Estimating Cellular Network Delays in Finnish Railways: A Machine Learning Enhanced Approach

There is growing interest in using public cellular networks for specialized communication applications, replacing standalone sector-specific networks. One such application is transitioning from the aging GSM-R railway network to public 4G and 5G networks. Finland is modernizing its railway communication system through the Digirail project, leveraging public cellular networks. To evaluate network performance, a nationwide measurement campaign was conducted in two modes: Best Quality and Packet Replication. However, Best Quality mode introduces artificial delays, making it unsuitable for real-world assessments. In this paper, railway network delays are modeled using machine learning based on measurements from the Packet Replication mode. The best-performing model is then employed to generate a dataset estimating network delays across Finland’s railway network. This dataset provides a more accurate representation of network performance. Machine learning based network performance prediction is shown to be feasible, and the results indicate that Finland’s public cellular network can meet the stringent performance requirements of railway network control.

💡 Research Summary

The paper investigates whether Finland’s public 4G/LTE and emerging 5G cellular networks can satisfy the stringent latency and reliability requirements of modern railway signalling systems, specifically the European Train Control System (ETCS) and the forthcoming Future Railway Mobile Communication System (FRMCS). As part of the Digirail project, a nationwide measurement campaign was carried out along the Finnish railway network using a multi‑SIM, multi‑channel router that simultaneously accessed the three major mobile operators (Elisa, Telia, DNA). Two measurement modes were employed: Best Quality (BQ) and Packet Replication (PR). BQ samples the strongest‑signal network every five seconds, which introduces artificial delays and fails to capture rapid network fluctuations. PR, by contrast, sends identical packets over all available radio links at the same time and records the earliest arrival, thereby providing a near‑real‑time view of the network’s true performance.

The dataset comprises 1‑second samples of GNSS position, train speed, radio‑link quality indicators (RSRP, RSRQ, SNR), and end‑to‑end one‑way delays for several traffic types: TCP, HTTP, DNS, Position Report, and Movement Authority (MA) requests (the latter sampled every 10 seconds). Approximately 2 000 km of track were measured in PR mode, while the remaining ~8 000 km were only covered by BQ measurements.

Because it is impractical to repeat PR measurements for the entire network, the authors frame delay estimation as a supervised regression problem. Input features consist of the three operators’ KPIs together with train speed, yielding a structured tabular dataset. Four tree‑based ensemble algorithms—Random Forest, XGBoost, LightGBM, and CatBoost—were trained under identical hyper‑parameter settings (100 trees/iterations, fixed random seed). Model performance was evaluated using Root Mean Squared Error (RMSE), coefficient of determination (R²), Mean Absolute Error (MAE), and, for safety‑critical assessment, precision and recall on “critical” events (delays > 500 ms for most services, > 1 000 ms for HTTP).

Results show that Random Forest delivers the best overall performance for Position Report, TCP, HTTP, and DNS, achieving RMSE values as low as 8.18 ms (Position Report) and R² up to 0.84. XGBoost excels on MA requests with an R² of 0.97, albeit with a slightly higher MAE. Precision/recall balances indicate that the models reliably capture critical delay spikes, often detecting them a few seconds earlier and extending the flagged interval, which is acceptable for worst‑case statistical analysis. Feature importance analysis highlights train speed as the dominant predictor, reflecting the impact of handovers and varying radio conditions at different velocities.

The best‑performing model for each delay type was then used to generate synthetic PR‑mode delay data for the BQ‑only segments. Statistical comparison (minimum, quartiles, median, mean, maximum) demonstrates strong alignment between the synthetic and the actual PR measurements for central tendencies, while the synthetic data exhibit tighter clustering at the extremes. This suggests the models are conservative in estimating extreme delays, a desirable property for safety‑critical planning.

Reliability analysis confirms that BQ mode dramatically over‑states critical events due to its artificial delay component, whereas measured PR mode and the ML‑generated PR data contain virtually no critical events (< 0.01 %). Consequently, the public LTE/5G infrastructure in Finland meets the latency and reliability thresholds defined for ETCS and appears capable of supporting FRMCS with minimal additional investment.

In summary, the study demonstrates that (1) PR‑mode measurements provide an accurate baseline for railway communication performance, (2) structured ensemble machine‑learning models can predict per‑service delays with high fidelity using readily available radio‑link metrics and train dynamics, and (3) such models enable the extrapolation of high‑quality delay estimates across an entire national rail network, reducing the need for exhaustive field campaigns. The work validates the feasibility of leveraging commercial cellular networks for railway control traffic, while noting that further validation on extreme delay scenarios and cross‑border deployments remains a future research direction.