A multi-model deep learning approach for proactive QoS prediction in 5G network slicing

Abstract Maintaining Service Level Agreements (SLAs) and enabling mission-critical applications in dynamic 5G environments depends on accurate, proactive Quality of Service (QoS) prediction. However, existing approaches often rely on static, reactive models that fail to capture temporal traffic dynamics and struggle with the severe class imbalance inherent in network anomalies. To address these gaps, this study proposes a robust dual-stream architecture trained on high-fidelity data generated via Digital Twin Network Emulation. The methodology decomposes the prediction task into two specialized streams: a Bi-Directional LSTM (BiLSTM) regressor that leverages temporal lag features to predict continuous Packet Loss Rate (PLR), and a Residual MLP (ResNet-MLP) classifier that predicts Packet Delay. To overcome the critical issue of minority class neglect, we implement dynamic K-Means binning for target definition and utilize the Synthetic Minority Over-sampling Technique (SMOTE) combined with Focal Loss during training. Experimental results demonstrate that the proposed framework significantly outperforms state-of-the-art baselines, including XGBoost and Random Forest, achieving an R^2 of 0. 91 for PLR prediction and an overall accuracy of 81. 1% for delay classification. More importantly, the framework addresses the blind spot to medium-latency events that were seen in previous studies, with a macro-recall of 0. 86. Robustness testing confirms the model's stability under simulated signal noise, and SHAP interpretability analysis validates its alignment with physical network parameters, providing network operators with a reliable, transparent tool for automated slice management.