Deep Learning–Based Speed Reduction Modeling for Independent Tangent–Curve Transitions on Four-Lane Divided Rural Highways

Abstract Speed reduction along successive highway geometric features, such as tangent-to-curve (T-C) transitions, is a key factor in assessing the geometric design consistency of rural highways. Design consistency is a surrogate safety strategy for rural highways. Speed reduction is a widely accepted measure for evaluating design consistency at T-C transitions, especially when the tangent is independent (i.e., an isolated curve). Therefore, it is essential to identify independent tangents and predict the speed reduction at T-C transitions. The overarching goal of this study is to develop a speed reduction model for independent T-C transitions on four-lane divided rural highways using instrumented vehicle data. As such, the study determined the criterion for identifying independent tangents and developed an 85th percentile individual drivers’ speed reduction model along the independent T-C transitions. For this analysis, the threshold tangent length was determined to identify the dependent (short) and independent (long) tangents using acceleration and deceleration values. Following that, a total of 32 independent T-C transition sets in flat terrain were used to model speed reduction using multiple linear regression (MLR) and deep neural network (DNN) techniques. Curve radius and deflection angle were found to significantly influence the extent of speed reduction. Overall, the DNN model ( R 2 = 0.957 ) outperformed the MLR model ( R 2 = 0.563 ) in predicting speed reduction on T-C transitions of four-lane divided rural highways. This study provides guidance on controlling geometric elements and identifying critical T-C transitions to establish appropriate posted speed limits for achieving the desired speed reduction.