TY - JOUR
T1 - Feature extraction and resident number prediction method using power consumption data
AU - Yoshida, Masahiro
AU - Imanishi, Tomoya
AU - Nishi, Hiroaki
N1 - Funding Information:
The prediction on the resident’s number can be applied, for example, to the targeting algorithm of the ads for several services and products in real estate and the construction industry, for example. Especially, since demanding services or products for people living alone is different from those who are not, the prediction performance for classification of “1 person” and “2 or more people” is important. In our additional evaluation, the maximum accuracy for the two-class classification of “1 person” and “2 or more people” was 78.6% with RF, which would be useful for the targeting ads. Acknowledgement This work was supported by Technology Foundation of the R&D project “Design of Information and Communication Platform for Future Smart Community Services” by the Ministry of Internal Affairs and Communications of Japan, grants from the Project of the Bio-oriented Technology Research Advancement Institution, NARO (the research project for the future agricultural production utilizing artificial intelligence), and MEXT/JSPS KAKENHI Grant (B) Number JP16H04455 and JP17H01739, respectively.
Publisher Copyright:
© 2019 The Institute of Electrical Engineers of Japan.
PY - 2019
Y1 - 2019
N2 - A large amount of power consumption information generated from private houses is being aggregated nowadays, particularly by the spread of smart electric meters. Applications that utilize these data are widely studied, and several services have been proposed. In order to utilize the power demand information effectively, an appropriate feature extraction and selection method is necessary. In this paper, a feature extraction method for power consumption information is proposed. Extracted features are used to predict the “number of household members (number of residents)” using typical machine learning algorithms, namely, Support Vector Machine (SVM), k-Nearest Neighborhood (k-NN), and Random Forest (RF). The number of residents represents significant information for the marketing departments of several industries such as real estate and construction industries. The proposed feature extraction method consists of two steps: feature variable generation and feature variable selection. In the feature variable generation step, we have used both fundamental statics and an ARMA model to generate 33 feature variables. In the feature variable selection step, the extracted feature variables are first ranked by applying Analysis of Variance (ANOVA). An appropriate feature variable set is selected by assessing several combinations of the 33 features, based on proposed extended algorithm of Recursive Feature Elimination (RFE). Our overall feature extraction method is evaluated based on the prediction accuracy using extracted feature variables. Compared with the accuracy using feature variables extracted by conventional methods, the accuracy is improved by 6.78%, 4.98%, and 8.11% for k-NN, SVM, and RF, respectively, and we have successfully proven validity of our proposition.
AB - A large amount of power consumption information generated from private houses is being aggregated nowadays, particularly by the spread of smart electric meters. Applications that utilize these data are widely studied, and several services have been proposed. In order to utilize the power demand information effectively, an appropriate feature extraction and selection method is necessary. In this paper, a feature extraction method for power consumption information is proposed. Extracted features are used to predict the “number of household members (number of residents)” using typical machine learning algorithms, namely, Support Vector Machine (SVM), k-Nearest Neighborhood (k-NN), and Random Forest (RF). The number of residents represents significant information for the marketing departments of several industries such as real estate and construction industries. The proposed feature extraction method consists of two steps: feature variable generation and feature variable selection. In the feature variable generation step, we have used both fundamental statics and an ARMA model to generate 33 feature variables. In the feature variable selection step, the extracted feature variables are first ranked by applying Analysis of Variance (ANOVA). An appropriate feature variable set is selected by assessing several combinations of the 33 features, based on proposed extended algorithm of Recursive Feature Elimination (RFE). Our overall feature extraction method is evaluated based on the prediction accuracy using extracted feature variables. Compared with the accuracy using feature variables extracted by conventional methods, the accuracy is improved by 6.78%, 4.98%, and 8.11% for k-NN, SVM, and RF, respectively, and we have successfully proven validity of our proposition.
KW - ARMA model
KW - Analysis of variance (ANOVA)
KW - Feature extraction
KW - Feature selection
KW - Machine learning
KW - Smart electric meter
UR - http://www.scopus.com/inward/record.url?scp=85062404260&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85062404260&partnerID=8YFLogxK
U2 - 10.1541/ieejeiss.139.227
DO - 10.1541/ieejeiss.139.227
M3 - Article
AN - SCOPUS:85062404260
VL - 139
SP - 227
EP - 236
JO - IEEJ Transactions on Electronics, Information and Systems
JF - IEEJ Transactions on Electronics, Information and Systems
SN - 0385-4221
IS - 3
ER -