The 2022-2023 winter holiday travel period was extremely tough for airline customers. Airlines dominated the news following adverse weather events accompanied by large technology and process failures. These failures led to multi-day cancellations and thousands of stranded customers long after the weather had passed. The flying public and regulators demand that airlines take action to resolve their enduring technology deficiencies.

There is, however, a silver lining in this renewed airline tech focus – an opportunity to take a closer look at machine learning when implementing new airline disruption management and recovery tools. 

Background of Airline Operations Control

Very few airline passengers (and even airline employees) fully comprehend the complexities and logistics involved to get a flight from A to B.    

Once airline revenue managers and network planners publish a flight schedule and make seats available for sale, there is massive work to ensure every airline’s flight is positioned for safety, success, and an on-time departure.

To execute that schedule, every airline has an Operations Control Center (OCC or IOC) to coordinate every flight component, including maintenance control, crew scheduling, aircraft routing, flight planning, customer service, and meteorology. Leading these OCCs are airline managers tasked with directing an airline’s network response to weather, air traffic control restrictions, and emergencies. OCCs for large global airlines are massive facilities with hundreds of employees at workstations analyzing data relevant to their job function. Imagine pictures of NASA’s Mission Control but up to 10 times the size.

Current State

As one could imagine, and as evidenced by recent news of airline operations breakdowns, massive amounts of data and processing are needed for airline managers to handle the complexities of aircraft and crew flows – especially when faced with delays and cancellations that result from weather or other irregularities. Here’s just a small sample of the information airline managers utilize to make network decisions:

• Flight times – both scheduled and actual
• Crew schedules
• Flight routes and air traffic control demand data
• Individual aircraft schedules and maintenance tracking
• Passenger bookings and future demand
• Baggage and cargo information
• Weather information

Airport and Gate Information

It’s been noted that a single flight can generate over one terabyte of data (Tableau Software, n.d.). Now couple that volume with a network of 5000 flights per day, and the data available to make optimal decisions to respond to weather and other challenges is immense.

However, in the interest of expediency and based on the longevity of legacy airline technology systems, large network decisions are made with just a small fraction of the data available. 

Machine Learning Opportunities

In the early 2000s, airline technology advanced to the point that airline OCCs could employ “operations solvers” to help airline managers cancel flights ahead of large weather events and then rebuild aircraft and crew flows – very complex tasks akin to 3-dimensional chess. The impact these technologies had was profound.  

Airline flight schedules could be rebuilt in hours when a snowstorm or hurricane forced the suspension of flights.

The software, however, is not a be-all, end-all solution. As evidenced by the recent holiday season airline failures, the data inputs into these solvers must be complete and timely. If the data is stale or processed when there are already disparities, the outputs can make schedule recovery nearly impossible. As a result, many airlines are now taking a fresh look at their data and software capabilities; the opportunity lies in leapfrogging from band-aid fixes to vast data machine learning.  

Airline Operations Decisions and Data Inputs

When faced with a weather challenge, airline managers must still make many educated guesses based on their experience and post-mortems of previous events to determine the percentage of flights to cancel or delay. Take a snowstorm, for example. Airline managers need a reliable weather forecast to know when and how many flights can safely operate, but they also need to understand how best to schedule aircraft flows to maximize passenger throughput. Then, after canceling flights, they need to evaluate how best to gradually increase an airport’s traffic volume back to a normal schedule. These are their core inputs in crafting a plan to handle bad weather. 

But imagine if they could then also understand historical airport runway snow plowing times and air traffic control’s typical traffic metering pattern. Or know historical deicing time throughput rates and how each flight’s deicing delay could translate into customer misconnections or delays cascading into the next day. These inputs could paint a more comprehensive picture of forecasting delays and how they may affect passenger and crew schedules when building recovery plans.  

As evidenced by the recent airline meltdowns, crew scheduling components were a major failure point. Flights were canceled, but crews could not be reassigned to different trips as the recovery plan needed to be modified mid-event.

What are the reasons for the recovery plan’s delay?

Weather, airport conditions, and aircraft schedules could have been mitigated by operations technology that models and forecasts impact using historical data.

Airlines are extremely data-rich due to the complexity of all the logistics involved in their operation. There’s an opportunity now for airlines to make band-aid fixes and leverage that data to provide airline OCC Managers with fully comprehensive operations modeling, planning, and recovery tools.    

 

This article was contributed by our expert Jim DeYoung

Frequently Asked Questions Answered by Jim DeYoung

Q1. How is technology changing the airline industry?

As evidenced by recent airline disruptions and media reporting, many large legacy carriers are taking a renewed look at their mainframe operations systems, which are 30+ years old. These systems have served airlines well up to this point. However, the expertise needed to support them (programming sometimes in FORTRAN or COBOL languages) is no longer there. We’re seeing the industry migrate to new Enterprise Data Warehouses and then building new decision support tools to tap the opportunities that refined and rapid data access can provide.

The industry is truly at an inflection point now – do we update our operations software to take advantage of database advances? Or do we take additional time to incorporate this data with machine learning? Machine learning can revolutionize the industry and enable airline leaders to make more informed tactical decisions with a flight delay or cancellation tradeoffs. 

Q2. What are four technologies used to predict the weather?

Almost 40% of an airline’s delays are due to weather and air traffic control (with the weather having the largest impact on air traffic control efficiency). Accurate weather modeling is vital for airlines to plan properly for snowstorms, cyclones, and thunderstorms. Most larger airlines employ in-house meteorologists to analyze multiple weather models (European, Canadian, US) to establish their confidence in the forecast to make cancellation decisions.

While working at United Airlines, we would wait to act on our operations strategy until we saw 2+ models align on a specific city’s forecast. We’d also leverage experience with different weather models to determine the most accurate model source. It can certainly vary by model, city, and even season. However, one important thing to note is that with climate change, the weather models also need to evolve.

Those models that will skew results with more recent outcome data are certainly at an advantage. Lastly, we’re starting to see the development of micro-forecasting. The proliferation of personal weather stations and new aircraft self-reporting of weather conditions fill the gaps in weather analysis, giving advanced models instantaneous data to tweak forecasts. 

Q3. What are the limitations of Deep Learning in predicting Weather Disruptions?

Deep learning has the potential to improve weather forecasting drastically. Legacy weather models rely on the historical surface, and upper data gathered from reporting stations, weather balloons, and some manual manipulation based on observed outcomes. Deep learning can improve weather modeling by considering smaller surface deviations – including sea temperatures, soil moisture content, and even low altitude variations in wind direction. The only existing limitation is historical data. Some weather models are starting to be supplemented by this additional data, which, while not fully evolved, could improve short-term accuracy within the next 1-2 years, and also, as the weather grid expands with more parameters, give us more accurate long-term forecasts.

Q4. What can airlines do to give better customer service?

From an operational perspective, customers demand predictability. We’ve all had that experience where we’re at the gate, posted departure time is in 15 minutes, and there isn’t even an aircraft at the gate. Customers understand weather constraints and even mechanical delays. However, the lack of information about why their flight is delayed or why delays may keep extending frustrates them.

Airlines can be prone to “operational optimism,” always erring on the best-case outcome (i.e., the weather may clear out early, or the plane could have a rapid fix). This optimism indeed mitigates the risk of an extended delay that may be unnecessary, but there is certainly a balance. Airline operations experts must better weigh the benefits and risks of an updated departure time. Accuracy and transparency are key. Communicate the reason for the delay, plus a realistic expectation on a new departure time or rebooking in the event of cancellation. Again, machine learning can certainly help here. And we’re seeing some airlines take advantage of this new data opportunity. 

Q5. What factors affect capacity and delays, and how can AI be used in these areas?

Weather is the #1 cause of airport capacity constraints that lead to delays. It’s similar to a highway driving experience in heavy rain – traffic slows down, and it will take longer to get to your destination. When clouds, snow, rain, or storms are in an airport area, air traffic control must expand their aircraft separation tolerances which will reduce the quantity of aircraft that can operate in/out of an airport or airspace.

Airlines and air traffic control leaders have heaps of historical data that they will use to gauge throughput against hourly flight demand. This information can usually get it correct within 3-5 flights per hour. However, there are often cases where airports will “under-deliver.” We certainly want to err on the side of conservatism, but a lot of emerging technology could be leveraged to increase that throughput by 1-2 flights.

It doesn’t sound like much, but if a weather forecast can show that storms will move a little more quickly, or air traffic control can plan some creative alternate routings based on a machine learning review of similar conditions, the added 1-2 flights could reduce all delays by up to 5 minutes. That is a massive reduction in aggregate delays where on average, a single minute of delay can cost an airline over $50. Spread that among 100 flights at an airport, and the savings can truly add up.