PoC: Forecasting PCE based on historical data: AI model comparison for Färjerederiet
2024 (English)Independent thesis Advanced level (degree of Master (Two Years)), 20 credits / 30 HE credits
Student thesis
Abstract [en]
Färjerederiet provides ferry operations on several locations around Swe- den. The decisions regarding vessel assignments to specific routes typically rely on subjective assessments and past experiences. Färjerederiet has col- lected data for several years from their ferry operations and seeks to explore the potential of AI/machine learning to enhance decision-making processes. Particularly in determining the optimal ferry size based on Passenger Car Equivalent (PCE) score per route. The aim of this study is to replace the past way of assigning vessels to a route with data-driven insights that stems from historical data and in the end optimizing ferry utilization. Färjerederiets technology provider Rely IT has for several years supported them with var- ious apps and solutions. This study holds significance for both Rely IT and Färjerederiet as it introduces them to AI/machine learning applications and generates value for their business. This study will result in a Proof of Con- cept (PoC) that will not only yield operational efficiencies but also contribute to environmental sustainability in ferry operations. This research will show- case the practical integration of AI in a marine domain, potentially serving as a catalyst for further research pursuits. By answering three key research questions, this research aspiration was to assess the accuracy of several ma- chine learning models in forecasting PCE scores for multiple ferry routes. A pre-study was conducted to narrow down the amount of models to three, XG- Boost, SARIMA and Prophet. The study also assessed the feature importance of the best performing model and explored potential variables for future PCE forecasting efforts, thereby enriching the existing knowledge base within the domain. The findings of this controlled comparison experiment showed that XGBoost outperformed the others with average R-squared score of 0.98. The most important variables/features in the dataset for predicting were "Car" and "Truck" converted to PCE score for the XGBoost model. Besides this, the importance of the size of data set, the season of the year and type of route holds importance. The XGBoost model showed great performance due to low dataset and could handle outliers well but take time and computational power to optimize.
Place, publisher, year, edition, pages
2024. , p. 32
Keywords [en]
PCE forecasting, XGBoost, Facebook Prophet, SARIMA, Machine learning, Färjerederiet, Rely IT
National Category
Other Electrical Engineering, Electronic Engineering, Information Engineering
Identifiers
URN: urn:nbn:se:lnu:diva-129589OAI: oai:DiVA.org:lnu-129589DiVA, id: diva2:1882265
External cooperation
Rely IT i Norrtelje AB; Trafikverket Färjerederiet
Educational program
Social Media and Web Technologies, Master Programme, 120 credits
Supervisors
Examiners
2024-07-052024-07-042024-07-05Bibliographically approved