Traffic forcast prediction using android application
Abstract : We consider the problem of forecasting high frequency sampled mobile cellular traffic starting from a lower frequency sampled time series. We use a dataset of real downlink/uplink traffic traces obtained from a mobile cellular network and apply different methodologies for performing forecasts at different sampling frequencies. Through extensive evaluation we show that such type of forecasting is possible and in some cases is also able to outperform forecast results obtained starting directly from the high frequency time series. The outcomes of this work can be used for several scenarios of cognitive networking, including prediction of data traffic requests in specific locations, as well as for data storage optimization and improvement of BBU clustering tasks.
? The existing works are described by considering five criteria; the used video preprocessing tools, counting approach, whether the algorithm is real-time or not (R-T), validation data size, and system accuracy.
? These factors motivate us to go further by proposing a new application that is real time, simple to use, automatic, and validated on long-lasting videos from different real scenarios, including, different smartphones, video duration, geographical position, intrinsic and extrinsic camera characteristics, and weather conditions.
? The problem of forecasting time series sampled at different frequencies has been already tackled in the past, especially in the field of economics.
? In that area it is common to encounter time series sampled at both high-frequencies (e.g. quarterly sampled GDP data) or low-frequencies (e.g., annual data), together with the need of forecasting one series given the other.
? Early works in the area approach the problem using simple interpolation or distribution approaches, where the latter is used when the high-frequency data is ought to sum to the value of the low-frequency data in an observation period.
• The developed app can be used by many public agencies, private sectors, governments for traffic flow analysis or for personal use, for example, to gather information about road traffic in the neighborhood or to present a quantitative evaluation of the traffic flow when selling a residence.
• To the best of our knowledge, such application does not exist. In the next section, we review the proposed system
• That makes the algorithm faster, requiring less memory space, minimizing the energy consumption, and real time on a smartphone. Let us detail the proposed counting approach.
? In particular, we focus on uplink/downlink video traffic demand traces obtained from a real country-wide mobile cellular network and analyze the performance of short-term/high-frequency forecasting (e.g., 5 minutes-ahead predictions) starting from input data which is sampled at a lower frequency (e.g., hourly).
? The application of the aforementioned forecasting methods requires training of the models and estimation of the corresponding parameters.
? The available one month data is therefore divided in two parts: two weeks of data are used as a training set to estimate parameters for all models.
? The remaining two weeks are used to evaluate each model performance.
|