A k-Nearest Neighbor Locally Weighted Regression Method for Short-Term Traffic Flow Forecasting

In this paper, a k-nearest neighbor locally weighted regression method (k-LWR) is proposed to forecast the short-term traffic flow. Inspired by k-nearest neighbor (k-NN) method, the traffic flows which have the same clock time with the current traffic flow are viewed as neighbors. The traffic flows which have the same clock time with the predicted traffic flow are viewed as the outputs of the neighbors. The neighbors most similar to the current traffic flow are viewed as nearest neighbors. It is observed that each nearest neighbor has different similarity with the current traffic flow, and the similarity is relevant to the contribution of the nearest neighbor's output to predicted traffic flow. The greater the similarity is, the greater the contribution is. These contributions of the nearest neighbors' outputs are obtained by the locally weighted regression (LWR) method. In this way, k-LWR uses less data, but uses it more effectively. We use the root mean square error (RMSE) between the actual traffic flow and the predicted traffic flow as the measurement. The proposed method is tested on the actual data from Xingye intersection and Feihu intersection in Jiangsu Province in China. The experimental results show that k-LWR has 20% and 24% improvement over the pattern recognition algorithm (PRA), 26% and 30% improvement over k-NN, for the two intersections, respectively.