STC-RL: A Spatiotemporal Control Framework for Intelligent Environmental Regulation in Automated Warehousing Using IoT and Deep Learning

Abstract

Intelligent warehousing environments require precise, energy-efficient control of temperature, humidity, and other environmental parameters. To address this, we propose STC-RL, a novel deep reinforcement learning framework that integrates Transformer-based temporal modeling, 3D CNN-based spatial field reconstruction, and graph neural network (GNN)-enhanced anomaly detection within a continuous- action reinforcement learning policy. Specifically, the Transformer captures long-range temporal dependencies from multi-sensor time series, while the 3D CNN generates a spatial thermal-humidity field via bilinear interpolation of sensor coordinates. The GNN encodes physical sensor topology to detect equipment failures through residual-based anomaly scoring. The RL agent operates in a continuous action space (e.g., setpoint temperature, humidifier output) and optimizes a multi-objective reward balancing environmental deviation, energy consumption, switching frequency, and anomaly alerts. Experimental results on a real 2,100 m² cold-chain warehouse show that STC-RL reduces energy consumption by 13.1%, achieves an average AUC-ROC of 0.936 for anomaly detection, and lowers temperature/humidity prediction RMSE to 1.02°C / 4.15%, outperforming six baselines. The system also cuts food spoilage by 66.7% and improves temperature compliance to 97.6%.