Large-scale parking data prediction is a key technical problem to be tackled in current citywide parking systems. Graph neural network is the main parking prediction approach at present. However, due to the large size of citywide parking networks, the time cost of training the entire network is unaffordable. This paper proposes a parking zone division(PZD) algorithm to accelerate the model training progress from the perspective of graph partition. PZD divides the entire parking network into multiple parking zones consisting of several parking lots according to the importance ranking of urban parking lots and cruising preferences for adjacent parking lots. The obtained parking zones are used for training and prediction via T-GCN in parallel to improve the model's performance on the entire network. The experimental results show that PZD is interpretable and reasonable as the partition conforms to the common experience of urban parking. In terms of training efficiency, it improves by 3.216-4 times compared to training the entire network, and in terms of accuracy, it improves by 48.90 % compared to predicting the entire network and by 19%-27% compared to the baselines of the referenced graph partition algorithms.