Sustainable Reinforcement Learning-Based Energy Optimisation for AHU-Controlled Potato Cold Storage in West Bengal
Sourish Mitra, Shrabani Sutradhar, Rajesh Bose, Sandip Roy, Arfat Ahmad KhanIntroduction:
Potato (Solanum tuberosum L.) storage conditions are essential for reducing post-harvest spoilage; however, conventional systems in West Bengal that rely on fixed-speed compressors and PID controllers consume excessive energy and struggle to maintain stable storage conditions. A typical 1000 MT cold storage plant loses 8% to 10% of stored potatoes annually, incurring high electricity costs. This paper proposes an IoT-enabled, reinforcement learning (RL)-based, return air (RA) dependent adaptive air handling unit (AHU) control system for energy-efficient cold storage.
Materials and Methods:
The system integrates TEC1-12706 Peltier modules, SHT-31/DHT11 sensors, ESP8266-12E IoT modules, Arduino-UNO, and DRV8873-Q1 motor drivers. Realtime monitoring of temperature, humidity, CO2, and power tariff serves as real-time inputs to the Q-learning RL algorithm. Control actions include adjusting fan speed, cooling intensity, and ventilation rate. The reward function penalizes energy expenditure and deviations from optimal storage conditions. Performance was evaluated over a 12-month cycle and compared against PID and ML-based baselines.
Results:
The proposed RL-based control system reduced annual energy consumption by 20%, from 1,60,000 to 128,000 kWh.Product quality improved markedly as spoilage decreased from 12% to 5%, and system efficiency increased significantly, as evidenced by a rise in the coefficient of performance(COP) to 4.1. Financial analysis indicates substantial economic benefits, with projected annual savings of Rs. 12-15 lakh from reduced electricity usage and post-harvest losses, and an estimated return on investment (ROI) of approximately 110% over seven years.
discussion:
Unlike PID systems, RL dynamically adapts to varying load, respiration rates, and tariffs, ensuring both efficiency and product quality. Federated learning enhanced scalability across heterogeneous racks.
Discussion:
This paper presents a tariff-aware, IoTenabled, reinforcement learning-based control system for potato cold storage, validated under real-world field conditions. A model-free Q-learning algorithm regulates AHU fan speed, humidification, and ventilation in real time based on temperature, humidity, CO2 levels, and electricity prices, in contrast to classical PID or offline ML approaches. The system simultaneously optimizes energy utilization, operational costs, and product quality while accounting for CO2 fluctuations arising from produce respiration. The 1000MT facility’s 12-month installation shows its flexible response to seasonal variations, scalability, and cost-effectiveness.
Conclusion:
The framework provides a sustainable and scalable cold storage model that addresses post-harvest losses. Moreover, it reduces energy cost and carbon footprint. Further, it is in line with India’s commitment to reducing its food loss by 50% by 2030.