Research On The Prediction Of Building Energy Consumption Based On Converted Prediction Space

With population growth and economic development,the energy consumption of the construction industry has increased year by year.Excessive building energy consumption not only exacerbates the energy shortage but also emits a large amount of carbon dioxide,which seriously hinders the goal of carbon peak and neutrality.Therefore,it is very important to realize building energy efficiency.As the essential foundation of building energy efficiency,energy consumption prediction plays an irreplaceable role in obtaining the optimal strategy of control equipment,planning energy dispatching,and evaluating the operation of the building system.Deep reinforcement learning(DRL),which combines the perception ability of deep learning(DL)with the decision-making ability of reinforcement learning(RL),can deal with complex problems in high-dimensional continuous state space,so it has attracted extensive attention in the field of building energy consumption prediction.When using DRL algorithms to predict building energy consumption,algorithms with discrete action space can take fewer computing resources and less time to converge,but obtain lower prediction accuracy than ones with continuous action space.One important reason is that the prediction accuracy of algorithms with discrete action space is closely related to the prediction space of building energy consumption.When the meaning of prediction space has not been changed,the algorithms with discrete action space is easy to fall into local optimum in large prediction space,which leads to lower prediction accuracy.Based on the Deep Q Network(DQN)in algorithms with discrete action space,this paper studies the building energy consumption prediction by changing the size and meaning of the prediction space.The main contents include three parts:(1)A shrunken-prediction-space DQN(S-DQN)algorithm predicting building energy consumption was proposed.In this algorithm,the original prediction space is divided into multiple subspaces.Deep forest(DF),which is employed to classifier states,is introduced to represent the different subspaces uniformly,then the shrunken prediction space is generated.Meanwhile,state probabilities obtained by DF and original states are employed to construct new states,which can improve the robustness of the S-DQN algorithm.The experimental results show that the building energy consumption prediction model,which is constructed by the S-DQN algorithm with 15 state classes,has the highest accuracy and outperforms other algorithms.(2)A transformed-prediction-space DQN(T-DQN)algorithm predicting building energy consumption was proposed.Firstly,the original prediction space is transformed based on the first order difference of energy consumption function.Then the treated space is transformed twice by using the principle of set closure and DF.Finally,the symbol set and transformed prediction space replace the original prediction space and are employed to construct the prediction model.The results show that the T-DQN algorithm can achieve higher prediction accuracy than the S-DQN algorithm.(3)A transformed and shrunken-prediction-space DQN(TS-DQN)algorithm predicting building energy consumption was proposed.In TS-DQN algorithm,the transformed prediction space is generated firstly based on the first difference method,the principle of set closure,and DF.Then the transformed space is divided into several subspaces,and DF was employed to denote these subspaces uniformly.Finally,the transformed and shrunken prediction space is generated and used to construct a model to predict the building energy consumption.The experimental results show that the TS-DQN algorithm with 8 state classes has the highest prediction accuracy among other algorithms such as S-DQN and T-DQN,which demonstrates its superiority in building energy consumption prediction.To overcome the defects of algorithms with discrete action space in building energy consumption prediction,this study constructs new models by shrinking the size of the original prediction space,transforming the meaning,and shrinking the transformed prediction space.The results verify the potential of discrete action space algorithms in the field of building energy consumption prediction.