FI-TW: An Open Train-Weather Dataset for Railway Delay Analysis in Finland

AI summary

Overview

Research area: Railway operations analytics and applied machine learning — specifically, dataset construction that fuses railway operational records with synchronized meteorological observations for delay analysis in a Nordic climate.

Technical level: Intermediate. The paper is a dataset/benchmark contribution rather than a new algorithmic method. It assumes familiarity with data preprocessing concepts such as cyclical encoding, missing-value imputation, rolling-window features, and a baseline gradient-boosting regression model (XGBoost), but explains each step in procedural detail.

Scope (one sentence): The authors build and publicly release FI-TW, a seven-year (January 2018–December 2024) machine-learning-ready dataset that joins Finnish railway traffic records with observations from 209 Finnish Meteorological Institute weather stations, and they characterize it through exploratory analysis and a baseline delay-prediction experiment.

What This Paper Is About

Train delays arise from a mix of technical, operational and environmental causes, and weather is a major driver in Nordic conditions. Existing publicly available railway datasets mostly omit weather entirely or include only a handful of meteorological variables, so researchers who want to study weather-driven delay lack a reusable, weather-integrated resource. The goal of this paper is to close that gap for Finland: the authors combine two Finnish open data sources into a single dataset that aligns each train event with the weather measured closest to it in both space and time, then document its structure, quality limitations and baseline usefulness for delay prediction.

Key Contributions

1. A first-of-its-kind integrated dataset for Finland. The authors state that FI-TW is, to the best of their knowledge, the first publicly available dataset that couples operational data with meteorological data for the Finnish railway network.

2. A documented, step-by-step processing pipeline. The paper lays out the full chain from data acquisition through spatial–temporal merging, missing-data mitigation, column pruning, imputation, cyclical temporal encoding and derived-feature engineering, so the dataset can be reproduced and reused for machine learning.

3. Broad spatial–temporal coverage across diverse climatic conditions. The data span January 2018 to December 2024, cover a 5,915-kilometer rail network, involve 549 Finnish railway stations and 209 weather stations, and total approximately 38.5 million observations.

4. Rich feature engineering and multiple target formulations. The released dataset contains 138 features spanning operational variables and meteorological measurements, including 13 underlying weather parameters, engineered weather-scenario indicators and multi-scale rolling-window aggregations, and it supports multiple target variable formulations for tasks beyond delay prediction.

Main Findings

• Seasonal delay structure is pronounced. Winter delay rates exceed 25%, compared with below 20% in summer. The cold months (January, February and December) show substantially higher delays across all weekdays, often exceeding 25%, while April, May, August and September frequently fall below 20%.

• June is an outlier within summer. June shows a relatively higher delay percentage of 21.8% compared to other summer months, which the authors suggest may relate to the increased number of trains in circulation and higher holiday passenger volumes.

• Medium delays dominate the delay-intensity distribution. Across the 2018–2024 period, medium delays (10–15 minutes) are the most prevalent with 1254 occurrences (49%), followed by high delays (15–20 minutes) with 597 cases (23.35%) and very high delays (more than 20 minutes) with 405 occurrences (15.84%). Low delays (5–10 minutes) are the smallest category with 300 entries (11.73%).

• Weekly patterns matter. Weekdays consistently exhibit higher delays than weekends. Fridays show particularly high delays throughout the year, reaching up to 31.6% in February, while Saturday has the lowest delay percentages, ranging between 12.2% and 25.5%.

• High-delay corridors cluster geographically. Exploratory analysis shows geographic clustering of high-delay corridors in central and northern Finland.

• The weather station network is unevenly equipped. Air temperature, relative humidity and dew-point temperature are measured at 206 stations (98.60% coverage), wind parameters at 166 stations (79.40%), and pressure at 153 stations (73.20%). Hydrometeorological parameters are far sparser: precipitation amount and intensity at 119 stations (56.90%), snow depth at 112 (53.60%), cloud amount at 111 (53.10%) and horizontal visibility at 110 (52.60%). Three EMS have no measurements at all (01.43%).

• Missing weather data after mitigation remains substantial for some variables. Precipitation amount has 33,506,086 missing values (86.91%) and cloud amount 13,048,473 (33.85%). Others are far lower, for example air temperature at 3,182,196 (8.25%) and gust speed at 2,153,067 (5.58%), against a total of 38,551,161 observations.

• A baseline model achieves minute-level accuracy. An XGBoost regression baseline reached a mean absolute error of 2.73 minutes for station-specific delay prediction.

• Station composition. Of the 549 Finnish stations present during 2018–2024, 340 (61.9%) are non-passenger stations for freight/cargo handling, docks or technical service points, and 209 (38.1%) are passenger stations.

Methodology in Plain English

The authors assembled the dataset from two open Finnish sources.

Operational data. Historical train records were pulled from the Digitraffic Railway Traffic Service API operated by Fintraffic, which supplies timetables, actual arrival and departure times, GPS coordinates and speed, train composition and station metadata. Fintraffic precomputes a differenceInMinutes field as actual time minus scheduled time; the dataset labels each record as on time (zero), delayed (positive) or ahead of schedule (negative), and encodes arrival/departure event types so dwell time and delay propagation can be tracked.

Weather data. Meteorological observations came from the Finnish Meteorological Institute through a Web Feature Service API, using the multipoint weather coverage query. Thirteen weather parameters serve as the basis, including air temperature, wind speed and gusts, wind direction, relative humidity, dew-point temperature, precipitation amount and intensity, snow depth, mean sea level pressure, horizontal visibility, cloud amount, and the WMO Code Table 4677 "present weather" code. Of the 209 weather stations used, 164 (78.47%) report at 10-minute intervals and 45 (21.53%) at 1-minute intervals.

Merging. Weather and train records were joined in two stages. Spatially, each train section was matched to the geographically closest weather station using the Haversine great-circle distance formula (with Earth's radius taken as approximately 6,371 km). Temporally, a left join was performed with the train dataset as the base, so no train records were lost; where timestamps did not correspond exactly, nearest-neighbor temporal matching was applied within a tolerance window.

Handling sparse sensors. Because not every station measures every parameter, a radial fallback algorithm was added: if the nearest station lacks a required measurement, the pipeline searches for alternative stations within a 50 km radius that do have it and takes the value from the nearest such station. If none is found within the radius, the value is left missing.

Cleaning and imputation. Identifier fields, duplicate features and irrelevant administrative metadata were dropped. Missing values were treated with four strategies: list-wise deletion when timestamps, delay targets or all weather features were missing; removal of weather columns exceeding 70% missingness; setting missing boolean flags (trainStopping, commercialStop) to False; and month-specific median imputation for the remaining weather features, chosen because the median is robust to sensor outliers.

Feature engineering. Hour, month and day-of-week were converted to sine/cosine pairs (periods of 24, 12 and 7 respectively, with day 1 = Monday) and the original hour column was removed, avoiding the wraparound problem where 23:00 and 00:00 appear numerically distant. Two families of derived weather features were computed at the station level before merging: rolling-window aggregations (minimum, maximum and mean of seven variables over 12 h, 24 h and 72 h, plus mean and cumulative precipitation amount) yielding 69 columns, and ten binary adverse-weather scenario indicators (such as Blizzard, Freezing_Rain, Black_Ice) evaluated instantaneously and over the same three windows, yielding 10 × 4 = 40 columns. The abstract additionally reports robust scaling of weather data to limit the effect of sensor outliers and duplicate removal; the truncated excerpt does not show those steps in detail.

Why This Matters

Impact on research. A systematic review cited by the authors examined 62 publicly available AI-oriented datasets for railway applications and found that operational data dominates the field (47 of 62 datasets) while the vast majority do not incorporate meteorological data. Existing weather-integrated alternatives are limited by short temporal spans or narrow meteorological coverage — for example, the Chinese high-speed dataset of 727 stations and 3,399 trains over a 16-week window, the Italian dataset of 1,397 stations and 3,324 trains over six months of 2024, and the RIDE benchmark linking 94.5 million Belgian train events with 35.7 million weather records. FI-TW's seven-year span, 13 weather parameters underlying 138 features, and sub-Arctic focus fill a specific niche, and the authors note that Nordic studies such as the analysis of Norway's Nordland Line (which found snow depth the strongest weather-related predictor of low punctuality) and Swedish infrastructure-failure classification work relied on proprietary or task-specific data rather than a reusable public resource.

Real-world applications:

• Delay prediction. The baseline XGBoost experiment reaching 2.73 minutes mean absolute error for station-specific prediction shows the dataset is directly usable for forecasting tasks.
• Weather impact assessment. The causes_related_to_weather feature and the ten weather-scenario indicators let analysts quantify how specific conditions such as blizzards and black ice relate to operational disruption.
• Seasonal reliability analysis. The documented month-by-month and weekday-by-week delay patterns support planning around the winter and June peaks.
• Infrastructure vulnerability mapping. The observed geographic clustering of high-delay corridors in central and northern Finland provides a basis for identifying where weather-sensitive assets are concentrated.

Industry relevance. Finnish long-distance services carried more than 15 million of the country's passenger journeys in 2024, and 86.28% of long-distance trains met the Finnish Transport Infrastructure Agency's five-minute punctuality standard in 2024. The paper also notes that Finland's extreme climate — winter temperatures reaching −40 °C — can cause mechanical failures in automatic doors, couplings and switching systems, while heavy snowfall disrupts signaling and autumn leaves reduce rail adhesion. Because the underlying sources are openly licensed (Digitraffic under CC BY 4.0, FMI data under GPL-3.0) and require no authentication, the resource is practical for operators, infrastructure managers and analytics teams rather than being locked behind proprietary access.

Future Directions

• Benchmarking and model comparison. The paper provides only a single XGBoost baseline; standardizing a broader evaluation protocol — as the RIDE benchmark does for Belgian data — would let competing delay-prediction architectures be compared on FI-TW. The paper does not report such a protocol.

• Resolving delay causes more finely. The authors note that although the dataset flags whether a delay's officially recorded cause is weather-related, the detailed underlying causes are not further resolved in the aggregated view and could stem from various factors such as weather conditions or passenger-induced delays at stations. Disaggregating these cause categories is an open direction.

• Extending the radial fallback strategy. Precipitation amount remains 86.91% missing after mitigation and cloud amount 33.85%, which constrains analysis of snow and rainfall-driven delay. The paper does not report alternative imputation or sensor-augmentation strategies beyond the 50 km radial search.

• Transferring the approach beyond Finland. Comparisons in the paper span Chinese, Italian and Belgian datasets, and Nordic work exists for Norway and Sweden. Whether the same collection, spatial matching and feature-engineering pipeline generalizes to a joint Nordic or pan-European resource is raised but not attempted here.

Target Audience

Machine learning researchers and graduate students working on spatio-temporal data, transport analytics or weather-sensitive infrastructure; railway operators and infrastructure managers in Nordic or other severe-winter regions interested in data-driven reliability analysis; and practitioners building delay-prediction, scheduling or maintenance-planning models who need a public, openly licensed dataset with operational and meteorological variables already synchronized. Readers primarily interested in new modeling architectures will find less novelty here — the value lies in the dataset, its documented construction pipeline and its baseline characterization.

Authors’ abstract

Train delays result from complex interactions between operational, technical, and environmental factors. While weather impacts railway reliability, particularly in Nordic regions, existing datasets rarely integrate meteorological information with operational train data. This study presents the first publicly available dataset combining Finnish railway operations with synchronized meteorological observations from 2018-2024. The dataset integrates operational metrics from Finland Digitraffic Railway Traffic Service with weather measurements from 209 environmental monitoring stations, using spatial-temporal alignment via Haversine distance. It encompasses 28 engineered features across operational variables and meteorological measurements, covering approximately 38.5 million observations from Finland's 5,915-kilometer rail network. Preprocessing includes strategic missing data handling through spatial fallback algorithms, cyclical encoding of temporal features, and robust scaling of weather data to address sensor outliers. Analysis reveals distinct seasonal patterns, with winter months exhibiting delay rates exceeding 25\% and geographic clustering of high-delay corridors in central and northern Finland. Furthermore, the work demonstrates applications of the data set in analysing the reliability of railway traffic in Finland. A baseline experiment using XGBoost regression achieved a Mean Absolute Error of 2.73 minutes for predicting station-specific delays, demonstrating the dataset's utility for machine learning applications. The dataset enables diverse applications, including train delay prediction, weather impact assessment, and infrastructure vulnerability mapping, providing researchers with a flexible resource for machine learning applications in railway operations research.