Bus arrival time prediction is an important basis for the decision-making assistant system of bus dispatching. It helps dispatchers to find late vehicles in time and make reasonable dispatching decisions. However, bus arrival time is influenced by traffic congestion, weather, and variable time when stopping at stations or travelling duration between stations. It is a spatio-temporal dependence problem, which is quite challenging. This study proposes a new algorithm called STPM for bus arrival time prediction based on deep neural network. The algorithm uses space-time components, attribute components and fusion components to predict the total bus arrival time from the starting point to the terminal. In this algorithm, time-dependence and space-time components are used to learn the internal spatio-temporal dependence. It uses attribute components to learn the influence of external factors, uses fusion components to fuse the output of temporal and spatial components, as well as attribute components, to predict the final results. Experimental results show that STPM can combine the advantages of convolutional neural network and recurrent neural network model to learn the key temporal and spatial features. The proposed algorithm outperforms existing methods in terms of the error percentage and accuracy of bus arrival time prediction.