Proxy-Cost-Aware Explainable Graph Neural Network for Multi-Domain O&M Decision Support in Digital Twin-Enabled Smart-City IoT Systems

INTRODUCTION: Digital twin-enabled smart-city IoT systems may interact with interconnected transportation, parking, environmental, communication, and edge-computing infrastructures. Failures or congestion in one domain may propagate across service chains and increase operational burden, yet existing spatiotemporal graph models rarely represent cost semantics, risk propagation, and maintenance actions jointly. OBJECTIVES: This study develops a cost-aware explainable graph neural network for multi-domain operation and maintenance decision support, with emphasis on proxy-cost prediction, risk identification, rule-based action ranking, and interpretable attribution. METHODS: A multi-layer digital twin graph integrates sensing, communication, computing, service, maintenance, and proxy-cost nodes. A state–risk–cost three-channel propagation mechanism separately models operational states, abnormal-risk diffusion, and cost-impact transmission. Rule-based graph interventions, including maintenance, task migration, link switching, and sampling adjustment, are used to estimate action-induced proxy-cost changes. CityPulse is used as a heterogeneous multi-domain proxy scenario, while Caltrans PeMS provides a homogeneous traffic-sensor reference. Performance is evaluated using MAE, F1, NDCG@5, and Fidelity, where NDCG@5 measures consistency with predefined action-ranking rules. RESULTS: On CityPulse, the proposed method achieved lower MAE and higher NDCG@5 and Fidelity than the strongest corresponding baselines, although its F1 was slightly lower than STAEformer. On PeMS, it achieved the highest Fidelity but remained below the best traffic-specific models in MAE, F1, and NDCG@5. Ablation results showed the respective contributions of cost nodes, explicit three-channel propagation, rule-based intervention evaluation, and explanation constraints. CONCLUSION: The method is more effective for heterogeneous multi-domain O&M scenarios than for homogeneous traffic forecasting. Because both datasets provide proxy rather than real financial, energy-system, and intervention records, the results support proxy-based methodological validation but do not constitute direct physical-energy-system validation or evidence of realized cost savings and causal intervention effectiveness.