Developing an artificial neural network for estimating road capacity values for weaving sections

Static macroscopic transport models are used for long-term projections on road networks and deliver travel times and congestion on road sections. Nowadays, the capacity of highway road sections within a transport model is usually set as a fixed parameter. In case, the (fixed) parameter of the capacity of a certain road section does not correspond with the actual capacity of the road section, implausible computations of traffic conditions on certain road sections are made, especially when not the traditional capacity restrained, but strict capacity constraint assignment procedures are used. This has also an impact on traffic states on adjacent road sections and as a results of route choices and possible other feedback loops to mode, destination and even trip production on the entire road network. Due to their ‘node’ function within a road network, weaving sections are determinative for network operations. In many cases, a fixed capacity based on some heuristic is used on weaving sections, which the authors prove to be inaccurate in several cases in practice. Furthermore, in reality, the capacity of weaving sections is not fixed (i.e. does not only depend of design), but also depends of the distribution of weaving traffic flows. Therefore, more accurate capacity values estimations for road (and especially weaving) sections as part of the assignment procedure is an important issue for improving accuracy of transport models. In this research an artificial neural network that estimates capacity values of weaving sections is developed. A neural network is chosen to be able to quickly estimate capacity values for weaving sections (because calculation times of microsimulations are too high), deliver accurate capacity values and because it is an improvement over the currently applied heuristic capacity estimation methods. The development of the neural network is an iterative process, composed by the pre-training phase, training phase and post-training phase.