Estimating the Location of Private Infrastructure for Delivery and Pick-Up Operations in Dense Urban Areas

The growth of home deliveries, lower inventory levels and just-in-time deliveries drive the fragmentation of freight flows, increased frequency, more delivery addresses and smaller volumes. This leads to trucks inefficiently loaded and consequently more trucks in the road contributing to the growing congestion in cities. According to a study by INRIX and the Texas Transportation Institute, travelers in the U.S. are stuck 42 hours per rush hour commuter in their cars in 2014, that is twice what it was in 1982 and the problem is four times worse than in 1982 for cities of 500,000 people or less [28]. At the same time, a historical lack of integration of the freight transportation system into city planning efforts has left local governments unprepared. Under these circumstances, there is growing need for best practices for freight planning and management in U.S. cities. There is anecdotal evidence that the lack of areas for trucks to park and load/unload freight is one of the main causes of an inefficient urban freight parking infrastructure that leads to illegal parking and more congestion. The problem of lack of parking for freight loading/unloading has been studied with a focus on on-street parking. Meanwhile, the contribution of areas out of the public right of way (i.e. private) such as loading bays in buildings has not benefited from research. More importantly, the location and features of private freight parking are often unknown by local governments due to their private character.

This thesis presents the first predictive tool to estimate the presence of private freight loading/unloading infrastructure based on observable characteristics of property parcels and their buildings. The predictive model classifies parcels with and without these infrastructures using random forest, a supervised machine learning algorithm. The model was developed based on a rich geodatabase of private truck load/unload spaces in the City of Seattle and the King County tax parcel database. The performance of the random forest model was evaluated through cross-validated estimates of the test error. The distribution of the outcome variables is unbalance with over 90% of parcels without private freight infrastructure. To consider the problem of unbalance sample, the optimum model was set to maximize the area under the ROC curve (AUC). The authors investigated the confusion matrix and the model classifier was design to balance the sensitivity and specificity of the model. Model results showed AUC of 81.5%, a true positive rate of 82.1% and a misclassification error of 22.5%.

This research provides an assessment tool that reduces the field work required to develop a quality inventory of private freight loading/unloading infrastructure by targeting the parcel stock and making data collection methods more effective. Local governments can use this research to inform efforts to revise and update delivery operations and regulations of truck parking and loading.