DPIR-GSky: An Integrated Digital Twin Framework for Streaming Multi-Entity Intervention Prioritization in Vegetable Cold Storage

Abstract

When several vegetable cold storage units deviate from safe environmental conditions simultaneously, deciding which unit needs intervention first is difficult. Multi-criteria ranking methods generally require criterion weights fixed in advance, which does not fit well with continuously changing streaming conditions. This study developed and evaluated a digital twin framework built around DPIR-GSky, an incremental, dominance-power-based group skyline algorithm that prioritizes intervention across multiple storage entities without manually elicited criterion weights. The framework integrates streaming sensor acquisition, a virtual twin layer, and adaptive pruning with dominance-power scoring for sliding-window operation. Readings for temperature, relative humidity, ethylene, carbon dioxide, and illuminance were aggregated into six-dimensional deviation vectors, one combining temperature and ethylene into a shelf-life-based urgency index. The framework was evaluated using seven days of real sensor data from twenty storage racks for calibration and benchmarking, together with five simulated 72-hour disturbance scenarios generated from an autoregressive model calibrated on the real data, scored against independently defined ground-truth labels and compared with two uniformly weighted multi-criteria baselines. The virtual twin achieved a modest average coefficient of determination of 0.258, reflecting low signal variance under stable operating conditions, and is therefore better suited to directional trend monitoring than precise point prediction. DPIR-GSky reached an average F1-score of 0.685, trailing the weighted baselines by 0.042 points overall, yet it matched or exceeded both baselines under a heterogeneous, multi-variable disturbance scenario, detected an ethylene-related disturbance 65 minutes earlier than static-threshold monitoring, and processed each streaming window in 0.32 milliseconds at a throughput above 37,000 tuples per second. These findings indicate that the proposed framework is computationally and functionally feasible for real-time intervention prioritization without manually elicited criterion weights, with its advantage concentrated where several variables deviate together rather than one variable alone; validation at optimal 0–2°C storage conditions and across additional facilities remains necessary before operational deployment.