Innodata's experts involved in implementing the project analyzed an array of historic flight data and a huge amount of unstructured aircraft loading data. Then metrics for assessing the efficiency of forecasts were defined. The results of the analytics enabled experts to develop the test technique and assessment criteria including the quality of forecasting. The research was done using machine learning technologies based on Big Data software (Hive, Spark ML).
‘To forecast cargo-mail capacities for passenger flights, it is often necessary to use information about the flights without any historical data for machine learning’, says Alexander Sergienko, Executive Director at Innodata. ‘Yet the project team designed the core of the system to enable it process both previously unrecorded information and take data on flight loading, the number of passengers, and the fuel level into account. Moreover, the system responses to changes in the timetable, replacement of aircraft types, and other important factors.’
Apart from dealing with vast amounts of unstructured data, the implementation of the project was complicated by deciphering special CPM and LDM telegrams that are created when an aircraft is getting ready to depart. They contain data about the aircraft loading throughout the flight and about loading of the cargo holds. As these telegrams are prepared manually, they often contain errors and can exist in several contradicting versions.
The developers consolidated historical data, built self-correcting prediction models, created functionality to adjust and optimize the system work algorithms, thus providing Aeroflot with an accurate tool for predicting the cargo-mail quota.
‘As a result of implementing the automated system for forecasting cargo-mail capacities of passenger flights for Aeroflot, the accuracy of predicted values for passenger craft loading increased by 20% in 6 months, while the accuracy of predicting the available free cargo quota reached up to 90%’, says Kirill Bogdanov, Aeroflot’s Deputy Director General for IT. ‘It helped us to significantly optimize the payload of flights within the entire Group.’