Research On Combinational Forecast Models For The Traffic Flow Based On Intelligent Theory

With the development of social economy, the progress of science and technology, traffic travel demand of the people has been rising and the number of motor vehicles is also increasing. Traffic congestion, road safety, environmental pollution, energy consumption and other issues caused by the rapid growth of the number of vehicles are becoming more and more serious. A large number of practical experience at home and abroad shows that the advanced intelligent transportation system is the most effective means to improve urban traffic conditions and to enhance the public transport service quality, whether from the perspective of social economy or from the angle of ecological environment it is of great significance. Traffic flow forecast is the premise and key of intelligent transportation system, how to accurately predict the traffic information has become the focus of scholars. Traffic flow forecast is the premise and key to achieve intelligent transportation system and how to accurately predict the traffic information has become a hot research for scholars.Traffic flow has a certain randomness and dynamics because it is influenced by many random interference factors. Therefore, traffic flow forecast is very complex. Although many scholars have conducted in-depth research and achieved some results, it has not formed a relatively complete and mature theory system of traffic flow forecast. Which prediction model should be chosen in the end and which methods should be used to improve the model are still worth studying.In this dissertation, taking the intelligent prediction theory as the premise, it discussed in depth about effective combination between different intelligence methodsand models to improve the prediction accuracy. It mainly involves the following aspects of the research work.1. Two combination forecast models, DGM-GRNN and DGM-SVM, are proposed based on grey system theory DGM(1,1). Gray system is a prediction model in the case of small sample and poor information and GM(1,1) model and BP neural network are often used in existing combination model. Since GM(1,1)model does not consider the impact of possible future disturbances, it is not suitable for long-term forecast. DGM(1,1) improves conventional GM(1,1) in the discrete form to make up for its shortcomings. In addition, BP neural network is easy to fall into local minimum, and the calculation results are strong randomness, so this dissertation puts forward two kinds of DGM prediction model on the basis of residual error correction. First, DGM(1,1) model is used to predict original data sequence. Then, GRNN and SVM model are respectively trained on the tail residual and residual prediction sequence are obtained. At last, the final prediction results are synthesized. In the actual traffic flow forecast experiment, compared with the traditional GM(1,1) and DGM(1,1), the combination model proposed in this dissertation has been significantly improved the prediction accuracy, and it verifies the effectiveness of the combination model.2. The prediction model of support vector machine (SVM) is proposed based on grey system theory DGM(1,1). On the basis of statistical theory and structural risk minimization principle, SVM can classify and forecast in the small sample learning environment so it has been the focus of intelligent model. After thoroughly discussing the impact of different parameters on SVM prediction, PSO-LSSVM and PSO-SMOSVM prediction models are proposed respectively to overcome the major drawbacks of its parameters setting. Firstly PSO algorithm is used to optimize the penalty parameter C and the kernel width a and then the optimal C and a are determined. By means of them, SMOSVM and LSSVM begin to predict the traffic flow according to cross validation. Compared with through trial and grid search, the model proposed in this dissertation can get better performance indicating that it can effectively predict the trend of real-time traffic flow.3. A novel extreme learning machine (GA-ELM) is proposed based on genetic algorithm. The BP neural network consumes a lot of time in the process of adjusting the various training parameters, and extreme learning machine greatly shortens the training time at the same time does not decrease the convergence ability. Traditional ELM randomly sets the connection weights and thresholds between the input layer and the hidden layer, and no longer re-adjusts the values of these parameters in the training process so that the effectiveness of the hidden layer nodes needs to be improved. To solve this problem, in the conditions of determined network structure, the genetic algorithm is proposed for the selection process of weights and thresholds. GA-ELM wins the better weights and thresholds than the random assignment through the use of objective function. Which also makes the connection weight matrix between the hidden layer and the output layer is more reasonable. By the comparison of BP, GA-BP and traditional ELM, GA-ELM does better in the prediction accuracy and running time.4. A fixed weight prediction model on traffic flow and an uncertain weight one are proposed in this dissertation. At present, the statistical theory, nonlinear theory and intelligence theory have different applications in the traffic flow. Each single model has a certain advantage, but there is certain one-sidedness. In order to make full use of various information of each model, combination forecast models about a fixed weight coefficient and a variable weight coefficient are proposed based on GM(1,1), ARIMA and GRNN. After the proposed definition of quasi prediction absolute error, a fixed weight combination model and Elman variable weight combination one are established on the basis of absolute error percent. The experimental results show that the combination models especially variable weight in relative error, root mean square error and equal coefficient are better than the single model or the combination model of any two models, which also verify their effectiveness and feasibility.5. Comprehensive experiments are done for comparing a variety of combination forecasting models. Using the same traffic flow data for different combined forecasting models, the applicability of various forecasting models is analyzed and discussed.The innovations in this dissertation are mainly reflected by the traffic flow forecast model with intelligent combination. Respectively, GRNN and SVM are used to amend the residual error of DGM so as to further improve the prediction accuracy of the model; bionics particle swarm algorithm and genetic algorithm are applied to SVM and ELM. The model gets better prediction effect through the effective optimization of the parameters; based on GM, ARIMA and GRNN, fixed weight combination and Elman variable weight combination are proposed for better accuracy. Experimental results show that the combination models can obtain better performance in evaluation index. In conclusion, the research results of this dissertation have certain theoretical and practical value, and provide new ideas and new ways for more accurate traffic flow forecast.