A Gaussian maximum likelihood formulation for short-term forecasting of traffic flow

Traffic counts are key data generated by traffic surveillance systems. In predicting traffic flows, it is commonplace to assume that traffic at a given location repeats itself from day to day and the change in traffic happens gradually rather than abruptly. Consequently, many existing models for short-term traffic flow forecasting use historical traffic information, real-time traffic counts, or both. This paper proposes a new model based on the Gaussian maximum likelihood method, which explicitly makes use of both historical information: and real-time information in an integrated way. The model considers flows and flow increments jointly and treats them as two random variables represented by two normal distribution functions. Each assumption made in the model is verified against the field data. The physical structure of the model is easy to interpret. Computationally, the model is simple to implement and little effort is required for model calibration. The performance of the proposed model is compared with four other models using field data. The proposed model consistently yields predictions with the smallest absolute deviance and the smallest mean square error.