Application And Research Of Neural Network In Traffic Prediction

With the development of global economy recent years, communication techniques have been improved dramatically, and telecom industry advances rapidly especially since 90s. Telecom industry in our country will keep the development trend. Size of network expands, new businesses emerge, new technology spreads widely, and service quality improves. Customer service centers as low-cost and convenient-service channels are welcomed by customers and paid increasing attention by operation business. So the centers become the key point in communication networks. Schedules are made manually in the past because of few workers of service centers. While the number of telephonists in large-scale centers is over thousands with the development of communication networks. So the traditional manually way can not deal with problems, and an automatic system is needed by which schedules are made automatically and modified by experts. Telephone traffic should be predicted before schedules to make schedules reasonable. Therefore, it becomes a requirement to design a traffic predicting system for users of customer service centers according with demands of the center.A brief introduce to schedule-making and traffic system is made, the system architectures and function modules are described, and basic knowledge on predict are explained in this paper. Two methods to predict are focused on: time series analysis method and artificial neural network method. A new method combined with neural network and genetic algorithm is proposed to predict phone traffic. The following six models are introduced in detail: exponential smoothing model, linear random model, neural network BP model, neural network RBF model, neural network random model and model combined with genetic algorithm and BP model. The applications with the above models to predict phone traffic are expatiated on, and effects of predictions are illuminated by figures. A phone traffic predict system for managers of telecom customer service center is designed with a friendly interface and high-credible practicability. The system is supported background by Sql Server 2000 containing the real traffic data of 2005. The main modules interacting with users are modules of traffic predicting, predict results searching, and affair processing. All the three modules interact with database real-time. The traffic predict module is a whole process of predicting containing selections of predict model and pace. The result searching module presents a friendly interface with crystal reports.This paper compares similarities and differences of the models, points out their advantages and disadvantages by predict results in real circumstances, analyze the characters of the results effect, and make a summarize. Exponential smoothing model can obtain predict results efficiently with low accuracy, and secondary exponential smoothing is needed to predict traffic with the property of bimodal curve. So it is capable for short-term predict. Linear random model has complex matching rules asking for a number of historic data as matching basis, and its predict accuracy is not satisfied, only can be used for short-term predict. Artificial neural network model is self-adaptive and can be trained, has ability of self-modify and parallel processing and inferring information. Method of traffic predict based on artificial neural network can be fitted well after repeating training and selecting until the network structure is suitable. However, there is no theory guiding the development of network model, and design of network structure will cost much time. The most widely applied among artificial neural network models is the BP model. It uses gradient method (steepest descent method or conjugate gradient method) to adjust weight by numerical iteration to make the mean square deviation of output and expected of multi-layers of neural network minimum. While with increase of complexity of neural network and size of training data, BP method converges slowly, always fall to local minimum points, and the numerical stability is bad, the parameters is difficult to adjust. RBF model can determine network structure while training networks, which avoids local optimization relatively, but the process of determining takes much time. The method improved by this paper decreases time of training networks, avoids local optimization without lowering accuracy of prediction. The model proposed by this paper combined with neural network and genetic algorithm avoids local optimization and slow convergence, and overcomes the exhaustive caused by genetic algorithm. It is a rapid and credible method.From the above analysis, time series analysis method is capable for short-term prediction, and artificial neural network method for both short-term and long-term prediction. Obviously, the latter one is better. The BP method has high accuracy with local local optimization from the average relative errors and cost time. So the method combined the above two is proposed. Simulation results show that average relative errors is decreased by 16%, which means the proposed model is better.Besides, the concept of events is introduced at the end of traffic prediction, which is a factor effecting traffic during holidays. The introduction of the events is made for holiday, because the holiday traffic and the normal traffic have a certain difference usually. For example, there are big difference between the traffic on the New Year's Day holiday and other days. Therefore, for the holidays, the result predicted using the predict model is unlike the true traffic obviously. In addition, this is a problem that can not be resolved only by improving prediction model's effection. At this point, the experience and advice provided by experts played the important role. This paper provides an event processing module in the prediction system. The module adjust the forecast result based on the thing factor given by expert .The laboratory data show that after adding the adjust factor, the forecast traffic for the holidays are better than before.Generally speaking, we research and compare the above discussed prediction models in theory aspects meanwhile design and implement predicting system. At present, the system still needs to be improved in next work in order to be adaptive to the incessant changing of users' business demand.