The Study Of Applications In Streaming Media Service About Predictions Based On Time Series

The streaming media system usually costs lots of resource like network bandwidth and energy like electricity, because of hosting a large number of instantaneous users. No matter to cut the system budget itself or for the world energy crisis, it is more and more necessary to seek solutions to optimize resource or energy efficiency. However, the expected schemes would have to be developed based on predictions of future bandwidth requirement and sever workload of the whole system. Therefore, this paper proposes and solves two topics as follows:1) Topic 1: How to predict the network bandwidth requirement based on video traces?2) Topic 2: How to predict the future connection number based on its log data?To address these two problems, this paper proposes one prediction scheme for each topic based on the modeling and forecast of Time Series. Each prediction schemes uses an ARMA fitting algorithm designed by us. And each time, the sample series is analyzed, plot, processed to be stationary without trend or seasonal components, and subtracted by its mean before running the algorithm. These operations are all reversible. In addition, after running the algorithm, each scheme predicts the future values based on the forecasts from the gained stationary model of the algorithm.Topic 2 is relatively simpler, and can be processed and solved according to the normal steps of Time Series Analysis. But topic 1 is harder, and a special, effective and reversible initial transformation is needed before the fitting algorithm, which improves step by step as the experiments proceed. At last, we get a multi-step dynamic prediction mechanism. We design several experiments for each prediction and the results show that our schemes are effective with a high accuracy, and could capture the characteristics and variation of history data. Especially, compared with related work, the scheme for Topic 1 needs no stationary assumption and doesn't have to separate the I, P and B frames, and thus achieves higher accuracy with a lower implementation complexity.