In this work, we assess the performance of data-driven approaches on jet fuel demand data related to the Danish market, gathered from DCC & Shell Aviation Denmark A/S. In particular, we compare the predictive performance obtained with traditional time series based models (SARIMAX), LSTM sequence-to-sequence neural networks, and hybrid models. While developing such models, the main challenges lie in the needed forecasting horizon and in the novelty of the approach. Indeed, to support the company's sourcing strategy, the fuel demand needs to be predicted for the upcoming month. To assess the reliability of the data-driven approaches, three different case studies are proposed and analyzed: Ryanair, Fly Emirates and the whole Copenhagen Airport.
The project demonstrated a highly commendable performance of the automated predictive models, along with a significant enhancement in performance (as shown in the image above) by employing a data-driven approach as opposed to the company's performance. Furthermore, the results of our case studies demonstrate the potential of novel, data-driven approaches to bolster forecasting accuracy and inform decision-making in industry.
Please refer to my short paper here for further explantions about the approach and the results.
As a result of concerns regarding security, privacy, and sensitive business-related data owned by DCC & Shell Aviation A/S, all values are presented as percentages or scaled to random factors. Furthermore, no source code is publicly accessible. For further inquiries or additional information, please contact me via email at alessandrocontini96@gmail.com.
The present project was employed as my thesis for the Master's Degree program in Human-Centered Artificial Intelligence at the Technical University of Denmark. Submitted on February 5th, 2023, and merited a grade of 12 out of 12.