An AI–Digital Twin Approach to Closed-Loop Risk-Aware Resource Orchestration

Abstract

Dynamic resource systems must respond to changing operating conditions, uncertain disturbances, and competing performance requirements. Static allocation and prediction-only strategies may fail to incorporate the risk implications of future system states into repeated resource decisions. This study develops an AI-Digital Twin approach to closed-loop risk-aware resource orchestration. The approach uses observation-based twin synchronization, a supervised multilayer perceptron for future-state prediction, an action-dependent risk operator, and a constrained utility-cost-risk objective to select resource actions over successive decision cycles. After each action is executed, the resulting system observation is used to update the twin state and inform the next decision. The approach is evaluated in a simulation comprising 120 decision cycles, four strategies, and 12 independent random seeds per strategy: static allocation, predictive allocation, risk-aware orchestration, and the proposed closed-loop strategy. Performance is assessed using utility, deployment cost, shortage, volatility, constraint violation, and composite risk. The proposed strategy achieves the highest mean utility and the lowest mean shortage, constraint violation, and composite risk among the four strategies. Relative to risk-aware orchestration, it improves utility by 2.34%, reduces shortage by 2.04%, constraint violation by 3.69%, and composite risk by 2.39%, while increasing deployment cost by 5.52%. The difference in volatility is not statistically significant. These results indicate that synchronized prediction, action-dependent risk evaluation, and feedback updating can improve selected risk-sensitive performance dimensions, although the improvement involves a measurable cost trade-off.