Implementation Of Methods Of Traffic Time Series Prediction

Along with the vigorous development of telecommunication network,the number of users of telephone, mobile phone and broad band increasesquickly. As a low cost and convenient channel, users and sellers attach moreand more importance to telecommunication call center. The number oftelephonists of big telecommunication center has already reached thethousands. In the circumstances, traditional scheduling will not solve theproblem effectively, and we prefer to automatic scheduling. Traffic predictionis the base of automatic scheduling. The result of prediction has directinfluence upon the scheduling.The telecommunication trade is a special trade in economy, and itsproduction and management is in accord with economic rules. Therequirements how to estimate the future trend of traffic in order to provideassistant decision for telecommunication management department becomeimperious. The existing predicting tools are not fit for generic users becauseof their high price and specialty. On the other hand, some special predictingsoftware just aims at the given domain or the appointed project, and most ofthem are integrated into MIS or ERP systems, which strongly depend on therunning environment. Accordingly, the requirement to produce a predictingsystem for telecommunication trade whose users are decision makers intelecommunication management department is brought forward.This paper briefly state the design thought of predict system,introducebasic knowledge, principle of predict and several measure standards ofpredict. This paper analyzes three kinds of frequently used predictingmethods: Times Series Analysis Methods, Case Based Nearest Neighbor andArtificial Neural Network. And analyzes several kinds of frequently usedpredicting models: Exponential Smoothing Model, Cycle Model, LinearRandom Model, K Nearest Neighbor, and Artificial Neural Network Model indetails. After studied predicting methods and models, we make use ofeconomic predicting technology and realize traffic prediction system. Withthe target of understandability and practicability, we realized a Telecommunication Predicting System, combined database and predictmethod. The database includes basic data. The method base is made up ofpredicting method algorithm. This system is an integrated predictenvironment, which supports customizing selected predict method, predictthe trend of the traffic. The whole process includes predict method's select,predict's execution and results analyzing of multi-scenarios. The friendlygraph user interface directs users to work and return the results back in anunderstandable way. There's no need for users to know much about thepredict technology to control and adjust the system easily, which allows thecomplex and synthetical science to serve for common users. So the system ismuch utilizable.This paper discusses the applicable surrounds and factors which affectresults, expatiate on the implement process of the above models, summarizethe similarities and differences, the advantages and disadvantages of theabove models.Exponential Smoothing Model predicts well, but lags in time.Exponential Smoothing presume upon mean value's conservation or localconservation in short time, so it is applicable to short prediction. ExponentialSmoothing Model is the betterment and development of Moving Average, itconquers disadvantage of Moving Average and is simpler. Linear RandomModel does not take other factors into account and is simpler. In a short time,economic value is relatively stable. So in a short time, Linear Random Modelcan get a high prediction precision and is applicable to smaller size andhigher predict frequent. Cycle Model is applicable to economic value whichis impacted greatly by cycle factor and less by other factors. If we select rightmethod to predict the trend of the time series, we can get a high predictionprecision. On the condition of right time series similarity measurement, thevalue of K, and the method of synthesizing K neighbors, K Nearest Neighborcan get a high prediction precision. Artificial Neural Network Model predictswell, but it needs debug time in order to get right Artificial Neural Networkstructure which is the key of success prediction and also a taking time period.Exponential Smoothing Model and Linear Random Model are applicable toshort prediction. Cycle Model, K Nearest Neighbor and Artificial NeuralNetwork are applicable to both short prediction and long prediction. Aiming at requirements of traffic trade, Cycle Model, K Nearest Neighbor andArtificial Neural Network are more applicable. The experiment demonstratesthat Cycle-model does better in predicting weekdays, and Case-based nearestneighbor algorithm does better in predicting weekends. So this paper putsforward the combination of Cycle-model and Case-based nearest neighboralgorithm, using Cycle-model to predict the weekday's traffic and usingCase-based nearest neighbor algorithm to predict the weekend's traffic. Theexperiment demonstrates that the combination of Cycle-model andCase-based nearest neighbor algorithm get a better predict result, at the besttime, about 19.7% prediction precision in this method is improved thanCycle-model, and that is about 48.8% improved than nearest neighboralgorithm.In a word, this paper not only discuss the above models in theory, butalso carry through the design and development of predict system, and payattention to the practicability and feasibility of translating theory into practice.This system is at the stage of prime research and trying out, and meets theneed mainly. Still there are lots of works about making the system moreperfect so as to meet the users' added need.