Estimating the travel time of any path (denoted by a sequence of GPS points) in a city is the key research problem of this work. It plays a vital role in traffic control, path planning, vehicle scheduling, and so on. Travel time prediction is a challenging proposition to predict the travel time of motor vehicles accurately. It is interfered with by many complex factors, such as Spatio-temporal correlation, traffic light influence, data sparse, etc. In most of the existing research, it estimates the travel time of motor vehicles in a single section or subpath, and it gets the result without considering the situation of intersections and traffic lights. It is difficult to predict accurately the travel time of a longer path. In this paper, the deep network based on stacked denoising autoencoder is proposed, which can directly estimate the travel time of any path and overcome the error accumulation problem of long path estimation. The experimental results of taxi trajectory data set show that the travel time prediction method proposed in this paper based on the stack noise reduction self-encoder has good prediction accuracy and algorithm stability.