*Important notice: This news reports on an unedited version of the paper which has been accepted. and is awaiting final editing. Scientific Reports sometimes publishes preliminary scientific reports that are not fully edited and, therefore, should not be regarded as conclusive or treated as established information.
Hybrid AI model combining Lasso regression and random forests improves appliance energy prediction using sensor data. This approach enhances accuracy, efficiency, and interpretability in low-energy smart building systems.
Study: Hybrid Lasso-random forest framework for energy prediction using wireless sensor networks in low-energy buildings. Image Credit: FOTOGRIN/Shutterstock
A recent study published in Scientific Reports presents a hybrid machine learning framework designed to improve appliance energy consumption prediction in low-energy residential buildings. The research combines Lasso regression and random forest algorithms into a unified model called Lasso-RF-Net. The study demonstrates that the hybrid model delivers higher prediction accuracy, lower computational cost, and better interpretability than several conventional machine learning methods.
Challenges of Smart Building Energy Prediction
Accurate prediction of appliance energy consumption plays an increasingly important role in the building and construction sector. Low-energy buildings depend on smart monitoring systems to improve energy efficiency, lower operational costs, and maintain occupant comfort. Wireless sensor networks continuously collect real-time information on indoor temperature, humidity, occupancy patterns, and outdoor weather conditions.
Traditional linear regression models cannot effectively capture complex interactions. Advanced machine learning approaches, including random forests and deep neural networks, can model nonlinear patterns more effectively, but they often require more computational power and offer limited interpretability.
Although several hybrid frameworks combine multiple machine learning techniques, most rely on model stacking rather than clearly separating linear and nonlinear learning processes. Researchers developed Lasso-RF-Net, a two-stage hybrid regression framework designed specifically for high-dimensional building energy datasets to overcome the limitations. The researchers evaluated the framework using a real-world low-energy residential building dataset collected in Belgium.
Hybrid Modeling Framework and Experimental Methodology
The proposed Lasso-RF-Net framework combines sparse linear modeling with nonlinear residual correction to improve appliance energy prediction in low-energy buildings. The researchers designed the model as a two-stage learning framework.
In the first stage, Lasso regression selected the most important variables by shrinking less relevant coefficients toward zero. In the second stage, the selected variables were passed to a random forest model, which captured the nonlinear relationships and residual patterns that the linear model could not explain.
This combination allowed the framework to balance prediction accuracy, computational efficiency, and model transparency. The study used the appliances energy prediction dataset collected from a passive low-energy residential building in Belgium.
Wireless sensors installed across different indoor zones, including the kitchen, living room, laundry, office, bathroom, and bedrooms, continuously recorded temperature and humidity data every 10 minutes. The dataset also included outdoor weather variables such as wind speed, humidity, pressure, visibility, and dew point temperature. In total, the monitoring campaign generated 19,735 data samples across 26 variables.
Researchers applied 10-fold cross-validation and used performance metrics including Mean Squared Error, Root Mean Squared Error, Mean Absolute Error, and R² to ensure reliable evaluation. Additional statistical analysis, residual diagnostics, correlation heatmaps, and partial dependence plots helped evaluate model behavior and nonlinear environmental relationships.
Improved Prediction Accuracy and Key Building Insights
The experimental results demonstrated that Lasso-RF-Net significantly improved energy prediction accuracy compared with conventional machine learning models. The hybrid framework achieved a testing Mean Squared Error (MSE) of 3999.6 and an R2 value of 0.620, higher than the standard lasso regression and exhaustive subset selection methods. These conventional linear approaches struggled to capture the complex nonlinear relationships present in the building energy dataset.
Although Deep Neural Networks produced slightly higher R2 values, Lasso-RF-Net delivered similar predictive performance with much lower computational cost. The hybrid model completed training in only 6.4 seconds, while the deep neural network required more than 32 seconds. This result highlights the framework’s ability to balance prediction accuracy, computational efficiency, and interpretability.
Feature importance analysis identified indoor humidity and temperature conditions as major drivers of appliance energy consumption. Among the outdoor weather parameters, humidity and wind speed showed the strongest impact on energy demand. These relationships align closely with established building physics principles. Higher wind speeds increase air infiltration and heat loss, while humidity levels directly influence heating, ventilation, and air conditioning (HVAC) loads and indoor comfort requirements.
Download the PDF of this page here
Partial dependence analysis further confirmed the physical relevance of the model outputs. Increased lighting usage consistently raised predicted appliance energy demand. Humidity variables displayed nonlinear U-shaped patterns, with both extremely low and high humidity levels increasing energy consumption due to additional HVAC operation. Temperature variables also showed curved response trends associated with heating and cooling loads.
Residual diagnostics showed clear improvements after integrating the random Forest residual-learning stage into the framework. The hybrid model reduced heteroscedasticity and minimized the systematic residual patterns observed in the standalone Lasso model. These findings confirmed that the Random Forest component successfully captured nonlinear interactions that the linear regression stage could not model effectively.
The researchers also evaluated Lasso-RF-Net on several benchmark datasets, including aircraft control systems, crime prediction, and Parkinson’s disease monitoring data. The framework consistently maintained strong prediction accuracy and computational efficiency across these diverse applications, demonstrating its ability to handle complex high-dimensional regression problems beyond building energy prediction.
Implications for Smart Construction and Sustainable Building Design
The study represents an important step toward integrating hybrid machine learning, wireless sensor networks, and predictive analytics into smart building energy management. The results confirms that hybrid machine learning frameworks can improve prediction accuracy while maintaining computational efficiency and model interpretability. It successfully captures both linear and nonlinear relationships framework within complex building datasets generated by indoor sensors and weather monitoring systems.
The results also identify key environmental variables, including kitchen humidity, laundry-room temperature, outdoor humidity, and wind speed, as major contributors to appliance energy consumption. These findings strengthen the connection between machine learning predictions and established building physics principles.
Future research should focus to explore localized weather sensing, longer-term monitoring, and improved sensor placement strategies to enhance prediction reliability further. Overall, the study presents a scalable and interpretable pathway toward AI-driven energy optimization and smarter building management systems for future sustainable built environments.
Journal Reference
Li, G., Luo, Y., et al. (2026). Hybrid Lasso-random forest framework for energy prediction using wireless sensor networks in low-energy buildings. Scientific Reports. DOI: 10.1038/S41598-026-47935-3, https://www.nature.com/articles/s41598-026-47935-3
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.