Dublin Bikes Data Analysis and Forecasting (R)

Exploratory analysis and short-term demand forecasting for Dublin's public bike share system, using four months of station-level availability data.

What this does

Given historical bike availability records across every Dublin Bikes station, this project explores usage patterns and builds a model to forecast bike availability an hour ahead. The pipeline:

1. Data loading. Four months of historical station data (January to April 2022) loaded and combined, alongside station metadata (location, capacity).
2. Feature extraction. Timestamps parsed into hour, weekday, and month components, with duplicates and missing values removed.
3. Exploratory visualisation. Hourly availability heatmaps, weekday distributions (ridgeline plots), monthly boxplots, a radial hour by weekday heatmap, and an interactive Leaflet map showing the busiest stations by average availability.
4. Forecasting. Hourly aggregation with lag features (1 hour, 2 hours, and 24 hours prior), then two models: a Random Forest (500 trees, via ranger) and a Linear Regression baseline, both predicting bike availability from time features and recent lags.
5. Validation. The Random Forest is trained on January and February, validated on March, then retrained on January through March to forecast April.

Results

Random Forest (per station, per hour):

Evaluation	RMSE	MAE	R-squared
March validation (trained on Jan-Feb)	2.527	1.525	92.4%
April forecast (trained on Jan-Mar)	2.607	1.554	91.9%

The model explains over 90% of the variance in hourly bike availability a month ahead of its training window, with a typical error of around 1.5 bikes per station per hour.

Linear Regression was also evaluated as a simpler baseline, but on a different level of aggregation (total predicted bikes across all ~113 stations, summed per hour, for a single day) rather than the per-station hourly level the Random Forest was scored on. On that basis it reached an R-squared of 80.5% for April 1st, noticeably weaker than the Random Forest, though the two numbers aren't directly comparable since they're measuring different things. The honest takeaway is that the Random Forest is the model actually validated at the level that matters (per station, per hour), and the linear model serves mainly to show that a non-linear approach adds real value over a simple baseline.

Dataset

633,460 records across 113 unique stations, covering January through April 2022, sourced from a public Dublin Bikes historical dataset (station availability, GPS coordinates, timestamps). The raw CSVs aren't included in this repo; download the dataset separately and update the file paths at the top of the notebook before running.

Tech stack

R, tidyverse (data wrangling and visualisation), lubridate (datetime handling), tidymodels and ranger (Random Forest), parsnip (linear regression), leaflet (interactive station map), ggridges and viridis (visualisation).

Running it

install.packages(c("tidyverse", "lubridate", "ggthemes", "scales",
                    "ggridges", "viridis", "leaflet", "tidymodels",
                    "ranger", "parsnip", "Metrics"))

Open main.ipynb in Jupyter with an R kernel, or import the cells into RStudio, updating the dataset file paths first.