Abstract
Traffic control systems, like elevator group control systems, have the aim of improving their performance within limits of their resources, and they have to take specific control actions under particular traffic conditions. The origin-destination (OD) matrix is usually employed to represent traffic flows. While we have practical means to measure traffic counts, we cannot obtain the OD matrix directly from accumulated counts. Therefore, conventional elevator group control systems detect typical traffic patterns instead of estimating the OD matrix.
In this paper, we have proposed a detection system of typical elevator traffic patterns with neural networks. Inputs of the system are traffic counts, and outputs are feature modes of traffic. The advantage of our system is easy to apply any buildings without specific adjustments. Furthermore our system has robustness for stochastic fluctuations of input data. We have made some simulations and experiments using real traffic counts, and their results show the successful performance of learning and detecting elevator traffic patterns.
