. Traffic flow forecasting is a popular research topic in the development of Intelligent Transportation System. There have been many forecasting methods used for traffic flow forecasting, such as Generalized Regression Neural Network (GRNN) which has a fairly good accuracy. One of the GRNN's characteristics is that the number of neurons in pattern layer increases as the number of training samples raise and this can cause overfitting problem. In this research, a hybrid method to predict traffic flow is proposed, that is K-means and GRNN algorithm. K-means method aims to solve overfitting problem in GRNN model by choosing training samples based on their similar characteristics. Leave One Out Cross Validation (LOOCV) is used to select an appropriate smoothing factor parameter at each GRNN's model. Mean Absolute Percentage Error (MAPE) is used as the evaluation criterion in the testing process. The results show that the proposed method could improve the accuracy of predictions by reducing the value of MAPE by 0.82-3.81%.