Large-scale prediction of trip production is essential for origin–destination (OD) demand estimation and prediction. One of the main challenges in predicting trip production patterns lies in addressing spatial-temporal correlations and variations. Whereas many studies focus on temporal correlations, very few consider spatial adjacency between traffic analysis zones (TAZ) as explanatory variables. This research proposes a method that integrates a graph convolutional neural network (GCN) into a long short-term memory network (LSTM) to do exactly that. By introducing a nationwide graph that encodes the adjacency of TAZs, spatial heterogeneity is considered in the prediction process, and a single prediction model is trained for the entire network, thereby avoiding the need to train multiple separate models and potentially reducing overall training overhead, while increasing the prediction accuracy. Moreover, with this model, we investigate the effect of spatial scale on spatial uncertainty and prediction accuracy and analyze prediction errors, residual patterns, and their associations with socio-spatial features at different spatial scales. The findings of this research have important implications for improving OD demand prediction models and provide valuable insights into the role of spatial scale and socio-spatial features in travel demand prediction.

Global System for Mobile Communications (GSM) data provides valuable insights into travel demand patterns by capturing people's consecutive locations. A major challenge, however, is how the polling interval (PI; the time between consecutive location updates) affects the accuracy in reconstructing the spatio-temporal travel patterns. Longer PIs will lead to lower accuracy and may even miss shorter activities or trips when not properly accounted for. In this paper, we analyze the effects of the PI on the ability to reconstruct an origin–destination (OD) matrix. We also propose and validate a new data-driven method that improves accuracy in case of longer PIs. The new method first learns temporal patterns in activities and trips, based on travel diaries, that are then used to infer activity-travel patterns from the (sparse) GSM traces. Both steps are data-driven thus avoiding any a priori (behavioral, temporal) assumptions. To validate the method we use synthetic data generated from a calibrated agent-based transport model. This gives us ground-truth OD patterns and full experimental control. The analysis results show that with our method it is possible to reliably reconstruct OD matrices even from very small data samples (i.e., travel diaries from a small segment of the population) that contain as little as 1% of the population’s movements. This is promising for real-life applications where the amount of empirical data is also limited.

This paper explores temporal patterns in travel production using a full month of production data from traffic analysis zones (TAZ) in the (entire) Netherlands. The mentioned data is a processed aggregated derivative (due to pr ivacy concerns) from GSM traces of a Dutch telecommunication company. This research thus also sheds light on whether such a processed data source is representative of both regular and non-regular patterns in travel production and how such data can be used for planning purposes. To this end, we construct normalized matrix (heatmap) representations of weekly hour-by-hour travel production patterns of over 1200 TAZs, which we cluster using K-means combined with deep convolutional neural networks (inception V3) to extract relevant features. A silhouette score shows that three dominant clusters of temporal patterns can be discerned (K=3). These three clusters have distinctly different within-day and day-to-day production patterns in terms of peak period intensity over different days of the week. Subsequently, a spatial analysis of these clusters reveals that the differences can be related to (easily observable) land-use features such as urbanization levels (i.e., Urban, Rural, and mixed-level). To substantiate this hypothesis and the usefulness of this clustering result, we apply an OVR-SMOTE-XGBoost ensemble classification model on the land-use features of the TAZs (i.e., to identify their cluster). The results of our clustering analysis show that given the land-use features, the overall production patterns are identifiable. Further analysis of the mixed-level areas shows a more complex relationship between temporal heterogeneity and spatial characteristics. Population density seems to impose additional uncertainty on the temporal patterns. All in all, feature selection and spatial and temporal discretization play essential roles in identifying the dominant trip production patterns. These findings are directly useful for data-driven estimation and prediction of demand time series. Furthermore, this study provides further insights into people's mobility, relevant for transportation analysis and policies.