Short-term traffic forecasting has long been a hot research topic in the field of intelligent transportation systems. The traditional traffic forecasting methods used mostly fixed traffic sensors. However, most of the sensors on the road networks are subject to bad conditions, which leads to noisy and insufficient raw data. Recent advances and applications of machine learning methods have gained great popularity in the transportation field and provide new opportunities for traffic prediction. For example, the transfer learning method can take advantage of data and models trained on one good dataset and transfer the knowledge to the other with some bad sensor data. Nonetheless, existing applications of transfer learning in short-term traffic forecasting do not consider the underlying data distributions sufficiently, which limits the prediction performance. This study aims to overcome these weaknesses, and we propose a transfer learning-based traffic flow prediction framework using Balanced Domain Adaptation (BDA) method, which jointly considers the marginal and conditional distributions by assigning weights to them. Various regression models are fed into the framework to evaluate if the proposed method can utilize the good source of data and to make predictions upon bad target dataset. A case study using data from the Highways England is conducted. The results show that the proposed BDA-based framework can match the distributions between traffic flow datasets and therefore significantly improve the prediction accuracy.
 

Machine learning techniques;Traffic flow prediction;Domain adaptation;Transfer learning