Predicting Floating Car Data (FCD) Penetration Rate for Urban Traffic Estimation

For urban mobility planning and management, large-scale traffic monitoring is critical. Fixed sensors offer high-precision measurements, but their installation and maintenance are expensive, limiting road network coverage. As a result, floating vehicle data (FCD) presents itself as an attractive alternative, since it can potentially provide greater geographic coverage and lower operating costs. However, the application of FCD for traffic volume prediction depends on the Penetration Rate (PR), which is the sampling rate of detected vehicles in relation to the total flow. Considering that PR varies both spatially and temporally, assuming fixed values can impair the accuracy of flow estimates. This work introduces a machine learning model, based on LightGBM, to anticipate PR in urban streets, aiming to improve the estimation of traffic volumes. The model was created based on data from eleven inductive loop sensors and their respective FCD segments from TomTom data, obtained in Bologna, Italy. For the analysis of spatial transfer in the network under investigation, the Leave-One-Group-Out validation approach was utilized, where the network was tested on one sensor location at a time. The results indicate a high predictive capacity, with SMAPE between 8.49% and 12.81%. The findings indicate that PR presents recurring and predictable patterns, allowing its estimation even in locations without fixed monitoring. Thus, the proposed approach reduces the uncertainty associated with traffic flow estimates based on FCD and expands the potential use of this data source for monitoring and managing urban mobility.