Real-Time Cold Storage Temperature Control via CFD–ML Hybrid Modeling

Abstract

Cold storage facilities play a critical role in maintaining food quality and safety throughout the supply chain, with temperature uniformity and stability being essential for preserving product integrity and minimizing spoilage. Traditional temperature control systems often rely on simplified thermal models that fail to capture the complex three-dimensional airflow patterns and heat transfer phenomena occurring within large-scale cold storage environments. This research presents a novel hybrid modeling approach that combines Computational Fluid Dynamics (CFD) simulations with Machine Learning (ML) algorithms to achieve real-time temperature control optimization in cold storage facilities. The proposed CFD-ML hybrid model integrates high-fidelity CFD simulations for spatial temperature prediction with ML-based predictive control algorithms that can adapt to varying operational conditions and product loads. Our methodology employs a two-stage approach: offline CFD simulations generate comprehensive training datasets capturing diverse operational scenarios, while online ML models provide real-time control decisions based on current sensor measurements and predicted thermal behavior. Experimental validation was conducted in a 2,400 cubic meter commercial cold storage facility over a six-month period, comparing the hybrid approach against conventional PID control systems. Results demonstrate significant improvements in temperature uniformity, with spatial temperature variations reduced by 47% (from ±2.1°C to ±1.1°C) and energy consumption decreased by 23% while maintaining target temperature ranges within ±0.5°C. The ML component achieved prediction accuracies of 95.3% for temperature forecasting up to 2 hours ahead, enabling proactive control adjustments that prevent temperature excursions. The hybrid system demonstrated robust performance across varying ambient conditions, product loading scenarios, and equipment configurations, with average response times of 3.2 minutes for temperature corrections compared to 8.7 minutes for traditional control systems. This research contributes to the advancement of intelligent cold storage management by providing a scalable framework for integrating physics-based modeling with data-driven control strategies.