The Prediction Method And Application Of Building Energy Consumption Based Transfer Learning

Currently,with the society rapidly developing,buildings have become the second largest main body of energy consumption in China,and the ratio of building energy consumption in total energy consumption is increasing year by year.In order to achieve the goal of carbon emission reduction in China in 2030,it is of great significance to promote energy conservation and emission reduction of buildings.Realizing the refined management of building energy and rational scheduling of building equipment is an important link to promote building energy conservation and emission reduction,among which accurate prediction of building energy consumption is one of the most critical tasks.At present,in the field of time series forecasting,researchers have proposed deep forecasting models such as recurrent neural networks and convolutional neural networks,which have been applied to energy consumption forecasting and achieved good results.However,the deep models have the characteristics of long training time,reliance on a large amount of historical data and complex calculation,which result in a acutely increase in the time and computing power cost of constructing a model for each building and training it separately.Therefore,a new training strategy needs to be proposed,which can not only reduce the training cost of the deep model,but ensure the prediction performance of the deep model.The application of transfer learning in building energy consumption prediction is studied in this thesis.The specific contributions of this thesis are as follows:（1）A method for predicting building energy consumption based on domain adaptation is presented.In this method,the feature extraction model is built with a 1-dimensional convolutional network and GRU network.In addition,the model parameters are optimized through adversarial training of inter-domain distribution loss and prediction loss to ensure the prediction performance of the model on target data.During the model training process,the inter-domain distribution loss is jointly measured by CORAL loss and MMD criterion,which improves the generalization ability of the reused model.In order to test the performance of the proposed prediction method,it is applied to the prediction of energy consumption for office buildings and commercial buildings,and compared with Da NN method and inter-domain loss measurement method which only uses CORAL loss.Experimental results show that the proposed method can reuse the model for the data with different distributions,and the reused model can more accurately predict the energy consumption according to a small amount of target data.（2）In order to further reduce the cost of model training,a method for predicting building energy consumption based model transfer is presented.This method first builds and pre-trains different energy consumption prediction models with long short-term memory network（LSTM）in the cloud,and the dynamic time warping similarity measurement method is utilized to analyze the similarity of inter-domain data with different data dimensions.The similarity analysis reduces the distribution difference between the data and ensures the prediction performance of the reuse model on the target data.On the basis of similarity analysis,on edge devices,the reused model is optimized using the fine-tuning method in model transfer.For verifying the advantages of this method,experiments are also carried out on two types of building energy consumption datasets,and compared with other methods.It is shown by the experimental results that this method alleviates the burden of network training and achieves more accurate predictions with a small number of samples.（3）To relieve the training burden of edge devices,the development of cloud-edge collaborative energy consumption prediction is implemented on Jetson Nano.This scheme stores the existing prediction model in the cloud,and realizes the reuse of the cloud model in the energy consumption prediction of the target building at the edge based the cloud-edge collaborative architecture and model transfer and fine-tuning method.In order to make real-time prediction of energy consumption data,FTP protocol is adopted to complete data transmission and reuse model download functions,and a display interface is developed to display target data information and prediction results.