A Transfer Learning-Based Approach to Estimating Missing Pairs of On/Off Ramp Flows

Each freeway stretch's traffic states are indispensable in freeway traffic modeling, surveillance, and control. However, the unmeasured ramp pairs always exist in real-world freeway systems, and how to estimate the flows of those ramps is a longstanding and tricky issue. Set the stretch with intact traffic states as Source Stretch while the stretch with the unmeasured ramp pair as Target Stretch; existing work tries to train the non-transfer machine learning model like Random Forest by Source Stretch and act on Target Stretch. However, the estimation accuracy of non-transfer machine learning models could not be guaranteed because the mainstream traffic state distributions of the above two stretches are not the same, and the model structure is too simple to capture traffic flow's temporal dependencies. Note the great success of transfer learning in distribution-changed situations; this paper addresses this issue via transfer learning and deep learning. First, the Gated Recurrent Unit-Based Ramp Flow Estimator is designed to establish the relationship between the mainstream traffic states and ramp flows in Source Stretch. Then, taking the trained estimator as the backbone, we propose the Deep Domain Adaptation to match the marginal distribution difference between Source Stretch and Target Stretch; design the Model Transfer to reduce the conditional distribution difference (i.e., estimator difference) between Source Stretch and Target Stretch. The two approaches both improve the performance of the Source Stretch's estimator in Target Stretch. Finally, we evaluate the processed approach in two real-world freeway traffic datasets and observe satisfactory results.