Season-Based Occupancy Prediction in Residential Buildings Using Machine Learning Models

Abstract

• Data mining-based framework to predict occupancy accounting variation corresponding to different seasons is proposed. • Prediction performances of six machine learning algorithms are studied in terms of accuracy and computational time. • The recursive feature elimination with cross-validation technique is used for selecting optimal inputs. • The developed framework is generic and can be extended to a wide range of residential buildings. A reliable occupancy prediction model plays a critical role in improving the performance of energy simulation and occupant-centric building operations. In general, occupancy and occupant activities differ by season, and it is important to account for the dynamic nature of occupancy in simulations and to propose energy-efficient strategies. The present work aims to develop a data mining-based framework, including feature selection and the establishment of seasonal-customized occupancy prediction (SCOP) models to predict the occupancy in buildings considering different seasons. In the proposed framework, the recursive feature elimination with cross-validation (RFECV) feature selection was first implemented to select the optimal variables concerning the highest prediction accuracy. Later, six machine learning (ML) algorithms were considered to establish four SCOP models to predict occupancy presence, and their prediction performances were compared in terms of prediction accuracy and computational cost. To evaluate the effectiveness of the developed data mining framework, it was applied to an apartment in Lyon, France. The results show that the RFECV process reduced the computational time while improving the ML models’ prediction performances. Additionally, the SCOP models could achieve higher prediction accuracy than the conventional prediction model measured by performance evaluation metrics of F-1 score and area under the curve. Among the considered ML models, the gradient-boosting decision tree, random forest, and artificial neural network showed better performances, achieving more than 85% accuracy in Summer, Fall, and Winter, and over 80% in Spring. The essence of the framework is valuable for developing strategies for building energy consumption estimation and higher-resolution occupancy level prediction, which are easily influenced by seasons.