Abstract: The data-driven model predictive control (MPC) has been widely applied in the field of energy-saving control. In addressing the issues of poor interpretability and slow optimization speed of pure data-driven temperature prediction models in mushroom house energy-saving control, this study proposed a data-physical hybrid model-based air conditioning energy-saving control method for mushroom houses. First, a basic prediction model was established by combining the gated recurrent unit (GRU) neural network with the attention mechanism. Considering factors such as heat transfer from fresh air, heat dissipation from mushroom respiration, radiant heat from envelope structures, and air conditioning cooling, the internal thermal balance equation of the mushroom house was incorporated into the physical loss term, which, along with the data loss term between the predicted and actual temperatures, formed the loss function. The prediction model and the loss function together formed a special gray-box prediction model, enabling the implementation of a mushroom house temperature prediction method based on a data-physical hybrid model. Then, a target function was established by considering the deviation between the predicted temperature and the reference trajectory as the prediction accuracy term, and the air conditioning operation time as the energy consumption term. Finally, based on the basic Adam algorithm, an improved Adam algorithm was developed by integrating the "belief" mechanism and the "look-ahead" mechanism. This algorithm was used to solve the target function quickly and obtain the optimal control sequence for the air conditioner over the control time domain. The first control value in the sequence was applied to the air conditioning system, achieving optimal control that simultaneously considered both accuracy and energy consumption for the mushroom house air conditioning system. Experimental results showed that the temperature prediction model based on the data-physical hybrid method outperformed various other prediction models, with a 18% improvement in prediction accuracy compared to the pure data-driven GRU model, the root mean square error (RMSE) controlled within 0.10℃, and the model training time was 16 seconds. The fitness value of the improved Adam algorithm was 0.751, with a computation time of 47 seconds. A comparative analysis with various optimization algorithms showed that, compared to the Adam optimization algorithm, the fitness value decreased by 6%, and compared to the elitist fast non-dominated sorting genetic algorithm (NSGA-II), the computation time was reduced by 81%. Compared to traditional threshold control, the proposed mushroom house air conditioning model predictive control method adapted well to the flexible changes in the reference trajectory. Through its unique prediction mechanism, it adjusted the control input in advance, demonstrating good reference trajectory tracking performance. The tracking accuracy improved by 63%, and during a 5-day simulation experiment, the control accuracy RMSE was reduced by an average of 73%, with air conditioning energy consumption reduced by an average of 12%. The mushroom house energy-saving control model developed in this study enhanced the interpretability of the prediction model, improved the computation speed and solution accuracy of the rolling optimization process, and increased robustness to changes in control parameters. It achieved precise temperature control and energy-saving control in the mushroom house, providing a reference for energy-saving control in industrialized edible mushroom cultivation.