Dataset and analytical code for “Predictive Fidelity and Explanatory Consistency in Explainable Surrogate Models for Building Energy: Comparison with Simulator-Derived Response Structure”
Description
This repository contains the dataset and analytical code supporting the study “Predictive Fidelity and Explanatory Consistency in Explainable Surrogate Models for Building Energy: Comparison with Simulator-Derived Response Structure.” The dataset comprises 15,360 unique EnergyPlus/Honeybee simulation scenarios for a low-rise office building in the hot-arid climate of Kerman, Iran. The complete factorial design includes seven early-design variables: aspect ratio, orientation, window-to-wall ratio (WWR), glass U-factor, solar heat gain coefficient (SHGC), lighting power density (LPD), and external wall insulation thickness. The primary response variables are energy use intensity (EUI) and cooling energy use intensity (Cooling EUI), with Heating EUI retained as a complementary diagnostic output. The accompanying analytical code supports predictive validation using XGBoost and Explainable Boosting Machine (EBM), TreeSHAP analysis, direct simulator-derived factorial effects, exact discrete functional ANOVA, and cross-method comparison. The repository supports reproducibility of the analytical pipeline from the final simulation dataset; limitations concerning the historical simulation provenance are documented in the accompanying materials.
Files
Steps to reproduce
Download and extract the repository package. Review the README and data dictionary before running the analytical workflow. Use the provided final dataset containing the complete 15,360-case factorial design. Recreate the analytical Python environment using the software versions documented in the repository. Run the supplied analytical code to reproduce the held-out predictive validation, full-data XGBoost and EBM explanatory models, TreeSHAP analyses, direct factorial effects, exact discrete functional ANOVA, and cross-method comparisons. Interpretation should remain restricted to the simulated discrete design space described in the manuscript and repository documentation.
Institutions
- Iran University of Science and TechnologyTehran, Tehran