Public Building Energy Consumption Prediction Modelling And Optimization Based On Data Mining

With the increasingly prominent problems of energy and environment,sustainable development has become a trend.Building energy consumption accounts for nearly one-third of the global total energy consumption,and currently extensive management of building energy consumption implies a great energy-saving potential in China.The prediction and analysis of building energy consumption is an important part of building energy conservation.Based on several public buildings in Shanghai,this paper summarizes the techniques and processes of energy consumption forecasting modeling for public buildings based on machine learning methods,and compares the forecasting results of models based on K-nearest neighbor,support vector machine,artificial neural network and random forest algorithms in different buildings and seasons.The modelling can be divided in into seven steps:collection of data,data processing and analysis,division of training and testing datasets,selection of input variables,adjustment of parameters,training and output and finally validation and evaluation of model.The test of model training set shows that the minimum training sample size required by different modeling methods is different,but in general,reasonable prediction of building energy consumption in shopping malls requires at least one to two weeks of historical data for training.In the process of sensitivity analysis of the model,it is found that the parameters have a significant impact on the performance of the model.For example,the penalty factor in support vector regression model can be reduced from 23.7% to 8.3% by changing the parameter only.In addition,the effect of time input on the performance of the model is also obvious.The comparative analysis of energy consumption of shopping malls with different locations shows that when historical energy consumption data are not taken as input variables,the model based on K-nearest neighbor method usually has better prediction results.Adding the energy consumption of the previous hour as the input of the model can significantly reduce the average absolute percentage error of building energy consumption prediction,which can be reduced by up to 7%.Support vector regression model is more suitable for short-term energy consumption forecasting,and its one-week energy consumption prediction error can be as low as 2.6%.Artificial neural network model is more advantageous in long-term energy consumption prediction,and the average absolute percentage error of twomonth hourly energy consumption prediction can be as low as 2.7%.At the same time,the accuracy of building energy consumption prediction will also be affected by the season and the energy consumption curve of the building itself.In addition,this paper proposes a method combining genetic algorithm with artificial neural network algorithm to optimize the parameters of the artificial neural network model,which further improves the accuracy of building energy consumption prediction.Taking an office building in Shanghai as an example,the optimized model reduces the daily hourly energy consumption prediction error from 5.1% to 4.8%,and the monthly overall prediction error from 7.7% to 6.8%.