Research Of Urban Traffic Flow Forecasting Based On Hybrid Intelligent Computing

In recent years, along with the development of Intelligent Transportation System (ITS), the intelligent transportation control and inducement system has become the popular problem of ITS. The key question of traffic flow inducement is the real-time accurate traffic flow forecasting, namely how to forecast the transportation condition several minutes later by the use of real-time transportation date message. With forecast time span reducing, its non-linearity, time-dependent nature and uncertainty become more and more strong. The orthodox forecasting models such as History Average Model, Time-Series Model, Nonparametric Regressive Model, Kalman Filtering Model, Neural Network Model and Combination Forecasting Model can not predict well in effect and precision,new model is need to construct to predict well and precise.Typical main multi-intersection of urban road is researched in this paper, according to the property of urban traffic flow we researched, we combines the two kinds of basic methods of intelligent computing, artificial neural network and evolutionary algorithms, a traffic flow forecasting model based on hybrid intelligent computing is constructed. On the base of detailed analyzing neural network, we propose single neural network forecasting model and carry on the optimization to it using the optimization algorithms in view of its structural features, and then complete the assignment of modeling forecasting model based on hybrid intelligent computing among which the traffic flow forecasting model of Elman neural network based on dissimilation particle swarm optimization has the best performance. Based on the analysis of the standard particle swarm optimization and the characteristic of typical main multi-intersection of urban road, the traffic flow forecasting model using dynamic recursion neural network is presented. The feature of the network is that the output of hidden layer connects to the input of itself by the delay and storage of the context layer which is called self-connect method. It is sensitive to the historical data. The added internal feedback layer enhances the network's capacity of dealing with dynamic information. Thus the purpose of dynamic modeling is achieved. The model can reflect the characteristic of the dynamic system livelier and more direct. Meanwhile the initial inputs of the context unit and all the weights of the model are optimized by dissimilation particle swarm optimization. The algorithm mutates the local optimum based on the judging of local optimal value or global optimal value, and helps it jump out local convergence, so as to overcome the shortcoming which the founded value are local optimal value and fall into the problem of prematurely of traditional particle swarm optimization. The sample data for simulation in the study are obtained from a road segment of Jingshi road, which is an arterial road of Jinan city. Segment of multi-intersection is taken as the researched object in this paper. Database contained traffic volume, speed and time-copy which are collected every 5minites is set up. We simulate the constructed model by programming on Matlab. We contrast the hybrid intelligent computing model to the single neural network model, hybrid neural network model based on genetic algorithm and hybrid neural network model based on particle swarm optimization. The simulation results show that the training speed and accuracy of the model using history data are improved consumedly. The algorithm is simple and the model can be applied to engineering practice as a kind of short-term traffic flow forecasting model.The combination of recurrent network and dissimilation particle swarm optimization is the main innovation point in this paper and a forecasting model based on hybrid intelligent computing is presented. The model has simple structure and is easier to realized especially being suitable for the forecasting of the short-term traffic flows. Compared with the feed-forward neural network the model has more capacity of dealing with dynamic information. Thus the purpose of dynamic modeling is achieved. Dissimilation particle swarm optimization arithmetic is used to optimize the network parameters. Compared with genetic algorithm it is easy to implement and rapid in training, and also solve the problem of prematurity.