The traffic on Swedish railways is increasing and punctuality is of important matter for both passenger and freight trains. The problem of modeling train delay evolution is complex since conflicts between trains can occur and since a delay can have a wide variety of causes. Swedish railways faces in addition harsh winter climate. Studies of railways in Scandinavia have shown that harsh winter climate decreases the punctuality of trains. This thesis work investigates the possibilities of modeling train delay evolution as continuous time Markov processes and which specific modeling choices are preferable. It also further assesses the impact of a harsh winter climate on the delay evolution. The studied segments are Stockholm - Umeå and Luleå - Kiruna. Both over one winter season. It was found that a change in the time schedule, which in a way redefines the delay, allows for a better fit and better prediction capabilities. It reduced the MSE of the prediction by 50 %. As for the weather variables, four variables were included together with their week long moving averages. Low temperatures were found to increase the risk of a delay (Hazard ratio of 1.10) as well as to decrease the chance of recovering from a delay (Hazard ratio of 0.91). No other significant weather impacts were found.

Die Methoden der Netzwerktheorie erfreuen sich wachsender Beliebtheit, da sie die Darstellung von komplexen Systemen durch Netzwerke erlauben. Diese werden nur mit einer Menge von Knoten erfasst, die durch Kanten verbunden werden. Derzeit verfügbare Methoden beschränken sich hauptsächlich auf die deskriptive Analyse der Netzwerkstruktur. In der hier vorliegenden Arbeit werden verschiedene Ansätze für die Inferenz über Prozessen in komplexen Netzwerken vorgestellt. Diese Prozesse beeinflussen messbare Größen in Netzwerkknoten und werden durch eine Menge von Zufallszahlen beschrieben. Alle vorgestellten Methoden sind durch praktische Anwendungen motiviert, wie die Übertragung von Lebensmittelinfektionen, die Verbreitung von Zugverspätungen, oder auch die Regulierung von genetischen Effekten. Zunächst wird ein allgemeines dynamisches Metapopulationsmodell für die Verbreitung von Lebensmittelinfektionen vorgestellt, welches die lokalen Infektionsdynamiken mit den netzwerkbasierten Transportwegen von kontaminierten Lebensmitteln zusammenführt. Dieses Modell ermöglicht die effiziente Simulationen verschiedener realistischer Lebensmittelinfektionsepidemien. Zweitens wird ein explorativer Ansatz zur Ursprungsbestimmung von Verbreitungsprozessen entwickelt. Auf Grundlage einer netzwerkbasierten Redefinition der geodätischen Distanz können komplexe Verbreitungsmuster in ein systematisches, kreisrundes Ausbreitungsschema projiziert werden. Dies gilt genau dann, wenn der Ursprungsnetzwerkknoten als Bezugspunkt gewählt wird. Die Methode wird erfolgreich auf den EHEC/HUS Epidemie 2011 in Deutschland angewandt. Die Ergebnisse legen nahe, dass die Methode die aufwändigen Standarduntersuchungen bei Lebensmittelinfektionsepidemien sinnvoll ergänzen kann. Zudem kann dieser explorative Ansatz zur Identifikation von Ursprungsverspätungen in Transportnetzwerken angewandt werden. Die Ergebnisse von umfangreichen Simulationsstudien mit verschiedenstensten Übertragungsmechanismen lassen auf eine allgemeine Anwendbarkeit des Ansatzes bei der Ursprungsbestimmung von Verbreitungsprozessen in vielfältigen Bereichen hoffen. Schließlich wird gezeigt, dass kernelbasierte Methoden eine Alternative für die statistische Analyse von Prozessen in Netzwerken darstellen können. Es wurde ein netzwerkbasierter Kern für den logistischen Kernel Machine Test entwickelt, welcher die nahtlose Integration von biologischem Wissen in die Analyse von Daten aus genomweiten Assoziationsstudien erlaubt. Die Methode wird erfolgreich bei der Analyse genetischer Ursachen für rheumatische Arthritis und Lungenkrebs getestet. Zusammenfassend machen die Ergebnisse der vorgestellten Methoden deutlich, dass die Netzwerk-theoretische Analyse von Verbreitungsprozessen einen wesentlichen Beitrag zur Beantwortung verschiedenster Fragestellungen in unterschiedlichen Anwendungen liefern kann.

Disturbances in the railway network are frequent and to some extent, inevitable. When this happens, the traffic dispatchers need to re-schedule the train traffic and there is a need for decision support in this process. One purpose of such a decision support system would be to visualize the relevant, alternative re-scheduling solutions and benchmark them based on a set of relevant train traffic attributes which quantify the effects of each solution. Currently, there are two research projects financed by the Swedish Transport Administration (i.e. Trafikverket) which focus on developing decision support to assist the Swedish train traffic managers: The STEG project and the EOT project. Within the STEG project, researchers at Uppsala University in co-operation with Trafikverket are developing a graphical user interface (referred to as the STEG graph). Within the EOT project, researchers at Blekinge Institute of Technology (BTH) are developing fast re-scheduling algorithms to propose to the Swedish train traffic dispatchers a set of relevant re-scheduling alternatives when disturbances occur. However, neither the STEG graph nor the EOT algorithms are at this point designed to evaluate, benchmark and visualize the alternative re-scheduling solutions. The main objective of this work is therefore to identify and analyze different train traffic attributes and how to use the selected relevant ones for benchmarking re-scheduling solutions. This involves enhancing an existing visual tool (EOT GUI) and using this extended version (referred to as the EOT GUI+) to demonstrate and evaluate the benchmarking of different re-scheduling solutions based on the selected train traffic attributes. The train traffic attributes found in the literature (foremost research publications and documents by Trafikverket) were collected and analyzed. A subset of the most commonly used attributes found were then selected and their applicability in benchmarking re-scheduling solutions for the Swedish train traffic system was further analyzed. The formulas for calculating each of the attribute values were either found in the literature and possibly modified, or defined within this thesis project. In order to assess the use of the attributes for benchmark solutions, experiments were conducted using the enhanced visual tool EOT GUI+ and a set of sample solutions for three different disturbance scenarios provided by the EOT project. The tool only performs a benchmark of two solutions at a time (i.e. a pair wise benchmark) and computes the attribute values for the chosen attributes. The literature review and attribute analysis resulted in a first set of ten different attributes to use including e.g. total final delay (with a delay threshold value of 1 and 5 minutes respectively), maximum delay, total accumulated delay, total delay cost, number of delayed trains and robustness. The formulas to compute these attribute values were implemented and applied to the sample solutions in the experiments. The first phase of the experiments showed that in one of the disturbance scenarios, some of the attribute values were in conflict and that none of re-scheduling solution was dominating the others. This observation led to that the set of attributes needed to be narrowed down and internally prioritized. Based on the experimental results and the analysis of what the research community and the main stakeholder (i.e. Trafikverket) consider are the most important attributes in this context, the final set of attributes to use includes average final delay, maximum delay of a single train, total number of delayed trains and robustness. The contribution of this thesis is primarily the review and analysis of what attributes to use when performing a benchmark of re-scheduling solutions in real-time train traffic disturbance management. Furthermore, this thesis also contributes by performing an experimental assessment of how the attributes and their formulas could work in a pair-wise, quantitative benchmark for a set of disturbance scenarios and which issues that may occur due to conflicting objectives and attribute values. Concerning the enhancement of the visual tool and the visualization of the re-scheduling solutions, the experimental evaluation and analysis shows that the tool would not fit directly to the needs of the train dispatchers. This work should therefore only be seen as a starting point for the researchers whom are working with the development of decision support systems in this context. Furthermore, several iterative experiments have been conducted to select the appropriate attributes for benchmarking solutions and suggesting the best re-scheduling solution. During the experiments, we have used a limited set of different problem instances (2+2+7) representing three different types of disturbances. The performance of the enhanced visual tool EOT GUI+ and its functionalities should ideally also be analyzed further and improved by experimenting with a larger number of instances, for other parts of the Swedish railway network and in co-operation with the real users, i.e. the dispatchers.

Train delays occur on a daily basis in the commuter rail of Stockholm. This means that the travellers might become delayed themselves for their particular destination. To find the most accurate method for predicting train delays, the machine learning methods decision tree with and without AdaBoost and neural network were compared with different settings. Neural network achieved the best result when used with 3 layers and 22 neurons in each layer. Its delay predictions had an average error of 122 seconds, compared to the actual delay. It might therefore be the best method for predicting train delays. However the study was very limited in time and more train departure data would need to be collected. / Tågförseningar inträffar dagligen i Stockholms pendeltågstrafik. Det orsakar att resenärerna själva kan bli försenade till deras destinationer. För att hitta den mest träffsäkra metoden för att förutspå tågförseningar jämfördes maskininlärningsmetoderna beslutsträd, med och utan AdaBoost, och artificiella neuronnät med olika inställningar. Det artificiella neuronnätet gav det bästa resultatet när det användes med 3 lager och 22 neuroner i varje lager. Dess förseningsförutsägelse hade ett genomsnittligt fel på 122 sekunder jämfört med den verkliga förseningen. Det kan därför vara den bästa metoden för att förutspå tågförseningar. Den här studien hade dock väldigt begränsat med tid och mer information om tågavgångar hade behövts samlas in.