Multi-step forecasting of online car-hailing demand is necessary for the long-term planning of traffic resources. However, accurate forecasting is very challenging, because it is difficult to extract the dynamic spatial-temporal dependencies from high-dimensional nonlinear data. This paper proposes a regional traffic demand forecasting framework (RTDFF) to predict the boarding and alighting demand of online car-hailing. The RTDFF can learn robust representations of traffic patterns with multi-resolution spatial dependencies. It uses a 3D deconvolution component to enhance the spatial resolution of intra-regions and a 2D residual component to connect the large-scale spatial dependence. This model investigates the effect of the 3D convolution in capturing the long-term temporal dependence. Hybrid modeling is conducted with multi-granularity temporal segmentation and fusion. The experiments were carried out using real-world datasets. The results verify that RTDFF is accurate and stable, and that it has slower deterioration than baseline models, with the raise of prediction steps.