K-nearest neighbor model of short-term traffic flow forecast

In order to accurately forecast the short-term traffic flow, a K-nearest neighbor (K-NN) model was set up. The time and space parameters of the K-NN model were analyzed. Based on four different combinations of state vectors, the time dimension model, upstream section-time dimension model, downstream section-time dimension model and space-time dimension model were proposed. The four different models were validated by using the GPS data from taxis of Guiyang. Analysis result indicates that the K-NN model with both space and time parameters has highest forecasting precision than the other three models, and its average prediction error is about 7.26%. The distance measuring mode with exponent weight has higher accuracy in choosing the nearest neighbors, and its average prediction error is about 5.57%. The predicting performance of improved K-NN model with exponent weight and space-time parameters is best compared with the artificial neural network model and the historical average model, and its average prediction error is only 9.43%. So the improved K-NN model is an effective way for forecasting short-term traffic flow.