Hybrid Model-Based and Memory-Based Traffic Prediction System

Abstract

Short-term traffic forecasting capabilities on freeways and major arterials have received special attention in the past decade primarily because of their vital role in supporting various traveler trip decisions and traffic management functions. A hybrid model-based and memory-based methodology used to improve freeway traffic prediction performance is presented. The proposed methodology integrates both approaches to strengthen predictions under both recurrent and nonrecurrent conditions. The model-based approach relies on a combination of static and dynamic neural network architectures to achieve optimal prediction performance under various input and traffic condition settings. Concurrently, the memory-based component is derived from the data archival system that encodes the commuters' travel experience in the past. The outcomes of the two approaches are two prediction values for each query case. The two values are subsequently processed by a prediction query manager, which ultimately produces one final prediction value by using an error-based decision algorithm. It was found that the hybrid approach produced speed estimates with smaller errors than those produced when the two approaches were used separately. The proposed prediction approach could be used to derive travel times more reliably for advanced traveler information systems applications.