Abstract: To address the significant sensing lag and response overshoot in existing pressure-equalized ventilation systems caused by manual adjustment or single-threshold control, a method for predicting and controlling oxygen concentration in the upper corner based on Long Short-Term Memory (LSTM)-Transformer was proposed. A hybrid model integrating LSTM and the global attention mechanism of Transformer was constructed to achieve accurate advance prediction of oxygen concentration in the upper corner using multi-source monitoring data. A feedforward-feedback composite control strategy was designed, in which the prediction output of LSTM-Transformer was used as the feedforward signal to compensate for time lag; a dynamic weight adjustment mechanism was introduced into the control strategy, and the limited-weakening integral method was used to improve the PID algorithm, eliminating steady-state error and improving dynamic response safety and stability. Numerical simulation based on computational fluid dynamics was used to analyze the nonlinear resistance characteristics of the louvered air window, and an opening-resistance mapping model was established. The target resistance output by the PID controller was inversely solved in real time into an accurate blade angle, thereby compensating for the nonlinear response characteristics of the air window and improving execution accuracy. Experimental results showed that compared with mainstream time-series prediction models such as random forest, BP neural network, support vector machine, convolutional neural network, and single LSTM, LSTM-Transformer performed best in root mean square error, mean absolute error, coefficient of determination, and mean bias error. Field test results showed that the LSTM-Transformer and PID composite control strategy shortened the overall system response time by 1-2 min, stabilized the fluctuation range of oxygen concentration in the upper corner in 18%-19.5%, effectively suppressed overshoot, and realized the transition of ventilation regulation from passive response to active intervention.