Abstract
Cold storage is essential in the perishable-goods supply chain, yet its HVAC systems remain highly energy-intensive. This study presents a machine learning (ML) metamodel trained on EnergyPlus simulations and integrated with reinforcement learning (RL) for adaptive, closed-loop control of HVAC–PV–battery interactions. Replacing static sizing methods, the metamodel explores system parameters and identifies optimal insulation levels and PV–battery configurations. It selects 25 south-facing, 10 east-facing, and 15 west-facing PV panels (17.5 kWp total), generating 8,163 kWh/year, and a 16-kWh battery that supports 7,342 kWh/year of solar self-consumption. The framework supports planning-based control over future horizons, maintaining simulation-level accuracy with a 60% reduction in computational time. It maintains optimal indoor conditions within allowed HVAC bounds (temperature: − 0.5–4 ℃, humidity: 75–90%, airflow: 1,000–8,000 m3/h) to prevent beef carcass spoilage and increase renewable energy use. The approach lowers annual HVAC energy consumption by 14% (from 11,591 to 9,968 kWh) and grid reliance by 38% (from 4,249 to 2,626 kWh/year). Using sequential predictions, the RL agent anticipates how present actions affect future energy flows, battery levels, and indoor conditions, supporting proactive control. Unlike reactive or MPC-based methods, this forward-looking framework can handle variability—such as weather shifts or irregular meat loading—improving energy efficiency, preserving food quality, and reducing CO₂ emissions. It offers a scalable and practical solution for real-world cold-storage management.