Research Of Generalized Neural Network And Its Application In Traffic Flow Forecasting

As an important aspect of Intelligent Transportation Systems (ITS), traffic flow guidance is considered as an optimum way to improve traffic efficiency and mobility. The essential of the Traffic Flow Guidance Systems (TFGS) are supplying real-time exact traffic information. Traffic flow is important information in urban traffic, so traffic flow forecasting has important significance. There are many factors that can influence the traffic flow, all of these results in the difficulties of real-time traffic flow forecasting.Owing to the good adaptability, neural network has become a common model for information forecasting. Based on traditional neural network, this paper presents a intelligent neuron model, which is composed of linearly independent functions and Sigmoid function with adjustable parameters. It is proved that the information storage ability of this intelligent neuron is greatly improved compared with traditional ones, consequently greatly improves the information processing ability of the whole neural network. Meanwhile, in order to reduce the size of the neural network's input, this paper uses the correlation theory to analyze the correlation between neighbor road sections, and choose the traffic flow of different road sections, which has strong correlation with the being forecasting one as neural network's inputs, and establish the traffic flow model based on generalized neural network. Experiment results show that, the generalized neural network converges faster than traditional BP neural network, and meet practical requirements well.In order to greatly improved the converge speed of generalize neural network, this paper designs a parallel training algorithm, which is based on training set decomposition. This parallel training algorithm uses a new communication profile. This profile greatly reduces the communication cost of the parallel algorithm. Experiment results show that, this parallel training algorithm is effective for reducing the training time of generalized neural network.