Traffic Prediction of Real Network Traffic Data with LSTM Neural Networks
Marleen Bahe, Kenji De Souza Yamane, Nina Großegesse, Matthias Herlich, Jia Lei Du, Christof Brandauer, Christian Ferenz (2024): Traffic Prediction of Real Network Traffic Data with LSTM Neural Networks In: Advances in Information and Communication. FICC 2024. Lecture Notes in Networks and Systems, vol 919. Springer.
The aim of this work is to make time series predictions for real network traffic data by using long short-term memory neural networks (LSTMs). The data used for this work was captured in an office environment by a company from the field of network security and monitoring. The trained LSTMs are compared to baseline methods which have lower training costs. It is shown that the relative performance of LSTM-based predictions depends on the composition of the data. For aggregated traffic, the analyzed time series contains regular patterns with no big gaps and the trained LSTM outperformed baseline methods with an NRMSE of 0.072. In case of non-aggregated traffic of single clients/services with more gaps, jumps or steep trends within the time series, baseline methods can outperform LSTMs. Out of the used baseline methods, the ARIMA method shows the best performance in most experiments. Finally, it is also shown how the traffic predictions can be used to statistically detect anomalies within networks by considering the difference between the predicted and the actual values.
