An extreme value based neural clustering approach for identifying traffic states

Traffic data are characterized by the occurrence of extreme events and frequent shifts to and from congestion. Current practice in traffic forecasting suppresses or disregards these features. But, indications suggest that these features may encompass useful information for modeling congested condition, as well as the transitions to and from congestion. This paper proposes a self-organizing approach to clustering traffic conditions based on information acquired from the 'peaking' behavior of traffic. Primary findings suggest that traffic has a strong transitional behavior that is reflected by frequent peaks detected in time series of volume and occupancy. The main finding is that traffic can be clustered into four distinct areas of traffic characteristics: (i) free-flow, (ii) medium flow states where traffic volume fluctuates in high values and occupancy is low, (iii) medium flow states where volume fluctuates in high values but occupancy increases sharply, and (iv) congestion. The main contribution of this approach is that it enables extracting a posteriori information from real-time traffic data as it pertains to boundary traffic conditions and it clusters traffic based on the occurrence of transitional movements.