Evaluating RNN Models for Multi-Step Traffic Matrix Prediction

Network traffic matrix prediction is used to estimate the patterns of future network flows before they are initiated. Traffic matrix prediction is needed by a wide range of network management functions such as network monitoring and it relies on historical data. In this paper, we address the task of multi-time step traffic matrix prediction using Recurrent Neural Networks (RNN). Our objective is to conduct a comparative study of different types of RNNs and to evaluate their ability to predict multi-time step Origin-Destination traffic matrices. Experiments show that RNNs are capable of predicting multiple steps of traffic matrices, however, the RMSE of the predictions increases as we move further away from the last true value. Our primary finding is that the RNN-GRU show has the best prediction ability in the very beginning steps with an RMSE of 0.048, while RNN-LSTM demonstrated higher capability with the last steps, having an RMSE value of 0.058.