A network traffic forecasting method based on SA optimized ARIMA–BP neural network

Abstract

Network traffic forecasting provides key information for network management, resource allocation, traffic attack detection. However, traditional linear and non-linear network traffic forecasting models cannot achieve enough prediction accuracy for future traffic prediction. In order to resolve this problem, a network traffic prediction method based on SA (Simulated Annealing) optimized ARIMA (Autoregressive Integrated Moving Average model)-BPNN (Back Propagation Neural Network) is proposed in this paper, which makes comprehensive use of linear model ARIMA, non-linear model BPNN and optimization algorithm SA. With enhancement of the BPNN global optimization ability, it can fully realize the potential of mining linear and non-linear laws of historical network traffic data, hence improving the prediction accuracy. This paper selects the historical network traffic data of two different sampling points in the WIDE project to predict, and utilizes the MAE(Mean Absolute Error), RMSE(Root Mean Square Error), and the MAPE(Mean Absolute Percentage Error) as the evaluation index of the prediction effect. Experimental results show that our proposed method outperformed traditional network traffic prediction model, with several improvements in network traffic prediction accuracy.

Introduction

Living in a society where digital transmission and mobile traffic pervade, we can get tons of information on the Internet through various mobile devices. The Internet has feed people with knowledge of the world around us, which in turn makes the network business developing rapidly [1]. It is inevitable that the network traffic volume has grown dramatically ever since [2]. Literally, network traffic refers to the total traffic of the network link per unit time. Network traffic is often collected within a certain time interval to obtain a time series in the process of network traffic collection [3]. The prediction of network traffic is of great significance nowadays to improve the network efficiency and optimize network resources allocation. After an accurate network traffic prediction, Network operators can adjust the network congestion control mechanism in a timely manner to reduce network delay and packet loss rate, cut the risk of network congestion, and improve user’s Quality of Experience (QoE). Meanwhile, network traffic prediction not only provides key information for network management, resource allocation, but also makes a huge difference against malicious network traffic attacks [4], with the potential of uncovering breaches vulnerable to network attacks and intrusions. It is a crucial technology to against numerous network security incidents and of great importance for maintaining the whole network performance [5], [6].

Network traffic prediction can be categorized into linear prediction and non-linear prediction [7], where linear prediction mostly includes autoregressive model (AR), moving average model (MA), autoregressive moving average model (ARMA) [8], differential integrated moving average autoregressive model(ARIMA). Non-linear prediction mainly contains the prediction models based on wavelet analysis [9] and neural network model. As a traditional neural network, BP neural network can map any complex non-linear relationships thanks to its non-linear fusion ability, which has been well studied in the field of traffic forecasting [10]. But the BP neural network model training algorithm has inherent characteristics, which causes the neural network model to have the disadvantages of slow convergence rate, and readily to fall into local minimums in practice [11]. The Simulated Annealing Algorithm (SA) is derived from the physical annealing process of solids, it is a stochastic optimization algorithm based on Monte Carlo iterative solution method. As a random search technology, SA has been widely used in various engineering problems [12]. The algorithm sets a certain initial temperature, then cools down continuously, combines the probabilistic jumping property to continuously search for the global optimal solution in the global solution space. Therefore, the global optimization ability of the simulated annealing algorithm (SA) can just solve the problem, which is the artificial neural networks that are easily to fall into local minimums.

Under the influence of many inherent factors, network traffic possesses the characteristics of sudden change, weak coupling and non-linearity [10]. Therefore, neither the single linear model nor the non-linear model can describe the characteristics and change laws of network traffic well. Analyzing linear and non-linear factors in network traffic accurately is the key to improve the network traffic prediction accuracy. Since a single model has limited ability to reveal linear and non-linear relationships in network traffic, these two kind of models can be combined to extract linear and non-linear relationships in network traffic prediction.

A network traffic forecasting method based on SA optimized ARIMA–BPNN (ARIMA–SA–BPNN) is proposed in this paper. For the ARIMA–BP model, ARIMA itself can only capture the linear features in the time series, without the ability to capture non-linear features, as well as the model requires more historical data beforehand. For non-linear features, utilizing learning and data processing ability of the neural network model can mine the non-linear information in the network traffic time series. Meanwhile, the simulation of annealing algorithm is used to improve the global optimization ability of the BPNN. During the process of ARIMA–SA–BP, at first, the ARIMA model is used to extract the linear features of network traffic. With strong processing abilities for time series-based data, ARIMA can still retain the original information of its historical data while extracting linear features.

Secondly, the network traffic residual predicted by the ARIMA model serves as an input in the BPNN, which is optimized by the SA algorithm. Non-linear factors in the residual data pool can be fully captured via its ability to perform non-linear fusion and global optimization. Finally, the final network traffic prediction value is to add the ARIMA model prediction results and the SA-BPNN residual prediction value. This module improves the accuracy of network traffic prediction in two ways. Adaptation of this module solves the problem of fully extracting linear and non-linear features in network traffic. Moreover, this module uses simulated annealing algorithm to make up for the shortcomings of BPNN, and enhance the BPNN global optimization ability. The main contributions of this paper can be summarized as follows:

• We proposed a network traffic prediction method based on SA optimized ARIMA–BPNN. This method can extract linear and non-linear features from traffic data, and can predict network traffic with more accuracy.

• Our work make advances in adjusting the network congestion control mechanism, optimizing the network infrastructure architecture with any other fields that require more accurate network traffic prediction results.

• Experimental results based on real-world network traffic data sets, which validated that our method outperforms current existing network traffic prediction models, like single prediction model and hybrid prediction model, such as LSTM, WNN, ARIMA–BPNN.

The rest of this article is organized as follows: The second part introduces existing studies done by researchers in the field of network traffic prediction. In the third part, we made some elaboration of our principle and analyzes the basis of the modules that comprise the method in this paper. Then we move on to discuss steps of our proposed approach. The fifth part uses real network traffic data sets to verify the effect of the model. Our paper closes with summary of the researches methods in the article and shed some light on future research directions.