Research On Method And Application Of Building Energy Consumption Prediction Based On Transfer Learning

Building energy conservation is an important part of energy conservation and emission reduction,and the accurate prediction of building energy consumption is a prerequisite for building energy conservation,the data-driven method(namely the black-box model)is an effective means of building energy consumption prediction.However,in actual engineering application and scientific research,on the one hand,the application of the black-box model is often restricted by insufficient energy consumption data of the target building.On the other hand,the historical energy consumption data of a large number of buildings has been shelved.In the absence of data in the target building,in order to make full use of the potential value of the historical energy consumption data of existing buildings to solve the problem of energy consumption prediction,this paper uses the idea of transfer learning to propose corresponding building energy consumption prediction method for three different task scenarios.On this basis,using historical energy consumption data of related buildings as a carrier,the proposed method is applied for research.First,in order to make full use of the historical energy consumption data of the same type buildings to predict the energy consumption of buildings lacking data,this paper combines deep learning with model-based transfer learning to propose a deep learning-based building energy transfer prediction method.With the help of measured energy consumption data of residential building,office,hotel and shopping mall,the method is applied to analysis of boundary condition and research of application.The results show that the longer the input supervision window and the more fine-tuning layers,the higher the model prediction accuracy will get,and that the application of this method increases the prediction accuracy of energy consumption of the target building by 3.44% to 74.19%.In addition,this paper also analyzes the prediction results based on the energy consumption characteristics of the building itself.It is found that the more relevant,the stronger the periodicity,and the more stable the trend of energy consumption between the target building and the relocated building,the higher the accuracy of the prediction.Secondly,for the problem of building energy consumption prediction without data at all,this paper integrates historical energy consumption data of the same type buildings,De ST software simulation data and deep learning algorithms,and proposes a gray-box model-based building energy transfer prediction method.And with the help of the measured data of residential buildings in Beijing and Shanghai,the method was verified.Results show that,compared with single physical modeling method,this method can effectively improve the prediction accuracy by 18.43% to28.11%.Finally,In order to further fully tap the value of existing building energy consumption data,the historical energy consumption data of different type buildings are used to predict the energy consumption of buildings lacking data.,This paper combines the gated recurrent unit algorithm with the domain adversarial neural network and proposes the GRU-DANN-R model.and based on this model,proposes a cross-building energy transfer prediction method.Then,the method was validated with the help of measured energy consumption data of office,hotel and shopping mall building.Studies have shown that this method can increase the prediction accuracy by 3.43% to 9.35%.In addition,this paper also analyzes the parameter sensitivity of the model.It is found that when the length of the data supervision moving window is set to 24,week label is added,the classification loss function weight coefficient is set to 0.2,and Adam optimizer and regularization are used,the model can obtain the best prediction results.