Long-term bike-sharing demand forecasting and interpretability analysis considering Urban Heat Island (UHI)

The Urban Heat Island (UHI) effect and deteriorating air quality are reshaping travel behavior in many large cities and pose new challenges for the operation of bike-sharing systems. This paper examines how UHI and related environmental conditions affect long-term shared bike demand and how these effects can be incorporated into operational planning, using Shenzhen as a case study. We compile an hourly multi-source dataset that links bike-sharing trips with UHI intensity, meteorological variables and air pollutant concentrations, and propose an Head-Gated Transformer model (HGT) – an improved Transformer with an adaptive multi-head attention mechanism – for 24–72 h ahead demand forecasting. Compared with representative deep learning (LSTM, CNN, standard Transformer) and tree-based (GBDT) benchmarks, Head-Gated Transformer consistently delivers the lowest mean squared and mean absolute errors, reducing 24–72 h prediction errors by around 1%–6% in MSE relative to the best Transformer baseline. Building on this, we combine SHAP analysis, attention visualization and (bi)variate partial dependence plots to characterize temporal dependence, nonlinearities and interaction effects. The study indicates that residents consider both short-term risks and long-term stability of environmental conditions when traveling, simultaneously reflecting certain rigid demands and adaptive responses. Meanwhile, variables such as UHI have significant interactive, nonlinear, and threshold effects on the travel demand for shared bikes. The research outcomes provide robust data support and policy recommendations to promote shared mobility.