Data-driven decision support for low-carbon retrofits: Balancing heat loss and indoor environmental risks using interpretable machine learning

Abstract

Real estate decarbonization mandates necessitate improved building envelope thermal efficiency. These physical retrofits inadvertently restrict natural ventilation, increasing the risks of indoor dampness and poor air quality. Addressing this operational trade-off, this research proposes a quantitative framework integrating IoT sensor telemetry with static asset data. Formulated as a multi-label classification problem predicting three concurrent risks, the study evaluates a Random Forest model on a pilot portfolio of 81 residential units, achieving an average localized test Area Under the Curve (AUC) of 0.751. Spatial cross-validation yields a baseline AUC of 0.62, reflecting the physical complexity of generalizing predictive models to unseen architectural archetypes lacking historical telemetry. Applying the SHapley Additive exPlanations (SHAP) algorithm interprets model predictions, distinguishing between physical envelope degradation and occupant ventilation behaviors based on feature contributions. The framework maps these predictive probabilities into a prescriptive four-quadrant intervention matrix, designed to inform targeted capital and operational budget allocations. Longitudinal emission projections to 2040 quantify the environmental impact of delayed funding. Asset managers can use these outputs to balance energy efficiency targets with indoor environmental quality standards and justify accelerated green financing.