Research On The Method And Application Of Public Building Energy Consumption Forecast Based On Machine Learning

With the progress of urbanization,the energy consumption of public buildings is increasing year by year.In order to achieve the energy-saving goals of public buildings,it is necessary to conduct intelligent management on the basis of mastering the operating characteristics of buildings.As an effective means to reflect the operating characteristics of buildings,building energy consumption prediction can provide effective support for the intelligent management of buildings.Traditional physical model methods have problems such as high model complexity and long modeling period,while machine learning methods can predict building energy consumption more quickly and accurately by mining the laws of historical energy consumption data.Therefore,based on machine learning methods and historical energy consumption data of public buildings,this paper has carried out the following research work:First of all,experimental research is conducted on three types of typical public building cases such as offices,malls,and hotels.Based on the characteristics of machine learning,a building energy consumption prediction method and process are proposed.The data is then sorted and pre-processed.The input variables are determined for each type of building through the univariate model checking method.After that,three months of historical data are used to train the model and grid search was used to determine the model hyper-parameters during the training process.The four machine learning models of support vector machine,multi-layer neural network,random forest,and gradient enhancement tree are compared.The prediction results show that the support vector machine model performs well under most working conditions;the prediction evaluation index of the shopping mall construction case is the best,three The quarterly average MAPE indicator was 4.24%.Furthermore,through the analysis of the characteristics of the two network layers in deep learning,a data structure and three model structures suitable for deep learning algorithms and building energy consumption prediction requirements are proposed.Through fitting experiments on the annual data of the office building case,it is found that the row-wised convolutional neural network has achieved the best predictive performance,and the MAPE on the test set is 10.49%.This paper proposes a CNN-RNN model that has no obvious over-fitting phenomenon while obtaining high prediction accuracy;the test results on hotel buildings show that this model has good feature extraction capabilities.After completing the study of prediction methods,combined with actual engineering problems,the application method of public building energy consumption prediction model based on machine learning is proposed.The prediction curve of the model is supplemented by the confidence interval to form an energy consumption interval forecast,which reflects the range of energy consumption changes caused by fluctuations in building operation.In view of the decrease in model prediction performance caused by seasonal changes,a joint training program based on historical and quarterly data is proposed and the recognition of seasonal features is realized based on the classification model;a single-layer neural network layer is proposed to learn long-term trends,and the model is trained after short-term data transformation.Incorporating long-term trends into short-term models,the R~2 index for office building forecasts was increased from 0.95 to 0.98,which effectively improved the performance of the model when changing seasons.The clustering method is used to identify similar building groups,and the building with the median energy consumption index per unit area is used as a reference,and the energy consumption prediction curve of this building is used to guide the operation of other buildings in the building group.Using the working condition settings,energy consumption data and cooling load data in the tuning process of an office building in Shanghai,the improvement of system energy efficiency is judged by the difference between the predicted value and the actual energy consumption.Regression analysis shows that the difference between each 1000kWh corresponds to 0.3 energy efficiency improvement.Based on the formation of a personalized shallow machine learning model for public buildings,this research further obtained a more versatile CNN-RNN combined model based on deep learning.The proposed prediction application oriented to the management platform can effectively improve the accuracy of energy consumption prediction and provide guidance for the operation of public buildings.