Bandwidth Usage Prediction on Enterprise Network Infrastructure Using Time-Series Forecasting

Abstract

The increasing dependence of enterprise operations on cloud applications and centralized information systems produces bandwidth demand that changes dynamically over time. This study develops and compares time-series forecasting models for predicting total bandwidth usage at the next monitoring interval in an operational enterprise network. Historical inbound and outbound traffic records were obtained from three separate internal routers and synchronized into 1,920 timestamped observations collected. Data preparation included standardization, unit conversion to Mbps, duplicate and empty-column removal, time-based interpolation, forward and backward filling, and aggregation of inbound and outbound traffic. The data analysis stage generated temporal variables, three lag features, and a three-interval rolling mean to predict one-step-ahead bandwidth usage. Linear Regression, Random Forest Regressor, and Multilayer Perceptron Regressor were evaluated using chronological 70:15:15 partitions and RMSE, MAE, MAPE, and R². Random Forest achieved the best average performance across the three routers, with an RMSE of 103.34 Mbps, MAE of 75.13 Mbps, and R² of 0.1171. Linear Regression produced an average RMSE of 108.62 Mbps and MAE of 93.89 Mbps, whereas MLP Regressor produced substantially higher errors and a negative average R². The results indicate that the tree-based ensemble model was the most suitable among the evaluated approaches for the observed short-term bandwidth patterns. Keywords: bandwidth forecasting, enterprise network, timeseries forecasting, Random Forest, network traffic.