Short-term traffic prediction with deep neural networks and adaptive transfer learning

Abstract

A key problem in short-term traffic prediction is the prevailing data missing scenarios across the entire traffic network. To address this challenge, a transfer learning framework is currently used in the literature, which could improve the prediction accuracy on the target link that suffers severe data missing problems by using information from source links with sufficient historical data. However, one of the limitations in these transfer-learning based models is their high dependency on the consistency between datasets and the complex data selection process, which brings heavy computation burden and human efforts. In this paper, we propose an adaptive transfer learning method in short-term traffic flow prediction model to alleviate the complex data selection process. Specifically, a self-adaptive neural network with a novel domain adaptation loss is developed. The domain adaptation loss is able to calculate the distance between the source data and the corresponding target data in each training batch, which can help the network to adaptively filter inconsistent source data and learn target link related information in each training batch. The Maximum Mean Discrepancy (MMD) measurement, which has been fully validated and applied in transfer learning research, is used in combination with the Gaussian kernel to measure the distance between datasets in each training batch. A series of experiments are designed and conducted using 15-minute interval traffic flow data from the Highways England, UK. The results have demonstrated that the proposed adaptive transfer learning method is less affected by the inconsistency between datasets and provides more accurate short-term traffic flow prediction.