The present study investigates the determinants of the volatility of passenger demand for paratransit services and explores the feasibility of a data-driven model for medium-term forecast of the daily demand. Medium-term demand forecasting is a significant insight to optimise resource allocation (staff and vehicles) and reduce operations costs. Using operational data from the reservation platform of the paratransit services in Toulouse, France, and enriching them with exogenous information, the study derives statistical and deep learning models for medium-term forecast. These models include a seasonal ARIMAX model with rolling forecast, a Random Forest Regressor, a LSTM neural network with exogenous information and a CNN neural network with independent variables. The seasonal ARIMAX model yields the best performance, suggesting that when linear relationships are considered, econometric models and deep learning models do not have significant differences in their performance. All the models show limited ability to grasp unique events with multi-day impacts such as strikes. Albeit a highly volatile demand and limited knowledge ahead of the forecast, these models suggest the volume of early reservations is a good proxy for the daily demand.

本研究调查了乘客对辅助运输服务需求波动性的决定因素，并探讨了以数据驱动模型进行每日需求中期预测的可行性。中期需求预测是优化资源分配（人员和车辆）并降低运营成本的重要见解。该研究利用法国图卢兹的辅助运输服务预订平台上的运营数据，并利用外来信息对其进行了充实，从而得出了用于中期预测的统计和深度学习模型。这些模型包括具有滚动预测的季节性ARIMAX模型，随机森林回归，具有外源信息的LSTM神经网络和具有自变量的CNN神经网络。季节性ARIMAX模型可产生最佳效果，这表明，当考虑线性关系时，计量经济学模型和深度学习模型的性能没有显着差异。所有模型均显示出在罢工等多日影响下把握独特事件的能力有限。尽管需求高度波动，并且预报前的知识有限，但这些模型表明，提前预订量可以很好地替代每日需求。