New Framework Shows No Single Model Predicts Demand for Flexible Bus Routes
Researchers tested four different approaches to forecasting passenger demand for on-demand transit systems and found each excels at different metrics—suggesting operators need hybrid strategies. The finding matters because cities increasingly deploy flexible transit to reduce costs, and picking the wrong demand model could waste both money and service quality.
Originaltitel: A Modular Synthetic Demand Generation Framework for Demand-Responsive Public Transport
<p>This study develops a modular framework for generating synthetic demand for demand-responsive public transport (DRT) using real service data. From 518 days of operational records (∼37,000 trips), we construct a simulated empirical demand (SED) baseline and four synthetic demand variants (V1-V4), each incorporating distinct aspects of travel behaviour: independent trip sampling, trip chaining, user-specific temporal preferences, and origin-destination (OD) memory. Thirteen performance indicators derived from DRT simulations, covering passenger experience and fleet efficiency, reveal that no single model outperforms the others across all metrics. V1, based on independent sampling, produces inflated waiting times in simulation, with a mean value +8-17% above SED. V2, which incorporates within-day trip chaining, is closest to SED on core passenger-experience indicators such as mean waiting time and in-vehicle time, with mean waiting increasing by ∼ +6% on weekdays and ∼ +15% on weekends. V3 yields efficiency indicators closest to SED in terms of passenger-kilometres per vehicle-kilometre (especially on weekends), but increases peak-hour concentration and consequently worsens waiting-time reliability relative to V2. V4 introduces OD memory and user-aware activation, achieving moderate performance across most metrics but not outperforming other variants in any specific dimension. These results highlight trade-offs between behavioural realism and data requirements, while the sensitivity analysis confirms that the main model-ranking patterns remain broadly stable across different demand-intensity levels. Our findings suggest that model selection should depend on the simulation objective: V2 for passenger-focused analysis, V3 for efficiency-oriented assessments, and V4 when user-level regularity and OD repetition are relevant.</p>