Wrapper feature selection based multiple logistic regression model for determinants analysis of residential electricity consumption

Abstract

Improvements of energy efficiency and reduction of electricity consumption can be promoted by growing knowledge on the determinants of residential electricity consumption level (RECL). Due to numerousness, complexity and multiple correlations among impact factors (IFs) of RECL, feature selection is an essential step to ensure the precision and stability of an explanatory model. However, the current linear feature selecting models such as stepwise regression are more likely to remove the determinants that have strong-nonlinear association with RECL, which is caused by the limited ability to merely capture the linear relationship between factors. To address this issue, wrapper feature selection that combines genetic algorithm (GA) and multiple logistic regression (MLR) based classifier is proposed in this paper to find out the optimal feature subset (FS). GA is applied to generate better FS from the current to the next iteration and the forecasting accuracy of classifier is utilized to further evaluate the newly reproduced FS. Furthermore, a new relationship measuring approach called maximal information coefficient (MIC) is introduced to verify the validity of feature selecting results. Finally, an explanatory model based on MLR is established to explore the influence mechanism of selected IFs on RECL. Case studies for the proposed methods were conducted using both smart metering and survey dataset of over 3300 households in Ireland. The results not only identified the significant determinants of RECL, but also provided detailed information about how these selected IFs affect residential electricity consumption behaviors.