Location-specific demand forecasting for park and ride lots: The use of models in practice

Abstract

The management of hundreds of Virginia park and ride lots requires substantial public investment, and forecasting their demand helps agencies use scarce resources to improve multimodal mobility. Unfortunately, direct application of existing demand estimation approaches yielded errors typically 14 to 141 times the mean occupancy of a lot. Accordingly, this paper reports on the development of 19 regional models that yielded an average demand forecasting error of 0.56 times the mean occupancy.Because of Virginia’s geographic diversity, the most appropriate forecasting technique varies by region: in some cases, the key step is to group lots by commuting boundary or population density; in other cases, diverse socioeconomic variables (e.g., population, commuting distance, rent relative to income); facility-specific variables (e.g., transit availability, lighting); or traffic characteristics (e.g., volume on the road serving the lot, highest roadway volume within 2.5 miles) were needed to improve model performance. Ultimately, such variables explained on average 76% of the variation in lot demand. In practice, the results demonstrated the necessity of regional calibration: in rural southwest Virginia, for instance, the only key determinants of demand were the number of commuters living near a lot and having a job more than 50 miles away. In the more urban area that encompasses the state capital (Richmond), the two determinants were traffic volume and congestion levels. Although the models themselves are not transferrable, the methodology underlying their development can be used by other agencies as the data and techniques are publicly available.