关键词: 交通运输系统工程, 交通流短时预测, 样本分类拟合, KSOM-BP神经网络, 动量-自适应学习速率

Abstract:

A short-term traffic flow prediction model was built based on KSOM-BP neural network. Using the kernel sample self-organizing map(KSOM) neural network, under the condition without a priori knowledge, the history samples with similar statistic character were classified into several categories by self-organizing and self-learning to enhance the statistic significance of the classified sample. A back-propagation(BP) neural network prediction model was built for the momentum-adaptive learning rate of every category sample to enhance the short-term traffic flow prediction accuracy, reduce the prediction time consumption. The model was validated using the practical road network data. The results show that the statistic index MARE of prediction error of KSOM-BP neural network is less than 7%, and less by 4% than that of BP network based on the whole sample without classification. The modeling time consumption of KSOM-BP neural network is also less than BP neural network without sample classification. All of this proves the validation and state-of-the-art of the proposed method.

Key words: engineering of communications and transportation system, traffic flow short-term prediction, sample classification fitting, KSOM-BP neural network, momentum coefficient and adaptive learning rate