Abstract
Postharvest losses during potato storage remain a serious problem, especially for small-scale storage facilities with limited automation, monitoring, and energy resources. This paper proposes a data-driven cyber-physical framework for sensor-based microclimate analysis and predictive management of small-scale potato storage facilities. The framework defines the integration of a physics-informed digital twin, a safe reinforcement learning controller formulated as a constrained Markov decision process (CMDP), and a multi-sensor anomaly detection module operating on multidimensional time-series sensor data. The proposed approach is aimed at analyzing microclimate dynamics, detecting abnormal changes in temperature, humidity, and gas composition, and supporting prediction-oriented control decisions under technological and energy constraints. The digital twin is intended to serve as a simulation environment for training and future evaluation of predictive control strategies without risk to stored products. The control formulation explicitly represents the trade-off between energy consumption and storage quality while considering constraints on temperature, relative humidity, oxygen, carbon dioxide, and spoilage-related gas indicators. A set of forecasting-oriented validation scenarios is defined, including normal operation, hypoxia, early biodegradation, sensor failures, and energy-constrained storage conditions. The work forms a methodological basis for subsequent quantitative simulation, sensor data analysis, and experimental validation on real potato storage facilities.