Research On Neural Network Algorithms For Short-term Traffic Flow Forecasting

The Intelligent Transportation System(ITS) alleviates traffic congestion by the control and guidance of traffic flow, and short-term traffic flow forecasting provides effective data support for such traffic flow control and guidance. At present, the Artificial Neural Networks(ANNs) has been widely used in short-term traffic flow forecasting, because of its superior capability in nonlinear mapping and self-learning, but the forecasting accuracy of the existing algorithms is usually hard to meet the demand of the ITS. This thesis focuses on the application of BP Neural Networks(BPNNs) and Convolutional Neural Networks(CNNs) algorithms in short-term traffic flow forecasting, and working to improve the accuracy of traffic flow forecasting and minimize the calculation. The main achievements of this thesis are summarized as follows:1) A short-term traffic flow forecasting algorithm is proposed based on the combination model of Fuzzy C-Means(FCM) clustering and BPNNs. The algorithm can take the precise division of traffic flow pattern by clustering, builds a corresponding forecasting model of BPNN for each pattern, and figures out the weighted sum of forecasting value of each forecasting model, as the forecasting results. The experiments show that the algorithm can take the reasonable division of traffic flow pattern, and its forecasting accuracy is higher than the traditional BPNN algorithm and the existing algorithm based on combination forecasting model, so it has a certain practicality.2) An improved algorithm is proposed for the above algorithm based on the combination forecasting model. The improved algorithm uses the Taguchi experimental design method to experimentally design, obtains the optimal values of model structure parameters through analyzing the experiment results, and uses the optimal parameter values to perform the above combination model-based algorithm to forecast short-term traffic flow. The experiments show that the forecasting accuracy of the improved algorithm is further enhanced, and the calculation amount of the parameter optimization process is small, and in some extent, the improved algorithm can solve the problem of large calculation when using Genetic Algorithm(GA) to optimize the parameters.3) A short-term traffic flow forecasting algorithm is proposed based on CNNs. The algorithm applies CNNs to traffic flow forecasting, comprehensively considers the forecasting road section with its upstream and downstream sections, makes the input data extended to be two-dimensional to meet the input format of CNNs, and designs the structure of the CNN model according to the feature of input data. The experiments show that the forecasting accuracy of the algorithm is higher than the BPNN algorithm based on the single forecasting model, and better than the existing algorithm based on the combination forecasting model in some extent.