Answers are grounded in this paper and cite their sources.
Purpose This study develops a self-healing digital-twin architecture for fault-tolerant building energy management in which anomaly detection is linked directly to control recovery. Design/methodology/approach A two-state RC thermal model, identified online by recursive least squares as an ARX approximation, is combined with an LSTM residual corrector. The hybrid-model residual drives a nine-state fault detection, isolation and recovery (FDIR) machine that reconfigures a stochastic model predictive controller (MPC). The architecture was tested in an eight-week residential pilot with 50 software-emulated faults and in a virtual hardware-in-the-loop test. Findings The architecture ran on a Raspberry Pi 4 (maximum MPC solve time 4.21 s). In the summary results of the original pilot analysis, 40 of 43 emulated non-cyber faults were correctly classified (93.0%; 95% CI 80.9–98.5%). As the original logs and data partitions were not retained, these results are descriptive, not held-out estimates. Research limitations/implications The evidence comes from one building and emulated faults and cannot be independently re-analysed; comfort is assessed with a model-based PMV estimate. A prospective study with pre-specified data partitions is required. Practical implications The design shows how diagnostic outputs can trigger explicit controller reconfiguration on low-cost edge hardware without redundant sensors. Originality/value The paper specifies an architecture that couples a hybrid digital twin, residual-driven FDIR and stochastic MPC in one control loop, and states transparently which parts of its pilot evidence can and cannot be verified.
No takes yet. Share an insight, caveat, or question.
Tsiakalos et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: