Accurate prediction of bus arrival time is critical to improving transit service reliability and scheduling efficiency. The estimated time of arrival (ETA) is jointly determined by two processes, namely inter-station travel and station dwell, which are dominated by different traffic mechanisms and thus follow essentially different operating patterns. Most existing methods model the entire route in a unified way, making it difficult to account for both processes simultaneously. This paper therefore proposes a feature-separation prediction method. For inter-station travel time, which is driven by the continuous evolution of traffic flow, we design a Dual-Branch Spatio-Temporal Graph Convolutional Network (DSTGCN); its spectral graph convolution and spatial self-attention branches work together to capture road-network topology and the spatial propagation of traffic flow. For station dwell time, which follows a long-tailed distribution, we design a Spatio-Temporal Graph Convolutional Network with Transformer (STGCN-Trans) and introduce a robust loss function to suppress the effect of extreme outliers. Comparative and ablation experiments show that both models achieve the lowest RMSE, reducing it by 21.7% and 6.7% over the second-best baselines, respectively. They also remain stable under disturbances such as peak hours and extreme weather.