Capacity prediction plays an important role in smart logistics, and its results are important for improving the accuracy of capacity scheduling and truck-cargo matching. Existing researches on capacity prediction in urban road networks aim to determine the number of available vehicles in future periods, while the problem of capacity prediction in bulk logistics aims at predicting the information on the trucks (e.g. the truck’s identity document (ID)) to carry certain types of goods for different flows, which is closely related to whether the trucks can return to the steel plant within the expected time (called capacity accessibility). In the case of bulk logistics, it is necessary to take into account the impact of the time spent on the two trips from the steel plant to the customer’s business and back to the steel plant. Since trucks need to stop several times in the long-distance transportation process but the length of stopping time varies, the uncertainty of stopping time makes the accurate prediction of transportation delivery time difficult. In addition, the freight platform only assigns capacity to one-way transport tasks (i.e. from the steel plant to the customer’s business), and the return trip (i.e. back to the steel plant) is determined by the truck drivers, which leads to the lack of return trajectory and poses a challenge to predict the return time of trucks to the steel plant. In order to solve the above challenges, based on the data sets of waybills, trucks, trajectories, and transport endpoints of logistics enterprises, we extract the stay behavior features, transport endpoint features, and environmental features. Then, the self-attention mechanism is introduced to obtain the weights of different features on the time consumption of two trips respectively to further improve the accuracy of capacity accessibility prediction. On this basis, a truck capacity prediction method based on self-attention mechanism is proposed, including capacity candidate set generation based on historical flow similarity, capacity accessibility prediction based on self-attention mechanism, and capacity carrier flow prediction based on long short-term memory (LSTM). Finally, the experimental results of comparison experiments on real logistics datasets show that the proposed method has higher prediction accuracy and can provide powerful decision support for the optimization of capacity scheduling in bulk logistics.