A Scenario-Based Digital-Twin Assessment of Vision-Language Models for Building Energy Demand Forecasting and Automated Asset Intelligence

Replication package for the article "A Scenario-Based Digital-TwinAssessment of Vision-Language Models for Building Energy Demand Forecastingand Automated Asset Intelligence" (Sustainable Energy Technologies andAssessments, Elsevier; article reference SETA 105238). The study evaluates whether vision-language models can decode 3D buildingvisualizations — multi-view geometric renders and simulated physicsoverlays — into features that improve building energy demand forecasting.On a benchmark of 250 U.S. DOE commercial reference archetypes simulatedwith EnergyPlus across five ASHRAE climate zones (2.19 million hourlyobservations), a LoRA-adapted LLaVA-1.5 pipeline fused with a TemporalFusion Transformer reduces forecasting error from 9.8% to 7.5% MAPE, a23.4% relative reduction confirmed by Diebold-Mariano testing. This deposit contains: - data/ — the results underlying the article's tables and figures: per-building forecasting results for all 250 buildings (MAPE with and without visual features, delta-MAPE, negative-transfer flags), example weekly forecast traces with 10-90% prediction intervals, t-SNE/PCA projections of the learned visual embeddings, LLM-judge vs. HVAC-expert reliability scores, and the metadata-attribution, component-ablation, typology/climate-zone, missing-data, and multi-seed analyses (CSV, plus a consolidated Excel workbook).- code/ — a Python script that regenerates the manuscript figures from the included data (600-dpi PNG and vector PDF).- figures/ — the manuscript figures as published.- supplementary/ — the article's supplementary material (PDF).- replication/ — the data-generation scripts for the simulated dataset (EnergyPlus and rendering pipeline), the model configuration files (LoRA/QLoRA/DoRA/QDoRA and TFT), and the evaluation templates for the LLM-as-a-judge screening. Data and figures are released under CC BY 4.0; code under the MIT License.See README.md for file-level documentation and reproduction instructions.