QUASI DYNAMIC O-D MATRIX ESTIMATION: PERFORMANCE ANALYSIS ON REAL DATA

The paper proposes the first results of a research project focused on the validation of the quasi-dynamic o-d matrix estimation framework, that is considering a number of sub-periods within the whole time horizon of estimation in which distribution shares across all time slices within each sub-period are assumed constant. Firstly, the paper checks whether real data confirm such assumption from an empirical perspective, and then compares the performances of the quasi-dynamic o-d matrix correction with both classical off-line dynamic estimators and with other possible evolution rules characteristics of on-line dynamic estimation. Experiments are carried out on the real test site of the motorways A4-A23 in North-East Italy. Firstly, the analysis of the true o-d matrices evidenced that the assumption of quasi-dynamic o-d matrix pattern is quite realistic, even under the hypothesis of constant distribution shares for the whole day. Then, estimations have been performed using the simultaneous estimator, the Kalman filter approach and the quasi-dynamic approach. The main finding is that the quasi-dynamic estimates outperforms the simultaneous estimates, even if a remarkable heterogeneity of results across days has been observed. Furthermore, the impact of the reliability of prior estimates on the performances of the Kalman filter approach have been tested as well. Interestingly, using the quasi-dynamic estimates as historical matrices leads to satisfactory performances for the Kalman filter, significantly better than the implementation based on the historical simultaneous estimates. Finally, the paper describes the experiments to be still carried out and points out future research directions.