Code · Zenodo · September 11, 2026
AARS: Simulation-Budgeted Few-Shot Adaptation of Building Energy Surrogates in Unseen Climates
Ozge Gunes Yavru, İsmail Burak YAVRU
Abstract
AARS v1.0.0 — Paper Submission Reproducibility Release This release provides the reproducibility materials accompanying the manuscript: "Simulation-Budgeted Few-Shot Adaptation of Building Energy Surrogates in Unseen Climates: A Physics-Guided Anchor-Relative Approach" Overview The study introduces the Adaptive Anchor-Relative Surrogate (AARS), a physics-guided few-shot adaptation framework for deploying building-energy surrogate models in previously unseen climates using only one or three target-climate EnergyPlus simulations. The experimental framework comprises: 1,000 paired classroom design alternatives 6 source climates and 6,000 source EnergyPlus simulations 8 held-out climates and 8,000 external EnergyPlus simulations 1-shot adaptation with 999 non-calibration test designs per climate 3-shot adaptation with 997 non-calibration test designs per climate strict exclusion of calibration anchors from predictive and decision-oriented evaluation Main reported results For total annual site energy across the eight held-out climates: | Setting | Mean MAE (kWh/m²-year) | R² | Top-5% recovery | |---|---:|---:|---:| | 1-shot AARS | 3.33 | 0.971 | 89.0% | | 3-shot AARS | 2.96 | 0.979 | 90.8% | On the aligned three-shot holdout, AARS reduced total-energy MAE by: 87.6% relative to zero-shot prediction 48.8% relative to three-shot global-bias calibration 38.7% relative to three-shot raw anchor-relative transfer Included materials This release includes: processed source and external EnergyPlus simulation datasets source-only anchor-ranking outputs selected calibration-anchor information authoritative strict external-validation results AARS and baseline evaluation code climate-level statistical analysis code publication-figure generation scripts software environment specifications data dictionary and provenance documentation reproducibility and validation reports Reproducibility The strict evaluation excludes the target simulations used for calibration from all reported test metrics. The distributed strict result set is the authoritative reproducible evaluation used in the manuscript. No external target result was used to tune the final method architecture, feature representation, anchor budget, or hyperparameters. A quick verification can be run with: python src/run_reproduction.py --quick-reference-check
More by İsmail Burak Yavru
- Turkish–Kazakh Sentiment Classification: Research Code and Aggregate ResultsZenodo · 2026
- PopRobust: Reproducibility Package for Multi-Protocol Popularity-Robust Recommendation BenchmarkingMendeley Data · 2026
- Robust and Explainable Wearable Stress Recognition under Physical Activity ConfoundingInternational Conference on Computer Science and Engineering · 2026
- SUPER: Smart User-Centric Popularity Exposure Reduction for Fair and Diverse RecommendationsIEEE Access · 2026 · Citations: 1
