Modelling and predicting Computer Network using ARIMA Models and Neural Network Method

Abstract

This study compares ARIMA-based statistical models and RNN-based neural models for forecasting hourly network traffic. Four datasets—simulated, daily-Dominant, Weekly-Dominant, and Balanced-Dominant—are analyzed using spectral diagnostics to identify daily and weekly periodicities. Three MSARIMA variants, Auto-ARIMA, a tanh-RNN, and a mean-activation RNN are evaluated using MAE, RMSE, MAPE, and R². MSARIMA models consistently achieve the highest accuracy, with R² above 0.97 in simpler datasets. RNN models severely underfit, producing “extremely high error rates exceeding 90% MAPE. The results show that MSARIMA is a reliable solution for short-term traffic forecasting, while simple RNNs fail to capture seasonal structure.