Forecasting the demand for different modes of transportation at gateway hubs, such as high-speed train stations and airports, plays a crucial role in urban transit ecosystems. One important characteristic of this problem is that some future exogenous information, such as passenger inflow and weather conditions influencing future demand, can be obtained in advance. Traditional time series forecasting approaches have not fully utilized this characteristic. To address this issue, we propose a novel Transformer architecture called FXFormer that utilizes future exogenous information. We first decompose the input into historical information containing demand time series and covariate time series, as well as future information containing future covariates, and apply different attention mechanisms to each part. To fully exploit the relationships between variables, we treat each variate as a token. After processing through different attention mechanisms, we design a gate mechanism to fuse historical and future information to enhance the model’s performance. Extensive experiments conducted using multi-mode demand datasets from a high-speed railway station and an airport in Chengdu City demonstrate that the proposed FXFormer outperforms state-of-the-art multivariate time series forecasting approaches.