Using improved support vector regression to predict the transmitted energy consumption data by distributed wireless sensor network

Abstract

Massive energy consumption data of buildings was generated with the development of information technology, and the real-time energy consumption data was transmitted to energy consumption monitoring system by the distributed wireless sensor network (WSN). Accurately predicting the energy consumption is of importance for energy manager to make advisable decision and achieve the energy conservation. In recent years, considerable attention has been gained on predicting energy use of buildings in China. More and more predictive models appeared in recent years, but it is still a hard work to construct an accurate model to predict the energy consumption due to the complexity of the influencing factors. In this paper, 40 weather factors were considered into the research as input variables, and the electricity of supermarket which was acquired by the energy monitoring system was taken as the target variable. With the aim to seek the optimal subset, three feature selection (FS) algorithms were involved in the study, respectively: stepwise, least angle regression (Lars), and Boruta algorithms. In addition, three machine learning methods that include random forest (RF) regression, gradient boosting regression (GBR), and support vector regression (SVR) algorithms were utilized in this paper and combined with three feature selection (FS) algorithms, totally are nine hybrid models aimed to explore an improved model to get a higher prediction performance. The results indicate that the FS algorithm Boruta has relatively better performance because it could work well both on RF and SVR algorithms, the machine learning method SVR could get higher accuracy on small dataset compared with the RF and GBR algorithms, and the hybrid model called SVR-Boruta was chosen to be the proposed model in this paper. What is more, four evaluate indicators were selected to verify the model performance respectively are the mean absolute error (MAE), the mean squared error(MSE), the root mean squared error (RMSE), and the R-squared (R2), and the experiment results further verified the superiority of the recommended methodology.