Improved k-nn for short-term traffic forecasting using temporal and spatial information

This paper proposes an improved k-nearest neighbor (k-nn) model for short-term traffic forecasting and examines its applicability to forecasting for different links and time periods. The traditional k-nn model is adapted by formulating the weighted distance metric and the state vector, which consider both the temporal and spatial information. The adapted model's performance is examined in a numerical test where the data are derived from global positioning system (GPS) devices in 180 taxis running in Guiyang, China. The test results demonstrate that the model that considers both the temporal and spatial information outperforms models that only consider temporal information and that adaptation of distance metrics could significantly improve the forecasting accuracy. The adapted model shows the promising performance in comparison with the historical average (HA) model and the artificial neural network (ANN) model. The test results also indicate that information from the upstream and the downstream links plays almost the same important role in predicting the traffic conditions at the target link.