ESTIMATING RAMP TRAFFIC STATES IN SENSOR-SPARSE FREEWAY ENVIRONMENTS USING TRANSFER LEARNING
The lack of accurate readings on on-ramps and off-ramps, where sensors are frequently missing due to installation, maintenance, and financial constraints, is a key issue in freeway traffic condition estimation. The accuracy of traffic monitoring, modelling, and control applications is thus limited, especially in complex highway networks, by the associated lack of ramp flow and speed data. This study proposes a transfer-learning-based method for estimating on- and off-ramp traffic conditions using only mainline readings to overcome this difficulty. Specifically, we design a Bidirectional Long Short-Term Memory (Bi-LSTM) neural network that uses data-rich freeway segments with both mainline and ramp measurements to learn spatiotemporal traffic patterns. Seven days' worth of traffic data from seven Dutch freeway segments representing various traffic situations, geometric patterns, and ramp configurations are used to train the model. The trained network is then applied to sensor-limited freeway stretches with missing ramp data by utilizing transfer learning, which makes it possible to infer ramp traffic volume and speed without the need for additional physical equipment. The findings show that under different demand levels and congestion regimes, the proposed technique accurately reconstructs ramp traffic conditions. This suggests that the learned temporal representations can successfully substitute for missing ramp measurements and generalize well across various highway conditions. The proposed system supports improved traffic management and control in real-world applications by providing a practical and affordable way to improve traffic state estimation in freeway networks with sparse sensing equipment.