Abstract: To address the problems in greenhouse production management, such as limited dimensions of heterogeneous data collection, lack of intuitive human-computer interaction, and over-reliance of environmental regulation on human experience, a digital twin intelligent greenhouse management system based on Django and UE5 was designed and implemented. The system constructed a closed-loop management architecture of "heterogeneous perception-model reasoning-bidirectional virtual-real driving", integrated the DeepSeek large model's logical reasoning with UE5 pixel streaming technology, and realized the synergy of data display and intelligent decision-making. The system hardware adopted the ESP32 distributed management mode under a Wi-Fi ad-hoc network, and relied on MQTT services to achieve message forwarding; combined the Django and Vue frameworks to construct the Web management layer; utilized UE5 to build the twin model, achieving low-latency and high-fidelity bidirectional virtual-real mapping through pixel streaming and Lumen illumination technology; and integrated DeepSeek, YOLO, and Bayesian networks to establish a "perception-reasoning-assisted decision-making" system. The system test results showed that the average end-to-end data latency of the system was less than 150 ms, the error of virtual-real synchronization accuracy was less than 5%, and the accuracy of auxiliary decision-making was more than 90%, which improved the accuracy by about 25% compared with traditional experience-based regulation. The research results could provide a reference scheme for constructing a low-cost, highly integrated, and highly visualized modern intelligent greenhouse management system, as well as the migration and application of UE5 digital twin technology in the agricultural field.

Key words: greenhouse, digital twin, decision support, Unreal Engine 5, Django, ESP32