Abstract: Aiming at the load uncertainty on the demand side of the building energy system and the randomness of renewable energy on the supply side, a building energy system management optimization strategy is proposed based on deep reinforcement learning. Firstly, a supply-demand side research framework for the energy system and device model is built. The building energy management problem under the real-time stage is constructed as Markov decision-making process, and the deep reinforcement learning theory is used to minimize the cost of electricity, ensure the indoor heat comfort level and maximize the consumption of renewable energy as the optimization goals, and the duel dual deep Q network algorithm is used for model training, and the trained model can make adaptive control decisions according to real-time environmental parameters. Finally, through the application in the building energy system case, the results show that the proposed optimization strategy reduces the cost of electricity by 11.03%, the duration of thermal discomfort by 89.62%, and the amount of unconsumed photovoltaic power generation by 10.43%, comparing with the traditional rule-based control strategy.