An Encoder-Decoder Deep Learning Approach for Multistep Service Traffic Prediction

The prediction of traffic in regards to data services can be leveraged by cloud and edge computing orchestration mechanisms in order to minimize the costly number of resources being utilized and guarantee that the specified Quality of Services (QoS) metrics are within acceptable ranges. The service traffic prediction has a long history dating back to the 1990s that started with point process, time series statistical models followed by Recurrent Neural Networks. Contemporary cloud resource management mechanisms can provide resource allocation policies by leveraging multi-step traffic prediction.In this research paper, we compare statistical time series with Deep Learning (DL) models. We propose an encoder-decoder DL approach for multi-step traffic prediction. We examined four encoder-decoder DL architectures i) Stacked LSTMs, ii) CNN-LSTMs, iii) Bidirectional LSTM and iv) an innovative Hybrid Unidirectional-Bidirectional LSTM. We conducted experiments using a TCP trace data set with a 5 minutes time-step. We predict the number of requests, the transmitted data and the duration of the sessions with multi-steps in a range of one to five steps, which corresponds to a time window that spans 25 minutes in total. The results show that the encoder-decoder architecture provides better accuracy results in regards to predicting the traffic and the duration of the sessions.