Railway Delays? It’s a Network Problem – Let Graphs Untangle It

Dispatchers, timetable planners, and performance managers are responsible for keeping railway traffic running by making the right decisions and continuously improving railway operations. However, due to the complexity of railway operations, analysing traffic data and deriving actionable insights is difficult. Missing actionable insights prevents making the right decisions and improving railway operations.

In general, delays in railway traffic operations are predictable because railway traffic follows strict constraints, such as trains running in sequence on dedicated tracks. Modern technology, such as graph models, enable easier traffic data analysis and allows for more accurate prediction of future train movements and delays. This provides better insights and supports traffic operation stakeholders, such as dispatchers or performance managers, in making better decisions.

Each train generates data events upon arrival and departure, which, across an entire country, accumulate into vast amounts of data. However, the aforementioned constraints between trains or between trains and infrastructure, are not easily identifiable from the raw data. For example it is not easy to identify from the raw data if a train needed to wait before entering a station, because another train was blocking the platform or crossing the track.

Understanding the Ripple Effect

When a train is delayed, it can trigger a chain reaction by disrupting the schedules of other trains. This ripple effect can quickly escalate, as each subsequent delay causes further conflicts, leading to a snowballing impact across the entire network. This leads to various problems that need to be solved.

Building a system that is capable of solving the traffic-specific problems above, requires a new foundation of data handling which we identified as follows:

Scalable platform to run AI-powered solutions:

• a platform that can deal with network-wide data to solve the problem at scale
• that provides contextual information and understanding of the data instead of viewing it in an isolated way
• that allows deploying, operating, monitoring, and iterating of ML/AI models efficiently

Many traffic management systems do not account for upcoming train conflicts for predicting future delays. There is also no decision support system for train prioritisation. Bringing operations back to the planned timetable is based solely on the dispatcher’s knowledge and experience.

The effort invested into traffic performance management depends on the regulations in a country. For example, in the UK, railway companies have to pay a fee to other railway companies if they are responsible for delay. This requires identifying the responsibility of delays back to the root cause (e.g. a defective train, or a defective switch). Attributing delay is done with high manual effort by hundreds of people (about 400 people according to page 10 here).

What the System Needs

Based on these challenges, we derived the following technical requirements to model railway traffic:

• represent railway traffic with different levels of complexity such as
• make use of railway network map information
• allow to identify conflicts between trains
• allow to identify connected conflicts as conflict chains

Train movements can be modelled by consecutive arrival and departure events. Interpreting these events as nodes allows to model the traffic behaviour as a graph. Arrival and departure events are nodes, connections and constraints between these events are edges. Our approach relies on the Ph.D. thesis “Models for Predictive Railway Traffic Management” by Pavle Kecman from 2014.

Building the Graph Model

At first, we interpret stations and tracks as nodes. Station nodes and track nodes are connected by edges which represent logical connections. Let’s now imagine that trains “move” over this network graph, as indicated in the first row of the attached picture table. The moving trains emit events when arriving at a station, or starting on a track. We interpret these events as event nodes in the graph. Consecutive events of a moving train are connected by edges as well (see figure in row 2). Events of network-node (station or track) are connected by constraint edges. Edges between events contain further information about the train’s movement such as run-time on a track.

Excluding the network-nodes leaves the logical representation of connected train-movement-events. Dependencies between two events are encoded in the edges (see figures in row 3).

Propagation of traffic updates through the graph

We can calculate timestamps of upcoming events based on the timestamps of preceding events and edge-process-times. Updated information of a train movement can thereby be propagated through the graph towards upcoming events. E.g., if a train picks up additional delay at a station, this updated information is propagated to all following events.

Conflicts, Conflict Chains and Dispatcher Decisions

Conflicts between trains can be identified based on the constraint edges in the graph. Connected conflicts make up a conflict chain, which is further analysed and investigated. Incorporating dispatcher decisions is possible by representing them as additional edges.

Visualisation of insights

At KONUX, User Experience drives every product we create, transforming complex data into digestible, actionable insights.

Visualising multiple train movements is challenging because trains move in time, move in space, and there are dependencies between train movements. Furthermore, different user personas have different requirements for a visualisation. Dispatchers are used to time-distance views or schematic representations of the network, while other user personas may prefer map-based visualisations. Conflict chains are logical connections of train movements and some users may want to see these connections and chains.

The image on the left shows train movements on a map as well as in a time-distance view. Dependencies between trains can be seen particularly well in the time-distance-view, while the map perspective adds a natural impression of where delays happen.

The image on the right shows train movements on a schematic representation of a train station on the left side. Moving trains may block each other, and the corresponding conflict chain is shown on the right part of the screen. Adding shading allows us to additionally visualize patterns of recurring conflicts or conflict chains.

Real-World Impact

We demonstrated the feasibility of the approach in a project with a large railway operator. The graph-based approach specifically allowed starting from a simple traffic model that contained various simplifications, to iteratively increase complexity and improve the accuracy of the predicted event timestamps and identified conflicts.

We see future potential in the automatic generation of conflict chains as this saves costs for manually producing them, and allows for a systematic analysis of problematic traffic areas. Executing What-If analyses such as “what if I had prioritised train B over train A” is also possible and allows finding solutions to the identified problematic traffic areas. Following the steps of identifying problems and identifying solutions for these problems leads to an online decision support tool for dispatchers in the long run.

Looking Ahead

Data handling for the traffic initiative showed that there is high potential in deriving insights from treating data in a purpose-centered way. This raised the necessity of a data platform that is capable of handling data efficiently and connecting data from various sources and of various types. Such a data platform allows to derive insights with a new level of quality and allows for further improvements such as running models as a service.

This piece has showed how to master railway traffic using graph models, and how these modern technologies enable a great improvement of traffic operations for various user personas. The main challenges for mastering railway traffic simulations and analyses are on improving accuracy and realism of traffic predictions, while keeping the computational complexity at a meaningful level for the intended use-case.

In order to do so, a scalable platform is required that can deal with network-wide data and run AI-powered solutions to solve the problem at scale. This platform could be the next big step in transforming railway operations for a sustainable future.