Dynamic, adaptive and modular Digital Twin framework for resource-efficient Controlled Environment Agriculture.

In Controlled Environment Agriculture (CEA), traditional fixed set-point control is replaced by dynamic control strategies. These strategies enable joint optimization of resource use efficiency and biomass output, leverage electricity price fluctuations to reduce energy costs, and employ targeted environmental stressors to enhance crop quality and physiological resilience. Implementation of dynamic control strategies, however, builds upon real-time monitoring, robust data integration and management, and high-fidelity predictive modeling. These capabilities can be effectively provided through a Digital Twin (DT). This study introduces a novel open-source DT framework designed to support dynamic control strategies in CEA, addressing challenges in scalability, generalizability and interoperability. The framework is centered on the IoT platform ThingsBoard, providing unified, scalable data acquisition and management across heterogeneous sensor and actuator networks through vendor-agnostic integration and standardized interfaces. A significant contribution is its physics-based modeling backend, built on ordinary differential equation models developed in Modelica and exported as Functional Mock-up Units (FMUs). To ensure model accuracy across varying biological conditions, a parameter estimation pipeline is developed to calibrate and adapt these FMUs against experimental data. Building on this, a dedicated simulation backend is implemented to leverage the calibrated FMUs, providing the dynamic predictive capabilities necessary for proactive system control. Furthermore, the framework incorporates a Multirate Moving Horizon Estimation (MMHE) state estimator to estimate critical unmeasured variables, such as plant biomass. This estimator is specifically designed to handle multirate data, maintaining continuous estimates even when certain sensors provide frequent data while others are sparse or infrequent. Demonstrated through a simulation-based case study modeling lettuce growth in a vertical hydroponic farm, the DT framework's architectural feasibility and virtual modeling capabilities are verified. Using synthetic data generated from a known true parameter set, the calibrated growth model achieved a low cross-validation prediction error, with an RMSE of 0.221g and an NRMSE of 6.62% on an independent test set. The MMHE-based state estimator effectively maintained continuous biomass estimates despite sparse synthetic measurements and model mismatch. These findings underscore the framework's potential as a robust and extensible foundation for future physical DT implementations in CEA, enabling a 31 transition toward dynamic, data-driven, and energy-aware operations.