Individual mobility is driven by activities and restricted geographically, especially for trip destination prediction in public transport. An individual may perform the same activity at different places (e.g., using different stations for shopping), which is not modeled in existing prediction studies. The paper proposes a deep learning model with activity, geographic and sequential (DeepAGS) information in predicting an individual's next trip destination. It uses word embedding, GCN, and an adaptive neural fusion gate to model activity representation (semantic and geographical features extraction and fusion). Then GRU model with an attention mechanism is adopted to extract the activity's temporal mobility patterns. The approach is validated using a large-scale farecard dataset in urban railway systems. Also, the working mechanism of DeepAGS is illustrated using synthetic data.