Answers are grounded in this paper and cite their sources.
With the increasing use of data-driven models in building energy applications, the trustworthiness and auditability of black-box models have become concerns, highlighting the importance of explainable artificial intelligence (XAI). Following the PRISMA protocol, this paper systematically reviews 98 journal articles that explicitly implemented XAI methods in building energy applications. The studies are classified into five life-cycle-based tasks: (1) energy demand prediction; (2) fault detection and diagnosis; (3) control and optimization; (4) occupancy and comfort; (5) retrofit planning decision support. Results show that post-hoc feature attribution methods, particularly SHAP, dominate the field. Task-dependent differences are observed in explanation scope and output: retrofit and benchmarking studies rely more on global sensitivity and marginal effects, whereas operational diagnosis and control place greater emphasis on instance-level evidence. Despite rapid progress, key challenges include limited data representativeness, simulation-to-reality gaps, sensitivity to feature correlation and background settings, insufficient benchmarks for faithfulness and stability, and computational constraints.
No takes yet. Share an insight, caveat, or question.
Geng et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: