Objective: High-density space launch operations require reliable environmental control in spacecraft testing facilities, where energy-intensive heating, ventilation, and air-conditioning (HVAC) systems account for more than 70.00% of the total power consumption in low-latitude launch sites. Conventional strategies based on proportional-integral-derivative (PID) control fail to address systemic inefficiencies, including frequent equipment switching, uncoordinated chiller and air handling unit (AHU) operations, and sensitivity to extreme outdoor conditions. This study aims to develop an integrated data-driven control framework that synergizes load prediction and dynamic optimization to achieve two goals: maintaining temperature stability within ±0.5℃ and relative humidity (RH) stability within ±3% and reducing annual HVAC energy consumption by at least 15.00% in large-volume industrial facilities exceeding 300 000 m³. To address the dual objectives of environmental stability and energy efficiency in green spaceport construction, this study proposes an advanced air-conditioning control framework integrating an improved long short-term memory (LSTM) prediction model with hybrid PID-model predictive control (MPC) algorithms. Methods: Focusing on the large-scale complex HVAC systems of spacecraft testing facilities, this study systematically tackles the high energy consumption caused by independent PID control strategies, frequent equipment switching, and insufficient coordination between chillers and AHUs. A crucial innovation lies in the development of an LSTM prediction model enhanced by temporal attention mechanisms to address the limitations of conventional LSTM models, such as overfitting, weak long-term dependency capture, and poor generalization. By analyzing six key load-influencing factors (i.e., ambient temperature, relative humidity, solar radiation, occupancy, equipment heat, and lighting) through Pearson correlation and weight allocation, the model prioritizes temperature (27% weight) and solar radiation (18% weight) as the dominant variables. Trained on 8 760 h transient system simulation data covering seasonal and diurnal variations, the enhanced LSTM architecture incorporates eight hidden layers, six-dimensional inputs, and a 0.2 dropout rate, achieving exceptional prediction accuracy. Results: Comparative evaluations against baseline LSTM models showed significant improvements: the explained variance score increased from 0.999 0 to 0.999 5, the mean squared error decreased from 5.163 0 to 2.086 1, and the prediction accuracy within ±3 kW error bounds improved from 83.74% to 96.76%. The proposed model excelled in tracking irregular load fluctuations and long-term trends, as visualized in multiscale prediction-output comparisons. Building on these predictions, a PID-MPC hybrid control system was designed to synergize chiller plant optimization and AHU dynamic adjustments. The PID controller regulated chilled water supply parameters based on real-time load predictions, whereas the MPC module compensated for air-conditioning process delays and disturbances from volatile outdoor conditions. Field tests at a representative testing hall with a volume exceeding 300 000 m³ revealed that the integrated strategy achieved a 15.00% annual energy reduction compared with conventional PID controls, with monthly peak savings reaching 17.55%. The system mitigated issues such as frequent compressor cycling, uneven load distribution, and excessive reheating energy consumption, maintaining temperature and humidity within ±0.5℃ and ±3% RH thresholds, respectively. Conclusions: This study introduces three pivotal contributions to intelligent HVAC systems. First, the attention-augmented LSTM model establishes a new benchmark for industrial load forecasting, particularly in capturing nonlinear interactions between external weather and internal thermal loads. Second, the PID-MPC hybrid strategy effectively bridges the local actuator responsiveness and the global energy optimization, outperforming single-method approaches in terms of stability and efficiency. Next, validated in ultralarge spacecraft test facilities, the framework offers a replicable solution for comparable high-bay environments, such as semiconductor fabrication facilities or aerospace assembly plants. Practically, the 15.00% energy reduction means that launch sites can save hundreds of millions of kilowatt-hours of electricity annually, significantly advancing the development of green launch facilities.