Traffic agencies need current and near-future speed estimates for every link in their freeway network, but fixed sensors are typically sparse. A new arXiv preprint proposes a graph convolutional recurrent network that uses probe vehicle data as supervision to fill in the gaps. The work studies joint nowcasting and next-hour forecasting of hourly speeds on all 4,127 directional links in the network.
According to the abstract, the key idea is to train the network with probe data, which provides broader coverage than fixed sensors. This allows the model to estimate speeds on links that have no fixed sensor of their own. The authors frame the task as both nowcasting (current speeds) and forecasting (one hour ahead), suggesting a single model for both.
Because the abstract is brief, details on the network architecture, training data, and evaluation results are not available from this source. The paper is available on arXiv under ID 2610.22630.