In this paper we address a challenging problem of predicting on-street parking occupancy based on sensors with occasionally missing data. These missing values are most likely not missing at random. However, the exact process behind it is unknown at the outset.
For example, an entry in our data stream might read: “at time t, blockface b, which has capacity 20, has 18 working sensors, and of those working sensors 10 indicate that a parking space is occupied.” The challenge is to infer as accurately as possible the actual occupancy of blockface b, i.e. the occupancy of all parking spaces, including the ones with inoperative sensors. The fact that sensors are inoperative only occasionally allows us to learn a demand model and a sensor noise model jointly, and use this to “fill in the blanks” in the best possible way. We introduce a series of sensor noise models of increasing complexity that are suited for several sensor characteristics. Cross-validation allows the selection of the suitable model for a particular application. The sensor noise models we introduce are flexible and can be combined with any probabilistic model for demand and hence have a very wide applicability. In our demonstration application we find that sensors are more likely to be inoperative when parking spaces are empty, in particular during busy traffic periods. This supports the hypothesis that drive-by-traffic is an important source of noise for the sensor. Compared to
baseline methods based on a missing-at-random assumption our method gives better predictive performance and avoids a systematic bias in the inferred occupancies.