In opportunistic networks characterized by dynamics of node topology and attribute evolution, developing adaptive routing algorithm with multi-attribute fusion is imperative to minimize overhead and enhance delivery probability. Existing approaches predominantly rely on stationary weighting schemes for neighbor evaluation or single-attribute thresholding for routing algorithm, which suffer from inadequate modeling of dynamic network dynamics and overlook inter-attribute dependencies. In this paper, we propose an adaptive topology-aware routing algorithm for multi-attribute fusion(ATR algorithm), which is based on the Q-learning algorithm enhanced by Graph Attention Network(GAT). ATR establishes joint representations of node characteristics and message attributes. By employing a multi-head GAT for adaptive attribute fusion, attention weights are computed to quantify node collaboration potential, which are then mapped to initialize Q-values in Q-learning. This guides the algorithm to prioritize high-potential paths. Simulation results show that the ATR algorithm performs better in terms of delivery probability, average delay, and overhead compared to the Epidemic, Prophet, KROP, and ORQLCI algorithms, among them, the delivery probability is increased by 21.16%, 11.89%, 12.16%, and 5.91%, the average delay is reduced by 95.64, 69.67, 173.45, and 55.40, and the overhead ratio is reduced by 187, 90.18, 137.07, and 70.43.