Abstract: The variable air volume (VAV) air conditioning system is influenced by uncertainties and random disturbances during the temperature control process. Furthermore, the VAV system exhibits characteristics such as nonlinearity, complexity, and time variability, which make precise temperature control under low energy consumption challenging for conventional control algorithms. To address these shortcomings in conventional approaches, we propose the application of a homothetic tube model predictive control (MPC) algorithm based on a constrained deep neural network (DNN) for room temperature control in VAV air conditioning systems. By analyzing the VAV system, a minimum energy consumption objective function is established. The optimal control problem of the actual system is transformed into an online learning problem, which is efficiently and rapidly solved using the constrained DNN optimization algorithm. The algorithm employs tubes of variable sizes in the optimization process to counteract uncertainties and random disturbances encountered during system operation. Simulation results indicate that the proposed control algorithm achieves higher precision in indoor temperature control, requires a lower time cost for optimization, and offers stronger adaptability to uncertain factors than the tube MPC (TMPC) and homothetic tube MPC (HTMPC) algorithms. The total energy consumed during the implementation of the proposed algorithm is 15.67% less than that of the TMPC algorithm, and 6.46% less than that of the HTMPC algorithm.