Multimodal feature fusion and ensemble learning for non-intrusive occupancy monitoring using smart meters

• The main contributions of this study are as follows: • Multimodal feature fusion for non-intrusive occupancy monitoring (MMF-NIOM). • Sequence-to-sequence-to-point (s2s2p) to capture household load events. • Deep, statistical, and datetime features integration for real-time occupancy detection. • Highly correlated devices for s2s2p enhance occupancy classification. • Novel NIOM-Net classification model within s2s2p. Non-intrusive detection of household occupancy status using smart meter data presents challenges due to the intricate relationship between user behavior, energy usage, and occupancy. Effective occupancy classification relies on real-world factors beyond load consumption. In this study, we introduce the multimodal feature fusion for non-intrusive occupancy monitoring (MMF-NIOM) framework, which leverages both classical and deep machine learning algorithms to achieve state-of-the-art occupancy detection performance using smart meter data. Three modes of feature extraction are employed within MMF-NIOM. Using sequence-to-sequence-to-point (s2s2p) learning, we capitalize on cascaded deep learning systems to associate load-switching events with occupancy changes. Additionally, statistical features from aggregated smart meter data help correlate load consumption magnitude and patterns with occupancy labels. To gain deeper insights into user occupancy patterns over time, we integrate datetime information into the multimodal feature set. We also demonstrate that eliminating redundant devices during s2s2p learning further enhances occupancy detection. We combine features from all three modes of MMF-NIOM to achieve a state-of-the-art non-intrusive occupancy classification performance of 91.5 % accuracy and 91.5 % f1-score, approximately, by an ensemble of fine-tuned classifiers on the electricity consumption & occupancy (ECO) dataset. The proposed method is sustainable, robust, adaptable to various households, and can be mass-implemented within smart meters at a much lower cost and effort compared to the traditional internet of things (IoT)-based intrusive systems.