nZEB beyond prediction in smart built environments: formalising engineering knowledge through modular explainable machine learning

Abstract

This paper demonstrates how explainable machine learning (XAI) can be operationalised as a methodological pathway for formalising engineering knowledge from high-frequency building operational data. We propose a modular pipeline that combines feature engineering, ensemble and sequence learners, SHAP attribution and uncertainty quantification to convert raw sensor streams into machine-readable knowledge artefacts (JSON schema) suitable for automation workflows such as fault detection and demand response. Using a monitored nearly Zero-Energy Building (nZEB) in Lisbon (12 months, 5-minute resolution), we (i) report model performance (LightGBM, Random Forest, SVR, Linear Regression, and LSTM) under time-aware 70/15/15 split and 5-fold temporal cross-validation; (ii) present SHAP-based global and local attribution analyses that identify stable seasonal drivers; and (iii) provide computational cost (training and inference times) and uncertainty quantification. Results show ensemble models achieve superior short-term forecasting accuracy while producing consistent, actionable attributions that can be encoded as reusable artefacts. We close by describing a JSON artefact schema and outlining how these artefacts could be integrated within digital twins and supervisory control systems.