Utvärdering av ankomstprognoser för tåg

Sammanfattning: In this master thesis, an evaluation of arrival forecasts for trains is made to see how various prediction sources compare to each other and how well these reflect reality. Furthermore, this thesis is a sprint of a research project that is carried out and conducted at Lund University of Technology on behalf of the Swedish Transport Administration. Unlike the research project, this thesis is based solely on a first preliminary data dump and not on data from a specific time period regarding a certain test track. The purpose of the thesis is to examine and evaluate arrival forecasts for trains. This is done by analyzing the ability of the underlying prediction sources to create reliable arrival forecasts. The main focus is to increase knowledge of both the strengths and weaknesses of the prediction sources, as well as to see if there are situations where one or the other seems particularly suitable. Currently, there are different ways to create arrival forecasts for trains. The most trivial is through manual calculations, but it can also be done based on automatic calculations. The automatic calculations use digital graphs and in recent years it has become increasingly common for these to be based on AI solutions, such as AIRT. The data that has been used for analysis in this master thesis is provided by the Swedish Transport Administration and is the same that they use when creating arrival forecasts. The data consists of nationwide data points regarding arrivals and departures during the time period 2022-10-01 – 2022-11-15. The analyzed data set consists of 5,530,048 lines of code and, with the help of SQL-Server Management Studio, has been used to structure and sort. In the master thesis, a test was made to see which of the evaluation methods Mean value, MAE and RMSE should be used when evaluating the arrival forecasts. The result shows that MAE is the best suited and has thus continued to be used in the analysis of subsequent problems. The Swedish Transport Administration currently uses five different prediction sources when calculating arrival forecasts for trains: Amber, Tågprognos, Manual, STEG and AIRT. When analyzing which prediction source provides the most reliable arrival predictions, it emerged that AIRT provides significantly better predictions than any of the other four. The recommendation is therefore to drop the prediction sources Amber, Tågprognos, Manual and STEG and use AIRT exclusively. This is because, on average, this prediction source creates about 2.3 times better predictions than the second-best prediction source, STEG. To answer the last problem and to follow up and see if the ranking of how well the different prediction sources performed was constant or random, a test was performed. The test was to see how the various prediction sources performed when the traffic situation on the track in the previous hour had been good, normal and bad. The results of the test showed that the ranking remained constant, with AIRT emerging as the best alternative.