PENERAPAN DEEP LEARNING DENGAN ATTENTION MECHANISM PADA PERAMALAN TRAFIK JARINGAN UNTUK MANAJEMEN QoS ADAPTIF

Application of Deep Learning with Attention Mechanism on Network Traffic Forecasting for Adaptive QoS Management

Abstract

Static Quality of Service (QoS) management in government network infrastructures often fails to dynamically adapt bandwidth allocation to fluctuating and non-linear traffic conditions. This study develops an adaptive QoS management system based on a Hybrid Long Short-Term Memory (LSTM)-Attention model for network traffic forecasting at the Communication and Information Technology Office (Diskominfo) of Lombok Barat Regency, serving approximately 200 Regional Government Organization (OPD) users with 700 Mbps dedicated 1:1 ISP capacity. A dataset of 43,488 data points was collected from the WAN interface of a MikroTik Router via RouterOS API during January–May 2026 at 5-minute intervals. The preprocessing pipeline included Z-Score outlier detection with capping, Min-Max Scaling normalization, and Sliding Window transformation (timestep = 60). The model architecture consists of two stacked LSTM layers (128 and 64 units) and one custom Attention Layer, totaling 122,241 trainable parameters, with a sequential 70/15/15 train-validation-test split. Evaluation on the testing set (May 2026) yielded RMSE of 43.66 Mbps and MAE of 27.96 Mbps for download traffic (RX), and RMSE of 30.72 Mbps and MAE of 19.66 Mbps for upload traffic (TX). MAPE values of 19.64% (RX) and 17.95% (TX) fall in the "Good" category. The model outperformed the ARIMA baseline on three of four evaluation metrics, with RMSE improvement of 16.7% and MAE improvement of 28.8%, while MAPE was marginally higher (+1.00 point) than ARIMA, a nuance attributable to the differing sensitivity of scale-dependent versus percentage-based metrics; an ablation study further confirmed the Attention layer's contribution is asymmetric, consistently improving accuracy on upload (TX) traffic and substantially reducing training variance on both signals (62.5-83.1% lower standard deviation across seeds) at a marginal parameter cost (+3.5%). Integration with MikroTik Queue Tree via RouterOS API successfully executed 603 automated adaptive cycles, achieving 26.7% throughput increase, 37.0% delay reduction, 59.3% packet loss reduction, and 49.6% jitter reduction compared to conventional static QoS management.