Daily collision prediction by SARIMAX and GLM models based on temporal and weather Variables

摘要

Short-term collision prediction is a relatively new area of research in the field of traffic safetydue to the high randomness of data and the methodological complexity. Motivated byrequirements from frontline traffic operations and enforcement services, this study wasconducted to develop models to predict daily total collisions. The study started with a time seriesdata decomposition analysis to determine trends, seasonality, and randomness of the dailycollisions before proceeding with an investigation of potential collision contributors. Temporalfactors (i.e., months, weekdays and holidays) and weather forecasts (i.e., daily mean temperature,amount of rainfall and amount of snowfall) were selected as predictive factors. Accordingly, theSeasonal Autoregressive Integrated Moving Average model with External Regressors(SARIMAX) was identified and a series of SARIMAX models with different orders wasestimated and diagnosed. A Generalized Linear Model (GLM) was also developed and comparedto the SARIMAX models by validation measures. Finally, a calibration mechanism wasrecommended to optimize predictions. Model validations provided evidence that bothSARIMAX and GLM models are adaptable; however, the SARIMAX models are a viable andpreferable option as they can gain better accuracy than GLM in terms of short-term collisionprediction. In practice, the models developed in this paper are now being applied to supportscheduling of traffic operations, maintenance and enforcement, dispatch of material andpersonnel resources, and to also provide situation awareness for all road users and stakeholders.