Electric-Vis: Integrating Wearable Sensing with Residential Energy Data for Improved Energy Forecasting

• Gopi Krishna Melam, Computer Science, University of Delaware
• Zahra Rezaei, Computer Science, University of Delaware

• Matthew Mauriello, Computer & Information Sciences, University of Delaware

Residential energy consumption continues to increase over time and is strongly influenced by everyday human behaviors and lifestyles. However, most existing residential electricity forecasting approaches primarily rely on past power usage and weather data, while largely overlooking wearable sensing data as a source of information about occupant behavior. Thus, the relationship between occupant behavior and household energy use remains underexplored, limiting opportunities to develop personalized feedback and intervention strategies. To address this gap, Electric-Vis is a web-based platform that integrates wearable sensing with residential energy data to make these relationships visible, interactive, and actionable, helping families better understand their energy use and make more sustainable decisions. We implemented a multi-user platform capable of collecting intraday wearable data to support the research goals and enable large-scale studies across multiple households. The platform preserves participant privacy while supporting real-world household scenarios. 

Additionally, the platform collects household energy data at 15-minute intervals from Emporia alongside wearable data from Fitbit, including steps, heart rate, and sleep, providing a detailed representation of occupant behavior. We created a comprehensive dataset by backfilling more than six months of historical wearable and energy data, providing the foundation for developing and evaluating machine learning models for residential energy forecasting. Using XGBoost with traditional forecasting features, including historical energy consumption and weather, together with wearable data, we are assessing the potential improvement in forecasting accuracy beyond the previously established predictors. Preliminary results indicate that approximately 25% of the model’s predictive importance is attributed to wearable-derived features, suggesting that human activity plays a meaningful role in forecasting residential energy demand. With a further optimized model and more feedback from participating families, we will focus on refining the energy dashboard with visualizations and intervention strategies most likely to encourage reductions in household energy consumption.