Causal Aware Feature Fusion DeepONet Mamba Model for eVTOL Vertiport Demand Forecasting

Accurate demand forecasting is a prerequisite for eVTOL vertiport dispatch, pad staffing, battery service planning, and access management, yet short-horizon demand is affected by confounded interactions among weather, road access, queue feedback, aircraft availability, and pricing interventions. This study evaluates the proposed model on a synthetic but engineering-constrained eight-vertiport network, so the reported results should be interpreted as controlled methodological evidence rather than direct commercial demand measurement. The paper proposes a Causal Aware Feature Fusion DeepONet Mamba model that learns a function-to-function mapping from urban mobility states to vertiport demand fields. Causal gates separate contextual variables, network states, and intervention-sensitive variables before a Mamba branch encoder and coordinate-aware DeepONet trunk produce interval demand forecasts. Experiments with 184 simulated operating days show that the proposed model achieves an MAE of 18.7 requests per 15-minute interval, reducing error by 15.6% relative to the strongest non-causal baseline. Peak interval error decreases to 54 requests, 0.9-quantile pinball loss decreases to 11.3, and scenario ranking error is reduced by 32.8% compared with Mamba without causal fusion. Spatial transfer tests, missing-sensor experiments, and capacity-restriction scenarios indicate that causal feature control improves robustness when operating conditions deviate from ordinary correlations. The results show that combining causal fusion, neural operator learning, and selective state-space memory provides a practical forecasting framework for early-stage vertiport network planning and tactical eVTOL operations.