Traffic flow prediction for the OHT system using graph convolutional network and gated recurrent unit

Abstract

As semiconductor fabrication facilities (fabs) become large-scale, the Automated Material Handling System (AMHS) of the fab, Overhead Hoist Transfer (OHT) system, become more complex. This makes it difficult to solve various control problems in fabs, so we want to solve the traffic flow prediction problem which is the key to solve these control problems. We first formulate the traffic flow prediction problem to the graph structure, because this problem has big size information to solve with past methods. Then, we design a graph processing method, Bi-GCN-GRU, to solve graph structure problem. Bi means bi-directional graphs that are normal direction and reverse direction. We use these two directional graphs for removing directional confusion. Graph Convolution Network (GCN) assumes spatial dependence to understand the graph structure of the OHT system, and Gated Recurrent Unit (GRU) assumes temporal dependence to extract temporal patterns in traffic flow. For validation our proposed method, we use data of OHT system simulator, Automod, and experiment with other methods. We can confirm our proposed method achieves high performance and discuss practical implications of our proposed method.