<p>Accurate prediction of origin–destination matrices, which indicate the number of trips between different areas, plays a crucial role in urban planning, traffic management, and infrastructure development. This research aims to design and implement a scalable model based on Graph Neural Network (GNN) to estimate origin–destination matrices in expanding networks, where new zones will be added in the future. In this study, the features of origin and destination nodes are extracted and combined using two separate GraphSAGE networks, a type of GNN, and then the combined features for each origin–destination pair are fed into a neural network to predict the number of trips. The output of this model is compared and evaluated against the gravity model and machine learning-based models. A case study was conducted using data from the Greater Toronto and Hamilton Area under three scenarios: no expansion, minimal expansion, and extreme expansion. Across all scenarios, the proposed model consistently outperformed the baseline models. Notably, in the extreme expansion scenario, where the task involved predicting trips to and from many newly added nodes, the proposed model achieved an R-squared of approximately 90%, which represents a significant improvement over the 78% R-squared achieved by other models. Furthermore, the model demonstrates high computational efficiency on standard hardware, highlighting its potential as a practical tool for urban planning tasks.</p>