Çok değişkenli ve bulanık yaklaşımlarla trafik akımının dinamik sınıflandırılması

In this thesis, performances of multivariate and fuzzy clustering methods in specifying flow state variations reconstructed by a macroscopic flow model are sought. In order to remove the noise in and the wide scatter of traffic data, raw flow measures are filtered prior to modeling process. Traffic flow is simulated by the cell transmission model adopting a two phase fundamental diagram. Flow dynamics specific to the selected freeway test stretch are used to determine prevailing traffic conditions. The classification of flow states over the fundamental diagram are sought utilizing the methods of multivariate and fuzzy cluster analyses by considering the stretch density. The fundamental diagram of speed-density is plotted to specify the current corresponding flow state. Both multivariate and fuzzy clustering analyses returned promising results on state classification which in turn helps to capture sudden changes on test stretch flow states. The perfomance analyses are focused explicitly on the application of clustering for partitioning states of traffic flow, using three clustering algorithms: K-means clustering by Square Euclidean Distance, K-means clustering by City Block Distance and fuzzy C-means. Clustering algorithms have the flexibility to specify the number of clusters. Categorization has been based on similarities and dissimilarities of traffic flow variables without specifying arbitrary values to bound states. Clustering methods are processed in a time-varying fashion to partition the fundamental diagrams at selected temporal resolutions. In order to comparatively evaluate the clustering performances of multivariate and fuzzy methods on lane-based densities relative to deterministic clustering, a number of statistical criteria, including the root mean squared error, the mean square error , the mean absolute error , the mean absolute percentage error and the coefficient of determination. The performances of K- means clustering and fuzzy c- means clustering, using both the Square Euclidean and City Block Distance measures, are evaluated in two cases, with two assumptions.The procedures followed by multivariate and fuzzy clustering methods are systematically dynamic that enables the partitions over the fundamental diagram match approximately with the flow states derived by the static partitioning method. It is shown that the comprasions presented that the K-means clustering by Square Euclidean Distance and fuzzy c-means methods perform better and appear to yield to classifications consistent with two types of level of service, which are calculated by using HCM-defined level of service.