Neural Networks to Estimate Crashes at Zonal Level for Transportation Planning



Neural Networks to Estimate Crashes at Zonal Level for Transportation Planning

Authors

V R Duddu, S S Pulugurtha, The University of North Carolina at Charlotte, US

Description

The focus of this research is to develop crash estimation models at traffic analysis zone (TAZ) level as a function of land use characteristics.

Abstract

The focus of this research is to develop crash estimation models at traffic analysis zone (TAZ) level as a function of land use characteristics. Crash data and land use data for 765 TAZs with a total of 9,799 crashes (includes 20 fatal crashes, 3,227 injury crashes and 6,552 property damage only crashes) in the City of Charlotte, Mecklenburg County, North Carolina were used in the development of statistical and neural network models. Negative binomial models (with log-link) were developed as data was observed to be over-dispersed while the neural network models developed are based on a multilayered, feed-forward, back-propagation design for supervised learning. Demographic / socio-economic characteristics (population, the number of household units and employment), traffic indicators (trip productions and attractions), and, on-network characteristics (center-lane miles by speed limit) were observed to be correlated to land use characteristics, and, hence were not considered in the development of TAZ level crash estimation models.

Models were developed to estimate the total number of crashes as well as crashes based on severity (total injury and property damage only crashes). Results obtained show that mixed use development area, urban residential area, urban residential commercial area, single-family residential area, multi-family residential area, business and office district area were observed to play a statistically significant role in estimating crashes in a TAZ. The results from comparison of models for performance evaluation indicate that back propagation neural network models ensured significantly lower errors when compared to statistical models. The neural network approach can be particularly suitable for their better predictive capability, whereas the statistical models could be used for mathematical formulation or understanding the role of explanatory variables in estimating crashes.

The developed methodology and results can be used to incorporate safety into long range transportation plans and land use decisions so as to minimize anticipated crashes in the future. The models developed using the methodology can also be used to examine the effect of changes in land use characteristics (new development or re-zoning) on safety. The methodology, findings and recommendations from the research will be presented.

Publisher

Association for European Transport