An Improved K-Nearest Neighbor Model for Road Speed Forecast Based on Spatiotemporal Correlation

Most k-nearest neighbor models only focus on single step short-term traffic forecasting and can't perform well when extended to multi-step forecast. To enhance accuracy, this paper presents an improved k-nearest neighbor model considering the spatiotemporal correlation. The proposed model defines the current conditions by the two-dimensional spatiotemporal state matrices, instead of the one-dimensional state vector of the time series. Moreover, this paper determines the weights by Gaussian function to adjust the matching distance and manage the data of the nearest neighbors. The original speed data used in this paper are normalized in order to apply the proposed model to different types of road segments. The case shows the improved model performs more desirable than the original k-nearest neighbor models and demonstrates more appropriate for multi-step forecast of the road speed.