GOSPACE LABS: Prediction of Parking Space Occupancy in Towns

How we approached it

We approached this by combining two complementary modeling tracks in one reproducible machine-learning pipeline:

We used DVC (Data Version Control) to manage both tracks, allowing preprocessing, training, prediction, and evaluation to be rerun reproducibly across dataset versions.

What we delivered

The delivered result was a full end-to-end forecasting framework for parking management: data processing, feature engineering, model training, prediction generation, and automated evaluation outputs. It provides occupancy forecasts at multiple levels (aggregate and slot-level), and produces decision-ready artifacts such as interval plots, performance metrics, and model visualizations. In practical terms, the partner received a maintainable system for parking predictions with transparent confidence estimates, ready for iterative improvement as new data arrives.

Partner’s perspective (GOSPACE LABS)

We supplied occupancy data from our IoT parking sensors and defined the questions the model had to answer. What we wanted was probabilistic: for a given hour and day, how many spaces are likely to be free, expressed with a confidence level rather than a single number, and projected a day ahead. We also asked for forecasts at the level of individual bays and for patterns in how our most active users behave. The reason is operational. Reliable occupancy forecasts let us share the same physical spaces more efficiently and support dynamic pricing, since demand is not constant. Occupancy climbs in bad weather, for example, when more people drive instead of walk. The forecasts also power what we call easy sharing: when a driver reliably commutes on the same days and time window, we can offer their bay to someone else while it sits empty, and still hold back enough capacity for the small share of drivers who arrive unpredictably. Knowing that probability tells us exactly how many spaces to keep in reserve.