Abstract: Short-term traffic flow prediction is an important part of traffic flow modeling,and it also plays an important role in urban road traffic management and control.However,the common time series model (e.g.,ARIMA) and random forest model (RF) are limited in the prediction accuracy due to the residuals generated by the model and the input variables.Aiming at this problem,a short-term traffic flow prediction model based on gradient boosting regression tree(GBRT) was proposed to predict the travel speed.The model (GBRT) first introduces the Huber loss functionto deal with residuals.Secondly,the spatial-temporal correlations are also considered in the input variables.The model adjusts the weight of the weak learners in the training process,and corrects the residuals of the model to improve the prediction accuracy.Experiment was conducted by using traffic speed data of a city expressway,and ARIMA model and random forest modle were compared with the proposed model by using MSE,MAPE and other indicators.Results show that the proposed model has the best prediction accuracy,and the validity of the model in short-term traffic flow prediction is verified.

Key words: Gradient boosting regression tree, Loss function, Short-term traffic flow prediction, Spatial-temporal corre-lations