The construction and maintenance of park and ride lots represents a substantial public investment that if used judiciously can reduce congestion and emissions through the use of transit or the sharing of vehicle trips. With 297 lots scattered throughout Virginia, the Virginia Department of Transportation (VDOT) needs an approach for forecasting demand for these lots so that investments can be made wisely. Unfortunately, direct application of an existing approach yielded absolute differences (between forecast occupancy and observed occupancy) that depending on the VDOT district were 14 to 141 times larger than the observed occupancy. Calibrating an existing approach to Virginia-specific traffic volumes for the roadway serving the lot and the highest volume roadway within 2.5 miles of the lot reduced the scale of this error but still yielded forecasts where the mean testing error exceeded the mean occupancy for a majority of models.
Accordingly, 19 Virginia-specific models were developed that reflected distinct regions in Virginia. These models followed VDOT district boundaries for three of VDOT’s nine districts (Lynchburg, Richmond, and Northern Virginia); planning district commission (PDC) or metropolitan planning organization (MPO) boundaries for four VDOT districts (Bristol, Culpeper, Salem, and Staunton); and urban/rural categorizations for two VDOT districts(Fredericksburg and Hampton Roads). A key finding was that determinants of occupancy varied by location. Statistically significant determinants included residents with a commute of 50+ miles (used in four models affecting10% of Virginia’s lots, such as those in the Lenowisco PDC in the Bristol District); the availability of transit service or the number of commuters who choose transit (used as a positive factor in seven models affecting more than one-half  of Virginia’s lots, such as those in the urbanized portion of the Culpeper District); amenities such as lighting (a variable in two models reflecting 15% of Virginia’s lots such as those in the low population density areas of the Fredericksburg District); traffic volume (a factor in five models representing 46% of Virginia’s lots, such as those in the Lynchburg District);and the provision of bicycle spaces (a factor in the model for 78 of the Northern Virginia District lots, or about 26%of the statewide total). Thus, the models can help forecast how key changes (such as an increase in traffic, the introduction of transit service, or the addition of lighting) may influence demand at an existing lot.
The median-adjusted R-squared value (coefficient of determination) for the 19 models was 87%. The Richmond District was representative: a model based on the average 24-hour traffic volume of all facilities within 2.5 miles of the lot and the nearest peak hour expansion factor explained 86.7% of the variation in occupancy for the 11 lots in that district. When the models were tested on a dataset not used to build the models, the median ratio of mean testing error to mean occupancy was 56%. A typical model in this regard was for the lots in the Roanoke Valley-Alleghany Regional Commission (in the Salem District) where occupancy was based on the presence of transit service and the proportion of nearby residents with commutes of 25 to 50 miles: the mean testing error was 14 compared to a mean lot occupancy of 25, for a ratio of 56%. The models thus explained a portion of the variation in demand and can inform forecasts for new lots, although these results demonstrated that additional site-specific factors not included in each model also influenced demand.