Integrating large language models for data-driven adaptive decision support in floriculture production: Managing contextual heterogeneity

Orchid production in greenhouse environments functions as a biologically adaptive form of digital manufacturing, where real-time data and predictive models are used to coordinate complex growth processes. The strong interactions among temperature, humidity, light, and developmental stage challenge traditional rule-based systems, which often lack sufficient flexibility needed for adaptive control, especially when confronted with heterogeneous data inputs. We introduce an LLM-enabled framework for lifecycle management in orchid cultivation that integrates structured sensor measurements with unstructured text such as cultivation logs. The architecture uses modality-specific encoders that maintain distinctions between numerical signals and language-based operational knowledge while mapping them into a shared representation space. This design supports context-aware and stage-aware prediction of plant states, including growth phase transitions and stress responses. We provide a theoretical analysis showing that the multimodal fusion mechanism preserves task-relevant information and allows efficient optimization with bounded approximation error. Empirical evaluation on longitudinal commercial greenhouse data demonstrates that the proposed model consistently outperforms traditional baselines in predictive accuracy and robustness. It is notably effective in capturing nonlinear growth dynamics and stress events, which are critical for reducing losses and improving product uniformity. These results indicate that the framework can support scalable, data-driven control in orchid production.