Parking facility performance prediction using multi-target conformal regression
This thesis proposes a multi-target conformal regression approach for estimating the performance of new parking facility locations to be acquired by Q-Park.
Such forecasts should eliminate the need for consultancy reports prior to the development, sale or lease of new car parks: the basic parameters can be inserted into the algorithm and the artificial intelligence does the rest.
The basic data for the machine learning model include the capacity of the new car park, other car parks within a 1 km radius and their capacity, the presence of a train station within 500m and the numbers of offices, shops, hotels, restaurants and bars, educational institutions, industry and other buildings within 350m derived from OpenStreetMaps. Data relating to the floor space of shops, numbers of rooms in hotels and the like was not available for this research.
Data was collected for 1,037 existing Q-Park parking facilities in seven different countries. For these car parks, data including the number of hours parked, access and exit times, average length of stay, average occupancy and parking turnover per day were entered into the system. In addition, distinctions were made per country.
Various artificial intelligence techniques were applied to this database to identify which self-learning computational method best approximates the data imported. During the study, a prediction technology emerged which gave the best results. However, further research with more deep learning would be valuable.
Furthermore, additional more detailed basic data, such as shop floor area, numbers of workstations in offices, and numbers of hotel rooms, as well as results from existing car parks would make the artificial intelligence results even more reliable.
The thesis identifies the configuration of the regression model best suited for the task and compares the performance of different combinations of single and multi-target regression and conformal prediction. It also suggests the conformal method resulting in the most informative prediction regions.