Digital twin (DT) technologies have recently been adopted to support adaptive control, yet most existing approaches rely on a single DT and a predefined control paradigm, assuming its global optimality across all operating conditions. This assumption becomes limiting in highly dynamic environments, where the effectiveness of a control strategy is inherently context-dependent. This article introduces a multi-DT control paradigm, where multiple DT instances, each embedding a different control strategy, operate in parallel to continuously evaluate system performance and dynamically select the most suitable control policy. Each DT integrates an optimization layer based on genetic algorithms and two reinforcement learning agents to search the optimal parameters of different controllers. The proposed framework is applied to greenhouse lighting control, jointly regulating photosynthetic photon flux density (PPFD), daily light integral (DLI), and photoperiod across multiple subzones and solar variability. A virtual sensing layer is integrated to enhance spatial observability while reducing physical sensor deployment. The approach is validated in a real greenhouse environment where the multi-DT system achieves accurate PPFD and DLI regulation while reducing energy consumption by up to 35% compared to conventional on/off controllers, demonstrating the effectiveness of adaptive paradigm selection in complex cyber-physical systems. 

DOI: 10.1109/TII.2026.3691992