1 – The A-CDM concept
Airports are busy places where different stakeholders have key roles and a common goal to manage safely and efficiently the flow of flights departing and arriving. However, airport infrastructures are not exploited in the most optimal manner and increasing traffic makes it difficult for operations to be proactive rather than reactive. This is due to the lack of good information sharing procedures, each of the stakeholders involved in operations has a piece of the information rather than the global picture.
With a 3 hour look-ahead, the airport collaborative decision making (A-CDM) process enables all stakeholders to benefit from sharing the same information as early as possible in order to take informed decisions. Milestones and relevant flight details are updated and shared in order to have accurate Off-Block Times (OBT) and Take-Off Times (TOT) for better situational awareness. This information sharing leads to better traffic flow management at the network level.
1.1 – The taxi time contribution
The taxi time is the time spent by an aircraft between the runway and the parking stand.
For a departure flight, it is called the Taxi-Out Time (EXOT), spent from the parking P to the runway take-off point D.
For an arrival flight, it is called the Taxi-In Time (EXIT), spent from the touch down on the arrival runway A to the parking stand P.
1.1.1 – Classical airport flow management
The Taxi-Out Time is an important part of the take off time time estimation. Indeed on the airport side, a set of operations (cleaning, catering, passenger boarding…) need to be done for an aircraft to be ready to start its flight. This time where the aircraft is targeted as ready to off-block (TOBT) it is associated to the Estimated Taxi-Out Time (EXOT) in order to estimate the target time of take-off (TTOT).
|TTOT = TOBT + EXOT|
The same concept applies for arriving flight, as the aircraft Estimated Landing Time (ELDT) is used to predict the Estimated In-Block Time (EIBT) using the Estimated Taxi-In Time (EXIT) from the runway to the parking. This IBT is important in order to organize the aircraft turnaround process.
|EIBT = ELDT + EXIT|
Taxi time accuracy is key in TTOT and EIBT estimation accuracy. It gets even more important when the pre-departure sequencing is part of the airport management process.
1.1.2 – Pre-departure sequence oriented airport flow management
Pre-departure Sequencing (PDS) enables reactive handling methods of flights by ATC on a “Best Planned Best Served” principle in order to improve push back sequence and reduce parking occupancy time.
Indeed, when the number of flights to take off is important in an airport, the runway throughput optimization allows a better use of airport runway and thus enhances the airport capacity.
PDS involves the computation of a sequence of off-block times based on a target optimized sequence at the runway:
As we can see the Taxi-Out Time is used twice in the sequence computation. An inaccurate taxi time leads to an inefficient runway sequence and impacts the global airspace traffic flow management.
According the European airline delay cost reference values, Updated and extended values, Version 4.1 (University of Westminster 24 December 2015) the costs per minutes of taxi delay goes from 12€ to 350€ depending on the aircraft type and the total delay.
The same document proposes a high level average delay cost of 100€ per minute.
An accurate taxi time estimation will prevent the aircraft to wait on the taxiway or worst, to delay other aircrafts due to congestion at the runway or on the taxiway. Accuracy results in delay reductions and thus great cost reductions.
1.2 – The taxi time estimation methods
Several models exist today to estimate taxi times. All of them tend to increase the number of parameters to take into account in estimating the taxi time. You will find hereafter a presentation of classic taxi time models.
1.2.1 – Variable taxi time
Variable taxi time is a model of taxi time estimation based on a table where several input variables are combined in order to estimate the taxi time. Commonly used input variables are :
- Runway used by the aircraft
- Parking stand the aircraft is associated to
- Flow of the aircraft (departure/arrival)
- Aircraft type
Based on historical actual taxi time (actual take-off/landing time regarding actual off-block/in-block), a taxi time estimation table is built.
This basic model fills the need for airport having simple topology. But it becomes too simplistic when the number of aircraft on the taxiway increases or when the airport process induces non-linear behavior on taxi-time.
1.2.2 – Dynamic taxi time
Dynamic taxi time is a model that tends to estimate the taxi time using more dynamic data, including the current airport load (taxi-way load) or the current airport topology (case of the closed taxiway). This is a computation based model that tends to represent real physical behavior of aircraft and surface management.
This model is similar to a “Waze” model of the airport taxiways. It predicts the taxi time based on the constraint of aircraft crossing, route availability and aircraft taxi performances.
This type of intensive computation is comparable to what is done for aircraft classical trajectory prediction, using the BADA model, where computation tends to represent physical behavior.
1.2.3 – Next-Gen taxi time ?
At the era of Artificial Intelligence (AI), the question is “can we use AI to compute accurate taxi-time estimates?” We can imagine a taxi time prediction that would be able to adapt to unexpected events, or to be more accurate depending on a great number of parameters, or even to enable events to adapt to trend evolution of the airport process…
2 – Artificial intelligence and machine learning
Artificial Intelligence is the ability of a digital computer or computer-controlled robot to perform tasks commonly associated with intelligent beings. The aim of AI is to improve computer functions which are related to human knowledge, for example, reasoning, learning, and problem-solving.
2.1 – Machine Learning concept
Machine learning is an application of Artificial Intelligence that provides systems the ability to automatically learn and improve from experience without being explicitly programmed. Machine learning enables analysis of massive quantities of data, with the objective to look for patterns and make better decisions in the future based on the provided examples.
Machine learning algorithms are often categorized as supervised or unsupervised:
- Supervised machine learning algorithms can apply what has been learned in the past to new data using labeled examples to predict future events. The history data labelling (supervision) is usually done automatically;
- Unsupervised machine learning algorithms are used when the information used to train is neither classified nor labeled. Unsupervised learning studies how systems can infer a function to describe a hidden structure from unlabeled data.
Supervised machine learning algorithms are good to learn from the past and to apply to predict the future. This kind of algorithm really fits the ATM world since:
- The amount of data is huge and accessible;
- Supervision is already done thanks to tracking and monitoring of flight by air traffic control.
2.2 – Machine Learning application steps
In order to use a supervised machine learning algorithm, the following steps are needed:
- Prepare labeled data from raw data: this process allows to classify the data by their characteristics and also to identify the associated supervised result;
- Split labeled data into two set: one training set that will be used by the machine learning algorithm and one test set that will be used to assess the created prediction model accuracy;
- Train the machine learning using the training data set (labeled data) and create the prediction model;
- Evaluate this prediction model against the test set that contains actuals of what happened.
Data preparation is an important part of the machine learning work. Data preparation is divided into several steps:
- Collect, discover and assess data: retrieve the data, learn to discover it and understand what has to be done for the data to become useful;
- Clean and validate data: remove wrong data (missing values, incorrect values), fill missing parts;
- Transform and enrich data (feature engineering): format data to reach an expected outcome, enrich data by adding extra computed information or by connecting data with other related information.
3 – Machine learning applied to taxi time estimates
Taxi time prediction is a good use case for machine learning. Indeed, thanks to data provided by our partner FlightAware, we have access to:
- A great number of data, since you can access information for each flight and there is tens of thousands of flights per day depending on airport size;
- The supervision of data, already done thanks to the actual flight positions that are monitored (ADSB, Space-ADSB, MLAT. Take a look at the map of their data coverage, it is impressive!).
Indeed, actual taxi-out times (AXOT) and actual taxi-in times (AXIT) can be computed thanks to actual landing times (ALDT), actual in-block times (AIBT), actual off-block times (AOBT), and actual take-off times (ATOT).
Moreover, machine learning is interesting for taxi time prediction since it is the natural extension of existing taxi time prediction models, Variable Taxi Time (VTT), that are based on a number of known and defined parameters:
- Runway QFU
- Parking stand
- Aircraft type
- Flow (departure/arrival)
Machine Learning will easily expand the number of parameters that can be used to make the prediction, and even more remarkably, it is able to integrate a new parameter and update its predictions in a quasi-automatic way.
3.1 – Additional data of interest for taxi time estimation
We can think of several characteristics of interest that can affect the taxi time of an aircraft. Below are some parameters that we used during our evaluation process:
- Taxiing traffic load information
- number of aircraft on the taxiway
- number of aircraft arriving/departing in the next 20 minutes
- Seasonal related information
- Hour of day
- Day of week
- Day of month
- Month of Year
- Procedure specific information
- Is deicing needed
- Company operating the aircraft
- List of closed taxiways
What is important to keep in mind is that characteristics shall be adapted to the Airport layout and constraints.
3.2 – Data preparation
Below is the data preparation performed for our taxi time prediction by machine learning model:
- Check the presence of AXIT or AXOT;
- Check data that allow computation of AXIT and AXOT;
- Check percent of presence of input attributes for computation.
- Decide on outliers removal if they bear no meaning (they could be the result of an entry error or of an event that can be monitored and could thus be predictable by the model);
- Fill missing data with appropriate completion strategy (default value, mean or median of the train set, etc…) when possible;
- Removal of data with missing attributes that cannot be completed.
- Removal of the data with no (computable) AXIT or AXOT
- Derive the OBT or IBT data to compute seasonal related information (hour of day, day of week, day of month, month of year…);
- Compute the taxiing traffic load information based on the list of flights;
- Format to machine learner algorithm data representation expectations.
The prepared data is then used in the model hyper-parameters (i.e. the model architecture and complexity) optimisation to yield a final trained model to test against the previously isolated test data.
3.3 – Result validation
We will compare our ML-taxi-time predictor to the table based computed taxi time (VTT) that serves as our baseline method here. We will focus on the comparison of the taxi time prediction to what actually happened. We will then try to cost the delays introduced by both methods to compare them.
3.3.1 – Estimates vs Actual analysis
First we compute the mean absolute error (MAE) and the mean squared error (MSE) for both the VTT and the ML-predictor:
|MAE = 1/n * ∑ |prediction – actual|
MSE = 1/n * ∑ (prediction – actual)²
Then we visualize the error distribution compared to actuals to get some accuracy figures of the methods.
3.3.1 – Cost saving evaluation
Finally we estimate the cost generated by taxiing induced delays with the support of the Westminster University study mentioned in §1.1.2 .
We will consider that the taxi time estimation produces a delay when it is too short, that is when the actual taxi time is greater than the estimated one.
The tables 27 and 30 of the study (seen in §1.1.2) enables us to compute the cost of the taxi time imprecision for both methods.
4 – Case study and results on a top-ten airport
4.1 – Context
We built a taxi time predictor using the ML method on a top-ten airport with 1350 aircraft movement per day. We then evaluated the ML Algorithm against the baseline VTT algorithm on 4815 departure and 4755 arrival flights randomly selected over a month of recorded traffic.
4.2 – Estimate vs Actual analysis
We obtained the following error metrics results (compared with the baseline VTT table prediction error):
We observe that the mean absolute error (in minutes) is more than halved by the ML model compared to the baseline VTT table prediction. Overall, the VTT table gives a mean absolute error of 4.49 minutes (over one month of data) while the ML model gives an error of 2.03 minutes.
Now let’s see how the taxi time prediction errors of both models are distributed. The x-axis of the following charts represents the prediction error in minutes compared to the actual taxi time and the y-axis represents the number of flights with a taxi time error of the x value. A positive error means that the prediction was greater than the actual value.
As expected, the predictions of the ML model are less sparse and are generally closer to the actual taxi time. The following table present the X-Minutes window based accuracy of the estimation using the VTT or ML models.
|Window accuracy definition||VTT Table||ML Algorithm|
|+/- 2 minutes window accuracy||36%||62%|
|+/- 5 minutes window accuracy||71%||92%|
As we can see estimates using the ML Algorithm are more precise than using the VTT Table.
4.3 – Cost saving evaluation using the ML predictor
We can then compute the cost generated from delays due to imprecision in taxi time estimation for both models. We consider that there is a delay when the prediction error is negative (i.e. the actual taxi time was greater than the estimation).
Using the same test dataset as before, we estimate a cost of 200 443 € per day generated by the VTT model and a cost of 66 977 € per day for the ML model. That represents a saving of 133 466 € per day for the airlines on the considered airport.
5 – Product availability
The VTT-ML predictor functionality is fully integrated into Innov’ATM’s AirportKeeperⓇ & SkyKeeperⓇ products. Feel free to contact us to get more information or even a demonstration of our product capability, using our Contact page!