車両軌跡情報を用いたニューラル常微分方程式に基づく高速道路の所要時間予測

抄録

Trip time prediction (TTP) have great importance in traffic analysys and management. Existing research related to traffic forecasting, including traffic speed, enable TTP. However, these studies mainly use aggregated data such as vehicle detectors which cause intergration error. Trip time is realised by the integration on predicted speed, but the accuracy of the actual trip time, which can be gathered by vehicle trajectory data, is not guaranteed. In this research, we propose the extension of Neural ODE which can minimise integration error. The proposed method can learn velocity field from ETC2.0 probe data that is a type of vehicle trajectory data. The experiment result using artificial dataset and large scale dataset shows superiority and learning stability of proposed method.