A method about energy-saving optimization operation of central air-conditioning system based on adaptive modeling and self-learning was proposed to solve the difficulties of mechanism modeling and parameters identification． The Markov decision process model of air-conditioning system was designed and the reinforcement learning algorithm with dual neural network structure was used to solve the curse of dimensionality and overestimation of value function during the learning process． A TＲNSYS simulation platform based on the central air-conditioning system of an office building in Guangzhou was built and the effectiveness of the algorithm was validated． The simulation results show that under the premise of meeting the indoor thermal comfort requirement，the energy-saving optimization operation of the system is realized with the goal of minimizing the energy consumption． Compared with PID control and single neural network reinforcement learning control，the total energy consumption of the system is reduced by 5. 36% and 1. 64%，the proportion of total uncomfortable time is decreased by 2. 32% and 1. 37%，respectively． The reinforcement learning controller proposed can effectively solve the overestimation problem． It has good robustness，self-adaption optimization capability and better energy-saving effect，and it can provide new ideas for building energy efficiency．