Conception, implémentation et évaluation d’un système de recommandation énergétique

Thesis Abstract

This study addresses the escalating energy consumption in multi-tenant buildings and the consequent need for personalized, actionable guidance to reduce carbon footprints without compromising occupant comfort. The problem centers on the low adoption rates of energy-saving practices due to suboptimal personalization and real-time feedback. The objective is to design, implement, and evaluate an intelligent energy-recommendation system that delivers adaptive, user-specific guidance for energy optimization in commercial buildings. Specific objectives include (1) to develop a data-driven recommendation engine that integrates building energy meters, occupant behavior signals, and weather data; (2) to implement a modular architecture enabling plug-and-play integration with existing Building Management Systems (BMS) and consumer devices; (3) to evaluate the system’s effectiveness in reducing energy intensity and peak demand; (4) to assess user acceptance and perceived usefulness to identify barriers and facilitators of adoption. The methodological approach combines design science research with empirical evaluation. The research adopts a mixed-methods design in three phases. Phase 1 involves a design and prototyping phase in which an energy-recommendation prototype is developed using a modular microservices architecture and deployed within a 12-floor commercial building pilot. Phase 2 entails a quasi-experimental field study over six months with two comparable buildings assigned as intervention (n=150 occupants) and control (n=150 occupants). Phase 3 comprises a follow-up qualitative evaluation to explore user experience and behavioral adaptation, using purposive sampling to conduct 20 in-depth interviews and 6 focus groups with occupants, facility managers, and system administrators. Data sources include smart meter data with hourly granularity, building occupancy logs, weather data from the local meteorological station, BMS events, and user interaction traces from the recommendation interface. Instruments include calibrated energy consumption dashboards, a validated user acceptance survey (Technology Acceptance Model adapted for energy guidance), semi-structured interview protocols, and a system log audit trail. Validity and reliability are ensured through triangulation, pretesting of instruments, pilot testing of the prototype, and inter-rater reliability checks for qualitative coding. Data analysis utilizes a combination of quantitative and qualitative techniques regression analyses (fixed-effects, panel data) to quantify energy savings attributable to the recommendations, time-series decomposition to identify trends and seasonality, and ANOVA to compare intervention and control groups. The recommender employs collaborative filtering augmented with content-based rules and a reinforcement-learning module to adjust recommendations based on user feedback and energy impact. The theoretical underpinning draws on the Technology Acceptance Model (TAM) to frame user adoption and the Theory of Planned Behavior (TPB) to elucidate behavioral intentions, complemented by the Information-Cowerdedness-Action framework to assess how information disclosure affects decision-making. A conceptual model links system inputs (metered energy, occupancy, weather), processing (prediction, optimization, advisories), outputs (personalized recommendations), and outcomes (energy savings, occupant satisfaction). Expected findings include statistically significant reductions in site-wide energy intensity (target ?12% annualized reduction) and peak demand (target ?8% during peak hours) in the intervention building relative to the control. It is anticipated that personalized, context-aware recommendations will yield higher perceived usefulness (mean TAM score ?4.0 on a 5-point scale) and greater behavioral intention to adopt energy-saving actions (TPB constructs showing positive shifts with p<0.05). Qualitative results are expected to reveal enablers such as clear feedback, trust in data provenance, perceived autonomy, and seamless integration with existing workflows, alongside barriers including notification fatigue and data privacy concerns. The study contributes to knowledge by operationalizing a scalable energy-recommendation architecture that blends machine learning with human-centered design, enriching the evidence base on real-time, personalized energy guidance in commercial settings. It advances understanding of how behavioral and technological factors interact to drive energy-saving practices and provides a tested blueprint for integrating recommender systems within BMS ecosystems. Policy and practice implications include recommendations for standardizing data interfaces, ensuring occupant privacy, and aligning incentive structures with energy-saving outcomes. Concluding, the research is expected to demonstrate that a modular, user-centric energy-recommendation system can achieve meaningful energy savings while maintaining occupant comfort and acceptance. Recommendations emphasize extending the prototype to different building typologies, exploring long-term behavioral adaptation, enhancing explainability of recommendations, and refining the governance framework for data management and system updates.