Efficient monitoring and prediction of appliance energy consumption in low-energy houses are important for optimizing building performance and advancing smart energy management. Wireless sensor networks (WSNs) provide high-resolution indoor environmental data; however, the resulting datasets are often high-dimensional, redundant, and nonlinear, posing challenges for conventional regression models. This paper proposes a hybrid regression framework, termed Lasso-RF-Net, which combines linear feature selection and nonlinear adaptability in a computationally efficient manner. The first stage identifies a sparse linear structure and reduces dimensionality, while the second stage captures nonlinear interactions through residual learning. The proposed model was evaluated using a real-world low-energy house dataset incorporating indoor WSN measurements and outdoor weather variables, and further validated on three benchmark regression datasets. Results show that Lasso-RF-Net achieves the lowest testing mean squared error compared with Lasso, Random Forest, Subset Selection, and Deep Neural Networks, while maintaining a moderate computational cost. Feature analysis indicates that kitchen humidity and laundry-room temperature are dominant indoor predictors, whereas outdoor humidity and wind speed are the most influential weather variables. Overall, the proposed framework provides an accurate, computationally efficient, and interpretable solution for high-dimensional nonlinear energy prediction problems.
Li et al. (Mon,) studied this question.