A recursive traffic flow predictor based on dynamic generalized linear model framework

Previous approaches used to perform online flow predictions included time-series models, nonparametric regression, simple moving average, and adaptive filtering. This paper explores the statistical nature of traffic flows aggregated at short time interval and investigate the potential of using the dynamic generalized linear model (DGLM) for traffic flow predictions. Specifically, recursive algorithms for weighted least square (WLS) estimators resulted from the generalized linear model family, including Poisson, binomial, and negative binomial distributions were derived based on the quasi-likelihood and prediction error minimization principles. The Kalman filter interpretation of the recursive WLS is also described. The resulted formulations are distinct from the existing DGLM and suitable for multivariate case implementation.