RuAviation

Vyacheslav Pachkov, a student of MISIS University, has created a smartphone application that predicts flight delays as his final qualification work. The results show that the model trained on big data made forecasts that differed from the true values of delays by 12 minutes, which indicates high accuracy and broad prospects for the development of this topic, reported the press service of NITU MISIS.

Flight delays are a serious problem for both airlines and passengers, as they lead to significant losses of time and money. The main factors that lead to delays in aircraft departures include weather conditions, airport congestion, type and age of aircraft, as well as maintenance problems.

Today’s neural networks can help people get a head start on flight delays or cancellations and plan their next steps. Various machine learning algorithms such as the multilayer perceptron model, Bayesian modelling, decision tree, cluster classification or random forest are suitable for predicting delays. These can be used to assess both the likelihood and severity of flight delays, which can prove invaluable for airlines to develop more efficient flight scheduling and maintenance strategies.

“Machine learning methods are doing a great job in the airline industry. Identifying the most important and informative attributes is a key step in developing an effective flight delay prediction model. It seems to me that if airlines or airports invest in this area to improve the accuracy and reliability of predictions – it will demonstrate care for customers, increase the positive reputation of a responsible carrier and, of course, allow choosing the most effective risk management strategy”, – Vyacheslav Pachkov commented on his development.

The development is based on an artificial neural network capable of processing complex input data and performing non-linear classification or regression, it is called Multilayer Perceptron (MLP) model. MLP can model more complex features and dependencies between input and output data. The input layer accepts the feature vector, hidden layers process the data, and the output layer generates predictions. Neurons between layers are connected by weights that determine the degree to which each neuron influences other neurons. The learning process continues until a certain stopping criterion is reached, such as the minimum value of the loss function or stabilisation of the error on a validation dataset.

The model uses nine input features:

• the time between arrival and departure from the departure airport;
• expected time of arrival at the destination airport;
• flight range;
• departure airport;
• arrival airport;
• type of aircraft;
• temperature;
• probability of precipitation;
• time of year.

In total, about one million records representing flight information for the last year were collected from FlightAware and FlightStats flight tracking service APIs from Russia, Canada, UK, France, Germany, Australia, Japan and USA, providing significant data volume and geographic diversity. WeatherAPI, OpenWeatherMap, and Weather Underground served as sources for meteorological data collection.

This data was used as the basis for training and subsequent testing of the machine learning model. After measuring performance on the test dataset, the converted model was integrated into an iOS app to demonstrate performance on real data. The app uses the developed model to perform predictions on the mobile device.

During further training at NITU MISIS Vyacheslav Pachkov plans to refine the application by increasing the accuracy of the model and optimising the layers of the neural network to speed up the work on weak mobile devices.