Comparative Analysis of Multi-Step Time-Series Forecasting for Network Load Dataset

Increased usage of telecommunication networks leads to errors and congestion problems. Prediction of network load well in time can greatly help network operators in avoiding network problems. However, little is known about how different multi-step ahead prediction approaches would perform in predicting long-range variables in telecommunication data sets. In this paper, we use statistical algorithm, auto-regressive integrated moving average (ARIMA) and machine learning algorithms, multi-layer perceptron (MLP) and long short term memory (LSTM) to perform multi-step ahead prediction of network bandwidth utilization. We propose to compare different approaches such as recursive, direct, multiple-in multiple-out (MIMO), and variations of these approaches to perform multi-step ahead prediction of the network bandwidth utilization 1-hour, 2-hour, and 3-hour ahead in time. Results revealed that the direct approach performed better for 1-hour ahead predictions and MIMO for 2 and 3- hours ahead predictions. Also, all three predictive algorithms benefited from the multi-step approaches. We highlight the implication of our results for long-range predictions in telecommunication network.