Smart fusion of sensor data and human feedback for personalized energy-saving recommendations

Despite the variety of sensors that can be used in a smart home or office setup, for monitoring energy consumption and assisting users to save energy, their usefulness is limited when they are not properly integrated into the daily activities of humans. Energy-saving applications in such environments can benefit from the use of sensors and actuators when data are properly fused with previous knowledge about user habits and feedback about current user preferences. In this article, we present an online recommender system implemented in the EM3 platform, a platform for Consumer Engagement Toward Energy-Saving Behavior. The recommender system uniquely fuses sensors' data with user habits and user feedback and provides personalized recommendations for energy efficiency at the right moment. The user response to the recommendations directly triggers actuators that perform energy-saving actions and is recorded and processed for refining future recommendations. The EM3 recommendation engine continuously evaluates the three inputs (i.e. sensor data, user habits, user feedback) and identifies the micro-moments that maximize the need for the recommended action and thus the recommendation acceptance. We evaluate the efficiency of the proposed recommender system, which is based on a stacked-LSTM for fusing multi-sensor data streams, in several scenarios, and the observed accuracy on predicting the right moment to send a recommendation to the user ranged from 93% to 97%.

Automated summary

This article presents an online recommender system that fuses sensor data with user habits and feedback to provide personalized energy-saving recommendations at the right moment for users to take energy-saving actions. The system continuously evaluates sensor data, user habits, and feedback to identify micro-moments that maximize the need for recommended action, thus increasing recommendation acceptance. The proposed system, based on stacked-LSTM for fusing multi-sensor data streams, achieved an accuracy range of 93% to 97% in predicting the right moment to send recommendations to users in various scenarios.