Efficient Graph Attention-based Learning for Traffic Prediction and Uncertainty-Aware Anomaly Detection in AI-driven O-RAN

Traffic load prediction to enable proactive congestion management and resource optimization in Open Radio Access Networks (O-RAN) is challenging, given exhibit distinct temporal patterns, dynamic ranges, and burst behaviors in multiple service slices. To address this challenge, we propose a novelly efficient traffic prediction architecture that processes sixteen radio-level features through three parallel branches: a per-feature shared BiLSTM for temporal dynamics, a Graph Attention Network (GAT) over a feature-as-node graph for cross-feature dependencies, and a dual Transformer over magnitude and phase spectra from the input window. The branch outputs are fused per node, mean-pooled across nodes, and trained per network service slice using a single-step regression objective on granted physical resource blocks. We apply mean squared error for the smoother eMBB slice and show that pure Huber loss is more effective for the burstier mMTC and uRLLC slices than both mean squared error and upper-tail weighted mean squared error. A post-hoc Chebyshev test calibrated on validation residuals converts predictions into a slice-aware anomaly flag without labeled anomalies. On the Colosseum O-RAN dataset, the model achieves promising performance, $R^2 = 0.9011$ on eMBB, $R^2 = 0.3002$ on mMTC, and $R^2 = 0.6624$ on uRLLC, outperforming the existing studies with Chebyshev flag rates below $3.5\%$ across all slices.