Continuous Deployment) process.Though we developed the process with KCLT and KDFW airports, we limited the features to those which will be readily available across the NAS through SWIM.The trained models have a 1 sigma uncertainty as measured with the median absolute deviation mostly between 1 and 1.5 minutes and an overestimating bias also between 1 and 1.5 minutes.In relative terms, the AMA and total taxi time have an uncertainty of about 10 to 15% whereas the ramp taxi time has an uncertainty of about 20%.These models are currently deployed in a shadow live system for the Dallas -Ft Worth TRACON airports and feed a flight scheduler.The scheduler predictions as well as the service predictions that feed it will be compared to the STBO system.This comparison will allow characterization of the accuracy of each ML service in real-time predictions, as well as to check whether theses service predictions are accurate enough to produce useful schedules for operational planning.A future parallel effort is to train these models for other airports in the NAS, especially for airports in the North East Corridor, where deploying an accurate flight scheduler could provide substantial value.Fig. 11Fig. 1 Diagram of the ML pipeline used to train the taxi-out time estimate (top) and diagram of the ML service implementation (bottom).Kedro constitutes the backbone of the training pipeline, scikit-learn serves as a model container, that will be deployed as a microservice.
Fig. 22Fig. 2 Example of the MLFlow User Interface (UI) on a web-browser, that allows easy access to previous runs/models and to compare several models.A user can access further information about each model run (e.g., model pickle, plots) using the UI hyperlinks.
Fig. 33Fig. 3 Left: KDFW hierarchical clustering on AMA taxi time to three runways.This approach can easily be extended to any airport in the NAS without a complex adaptation.Terminal names are written in black, runways are written in blue.Right: Same color-coded clusters distributed in the 3D AMA taxi time space (spatial 2D and mark size 1D).The blue cluster seems to have the shortest AMA taxi to 17R, green cluster to 18L and the red cluster to 36R.
Fig. 4 Fig. 5 ML45Fig. 4 Histograms (in blue) of the residuals (in seconds, for ramp, AMA, total taxi from left to right) as defined as the STBO estimated unimpeded taxi time (our truth values) minus the predictions from our ML model.As a comparison, the residuals of the baseline model estimate is overplotted as the orange histograms.These distributions include only the KDFW test dataset residuals.
Fig. 66Fig. 6 Evolution with time (datasets from 3 different years) of the median of the residuals with ±1 MAD error bar for KDFW and KCLT using predictions from the 2019 trained model.
Table 1 Metrics of accuracy, median and MAD of residuals as compared to STBO predictions, as well as the median taxi time from STBO for KDFW and KCLT airports. Negative median of the residuals indicate that the estimates are smaller than the STBO predictions.1KDFWKCLTMedian [sec] MAD [sec]Median Taxi [sec]Median [sec] MAD [sec]Median Taxi [sec]ramp-6±280±240162±398±5522AMA61±152±135843±227±2296total95±3113±478693±491±4808
§ https://flask.palletsprojects.com/en/2.0.x/
¶ https://xgboost.readthedocs.io
* https://github.com/quantumblacklabs/kedro† https://mlflow.org‡ https://scikit-learn.org/stable/
SAEngelland
ACapps
KDay
MSKistler
FGaither
GJuro
Precision Departure Release Capability (PDRC) Final Report
2013
Engelland, S. A., Capps, A., Day, K., Kistler, M. S., Gaither, F., and Juro, G., "Precision Departure Release Capability (PDRC) Final Report," 2013.