- The paper presents the first publicly available dataset integrating Finnish railway operational records with synchronized meteorological observations from 2018 to 2024, covering 38.5 million observations.
- Exploratory analysis shows pronounced seasonal and weekly delay patterns, with winter months and Fridays having the highest delay rates, and highlights major multicollinearity among weather features.
Motivation and contribution
Train delays arise from the interaction of operational, technical, and environmental factors, yet publicly available railway datasets rarely integrate meteorological observations with operational records. A systematic review of 62 AI-oriented railway datasets found that although 47 contain operational data, the overwhelming majority omit weather information (2601.16592). Prior exceptions, such as RailSem19 and the Indian Metro Dataset, incorporate only limited weather parameters. This paper addresses that gap by releasing the first publicly available dataset combining Finnish railway operations with synchronized meteorological observations from January 2018 to December 2024.
The dataset merges two open sources: the Digitraffic Railway Traffic Service operated by Fintraffic (train timetables, actual arrival/departure times, pre-computed differenceInMinutes delay values, cancellation flags, and station metadata) and the Finnish Meteorological Institute (FMI) observation network of 209 environmental monitoring stations. The result spans approximately 38.5 million observations across Finland's 5,915-kilometer network, with 28 engineered features covering operational variables (schedules, train types, temporal encodings) and eleven meteorological measurements (air temperature, wind speed and gusts, humidity, dew point, precipitation intensity, snow depth, pressure, visibility, cloud amount). Finland is a particularly informative setting: roughly 5,200 km of the network is single-track, winter temperatures reach −40°C, and long-distance punctuality in 2024 was 86.28% under the national five-minute threshold at final destination.
Dataset construction and preprocessing
Integration proceeds in two stages. Spatial matching associates each train station with its nearest environmental monitoring station via Haversine great-circle distance over GPS coordinates. Temporal alignment performs a left join on the train records, retrieving weather features from the assigned station at each timestamp, with nearest-neighbor matching within a tolerance window when exact correspondence is absent.
Because sensor deployment across FMI stations is heterogeneous—precipitation amount is missing in 86.91% of merged records, cloud amount in 33.85%, while air temperature, humidity, and dew point are missing in only about 8%—the authors implement a spatial fallback algorithm that searches radially within 50 km for an alternative station measuring the missing parameter before designating a value as missing. Missing-data handling then follows a hierarchical strategy: list-wise deletion when timestamps or targets are absent; deletion of weather columns exceeding 70% missingness; constant imputation of boolean stop flags; and month-specific median imputation for surviving weather features, chosen for robustness against sensor outliers. Temporal features are encoded cyclically using sine-cosine pairs for hour (P=24), month (P=12), and day-of-week (P=7) to resolve time wraparound; raw columns are retained alongside the encodings since tree-based models may exploit them directly. Duplicates, which constituted 23.24% of records, were removed entirely. Scaling uses scikit-learn's RobustScaler computed exclusively on the training partition after an 80/20 split, preventing data leakage.
A notable methodological choice is the definition of multiple target formulations beyond the raw cumulative differenceInMinutes: a first-station-offset variant and a per-station offset (differenceInMinutes_eachStation_offset) that subtracts propagated delay from previous stops, isolating the incremental delay introduced on each segment. At Oulu station over seven years, the cumulative formulation yields 72.6% of observations showing some delay and 27.1% exceeding five minutes, whereas the per-station formulation yields only 51% and 14.2% respectively. This contrast indicates that most observed delay results from propagation rather than station-specific causes—an implication directly relevant to how researchers should frame prediction tasks, since target choice materially changes class imbalance and modeling difficulty.
Exploratory findings
The exploratory analysis reveals pronounced seasonal structure. Winter months (December–February) exhibit delay rates exceeding 25%, compared with below 20% in late spring and early autumn; June is anomalous among summer months at 21.8%, plausibly reflecting increased summer traffic volume. Weekly patterns show Fridays worst (up to 31.6% average delay in February) and Saturdays best (12.2–25.5%). Among days classified by delay intensity, medium delays of 10–15 minutes dominate at 49% of occurrences, with very high delays (>20 minutes) at 15.84%. High-delay corridors cluster geographically in central and northern Finland. One caveat stated plainly in the paper: the dataset does not attribute causes, so these seasonal and geographic patterns cannot be decomposed into weather versus passenger-induced or technical contributions without additional analysis.
Correlation analysis identifies substantial multicollinearity among weather features—gust speed and wind speed correlate at 0.946, dew-point temperature with air temperature at 0.898—which users must address through feature selection depending on the model family employed.
Baseline validation
To demonstrate utility, the authors train XGBoost regressors on long-distance trains passing through Oulu station (101,146 observations), predicting the three delay targets with 19 features spanning cyclical temporal encodings, train identity, and ten weather variables. Hyperparameters were tuned via randomized search with 5-fold cross-validation. The per-station offset target achieves a test MAE of 2.73 minutes, substantially better than cumulative delay prediction (differenceInMinutes, 4.21 minutes) and first-station offset (4.81 minutes). This gap confirms quantitatively that station-local delays are more predictable than network-propagated accumulation, validating the value of the offset target formulation introduced in the dataset.
Cross-system comparison is acknowledged as difficult: FCLL-Net on Chinese high-speed rail achieved 1.87 minutes MAE and Random Forest/XGBoost on Dutch conventional rail achieved approximately 1.71–1.72 minutes, but both systems operate predominantly double-track infrastructure under milder climates, whereas Finland combines extreme seasonal variation with mostly single-track operation. The Finnish baseline is therefore not directly comparable, and the reported MAE should be read as a reference point rather than evidence of model superiority.
Limitations and open questions
The paper concedes several constraints. Weather attribution is absent from the underlying operational data, so causal claims about meteorological impact require dedicated methods not exercised here. Sensor coverage is uneven—precipitation amount is unavailable at nearly 87% of merged records even after spatial fallback—and the 50 km fallback radius trades spatial representativeness for completeness, an assumption justified only by appeal to regional-scale weather continuity. Duplicate removal eliminated nearly a quarter of raw records, whose provenance (e.g., API versioning artifacts) is not analyzed. The baseline experiment covers a single station (Oulu) and one algorithm; no sequence-aware or network-wide models are evaluated, leaving open whether LSTM-, Transformer-, or GNN-based architectures can exploit the journey's temporal structure and inter-station dependencies to improve on the 2.73-minute MAE. Extending the static dataset with streaming pipelines and forecast (rather than observed) weather for real-time prediction also remains unaddressed.
Conclusion
This work delivers a large-scale, openly licensed integration of Finnish railway operations and meteorology—38.5 million observations, 28 features, seven years of coverage—with careful attention to leakage prevention, seasonal imputation, and multiple delay-target formulations. Its exploratory statistics quantify strong winter and Friday delay effects, and its decomposition of cumulative versus station-specific delay provides a concrete analytical insight: propagation dominates observed lateness. The XGBoost baseline establishes a reproducible reference point for subsequent machine learning research on Nordic railway reliability, weather-impact assessment, and infrastructure vulnerability mapping.