Building Energy Consumption Prediction Based On Ensemble Learning

With the further deepening of climate change and environmental issues,more and more attention has been paid to the current trend of energy consumption.Construction-related activities accounts for more than one-third of the global end-use energy consumption.With the urbanization of developing countries,it is expected that annual global energy consumption will continue to increase for some time in the future.In order to achieve the goal of "carbon neutralization" in 2060,a series of energy-saving measures are needed to improve the energy efficiency of buildings.Accurate and reliable prediction of building energy consumption is the most important part of these energy-saving measures.It is not only an important tool to assess energy-saving potential in the process of building design and renovation,but also an important component of intelligent buildings.With the wide application of BAS and IOT technology in buildings,a large number of smart-meters and sensors are deployed in smart buildings,and the data volume and dimension of building are increasing,which bring great challenges to building energy consumption prediction.At present,in the field of building energy consumption prediction and time series prediction,ensemble learning method and RNN are popular.The research work of this thesis will be carried out based on these two kinds of learning algorithms,mainly including the following contents:Firstly,due to the lack of comparative research on the prediction performance of different ensemble model algorithms in the current research on building energy consumption prediction based on ensemble model algorithms,this thesis uses the stacking ensemble to integrate five base model algorithms including Bagging and Boosting.The prediction performance of Stacking,Bagging and Boosting are compared through experiments.The experimental results show that the prediction performance of Stacking is better than Bagging and Boosting.In order to improve the prediction performance of the building energy consumption prediction model,this thesis proposes Stacking-LSTM,which uses LSTM to learn the meta features extracted from five base models（including KNN,SVR,MLP,RF,XGBoost）,mines the time-series features in the meta features,and use the time-series features to achieve better prediction effect.In order to verify the prediction performance of Stacking-LSTM,the prediction model is established on the energy consumption simulation data sets of three different types of buildings.The prediction accuracy of the model is evaluated through the comparison experiment with four baseline models,and the prediction stability of the model is evaluated through repeated experiments.The prediction performance of the model is evaluated by comprehensively considering prediction accuracy and prediction stability.The experimental results show that the prediction accuracy of Stacking-LSTM is the highest on the three data sets,and the prediction stability is significantly improved compared with the three recurrent neural network structures,which can meet the building energy consumption prediction tasks of hospitals,hotels and office buildings.