Comparative Study of Energy Storage and Supply Strategies for Enhancing Low-Income Community Resilience in Texas During Cold Snaps
Extreme winter events in Texas have revealed major vulnerabilities in low-income housing, particularly in communities with minimal insulation and limited access to backup energy. This study evaluates the resilience of such households in South Texas under prolonged grid outages through three scenarios: (1) passive survivability of poorly insulated homes, (2) decentralized rooftop photovoltaic (PV) systems with battery storage, and (3) community-scale ground source heat pumps (GSHP), centralized batteries, and solar within a microgrid. Results show that Scenario 1 homes could not maintain habitable conditions, with indoor temperatures dropping below 10 °C(50°F), far under the 18 °C (64.4°F) health threshold. Scenario 2 improved resilience, especially when even minimal insulation was added, which reduced heat loss, improved comfort, and lowered PV and battery system costs. Scenario 3 provided the most robust outcome, sustaining indoor temperatures near 18 °C (64.4°F) and enabling peer-to-peer energy sharing across households. While requiring higher upfront investment, the community-scale system demonstrated greater long-term cost-effectiveness through economies of scale and equitable distribution of resilience benefits. The findings suggest that while household-scale PV and storage paired with insulation can deliver short-term improvements, integrated community-scale systems represent the most sustainable and equitable pathway to protect vulnerable populations during extreme winter events.
Privacy-First System Identification for Smart Homes Using Fully Homomorphic Encryption
The widespread use of smart thermostats has enhanced comfort and energy efficiency in residential environments but also created significant privacy risks due to the continuous collection of sensitive behavioral and environmental data. To address these concerns, this paper investigates the use of Fully Homomorphic Encryption (FHE) to enable privacy-preserving data modeling for predictive control in smart home systems. We adopt the XGrad optimization method to mitigate the high computational costs of encrypted learning and make FHE feasible for resource-constrained environments. Using real building data from the UMAR unit at Empa’s NEST facility, we show that encrypted system identification achieves near-perfect predictive accuracy (R2 = 0.98 training, 0.97 testing) with substantial efficiency improvements compared to standard encrypted gradient descent. For short-term predictions (15 minutes), encrypted and unencrypted models perform almost identically, while longer horizons show moderate error increases that remain within practical limits. These findings demonstrate that FHE, when combined with advanced optimization techniques, can deliver privacy-preserving smart thermostat control without sacrificing accuracy or responsiveness. The framework ensures end-to-end data confidentiality and directly addresses risks such as smart home-facilitated abuse, while remaining computationally efficient for real-world deployment in IoT ecosystems.
Optimizing Privacy-Utility Trade-offs for Safer Environmental Data Sharing Using Noise Injection and NSGA2
Smart thermostats and other Internet of Things (IoT) devices enhance energy efficiency and comfort but also raise serious privacy concerns by exposing sensitive information about occupants’ routines and behaviors. This study proposes a privacypreserving framework that combines calibrated noise injection with multi-objective optimization using NSGA-II to balance privacy and utility in environmental data sharing. Using the Occupancy Detection Dataset, the framework evaluates featurespecific noise sensitivity and identifies Pareto-optimal trade-offs between privacy budgets and model performance. Results show that data anonymization can be achieved with moderate accuracy costs, while balanced configurations maintain over 89% occupancy detection accuracy. These findings demonstrate that privacy and utility are not mutually exclusive, and that careful design enables safer data sharing in smart home environments.
From Energy Use to Building Physics: A Feasibility Study for Scalable Digital Twin Augmentation
This study presents an interpretable machine learning framework for predicting key building characteristics from dynamic energy simulation data. Using CESAR-P, a set of synthetic building profiles was generated by systematically varying year of construction (YOC), window-to-wall ratio (WWR), aspect ratio, orientation, and floor area. Linear autoregressive models with exogenous inputs (ARX) were then fitted to each energy time series, producing 36 lagged coefficients per building that capture the dynamic influence of past energy use, outdoor drybulb temperature, and solar irradiance. From these weight vectors, statistical summaries, including mean, variance, slope, centroid, and concentration metrics, were extracted to form compact feature sets. CatBoost models were trained to predict building labels, including YOC, WWR, and aspect ratio. Model performance demonstrated strong predictive capability, with the highest accuracies achieved for YOC and somewhat lower but still informative accuracy for WWR. SHAP analysis identified a small set of dominant features driving predictions, providing insight into how dynamic response characteristics encode envelope insulation and façade glazing. To enhance interpretability, surrogate decision trees were trained on SHAP-selected features, yielding compact rule sets that align with physical reasoning: older buildings are characterized by high solar variability, concentrated dynamic responses, and unstable correlations, while newer buildings display lower variability, delayed solar impacts, and more coherent thermal–power alignment. For WWR, the models distinguished highly opaque envelopes, with quiet and delayed dynamics, from highly glazed façades, with spiky and persistent temperature-driven responses, while mid-range classes remained more ambiguous. Counterfactual reasoning further showed how small shifts in key descriptors, such as lowering solar jitter or increasing stability in the power channel, could drive transitions between classes. The results demonstrate that simulated energy dynamics can be distilled into transparent, physics-aligned rules capable of distinguishing building stock attributes. This approach bridges physics-based simulation and interpretable machine learning, offering a scalable pathway for digital twin augmentation