Amid growing global population and climate challenges, traditional agricultural models—with low resource efficiency and high environmental cost—fall short of meeting precision and sustainability goals. This study proposes an intelligent control architecture that integrates edge semantic computing and deep learning within a heterogeneous end–edge–cloud framework. The system utilizes an STM32H7 microcontroller for real-time data acquisition, deploys an LSTM-GRU model via TensorFlow Lite for edge-side time-series prediction, and enhances cloud-level decisions using a fine-tuned Large Language Model (LLM) with agricultural domain knowledge. A hybrid control strategy combining a PID-based kernel, LSTM predictive compensation, and LLM-driven dynamic correction significantly improves greenhouse regulation accuracy (MSE = 2.746). Experimental results show that, compared to a traditional PID system, the proposed LSTM-LLM scheme increases tomato growth rate by 51.3% and boosts decision frequency by 12-fold, validating its efficiency and practicality in smart agriculture.