EcoTrust-Net: A Causal, Explainable, and Energy-Aware Graph Transformer and Multi-Agent Reinforcement Learning Framework for Intelligent Network Traffic Prediction and Optimization

Abstract

The increasing complexity of sixth-generation (6G) communication systems, Software-Defined Networking (SDN), edge intelligence, cloud-native infrastructures, and Internet of Things (IoT) environments has intensified the need for intelligent network traffic prediction and optimization capable of supporting low latency, high throughput, cyber-resilience, explainability, and energy-efficient operation. Existing approaches often address traffic prediction, routing optimization, explainability, or energy efficiency independently, resulting in limited topology awareness, weak causal interpretation, and reduced adaptability in highly dynamic communication environments. To address these limitations, this study proposes EcoTrust-Net, a causal, explainable, and energy-aware framework that integrates Graph Transformer-based topologyaware traffic prediction, Multi-Agent Reinforcement Learning (MARL) for adaptive routing and congestion control, Structural Causal Modeling (SCM) for causal inference and counterfactual reasoning, SHAP- and attention-based explainability, and multi-objective energy-aware optimization within a unified architecture. The framework was evaluated using heterogeneous datasets and environments comprising CAIDA backbone traffic traces, CIC-DDoS2019 attack traffic, Alibaba Cluster Trace workload data and NS-3/Mininet SDN simulations. Comparative experiments against conventional methods and recent state-of-the-art models published between 2024 and 2026, together with component-level ablation analysis, were conducted to assess prediction accuracy, Quality of Service (QoS), explainability, cyber-resilience and energy efficiency. Under the experimental conditions evaluated, EcoTrust-Net achieved 97.4% prediction accuracy, 96.8% precision, 96.1% recall, 96.4% F1-score, and an RMSE of 2.1. The framework also achieved 781 Mbps throughput, 12.4 ms latency, 1.8% packet loss, 97.1% attack detection, a 2.3% false-positive rate and 187 W energy consumption, outperforming the selected baseline models. These findings suggest that integrating topology-aware graph learning, cooperative reinforcement learning, causal reasoning, explainable artificial intelligence, and energy-aware optimization provides a promising framework for scalable, interpretable, and sustainable intelligent network traffic prediction and optimization in next-generation AI native communication networks.