Research On Building Energy Consumption Forecast Based On Dueling Deep Reinforcement Learning

Now,the economy of various countries continues to develop rapidly which make energy consumption increasing day and day,and global climate and environmental problems are becoming increasingly prominent.Building energy consumption plays an important role in the energy consumption of the entire society.Therefore,strengthening the energy consumption control of various buildings is an inevitable requirement to alleviate the energy consumption problem of the whole society.Building energy consumption is dominated by the energy consumption of the air-conditioning system.This paper proposes a competitive deep reinforcement learning method based on sensitivity strategies and buffer sampling prior to experience playback,and uses this to predict building air-conditioning energy consumption.Compared with the traditional depth Reinforcement learning methods have achieved better prediction results.This article has conducted an in-depth study on the following issues:(1)In the traditional deep reinforcement learning process,the actual effect of the action selection strategy is relatively general,the sample utilization rate of the experience pool is low,and the network generalization ability is poor.This paper proposes a strategy based on sensitivity and buffer sampling.Competitive deep reinforcement learning method(SA-BDDQN,Sensitive Action-Buffer Dueling DQN)that prioritizes experience playback.Competitive deep reinforcement learning methods focus on the value of the state itself,which helps to objectively evaluate actions.Therefore,this article adds sensitivity analysis to each state,establishes the relationship between state sensitivity and ε-greedy action selection strategy,and adjusts the intensity of ε exploration and utilization according to the sensitivity.Thereby enhancing the quality of the model’s action exploration and improving the training efficiency of the competitive deep reinforcement learning model.Set up a virtual experience pool to cache high-priority samples with too high sampling times,and give lowpriority samples a greater probability of being selected for training,which improves the utilization of training samples and strengthens the generalization performance of the model.(2)Apply the above-mentioned competitive deep reinforcement learning method based on the sensitivity strategy and buffer sampling priority experience playback method to the air-conditioning energy consumption prediction of an office building in Bangkok,and select relevant factors related to the air-conditioning energy consumption as the model input.And perform outlier processing on the data.Experimental results show that this method has higher prediction accuracy than traditional DQN and Dueling DQN.Finally,by adjusting the action interval of the model,the influence of the action interval on the prediction result of the model is analyzed.(3)Aiming at the problem that the traditional building energy consumption prediction model is small in use and extremely dependent on re-modeling,this paper combines the transfer reinforcement learning algorithm with the competitive deep reinforcement learning method based on the sensitivity strategy and the buffer sampling priority experience playback method.The model used to predict the air-conditioning energy consumption of an office building in Bangkok was migrated to an office building in Singapore and New York.The results show that the migrated model can have a good prediction accuracy without a lot of re-training.This method will provide ideas for the universal application of building energy consumption prediction models.In general,on the basis of existing deep reinforcement learning,this paper introduces a sensitivity strategy for competitive deep reinforcement learning network models and prioritizes buffer sampling experience playback,and proposes a deep reinforcement with more stable and faster convergence performance.Learning algorithm SA-BDDQN.Applying it to the prediction of building air-conditioning energy consumption,the results show that the method in this paper can accurately realize the prediction.Finally,by transferring the method to the target task,the superiority of the migration model is proved.