Predicting Train Delays in Finland Using Machine Learning and Weather Data

Vinicius Pozzobon Borin; Jean Michel de Souza Sant'Ana; Nurul Huda Mahmood

Source-linked AI summary

Predicting Train Delays in Finland Using Machine Learning and Weather Data

Vinicius Pozzobon Borin, Jean Michel de Souza Sant'Ana, Nurul Huda Mahmood

arXiv:2609.11277v1Sep 10, 2026cs.AIcs.LG

TL;DR

Train delays are difficult to predict because many contributing factors challenge conventional models, especially under Finland’s extreme weather conditions. This paper evaluates XGBoost with weather-integrated railway data and finds that derived weather categories outperform raw meteorological features at Oulu central station.

Problem
Train delays involve complex contributing factors, while Finland’s railway network faces extreme weather that can cause mechanical failures, signal disruptions, and reduced adhesion.
Method
The paper evaluates XGBoost on the FI-TW dataset, combining railway records with FMI sensor observations and comparing full weather, instant observations, and derived weather-category features.
Results
Derived weather categories consistently outperform the alternative configurations, reaching R2 0.78, RMSE 8.5 minutes, and MAE 3.7 minutes at convergence.
Takeaways & Limitations
Derived weather categories provide operationally relevant, bandwidth-efficient features for transportation deployment over wireless infrastructure.
Takeaways & Limitations
Validation is restricted to Oulu central station, only XGBoost is benchmarked, nearest-station matching ignores micro-climatic variability, and reliable wireless backhaul is assumed.

Abstract

from arXiv · show

Reliable railway operations depend increasingly on real-time environmental intelligence delivered through wireless sensor infrastructures, a capability that 6G networks will substantially enhance through integrated sensing and edge computing. Adverse weather, particularly in Arctic regions with extreme temperatures and heavy precipitation, remains a leading cause of train delays, yet most prediction approaches rely on raw meteorological inputs without exploiting domain-informed feature engineering. This paper investigates machine learning for train delay prediction using the Finland Integrated Train-Weather (FI-TW) dataset, which fuses railway operational records with observations from the Finnish Meteorological Institute's nationwide sensor network of approximately 200 stations communicating over wireless links. We evaluate three feature configurations using XGBoost at Oulu central station (101,146 observations): full weather features, instant weather observations only, and derived weather category scenarios. The category-based approach, employing hierarchical classifications such as Blizzard, Heavy Snow, and Extreme Cold, achieved an R^2 of 0.78, root mean squared error of 8.5 minutes, and mean absolute error of 3.7 minutes, representing an 11% R^2 improvement and 10% error reduction over alternative configurations. These results demonstrate that compact, domain-informed features derived from sensor streams outperform raw meteorological observations, offering bandwidth-efficient representations suitable for edge deployment over current and emerging wireless infrastructures.

I. INTRODUCTION

Train delays are difficult to predict because environmental conditions, especially Arctic weather, interact with operational and infrastructure factors. The paper addresses this challenge by fusing railway records with FMI weather observations and engineering compact weather categories for machine-learning prediction.

I. INTRODUCTION: Conventional model-based approaches are insufficient for the complex combination of technical, operational, infrastructure, and environmental causes of delay.Machine learning is presented as a way to identify complex patterns from operational and environmental data.
I. INTRODUCTION: Adverse weather is especially challenging in Arctic regions because it can disrupt service reliability through extreme environmental conditions.Finland experiences temperatures as low as −40°C, with reported effects including mechanical failures, signal disruptions, and reduced adhesion.
I. INTRODUCTION: The paper uses the FI-TW dataset to combine Finnish railway operational records with environmental observations from the FMI sensor network.The network comprises approximately 200 stations communicating over wireless links.
I. INTRODUCTION: The study evaluates weather-integrated machine-learning approaches and transforms raw sensor measurements into hierarchical weather categories designed for operational relevance and bandwidth efficiency.The categories are intended for edge deployment over current and emerging wireless infrastructures.

II. DATASET DESCRIPTION

The FI-TW dataset integrates Finnish railway operations with meteorological observations collected by FMI stations. Temporal and geographic matching aligns each train observation with the nearest and closest-in-time weather record.

II. DATASET DESCRIPTION: FI-TW combines Digitraffic railway records with FMI meteorological observations spanning Finland’s railway network.The dataset contains about 38.5 million records from January 2018 through December 2024.
II. DATASET DESCRIPTION: The railway component includes train timetables, actual departure and arrival times, and station infrastructure metadata across 549 Finnish stations.Stations are divided between passenger traffic and freight, dock, or technical-service functions.
II. DATASET DESCRIPTION: Meteorological observations come from 209 FMI monitoring stations operating at 10-minute or 1-minute measurement intervals.Most stations use 10-minute intervals, while 45 provide higher-resolution 1-minute measurements.
II. DATASET DESCRIPTION: Timestamp and location matching associates each train observation with the geographically nearest and temporally closest weather record.This produces an aligned dataset linking railway events to corresponding meteorological observations.

B. Dataset Features

The dataset organizes target, operational, and weather variables, combining continuous meteorological measurements with engineered weather indicators. Its hierarchical classification gives disruptive weather conditions precedence when multiple criteria overlap.

B. Dataset Features: The final dataset contains 39 features organized into target, operational, and weather categories.The feature overview distinguishes engineered variables derived through mathematical transformations or aggregations.
B. Dataset Features: The primary delay target, differenceInMinutes, measures cumulative delay from journey start to the current station.Related target variants remove initial or intermediate delay propagation effects.
B. Dataset Features: Operational variables encode temporal patterns such as hour, month, and weekday using sine-cosine transformations.This preserves cyclical relationships across temporal boundaries.
B. Dataset Features: Weather features include 11 continuous FMI measurements and 10 additional binary indicators derived through hierarchical classification.The classification covers operationally relevant conditions such as blizzards, heavy snow, and freezing conditions.
B. Dataset Features: The hierarchy prioritizes the most operationally disruptive condition when several weather thresholds are met simultaneously.This prevents overlapping criteria from producing conflicting classifications.

III. MACHINE LEARNING TRAINING SETUP

The training setup uses XGBoost and defines three feature scenarios informed by correlation-based feature analysis. The analysis examines feature redundancy before scenario construction.

III. MACHINE LEARNING TRAINING SETUP: The study uses correlation-based feature analysis to inform feature selection and defines three XGBoost training scenarios with different feature configurations.The setup includes model selection, experimental design, and correlation analysis.

A. Model Selection: XGBoost

The study uses XGBoost to predict train delays from mixed operational and weather data, with sequential decision trees suited to nonlinear relationships and missing values.

A. Model Selection: XGBoost: XGBoost is the primary algorithm for train delay prediction.It constructs an ensemble of decision trees sequentially, with later trees correcting earlier residual errors.
A. Model Selection: XGBoost: XGBoost handles mixed feature types and missing values while capturing nonlinear interactions between weather and operational variables.Its built-in feature importance metrics also support interpretation of influential predictors.

B. Experimental Setup

Experiments predict cumulative train-event delay at Oulu central station using a chronological evaluation design and tuned XGBoost models.

B. Experimental Setup: The experiments use 101,146 long-distance-service observations from Oulu central station during 2018–2024.Oulu is identified as a high-delay node with a 19.0% delay rate, and the target is delay in minutes per train event.
B. Experimental Setup: A chronological 80/20 split holds the latest observations for final testing, while five-fold expanding-window validation selects models within the development set.This design prevents future information from leaking into training because railway and weather data are temporally autocorrelated.
B. Experimental Setup: The hyperparameter search varies estimators, tree depth, learning rate, subsampling ratios, and scale_pos_weight.The search covers 100–400 estimators and maximum depths of 4–8.
B. Experimental Setup: The base operational feature set includes train indicators and cyclical encodings for month, hour, and day of week.These features are designed to represent scheduling and temporal characteristics independently of weather.

C. Feature Analysis and Selection

Correlation analysis identifies redundant weather variables, seasonal relationships, and complementary information from weather features, informing feature selection.

C. Feature Analysis and Selection: Gust speed and wind speed are nearly redundant at r = 0.977, while dew-point and air temperature correlate at r = 0.966.These relationships motivate excluding gust speed and dew-point temperature from training scenarios.
C. Feature Analysis and Selection: Seasonal correlations link snow depth positively with month sin at r = 0.805 and negatively with air temperature at r = −0.658.Air temperature also correlates negatively with month sin and month cos at r = −0.549 and r = −0.678.
C. Feature Analysis and Selection: Pressure correlates negatively with wind and gust speed, while visibility correlates negatively with relative humidity at r = −0.246.The visibility relationship reflects reduced visibility during humid conditions such as fog or precipitation.
C. Feature Analysis and Selection: Figure 2 visualizes feature correlations using dark red for strong positive relationships and dark blue for strong negative relationships.The heatmap excludes cloud amount and derived weather scenarios from the correlation analysis.
C. Feature Analysis and Selection: Weather and operational features are mostly weakly correlated, with most operational-weather coefficients below |0.1|.This suggests weather variables provide information not captured by operational and temporal predictors.

D. Training Scenarios

The experiments compare three feature configurations that combine operational predictors with either all weather inputs, instant observations, or derived weather categories.

D. Training Scenarios: Highly correlated variables are removed before scenario construction to reduce redundancy while preserving predictive information.CommercialStop is removed for redundancy with trainStopping; gust speed and dew-point temperature are excluded because of their correlations with wind speed and air temperature.
D. Training Scenarios: Three scenarios compare full weather features, instant weather observations only, and weather category scenarios only.Each configuration combines weather inputs with operational features, while the category-only scenario uses 10 derived weather categories.

IV. TRAINING RESULTS AND DISCUSSION

The category-based weather scenario consistently outperforms full-weather and instant-observation configurations, with performance improving as training iterations increase before convergence.

0.78 R2 at 100 iterations versus roughly 0.70 for the other scenarios, indicating higher explanatory power for category-based weather features.The comparison is reported for the three feature scenarios in Figure 3.
8.5 minutes RMSE, 3.7 minutes MAE, and 50.5% WMAPE are achieved by Scenario 3 at convergence, compared with approximately 9.9 minutes, 4.1 minutes, and 56.5% for the alternatives.
Scenario 3 improves from approximately 9.9 minutes RMSE at 10 iterations to 8.5 minutes at 100 iterations, while the baseline scenarios converge near 9.9 minutes.
All scenarios improve rapidly initially; Scenarios 1 and 2 plateau around 50 iterations, whereas Scenario 3 stabilizes after approximately 80 iterations.The 50–80 iteration range is identified as a practical stopping point because further gains are marginal.

A. Comparison with Related Work

The comparison with related work is constrained by differences in delay definitions, route variability, and operating conditions, while the proposed model is evaluated on Finland’s full continuous delay distribution.

A. Comparison with Related Work: The related-work evaluation separates delays into short and long categories, whereas this study trains and evaluates on continuous delay values.
A. Comparison with Related Work: 27.5% of trains are delayed in winter versus 18% in summer, reflecting substantially higher seasonal variability in Finland’s network.
B. Generalization Beyond Oulu Station and Limitations: Validation is restricted to Oulu central station, although the authors argue that physically motivated weather thresholds can remain operationally meaningful across climates.

V. CONCLUSION

The study finds that derived weather categories outperform raw-weather configurations in XGBoost delay prediction at Oulu and offer a compact representation for deployment, while broader validation remains future work.

0.78 R2, 8.5 minutes RMSE, and 3.7 minutes MAE are achieved by derived weather categories, with approximately 14% RMSE and 10% MAE improvement over alternative configurations.
Derived weather categories consistently outperform the full feature set and instant-weather observations across all evaluation metrics.
Binary weather indicators reduce memory and storage requirements compared with continuous floating-point features, supporting processing of large observation volumes.
Future work should evaluate additional stations, compare deep-learning architectures with XGBoost, and model meteorological inputs jointly with wireless-backhaul reliability.