An energy-saving control strategy based on data-driven and self-learning mechanism was proposed to solve the problems of complex mechanism modeling, lack of self-learning ability and slow optimization speed of traditional energy-saving optimization methods for cold source system. The Markov decision process model of cold source was designed and the DDPG algorithm from policy gradient was used to solve the problem of dimensionality curse and can avoid discretization of control actions. In this paper, the central air conditioning cold source system of a large office building in the hot summer and warm winter area was selected as the research object and the control strategy of cold source system was optimized, the results show that under the premise of meeting the indoor thermal comfort requirement， the energy-saving control strategy of the system is realized with the goal of minimizing the energy consumption. In the comparison experiment, the total energy consumption of cold source system under DDPG control strategy is reduced by 6.47% and 14.42% than PSO control strategy and rule based control strategy, the average indoor thermal comfort is increased by 5.59% and 18.71%, the proportion of total uncomfortable time is decreased by 5.22% and 76.70%, respectively. The simulation results show that the control strategy we proposed has effectiveness and practicality, which has obvious advantage in energy-saving optimization compared with other control strategies.