Context-Driven Adaptive Representation Learning For Intelligent Decision-Making In Smart City Traffic Management Applications

Abstract

Intelligent transportation systems today are increasingly using graph neural networks in order to capture spatial dependencies in road networks and predict the state of traffic flow in a real-time manner. Most of the existing spatio-temporal graph learning models use a static representation of the graph structure, which fails in situations where the traffic dynamics are affected by contextual factors such as weather conditions, incidents, or any special events, where the historical spatial dependencies do not hold true anymore. This paper introduces a Contextual Adaptive Representation Learning (CDARL) framework, which allows for an adaptive representation conditioned on the current context in terms of weather, calendar, incidents, and events. The context-dependent spatial encoder is coupled with a temporal attention block to create an adaptive representation that is used by a downstream decision layer for signal-timing recommendations on the fly. We evaluate the effectiveness of this framework on a simulation-based 50-sensor urban arterial network with statistical properties based on the PeMS and METR-LA datasets, using four distinct scenarios—normal, rainy, incident, and special-event—to show how the proposed framework adapts. Compared to the standard LSTM and a context-independent static graph neural network model, our framework outperforms both methods in speed-forecasting performance, with an average absolute error of 7.75 km/h and 6.10 km/h reduced to 4.08 km/h; we see the largest improvement in speed forecasting during incident and special event scenarios. In the context of signal timing, our representation-based approach results in a 34.1 percent reduction in intersection delay compared to 12.4 percent and 21.3 percent reductions for the baseline models.