A new academic model forecasts long-term traffic on roads that have no sensors installed at all.
Researchers published a system called SLPF, short for Spatio-temporal Long-term Partial sensing Forecast, aimed at a real gap in traffic prediction: most cities only have sensors at some intersections, and existing tools either assume full coverage or only predict a few minutes ahead. SLPF instead predicts further into the future using data from a partial sensor network. It uses a rank-based embedding to filter noisy readings, a spatial transfer matrix to estimate conditions at unsensed locations from data recorded at sensed ones, and a multi-step training process that extracts more signal from the available data to refine its predictions. The team tested it on several real-world traffic datasets and reported better accuracy than existing methods, with code posted on GitHub.
Most traffic forecasting research assumes a tidy, fully-sensed network, which is not how actual cities work. Sensor coverage is patchy and costly to expand, so a model that infers conditions on unsensed roads from neighboring sensor data is directly useful for traffic management systems and navigation apps that have to work with incomplete data.
It is a lab result, not a shipped product, so the real test is whether any transportation department ever plugs this into a live system instead of a paper benchmark.