Network Traffic Statistical Upper Limit Prediction From Flow Features for Traffic Fluctuations

The network traffic prediction is essential for network operators in cost-effective link capacity planning. This paper proposes a method to predict traffic statistical upper limits (TSUL) as the required link capacity for a planning period of several months. Accounting for the large and non-stationary traffic fluctuations due to changes in flow numbers and bandwidth needs during the planning period is crucial. To adapt these fluctuations, we presents a network traffic prediction method for link capacity planning with two novelties. First, it uses flow features such as the number of flows and bandwidth requirements to adapt TSUL prediction to changing factors. Second, it utilizes a Supervised Variational Auto-Encoder (SVAE) model to estimate the traffic probability density function given flow features, and the predicted TSUL is derived from the upper limit of its confidence interval. In experiments using traffic data with large and non-stationary fluctuations, we evaluate the prediction accuracy of the proposed method in comparison with a linear probability density function estimation method (Linear) and the time-series prediction model (Long Short-Term Memory: LSTM and Temporal Fusion Transformer: TFT). Experimental results show that the proposed method is superior to Linear and LSTM in terms of TSUL root mean square error and log likelihood, and is inferior to TFT in terms of prediction accuracy but superior in terms of computation time. The experimental results suggest that the proposed method is suitable for link capacity planning targeting traffic with large and non-stationary fluctuations, considering not only the prediction accuracy but also the computation costs.