Bayesian combined neural network for traffic volume short-term forecasting at adjacent intersections

Abstract

Undoubtedly, one of the greatest issues nowadays is congestion. To face such problem, forecasting of traffic is required. Bayesian combined neural network (BCNN) is applied to four different locations in Kuwait (Cairo Street, Riyadh Street, Maghreb Road and Istiqlal Road) to predict the short-term traffic volume at the middle section due to traffic flow from adjacent intersections. All data were collected for a period of 1 week over 15-min observation intervals using loop detectors. In addition to time-series responses and regression plots, mean square error (MSE) has been used to validate the network performance after data normalization. In comparison with MSE and R values, both values were slightly less precise during weekdays compared to weekends. After standardizing, the average MSE during weekdays was 0.003468 and regression (R) was 0.98113 for the four streets. For weekends model, the average MSE was 0.003563 and regression (R) was 0.97374 for the four streets. Istiqlal Street weekday model was the best model that fits the information among all the four models; as it has the smallest MSE value equivalent to 0.0010087 and the highest R value of 0.9959. BCNN model has achieved outstanding prediction performance with great potential to be generalized for various locations at different times of the day. These results can allow transportation planners to forecast traffic congestions and take prior measures to avoid them. Further modeling can assist in studying factors causing intersection congestions.