Enhancing bike-sharing demand forecasting with spatio-temporal learning: A comparative study on a local dataset

Creative Commons License

ÇAKIROĞLU M. A., Ozdemir S.

Journal of Construction Engineering, Management & Innovation (Online), cilt.9, sa.2, ss.1-12, 2026 (ESCI, TRDizin)

Özet

Bike-sharing systems rely on accurate short-term demand forecasts to prevent shortages, surpluses, and costly rebalancing operations. Accurate predictions are also essential for enhancing the environmental, operational, and social sustainability of these systems, as improved forecasting may help reduce truck-based redistribution, potentially lowering emissions and supporting more efficient resource usage and more reliable urban mobility services. In this study, we evaluate a wide spectrum of forecasting approaches—ranging from classical time-series models (ARIMA, Prophet) and ensemble learners (Random Forest, XGBoost) to spatio-temporal deep learning models (LSTM variants, ST-GCN, GraphWaveNet, and GNN)—using a local station-level dataset. We incorporate both temporal history and spatial dependencies among stations under a unified evaluation protocol (24-hour look-back, one-hour-ahead prediction). Our findings show that integrating spatial context consistently improves accuracy. The spatio-augmented Random Forest (RF-ST) achieves the best performance, reducing error rates by over 10% compared to its temporal- only counterpart and by more than 35% relative to ARIMA and Prophet. Graph-based neural models (e.g., GNN) deliver comparable accuracy, further confirming the benefits of explicit spatial modeling. These results highlight the potential sustainability implications of spatio-temporal forecasting, suggesting that more accurate station-level predictions may support more sustainable rebalancing strategies, potentially help reduce operational inefficiencies, and strengthen the long-term viability of bike-sharing systems as a sustainable transportation mode.