This study constructed a building-PCM-environment dynamic coupling thermodynamic model, developed a multi-objective adaptive optimization algorithm based on deep reinforcement learning, and built a micro-meso-macro multi-scale experimental simulation platform to enhance the application effectiveness of PCM in building thermal energy storage. Through full-scale experimental chamber testing and multi-scale simulation, the results showed that the optimized PCM group (1.0% Al2O3 composite predicted mean vote, 7 mm thick) reduced energy consumption by 34% compared with the blank control group under 72 hour summer conditions, and the temperature fluctuation range was reduced to 1.8∘C. The average predicted mean vote value was 0.3, which falls within the thermal comfort range. The convergence speed of the deep reinforcement learning algorithm was more than 30% higher than that of the genetic algorithm, and the optimization function value was also improved. This study achieved the dynamic optimization of PCM design parameters and operation strategies, providing theoretical and engineering support for the low carbon transformation of buildings.