Periodic Reporting for period 1 - ASTRA (AI-enabled tactical FMP hotspot prediction and resolution)
Berichtszeitraum: 2023-09-01 bis 2024-08-31
Air Traffic Management plays a crucial role in the safety and efficiency of aviation, and one of its functions is Demand-Capacity Balancing (DCB). Sector overloads occur when demand exceeds capacity and measures need to be taken to restore balance. These include actions by the Network Manager at flight planning stage, as well as Short-Term Air Traffic Flow and Capacity Management (ATFCM) Measures (STAM) by Flow Management Positions (FMPs) at a later stage – before or during flight. Such measures are based on predictions and do not consider traffic conflicts and complexity in a sector.
In flight, actual conditions may differ from the predictions, and aircraft end up deviating from their flight plan due to weather, tactical Air Traffic Control (ATC) interventions (e.g. to deconflict), etc. Thus, demand-capacity imbalances can still occur and are likely to create areas of high complexity which have to be resolved tactically by Air Traffic Controllers (ATCOs), resulting in higher workload. Even in sectors that operate below declared capacity, complexity varies dynamically as it depends on many factors beyond sector occupancy.
ASTRA is developing an AI solution to address this issue by predicting 4D Areas of Relatively High ATC Complexity (aka 4DARHACs) before the concerned traffic enters the area of responsibility of an ATCO. In addition, the solution suggests actions that can be taken by FMPs to resolve 4DARHACs without requiring re-sectorisation. The solution focuses on en-route traffic and is intended for use by FMPs in coordination with ATCOs, thus bridging the gap between ATFCM and ATC.
The objectives of ASTRA are to:
- Develop an FMP function to predict and resolve 4DARHACs one hour in advance
- Develop novel Human Machine Interface (HMI) concepts to allow interaction between FMPs and the FMP function
- Demonstrate and validate the FMP function and HMI.
The benefits of the solution are: lower ATCO workload and increased capacity at ATC unit level; increased cost-efficiency and operational efficiency; more fuel-efficient trajectories; and improved safety.
In flight, actual conditions may differ from the predictions, and aircraft end up deviating from their flight plan due to weather, tactical Air Traffic Control (ATC) interventions (e.g. to deconflict), etc. Thus, demand-capacity imbalances can still occur and are likely to create areas of high complexity which have to be resolved tactically by Air Traffic Controllers (ATCOs), resulting in higher workload. Even in sectors that operate below declared capacity, complexity varies dynamically as it depends on many factors beyond sector occupancy.
ASTRA is developing an AI solution to address this issue by predicting 4D Areas of Relatively High ATC Complexity (aka 4DARHACs) before the concerned traffic enters the area of responsibility of an ATCO. In addition, the solution suggests actions that can be taken by FMPs to resolve 4DARHACs without requiring re-sectorisation. The solution focuses on en-route traffic and is intended for use by FMPs in coordination with ATCOs, thus bridging the gap between ATFCM and ATC.
The objectives of ASTRA are to:
- Develop an FMP function to predict and resolve 4DARHACs one hour in advance
- Develop novel Human Machine Interface (HMI) concepts to allow interaction between FMPs and the FMP function
- Demonstrate and validate the FMP function and HMI.
The benefits of the solution are: lower ATCO workload and increased capacity at ATC unit level; increased cost-efficiency and operational efficiency; more fuel-efficient trajectories; and improved safety.
Work performed:
- Literature review: This focused on work related to complexity prediction and resolution, with a focus on en-route airspace and tactical operations. It included a discussion of complexity indicators, data sets, ML techniques and AI explainability, and its purpose was to guide the development of complexity metrics and algorithms for the ASTRA solution.
- Definition and validation of operational concept, use cases and requirements: The operational concept and use cases were defined in a workshop with operational experts. These were then used as a basis for requirements relating to trajectory prediction, 4DARHAC detection and dissipation, and user interaction. The operational concept, use cases and requirements were validated in a workshop with ASTRA’s advisory board.
- Gathering and pre-processing of data: Historical data - containing flight plans, etc. - was downloaded from sources such as EUROCONTROL's Demand Data Repository (DDR). An existing data pipeline was tuned to preprocess this data.
- Design and development of ML algorithms: Algorithms were developed to generate synthetic flight plans; measure complexity; and identify, track and resolve 4DARHACs. Several techniques were used for this purpose, including supervised AI, generative AI, clustering, and Reinforcement Learning (RL).
- Deployment of a simulation platform: A platform was set up to run different traffic scenarios and to train/test the ML models. This platform includes trajectory prediction and conflict detection tools, and can handle historical and synthetic data.
- Design of a standalone HMI for FMPs: The HMI will allow an FMP to be notified when a 4DARHAC is predicted; and to analyse and implement the solutions proposed by ASTRA. HMI requirements were gathered in a workshop with end users, and HMI mock-ups were validated by FMPs in October '24.
Main achievements:
- Operational concept, use cases and requirements for an FMP tool to predict and resolve 4DARHACs
- ML models to generate synthetic data
- An algorithm to measure complexity
- A clustering algorithm to identify and track 4DARHACs
- An algorithm to resolve 4DARHACs
- HMI mock-ups for the proposed FMP tool
- Literature review: This focused on work related to complexity prediction and resolution, with a focus on en-route airspace and tactical operations. It included a discussion of complexity indicators, data sets, ML techniques and AI explainability, and its purpose was to guide the development of complexity metrics and algorithms for the ASTRA solution.
- Definition and validation of operational concept, use cases and requirements: The operational concept and use cases were defined in a workshop with operational experts. These were then used as a basis for requirements relating to trajectory prediction, 4DARHAC detection and dissipation, and user interaction. The operational concept, use cases and requirements were validated in a workshop with ASTRA’s advisory board.
- Gathering and pre-processing of data: Historical data - containing flight plans, etc. - was downloaded from sources such as EUROCONTROL's Demand Data Repository (DDR). An existing data pipeline was tuned to preprocess this data.
- Design and development of ML algorithms: Algorithms were developed to generate synthetic flight plans; measure complexity; and identify, track and resolve 4DARHACs. Several techniques were used for this purpose, including supervised AI, generative AI, clustering, and Reinforcement Learning (RL).
- Deployment of a simulation platform: A platform was set up to run different traffic scenarios and to train/test the ML models. This platform includes trajectory prediction and conflict detection tools, and can handle historical and synthetic data.
- Design of a standalone HMI for FMPs: The HMI will allow an FMP to be notified when a 4DARHAC is predicted; and to analyse and implement the solutions proposed by ASTRA. HMI requirements were gathered in a workshop with end users, and HMI mock-ups were validated by FMPs in October '24.
Main achievements:
- Operational concept, use cases and requirements for an FMP tool to predict and resolve 4DARHACs
- ML models to generate synthetic data
- An algorithm to measure complexity
- A clustering algorithm to identify and track 4DARHACs
- An algorithm to resolve 4DARHACs
- HMI mock-ups for the proposed FMP tool
Techniques to create synthetic flight plans: The generation of synthetic data is an open area of research and is relevant to ATM because of the scarcity of ATM data. In ASTRA, synthetic data has two objectives: it overcomes the absence of detailed environmental data in the historical dataset; and it enriches the dataset with complex situations by controlling the intensity of the generated traffic.
Metrics to measure complexity: Measuring complexity is nontrivial and depends on many factors. Historically, an FMP assesses sector complexity based on an occupancy metric. In contrast, several metrics are defined in ASTRA; some are defined at sector level, while others characterise a cluster of aircraft. To have a unique complexity unit among the complexity metrics, each metric is scaled such that the complexity is expressed as a percentage. The metrics are then combined into a global complexity using a statistical model.
A complexity-based clustering algorithm to identify and track 4DARHACs: At each time step, the last known position of each aircraft, and its flight plan, are used to predict its future trajectory. Then, for each time step in the predicted future (up to one hour ahead), a clustering algorithm groups flights together and identifies 4DARHACs. A tracking algorithm then identifies the clusters found in the current time step that correspond to those in the previous time step.
A deep RL algorithm to resolve 4DARHACs: An RL agent is implemented to take 4DARHACs as input and to output actions representing solutions for each 4DARHAC. Proposed actions can be lateral, vertical or longitudinal clearances. These are ranked by probability, thus aiding the interpretability of the RL agent. Apart from the dissipation action itself, the agent also determines which aircraft should take the proposed action.
An HMI design for user interaction: This HMI allows FMPs to interact with the ASTRA solution. Amongst other things, it allows the FMP to get alerts; view and simulate the proposed dissipation actions; and select actions for implementation. The HMI is designed in a way that gives the user an insight into why the RL agent proposes certain clearances, and the impact (on complexity) of implementing a certain clearance.
These results will lead to improved predictability at local ATFCM level and improved confidence in sector DCB management.
More work will be needed to develop the solution beyond TRL2. Also, better regulatory standards are required to develop AI aspects of the solution. Future research includes:
- The application of ASTRA network-wide as a service that is offered by NM to local ATFCM. In this scenario, NM would identify 4DARHACs and send information about these events, together with proposed solutions, to local ATFCM;
- The reduction of the coordination required to carry out dissipation actions e.g. by automating coordination tasks between various stakeholders;
- The integration of the ASTRA HMI with existing FMP tools.
Metrics to measure complexity: Measuring complexity is nontrivial and depends on many factors. Historically, an FMP assesses sector complexity based on an occupancy metric. In contrast, several metrics are defined in ASTRA; some are defined at sector level, while others characterise a cluster of aircraft. To have a unique complexity unit among the complexity metrics, each metric is scaled such that the complexity is expressed as a percentage. The metrics are then combined into a global complexity using a statistical model.
A complexity-based clustering algorithm to identify and track 4DARHACs: At each time step, the last known position of each aircraft, and its flight plan, are used to predict its future trajectory. Then, for each time step in the predicted future (up to one hour ahead), a clustering algorithm groups flights together and identifies 4DARHACs. A tracking algorithm then identifies the clusters found in the current time step that correspond to those in the previous time step.
A deep RL algorithm to resolve 4DARHACs: An RL agent is implemented to take 4DARHACs as input and to output actions representing solutions for each 4DARHAC. Proposed actions can be lateral, vertical or longitudinal clearances. These are ranked by probability, thus aiding the interpretability of the RL agent. Apart from the dissipation action itself, the agent also determines which aircraft should take the proposed action.
An HMI design for user interaction: This HMI allows FMPs to interact with the ASTRA solution. Amongst other things, it allows the FMP to get alerts; view and simulate the proposed dissipation actions; and select actions for implementation. The HMI is designed in a way that gives the user an insight into why the RL agent proposes certain clearances, and the impact (on complexity) of implementing a certain clearance.
These results will lead to improved predictability at local ATFCM level and improved confidence in sector DCB management.
More work will be needed to develop the solution beyond TRL2. Also, better regulatory standards are required to develop AI aspects of the solution. Future research includes:
- The application of ASTRA network-wide as a service that is offered by NM to local ATFCM. In this scenario, NM would identify 4DARHACs and send information about these events, together with proposed solutions, to local ATFCM;
- The reduction of the coordination required to carry out dissipation actions e.g. by automating coordination tasks between various stakeholders;
- The integration of the ASTRA HMI with existing FMP tools.