Explainable Machine Learning for Sustainable Urban Mobility: Capital Bikeshare Demand under Weather and Seasonal Conditions

Bike-sharing systems generate fine-grained ridership records that make weather- and season-sensitive demand visible, yet many prediction studies emphasize predictive accuracy over the transparency that operators need for planning. This study develops and validates explainable machine learning models for day- and hour-level bike-sharing demand using the publicly available Capital Bikeshare dataset (Washington, D.C., 2011-2012). Following a data-quality screening stage —which retained 165 genuinely absent hourly records without synthetic imputation, corrected a single humidity-sensor anomaly, and excluded casual and registered ridership counts as predictors to prevent target leakage— day- and hour-level models were compared under a strictly chronological validation scheme, using five expanding calendar-based cross-validation folds and a final-quarter 2012 hold-out. A Ridge regression with a linear day-trend term was selected at day level (cross-validation RMSE = 979.4; hold-out R² = 0.458), while an XGBoost model with the same trend term was selected at hour level (cross-validation RMSE = 73.4; hold-out R² = 0.757). SHAP-based explanation identified the day trend, year, and temperature as the leading day-level demand drivers, and hour-of-day, the day trend, and working-day status as the leading hour-level drivers. A sensitivity analysis showed that a single extreme low-demand day (29 October 2012, coinciding with Capital Bikeshare’s service suspension ahead of Hurricane Sandy) accounted for 82% of total day-level absolute percentage error, while scale-robust error measures (sMAPE, WAPE) remained comparatively stable. Variance-inflation diagnostics required excluding apparent temperature from linear model specifications due to strong collinearity with temperature. These findings support weather- and season-aware, chronologically validated demand forecasting as a transparent decision-support tool for sustainable urban mobility planning, while identifying service-disruption events as an unmodeled source of error.