Probabilistic Bike-Sharing Demand Forecasting under Changing Weather and Seasonal Regimes with Transformer-Based Models

Bike-sharing systems must plan fleet and rebalancing under changing weather and seasonal regimes. We conduct reproducible probabilistic one-hour-ahead demand forecasting on two public datasets (UCI Bike Sharing hourly data and Seoul Bike Sharing Demand) under forward-chaining temporal splits. We compare tree-based Random Forest and XGBoost quantile baselines with N-BEATS, a Transformer encoder, an Informer-lite encoder, and a Temporal Fusion Transformer (TFT). Across both cities, N-BEATS and tree models yield the lowest MAE, while TFT produces the best calibrated 80% prediction intervals (81.1% coverage in Seoul). Error and coverage degrade under extreme weather, motivating uncertainty-aware operations.