Data-driven aircraft performance models for improving baseline fuel estimations

In this era, modern life in globalized World forces people to live faster compared to previous times. The need which is caused to satisfy this pace incurs some industries a huge responsibility. Aviation industry is one of these main industries. Therefore, there is an enormous cash flow to the industry. As a result of this massive influx of cash even a minor cost reduction in aviation makes a great difference in Global perspective. This expected growth of aviation industry in Globe, force researchers to design improved Air Traffic Management (ATM) technologies and procedures. In order to increase capacity of the system and its efficiency while reducing environmental impact and maintaining safety regulations. To overcome these challenges, new automation and information systems are being developed and this new global operational concept is called Trajectory Based Operations (TBO). In aviation, there are some associations such as, EUROCONTROL, IATA, ICAO, ATCA, which focus on Air Traffic Management (ATM). These associations have partnership with operating airlines and aircraft manufacturers such as, Boeing, Airbus, Cessna on projects to reduce fuel consumption. These partnership relations provide Aircraft Performance Model (APM) documents like BADA to researchers which is an input of TBOs. The BADA APM is designed to simulate and predict the trajectories of aircraft in order to ATM research and operations. The parameters of aircraft performance and trajectories can be calculated using information and data contained in BADA. In order to simulate reality for both operational and research modeling each and every real aircraft needs a corresponding aircraft model and due to experimental study BADA also provides models. Nowadays in every single moment of 24 hours of a day, there are at least 10000 commercial aircraft cruise in Worldwide. This situation causes to a huge fuel cost in aviation industry and every year there are hundreds of studies to reduce this cost in global. Via new techniques in computer science, aircraft can be modeled better compared to past and as a result of these better models we can get more precise results in our studies. Undoubtfully, Earth does not have infinite sources. Therefore, the motivation of this study focuses to determine fuel consumption of airliners more precisely by using artificial neural network techniques due to reduce the cost which is caused by unnecessary additional fuel. In trajectory side both deterministic and probabilistic methods are tried but with the improvements in computation science World turn its face to the machine learning algorithms. The motivation of this thesis is improving baseline estimations of aircraft performance models with data-driven methods. In computer science, python is mostly preferred due to its more understandable nature compared to other coding languages. It is also one of the most effective software language for complicated systems such as aircraft. To be clear, modeling and simulating an aircraft is a very complex interdisciplinary profession. Therefore, python is preferred in this study due to its advantages. In this study Keras, Pandas, Numpy, Scipy, Mxnet, Matplotlib, Bokeh, Glob and Sklearn libraries are used. Firstly, formulas and technical parameters in BADA document are coded into computer environment. Quick Access Recorder (QAR) is used to store all information for flight in aircraft. It is also known as black box data. Secondly, for this thesis QAR data for the given aircraft which are B738 AC#1 (219 flights), B738 AC#2 (563 flights). QAR data for all flights are modified due to necessity and merged according to their tail number and separated to 3 according to flight phases such as, climb, cruise and descent. After that, these 3 modified and merged csv files are converted to binary to be used in machine learning. Secondly, artificial neural network algorithms such as Adam algorithm are applied to these binary files. In order to achieve a better learning a lot of method is used in thesis. Black-box model is created in order to control accuracy of data-driven APM. By the results of Black-box model it is easily seen that Black-box model and actual QAR data matches more than %99. After that Throttle/Auto-Throttle model which is the main aim of the thesis study is created. However, standing alone APM model is not sufficient. Thirdly, to obtain more healthy comparison, these trained models integrated to belonged flight plans and fuel results are compared with each other. In this phase, several data process is applied to the flight plan data. The main reason of data process is originated to the sensor characteristics. On the other hand, it is obvious that, following flight plan completely is impossible due to several reasons such as unexpected weather conditions or high traffic density which is managed by air traffic control (ATC). To overcome these situations an implementation algorithm is used in thesis.