Abstract: As energy-efficient vegetable production facilities in winter, solar greenhouses face significant challenges in internal temperature control due to high thermal inertia, strong nonlinearity, and large external disturbances．Traditional ventilation control strategies generally suffer from response delays and insufficient accuracy, making it difficult to meet environmental requirements for stable crop growth．To improve intelligence and real-time performance of greenhouse temperature regulation systems, a hybrid model combining long short-term memory（LSTM）and gated recurrent unit（GRU）optimized by firefly algorithm（FA）was proposed, referred to as FA-LSTM-GRU, for temperature prediction and ventilation control．First, model integrated LSTM and GRU structures, a multi-head attention（MHA）was incorporated to enhance temporal feature extraction, and FA was employed to optimize hyperparameters．Then, a model predictive control strategy based on predicted values was designed, in which ventilation behavior was proactively adjusted using proximal policy optimization（PPO）．Finally, a control system was implemented on cloud server and Arduino platforms to achieve closed-loop integration．Experimental results showed that FA-LSTM-GRU model achieved R²=0.9769 and root mean square error of 0.7708 °C．Control strategy stabilized temperature fluctuations within ±0.6 °C, demonstrating good accuracy and system robustness．