Full-Feature versus Limited-Input Machine Learning for Residential Energy Estimation: A Comparative Analysis of RECS and ResStock Under Realistic Input Constraints

This study evaluated residential energy estimation under two contrasting information conditions: information-rich modeling using detailed dataset-specific features and deployment-oriented modeling restricted to a predefined set of ten low-burden inputs. Using the 2020 Residential Energy Consumption Survey (RECS) and the 2024.2 ResStock dataset, machine-learning models were developed for annual total energy use, space heating, and space cooling. Under information-rich conditions, CatBoost provided the strongest overall performance among the evaluated algorithms, achieving an R2 of 0.726 for total energy consumption in RECS and 0.905 in ResStock, with similarly strong performance for heating and cooling. The higher performance obtained with ResStock is consistent with its physically structured, high-resolution simulation inputs, whereas RECS reflects the greater variability and uncertainty associated with empirical household survey and billing data.

The principal finding of this study, however, concerns what happens when this information advantage is removed. When both datasets were restricted to the same ten practically obtainable variables, including degree-day indicators, housing type, number of bedrooms, heating and water-heating fuels, cooking fuel, and air-conditioning equipment, the difference between the two datasets largely disappeared. Total-energy estimation performance declined from R2 = 0.726 to 0.61 for RECS and from R2 = 0.905 to 0.62 for ResStock. The substantially larger reduction for ResStock indicates that much of the predictive advantage observed under information-rich conditions is associated with access to detailed physical, system, and operational descriptors that are unlikely to be available in many real-world deployment settings. Thus, performance obtained from comprehensive research datasets should not be assumed to persist when models must operate using only readily obtainable dwelling information.

The results therefore frame residential energy estimation as an accuracy–accessibility trade-off rather than solely an algorithm-selection problem. Detailed building information can support high-fidelity prediction, but collecting such information may require audits, extensive records, monitoring, or detailed simulation inputs. Conversely, low-burden models are substantially easier to deploy at scale but sacrifice household-level predictive accuracy. This distinction is particularly important for applications such as preliminary screening, portfolio-level assessment, benchmarking, and utility-allowance estimation, where detailed envelope, equipment-efficiency, infiltration, occupancy, and operational information may not be available for every dwelling. The purpose of the limited-input model is therefore not to replace detailed audits or calibrated simulations, but to quantify the level of predictive capability that can reasonably be retained when data collection itself is a practical constraint.

The cohort-specific analysis further demonstrated that part of the performance lost under limited information may be recovered when the model is applied to a more homogeneous residential population. For the selected ResStock cohort consisting of single-family detached homes in Climate Zone 6A, heated by natural gas and constructed between 2000 and 2010, the reduced-input analysis produced an R2 of 0.849 with Lasso regression and 0.829 with CatBoost, compared with R2 = 0.62 for the reduced-input national ResStock model.

Several limitations should be considered when interpreting the results. RECS and ResStock represent fundamentally different forms of evidence: RECS contains observations from actual households and therefore includes behavioral variability, reporting uncertainty, and billing-related noise, whereas ResStock provides a statistically representative but simulation-based characterization of the U.S. residential building stock. Models developed from either source may therefore inherit limitations associated with the underlying dataset. In addition, the present analyses are based primarily on cross-sectional annual characteristics and do not explicitly capture temporal changes in occupant behavior, equipment operation, renovations, or weather-dependent behavioral responses. These uncertainties are particularly important when limited-input models are used for consequential applications, where prediction uncertainty, population-specific bias, and empirical validation should accompany point estimates.

Overall, the study demonstrates that the information available at the time of deployment is a central determinant of achievable residential energy-estimation performance. The comparison between RECS and ResStock shows that substantial differences in predictive accuracy under information-rich conditions can narrow considerably when both models are subjected to the same practical information constraint. At the same time, the cohort analysis indicates that targeted modeling may recover meaningful predictive capability without requiring fully detailed building descriptions. Future work should therefore focus not only on improving modeling algorithms, but also on identifying efficient combinations of accessible variables, developing systematic cohort-selection strategies, quantifying prediction uncertainty, and validating reduced-input models against independent empirical datasets. Such efforts could support residential energy-estimation tools that explicitly balance predictive fidelity against the cost and feasibility of obtaining the information required for real-world deployment.

Improvements to AI systems:

1. Uncertainty-aware performance calibration: Implement a meta-model that predicts when a trained energy-estimation model’s accuracy will degrade based on the overlap between training feature richness and deployment feature availability. This system would output both a point estimate and a confidence interval, explicitly flagging cases where information-rich training data (e.g., ResStock) is used in low-information deployment, preventing overconfident predictions.

2. Feature-accessibility-aware model selection: Build an AI system that automatically selects between a high-fidelity model (using detailed physical descriptors) and a low-burden model (using only the ten predefined variables) based on a user-specified data-collection budget. The system would estimate the expected R2 loss from using limited inputs and recommend whether to invest in additional data collection or accept reduced accuracy, optimizing the accuracy–accessibility trade-off in real time.

3. Cohort-adaptive transfer learning: Develop a domain-adaptation layer that, given a limited-input model trained on national data, automatically identifies homogeneous subpopulations (e.g., by climate zone, housing type, fuel mix, construction era) using only the ten low-burden variables. The system would then fine-tune predictions for that cohort, recovering a significant portion of the lost accuracy (as demonstrated by the R2 jump from 0.62 to 0.849 in the paper) without requiring new data collection.

4. Bias-detection module for simulation vs. survey data: Create an AI system that, when trained on either empirical survey data (RECS) or simulation data (ResStock), automatically quantifies and reports the expected bias due to behavioral variability, reporting uncertainty, and billing noise. This module would adjust predictions with a correction factor when deploying models across data sources, and would refuse to provide point estimates for consequential applications unless accompanied by population-specific validation.

5. Temporal robustness augmentation: Extend the model to incorporate a lightweight temporal correction layer that uses only the ten low-burden inputs plus publicly available weather and occupancy proxies (e.g., degree-day anomalies, time-of-year) to adjust annual estimates for interannual variability. This addresses the paper’s limitation of cross-sectional-only analysis, enabling the system to provide more reliable predictions for multi-year portfolio assessments or utility-allowance adjustments.

What the improved AI system can do:

• Deployable energy-estimation tool: Given only ten easily obtainable dwelling characteristics (e.g., housing type, bedrooms, fuels, AC equipment, degree-day indicators), the system predicts annual total energy, heating, and cooling use with a known, calibrated accuracy range (e.g., R2 ≈ 0.61–0.62 nationally, but up to 0.85 for targeted cohorts). It explicitly warns the user when accuracy will be lower than if detailed audit data were available.

• Data-collection advisor: Before a user invests in a full energy audit or simulation, the system estimates the expected R2 improvement from collecting additional features (e.g., envelope efficiency, infiltration, occupancy patterns). It then recommends whether the cost of data collection is justified for the intended application (e.g., benchmarking vs. utility allowance).

• Cohort-specific recommender: For a portfolio of buildings, the system automatically clusters dwellings into homogeneous groups using only the low-burden variables, then applies cohort-tuned models to each group. This recovers predictive accuracy (e.g., from 0.62 to 0.85) without requiring any new data, enabling more precise portfolio-level assessments and preliminary screening.

• Source-aware predictor: The system distinguishes whether its training data came from surveys or simulations. When deployed on real households, it applies a bias-correction term for behavioral variability and reporting noise, and it flags predictions as “uncertain” if the deployment population differs significantly from the training population, prompting the user to collect a small validation sample before using results for consequential decisions.

• Temporal updater: For multi-year applications, the system adjusts annual predictions using only degree-day deviations and seasonal patterns, allowing utilities or policymakers to track energy-use changes over time without re-collecting building data, while explicitly quantifying the added uncertainty from unobserved behavioral or renovation changes.

Residential energy estimates are often needed before detailed envelope characteristics, equipment efficiencies, infiltration, sensor, or billing data are available. This study quantifies the trade-off between predictive accuracy and input accessibility using two nationally representative U.S. residential-energy datasets: the survey-based Residential Energy Consumption Survey (RECS) and the simulation-based ResStock dataset. Full-feature models were first used to establish dataset-specific performance benchmarks. For total-energy estimation, the models were subsequently restricted to ten low-burden variables obtainable from occupants, administrative records, or location-based weather data without an on-site energy audit. Among CatBoost, XGBoost, LightGBM, Random Forest, and Neural Networks, CatBoost consistently achieved the highest predictive performance for the full-feature analysis, reaching R2 = 0.90 for ResStock and R2 = 0.73 for RECS. When the feature set was restricted to ten homeowner-accessible inputs to simulate realistic deployment conditions, model performance converged to R2 = 0.61 for RECS and R2 = 0.62 for ResStock, showing that algorithmic complexity cannot fully compensate for missing physical and behavioral information. However, for a more homogeneous ResStock cohort consisting of single-family detached, natural-gas-heated homes in Climate Zone 6A constructed between 2000 and 2010, a reduced-input model improved accuracy to R2 = 0.85, demonstrating the value of targeted modeling for homogeneous populations. The results indicate that tree-based ensemble models can serve as high-fidelity emulators of national-scale residential energy datasets. However, careful consideration of feature availability, dataset origin (empirical vs. synthetic), and applicable use cases are also important.