Digital twins have attracted growing attention in building energy research as tools for monitoring, calibration, fault detection, and performance assessment. Their conceptual foundations are closely linked to earlier work on simulation-based optimization and integrated energy system modeling (Nguyen et al. 2014), as well as broader advances in energy system simulation and analysis (Subramanian et al. 2018). Early building digital twins were predominantly used for offline calibration of physics-based models using measured operational data, with the primary objective of improving predictive accuracy.

Subsequent developments extended digital twin concepts toward real-time monitoring and short-horizon forecasting by integrating live sensor data with computational models. These capabilities enhanced situational awareness during building operation but generally remained decoupled from control decision-making. More recent studies have investigated digital twin frameworks for anomaly detection and operational monitoring in the built environment (Lu et al. 2020), including applications to multi-zone building systems (Hodavand et al. 2025).

Digital twins have also been applied to fault detection and diagnosis in HVAC systems through analysis of discrepancies between predicted and observed behavior (Xie et al. 2023). Systematic reviews of digital twin applications in building operations highlight their rapid adoption while emphasizing that most existing implementations function as passive analytical tools rather than active control agents (Hodavand et al. 2023). Broader surveys of digital twins in the built environment similarly conclude that current frameworks emphasize visualization, monitoring, and assessment rather than diagnosis-informed control adaptation (AlBalkhy et al. 2024). Despite advances in enabling technologies and digital platforms (Tuhaise et al. 2023), digital twins rarely close the operational loop by embedding diagnostic reasoning directly into control logic, particularly under climate-extreme conditions (Hosamo et al. 2022).

Performance drift in building energy systems has been extensively studied in the context of equipment degradation, sensor bias, and operational faults. Classical fault detection and diagnosis frameworks rely on residual analysis, statistical inference, and rule-based reasoning to identify abnormal system behavior (Liu et al. 2025). Comprehensive reviews of fault detection and diagnostic methods for building systems demonstrate that these approaches can be effective under controlled or quasi-stationary operating conditions (Katipamula and Brambley 2005).

More recent systematic studies have expanded diagnostic taxonomies and algorithmic libraries for building fault detection, providing detailed classifications of fault mechanisms and detection approaches (Melgaard et al. 2022). Data-driven techniques such as principal component analysis have been successfully applied to detect sensor and component faults in air handling units and variable air volume systems (Wang and Xiao, 2004; Du et al., 2007), as well as sensor faults across broader HVAC subsystems (Du and Jin, 2007).

Under climate-extreme conditions, however, diagnostic reliability deteriorates. Elevated thermal loads, increased variability, and interacting stressors can obscure fault signatures and mask gradual degradation. Reviews of automated fault detection and diagnosis systems indicate that most existing approaches implicitly assume stable environmental conditions and do not explicitly account for climate-induced stress (Chen et al. 2022). Modeling studies further demonstrate that HVAC operational faults may be compensated by control actions, delaying detection and exacerbating inefficiencies during prolonged heat events (Zhang and Hong 2017).

Model predictive control has been widely adopted in building energy systems due to its ability to anticipate future disturbances and balance energy efficiency with thermal comfort constraints (Lee and Braun 2008a). Robust and stochastic extensions of MPC have been proposed to address forecast uncertainty and model mismatch, improving disturbance handling under variable conditions.

Despite these advances, most MPC formulations assume fixed system parameters and reliable sensing. During extended periods of climatic stress, such assumptions are frequently violated. Predictive controllers may compensate for degradation or sensor bias through increasingly aggressive control actions, potentially masking underlying issues while increasing energy use or comfort violations (Lee and Braun 2008b). Demand-limiting and pre-cooling strategies have demonstrated effectiveness in peak mitigation (Sun et al. 2012), yet their performance depends strongly on accurate system representation and stable operating conditions.

More recent studies have explored demand response strategies combining optimal pre-cooling with adaptive temperature reset, reporting improved peak reduction in office buildings (Naderi et al. 2022; Wang et al. 2023). However, these approaches generally treat system degradation and climatic stress as external disturbances rather than endogenous processes requiring diagnosis-informed and adaptive control.

The increasing frequency and intensity of heatwaves have highlighted the vulnerability of building energy systems to climate extremes. Reviews of urban heat island effects and global warming impacts consistently report substantial increases in cooling demand and peak electricity consumption under extreme conditions (Santamouris et al. 2015). At the same time, maintaining thermal comfort becomes increasingly challenging, particularly in naturally ventilated and mixed-mode buildings (de Dear and Brager 2002).

Digital twins have been proposed as potential enablers of climate-resilient building operation, and recent frameworks emphasize user-centered and climate-aware approaches for the built environment (Osama 2024). Nevertheless, most existing implementations remain limited to offline analysis or short-term prediction. The integration of real-time diagnosis, uncertainty-aware reasoning, and adaptive control within a unified operational framework therefore remains a critical and unresolved research gap.

The proposed framework is conceptually aligned with the ISO/IEC 30173 Digital Twin reference architecture, which defines a digital twin as a dynamic virtual representation synchronized with a physical entity through bidirectional data exchange across its lifecycle. In accordance with this standard, the present framework integrates sensing, state estimation, model synchronization, and decision support within a unified operational loop.

While ISO/IEC 30173 provides a structural and interoperability-oriented definition of digital twin systems, it does not prescribe specific mechanisms for adaptive reasoning or resilience enhancement under climate-induced stress. The cognitive digital twin introduced in this study extends the standard architecture by embedding diagnostic attribution, episodic performance memory, and uncertainty-aware control within the operational layer. In this sense, the proposed framework remains structurally compliant with international digital twin principles while advancing toward resilience-oriented intelligence under extreme climate conditions.

Table 1 provides a structured comparison between existing approaches in building energy operation and the proposed cognitive digital twin framework. As shown, prior research has largely addressed individual aspects of building energy management in isolation. Conventional digital twins have focused primarily on real-time state updating, monitoring, and performance visualization, offering limited capability for diagnosing performance drift or adapting control strategies during abnormal operating conditions. Similarly, MPC-based control frameworks excel at anticipatory control and constraint handling but typically assume fixed system characteristics and reliable sensing, thereby lacking mechanisms for explicit fault attribution or degradation-aware decision-making.

Fault detection and diagnosis methods, on the other hand, have demonstrated effectiveness in identifying performance degradation and sensor or component faults. However, these approaches are commonly implemented as stand-alone diagnostic layers that operate independently of control logic. As a result, diagnostic insights are rarely translated into adaptive control actions, particularly under non-stationary conditions such as sustained climate extremes.

The comparison highlights a fundamental gap in current research: while digital twins, predictive control, and diagnostic frameworks each contribute valuable capabilities, none provide an integrated solution that supports diagnosis-informed and adaptive control during climate-extreme operation. In particular, the absence of root-cause attribution, uncertainty-aware reasoning, and closed-loop control adaptation limits the resilience of existing systems when confronted with prolonged heatwaves and evolving system degradation.

The proposed cognitive digital twin framework addresses these limitations by unifying real-time state estimation, performance drift detection, causal attribution, and adaptive control within a single operational loop. By embedding diagnostic reasoning directly into the control process, the framework enables proactive intervention before comfort violations or operational instability occur. This integrated approach positions cognitive digital twins as a scalable pathway toward resilient, climate-aware building energy systems capable of sustaining performance under extreme and uncertain operating conditions.

To evaluate cognitive operation under sustained climate-induced stress, a representative medium-scale commercial office building is selected as the reference system. The building configuration is intentionally chosen to reflect common design and operational characteristics of contemporary office buildings, thereby ensuring practical relevance while maintaining transparency and reproducibility for benchmarking.

The reference building has a total conditioned floor area of 6000 m² distributed across four floors. Cooling is provided by a centralized variable air volume (VAV) HVAC system serving all thermal zones. Envelope thermal properties, internal heat gains, and ventilation requirements are defined in line with standardized commercial building practice. This design choice allows the impacts of climate extremes and operational drift to be examined without confounding effects arising from atypical architectural or system features.

The building HVAC system is modeled as a variable air volume system with centralized mechanical cooling and zone-level reheat. This configuration represents a widely deployed solution in commercial office buildings and provides a neutral basis for comparison across rule-based control, conventional model predictive control, and digital twin–based strategies.

System capacities and baseline control parameters are selected to satisfy design cooling requirements under typical operating conditions. At the same time, sufficient operational flexibility is retained to allow degradation, sensing bias, and climate-induced stress to be introduced explicitly in subsequent scenarios without violating basic feasibility constraints.

To examine robustness across diverse climatic conditions, the proposed framework is evaluated under three representative climate regimes exhibiting distinct thermal stress characteristics. Typical Meteorological Year (TMY3) weather data are used as the baseline for each location, ensuring standardized and reproducible climatic inputs.

The selected climates span hot-dry, hot-humid, and temperate regimes, capturing a broad range of thermal demand profiles and heatwave dynamics relevant to commercial building operation.

Baseline weather files are systematically modified to construct controlled heatwave events, enabling consistent stress testing across climates while preserving realistic diurnal temperature patterns.

Heatwave events are defined using percentile-based thresholds to represent extreme yet climatologically plausible conditions. A heatwave is identified when daily maximum temperatures exceed the 95th percentile of historical monthly values and nighttime minimum temperatures exceed the 90th percentile for at least three consecutive days.

To enable systematic comparison across stress levels, three heatwave intensity tiers are constructed through deterministic temperature offsets applied during peak daytime hours (12:00–18:00):

Nighttime minimum temperatures are simultaneously elevated by +1°C, +2°C, and +3°C for L1, L2, and L3, respectively. This formulation preserves realistic diurnal temperature structure while intensifying sustained thermal stress.

Figure 1 shows deterministic heatwave construction applied to baseline TMY3 weather data. Daytime temperature offsets are imposed during peak hours (12:00–18:00), while nighttime minimum temperatures are elevated to preserve realistic diurnal structure across moderate (L1), severe (L2), and extreme (L3) heatwave levels.

A structured set of eight climate-extreme stress scenarios is defined to evaluate system behavior under progressively challenging operating conditions. The scenarios combine heatwave intensity with operational constraints, uncertainty, and drift mechanisms to expose the limitations of conventional control strategies and to test the adaptive capabilities of the proposed cognitive digital twin.

Cooling capacity degradation is modeled as a gradual linear reduction in effective cooling capacity at a rate of 0.5% per day, representing fouling and aging effects. Sensor drift is introduced as a biased temperature offset increasing linearly to +1.0° Cover a five-day period. Uncertainty levels correspond to increasing variance in internal heat gains, weather forecast errors, and sensor measurement noise.

Demand response constraint. During demand response events (weekday 14:00–18:00), total building electric demand is constrained to remain below , where denotes the peak 15-min demand observed under baseline rule-based control (B1) during the corresponding non-DR heatwave period. The same demand cap is applied across all controllers.

Performance drift is defined as a persistent deviation between expected and observed system behavior that exceeds normal stochastic variability. Within the defined scenarios, drift emerges through gradual equipment degradation, biased sensing, and altered internal load conditions, and is amplified by sustained climate extremes.

The interaction between climatic stress, operational constraints, and drift mechanisms creates a decision environment in which purely reactive controllers are insufficient. This environment provides a stringent and reproducible benchmark for evaluating diagnosis-informed reasoning and adaptive control within the proposed cognitive digital twin framework.

The proposed cognitive digital twin is formulated as a closed-loop operational framework that continuously couples the physical building with a physics-informed virtual counterpart. Unlike conventional digital twins that primarily support monitoring, visualization, or short-term prediction, the proposed framework embeds diagnostic reasoning and adaptive decision-making directly within the operational control loop.

The cognitive digital twin operates through four tightly integrated functional layers:

This hierarchical structure distinguishes the framework from conventional estimation-based digital twins. Estimation provides perceptual awareness, memory encodes temporal performance signatures, and causal attribution enables reasoning prior to control adaptation.

At each control interval, the physical building system is coupled to its digital twin through a data assimilation interface that integrates measured observations with model-based predictions. Let denote the measured output vector at time step , and let represent the latent system state vector, including both thermal states and performance-related parameters.

Based on the prior state estimate the digital twin generates a one-step-ahead prediction . Sequential state updating is then performed according to the following relationship:

Where is the assimilation gain that balances the relative confidence assigned to model predictions and sensor measurements. The innovation term captures discrepancies arising from modeling error, sensor noise, forecast uncertainty, or evolving system conditions.

State estimation is implemented using an Extended Kalman Filter (EKF), which enables joint estimation of thermal states and latent performance states within a unified framework. The process noise covariance Q and measurement noise covariance R are selected based on sensor specifications and calibration experiments conducted under nominal operating conditions. Initial state means are obtained from a 48-hour warm-up simulation using baseline weather data, ensuring stable filter initialization prior to closed-loop operation.

In the extended Kalman filter, measurement and process noise covariance parameters are selected based on sensor specifications and calibration tests conducted under nominal operating conditions. The measurement noise covariance for zone air temperature is defined as = . The airflow rate measurement noise covariance is defined as =

Process noise terms are introduced to represent unmodeled dynamics and gradual system degradation. The process noise covariance associated with thermal state evolution is set to = . Cooling capacity degradation is modeled using a latent state with process noise covariance = , while system efficiency degradation is represented with = . Temperature sensor bias is captured through an additional latent state with process noise covariance = .

These parameter values are chosen to balance estimation responsiveness and numerical stability, allowing the filter to track slow performance drift without amplifying measurement noise or introducing spurious state fluctuations.

This assimilation mechanism enables continuous synchronization between the physical building and its virtual representation under sensor noise, forecast uncertainty, and modeling mismatch.

The nonlinear COP() formulation is handled within the EKF framework through local linearization at each time step, ensuring numerical stability while retaining temperature-dependent efficiency effects. Adaptive covariance inflation is constrained within bounded limits to prevent filter divergence and maintain stable recursive estimation.

The building energy system is modeled using a hybrid state-space formulation that captures both thermal dynamics and latent performance characteristics. The system state vector is defined as:

where denotes the zone air humidity ratio, enabling explicit representation of latent heat dynamics under hot-humid climate conditions, denotes zone air temperatures, represents envelope-related thermal states, is the supply air mass flow rate, denotes the effective cooling capacity, represents system efficiency, and captures temperature sensor bias.

The state evolution of the system is governed by:

where denotes the control input vector and represents external disturbances associated with weather variability, internal heat gains, and unmodeled dynamics.

Performance drift is quantified based on residual analysis between measured and predicted system outputs. Residual evaluation incorporates both temperature and humidity prediction errors, ensuring that latent comfort degradation is not misinterpreted as purely sensible thermal drift. At each control interval, the residual vector is computed as the difference between the measured output and the one-step-ahead prediction,

where denotes the measured output vector and represents the predicted output obtained from the digital twin prior to measurement assimilation.

It is noted that the latent cooling capacity state inherently reflects the nonlinear formulation introduced in Section 3.3.1. Consequently, performance drift detection is evaluated relative to temperature-dependent efficiency limits rather than fixed nominal capacity assumptions.

To distinguish transient fluctuations from persistent performance degradation, a drift metric is evaluated over a sliding window of length . The drift metric is defined as:

where || · || denotes the Euclidean norm of the residual vector. This aggregation suppresses short-term noise effects while emphasizing sustained deviations indicative of drift.

Performance drift is declared when the drift metric exceeds a predefined threshold:

where the threshold is derived from residual statistics obtained under nominal, fault-free operating conditions.

In this study, the drift detection window length is set to = 96 samples, corresponding to a 24-hour rolling window at a 15-minute control interval. The detection threshold is defined as: , where and denote the mean and standard deviation of the drift metric under nominal operation. This choice balances sensitivity to gradual degradation against robustness to false alarms. Beyond short-term drift detection, the framework maintains an episodic performance memory that archives seasonal performance baselines and prior degradation trajectories. Let denote stored state evolution signatures under distinct operational regimes. During diagnosis, current residual trajectories are compared against archived signatures using similarity measures, enabling contextualized interpretation of performance deviations.

Root-cause attribution is performed by analyzing the temporal evolution of latent performance states estimated by the cognitive digital twin. Sustained reductions in or indicate equipment-related degradation, persistent non-zero values of indicate sensor bias, and elevated thermal demand without a corresponding loss in system efficiency is interpreted as load-induced drift. Diagnostic attribution is completed prior to control adaptation to ensure causal consistency between the identified degradation mechanism and subsequent corrective action.

To formalize diagnostic reasoning beyond residual thresholding, causal attribution is framed as a probabilistic inference problem over latent degradation hypotheses. Let denote candidate degradation mechanisms corresponding to capacity loss, efficiency loss, sensor bias, and load-induced drift.

The posterior probability of each hypothesis is updated according to

where denotes the residual structure over the detection window and is evaluated based on the temporal evolution of the associated latent state.

For example, sustained monotonic decline in increases , while persistent non-zero increases . This probabilistic framing enables structured reasoning over competing degradation explanations rather than deterministic rule switching.

Uncertainty in sensing, forecasting, and internal loads is explicitly represented within the cognitive digital twin. Sensor noise and weather forecast errors are treated as stochastic disturbances, while variability in internal gains and occupancy is bounded using predefined ranges. Uncertainty is propagated through the estimation process, yielding confidence bounds on predicted system behavior that directly condition subsequent control decisions during climate-extreme operation.

Uncertainty in sensing, forecasting, and internal loads is explicitly represented through structured stochastic models with three intensity levels. Zone air temperature sensor noise is modeled as a Gaussian process with standard deviation equal to 0.10°C, 0.25°C, and 0.40°C for low, medium, and high uncertainty levels, respectively, based on 15-minute sampling intervals.

Outdoor air temperature forecast uncertainty is represented using a first-order autoregressive process. For low, medium, and high uncertainty cases, the standard deviation of the forecast error is set to 0.8°C, 1.5°C, and 2.5°C, respectively, with corresponding temporal correlation coefficients of 0.6, 0.7, and 0.8. These parameters are applied to the short-term weather forecasts used within the predictive control layer.

Variability in internal heat gains arising from lighting and plug loads is modeled using a uniform distribution with bounds of plus or minus 5 percent, 10 percent, and 20 percent for low, medium, and high uncertainty levels, respectively. Occupancy uncertainty is modeled using a Poisson process with a nominal mean equal to the baseline occupancy level. To account for stochastic variation, the effective occupancy rate is scaled within the ranges of 0.95 to 1.05, 0.90 to 1.10, and 0.80 to 1.20 of the nominals mean for low, medium, and high uncertainty scenarios, respectively. The selection of uncertainty levels is scenario-dependent and aligned with the stress intensity defined in the experimental design.

Control actions are optimized over a finite prediction horizon H by solving a constrained optimization problem that balances energy efficiency, thermal comfort, and control smoothness. The objective function is formulated as minimize over

where denotes electric energy consumption, represents thermal comfort violations, and penalizes excessive control variability between successive control actions. Thermal comfort violations are evaluated based on combined temperature and relative humidity constraints to reflect both sensible and latent discomfort effects.

Crucially, diagnosed system states explicitly condition the optimization process. When sensor bias is identified, control actions compensate for biased measurements rather than altering physical setpoints. When capacity degradation is diagnosed, feasible control policies are computed under reduced effective capacity, avoiding infeasible or unstable operation.

A consistent set of objective weights is used across all optimization-based controllers to ensure a fair comparison. For the conventional model predictive control baseline (B2), the energy term weight is set to , the thermal comfort penalty weight is set to , and the control smoothness weight is set to . These weights are selected through a grid-search tuning procedure under nominal, non-extreme operating conditions.

The same objective weights are applied to the non-cognitive digital twin controller (B3) to isolate the effect of state estimation and monitoring from tuning-related advantages. For the proposed cognitive digital twin controller, the objective weights are frozen at the values obtained for the MPC baseline (B2) and are not re-tuned under climate-extreme scenarios.

The rule-based control strategy (B1) does not rely on an explicit optimization objective and therefore does not require weight tuning.

The same objective weights and operational constraints are applied across all optimization-based controllers to ensure fairness. Weights are tuned once under baseline conditions and then held fixed across all climate-extreme scenarios.

At each control interval, operational data are acquired, states are updated via data assimilation, performance drift is diagnosed and attributed, uncertainty is quantified, and diagnosis-conditioned control actions are optimized and applied. This closed-loop cognitive operation enables proactive intervention before performance degradation escalates into comfort violations or operational instability during prolonged heatwave events. Actuation smoothness and wear proxy metrics are tracked online to ensure that adaptation remains operationally sustainable and does not induce excessive damper or valve oscillations. Cognition in this framework therefore emerges from the integration of probabilistic perception, episodic memory, and causal inference prior to control adaptation.

All scenarios are evaluated using a time-domain co-simulation framework that couples the physical building model with the digital twin in a closed-loop configuration. At each control interval, the building model advances one simulation step, operational data are exchanged with the digital twin, and control actions are computed and applied before the next step. This step-synchronized execution ensures temporal consistency between sensing, estimation, decision-making, and actuation.

Identical disturbance realizations are applied across all controllers to enable paired comparison and to isolate the effect of control logic from stochastic variability.

All experiments are conducted using EnergyPlus v9.6 as the building performance simulation engine. The building model is implemented as a detailed EnergyPlus IDF file, coupled with a supervisory control layer written in Python 3.10 via the EnergyPlus Python API. Controller logic, state estimation, and scenario orchestration are executed externally at a 15-minute control timestep.

Weather inputs are based on TMY3 EPW files for each climate, with heatwave perturbations applied programmatically as described in Section 2.4. Simulations are executed on a Linux workstation (Ubuntu 22.04, Intel Xeon 3.6 GHz, 64 GB RAM).

To ensure reproducibility, all controller scripts, scenario configuration files, random seeds, and post-processing routines used to generate the reported results are fixed and logged. The full reproducibility package, including the IDF model, modified EPW files, controller implementations, and result aggregation scripts, will be released upon publication.

Reproducibility statement. All simulations are generated from a single EnergyPlus IDF model (E+ v9.6) coupled to Python 3.10 controller scripts through the EnergyPlus Python API. The full experiment is reproducible from:

A private repository containing these artifacts and a one-command execution script is archived for peer review and will be made public upon acceptance. The repository includes a frozen environment specification (conda YAML) and exact random seeds used in this study.

Three baseline controllers are implemented to isolate the contribution of diagnostic reasoning and adaptive decision-making introduced by the proposed cognitive digital twin. Each baseline represents a realistic operational strategy commonly encountered in commercial building practice.

The proposed cognitive digital twin integrates joint state estimation, explicit drift diagnosis, uncertainty-aware reasoning, and diagnosis-conditioned control adaptation within a unified operational loop. Control policies are updated at each interval based on diagnosed system states rather than purely predictive or reactive signals.

All controllers operate under identical building models, climate scenarios, sensing availability, actuation limits, and evaluation windows. Prediction horizons, control horizons, and operational constraints are aligned where applicable. Objective weights for optimization-based controllers are tuned once under baseline non-extreme conditions and then frozen across all scenarios. This protocol ensures that observed performance differences arise from cognitive reasoning and adaptive decision-making rather than modeling, tuning, or implementation artifacts.

Performance is evaluated using a consistent set of energy, comfort, and resilience metrics designed to capture both average efficiency and extreme-event behavior.

Each scenario is evaluated over 10 stochastic realizations, with identical disturbance sequences shared across controllers to enable paired statistical comparison. Reported results correspond to the mean and standard deviation across realizations.

Statistical significance is assessed using paired t-tests when normality holds (Shapiro–Wilk test, α=0.05) and Wilcoxon signed-rank tests otherwise. Effect sizes are quantified using Cohen’s d for paired samples.

Table 13 summarizes system performance under purely climate-driven stress without induced faults. Under both moderate (S1) and severe (S2) heatwave conditions, the cognitive digital twin consistently achieves lower annualized energy use intensity (EUI_yr), reduced peak demand, and fewer thermal discomfort hours compared with all baseline strategies, including B1, B2, RMPC, and the non-cognitive digital twin.

Figure 3 presents the comparison of discomfort hours under moderate (S1) and severe (S2) heatwave conditions for all control strategies. The cognitive digital twin maintains the lowest discomfort levels across both scenarios, with increasing relative benefits under more severe climatic stress.

All energy metrics are evaluated over a 28-day simulation window. To ensure benchmarking consistency across scenarios and climate zones, energy use intensity (EUI) is annualized and reported in kWh/m²·yr. Annualization is performed by scaling the total HVAC electricity consumption observed during the evaluation window to an equivalent yearly value, normalized by the conditioned floor area (6000 m²).

For transparency, total HVAC electricity consumption and the corresponding annualized EUI are internally cross-verified for representative scenarios. Under scenario S1, rule-based control (B1) yields a total HVAC energy consumption of 950.4 MWh over the evaluation period, corresponding to the annualized EUI reported in Table 13. The proposed cognitive digital twin reduces total energy consumption to 847.2 MWh under the same conditions, resulting in a proportionally lower annualized EUI.

Similarly, under scenario S2, rule-based control records 976.2 MWh total consumption, while the cognitive digital twin reduces total energy use to 867.0 MWh. These values are consistently scaled to annualized EUI metrics to enable fair comparison and cross-climate benchmarking. The cross-verification confirms numerical consistency between aggregated energy values and reported annualized performance indicators.

Under demand-response conditions (S2), baseline controllers reduce peak demand primarily through reactive setpoint shifts, resulting in increased discomfort. In contrast, the cognitive digital twin preserves comfort by adapting control actions based on diagnosed system states and uncertainty, indicating that performance gains extend beyond predictive control alone.

The evaluation scenarios are defined with consistent heatwave durations and scenario-specific analysis windows to ensure fair comparison across control strategies. Scenarios S1 and S2 are simulated over a total duration of 28 days, including 7 heatwave days followed by a 21-day post-event period. Performance metrics for these scenarios are computed over the full simulation window.

Scenarios S3, S6, S7, and S8 are simulated over 21 days, with 7 heatwave days and a 14-day post-event period. For scenario S3, performance metrics are evaluated starting from the onset of performance drift to isolate diagnostic and adaptive behavior. In scenario S6, metrics are computed over the full simulation window. Scenario S7 focuses on post-event resilience, and metrics are evaluated over the recovery phase. Scenario S8 uses a comparable evaluation window to enable cross-climate comparison.

Scenarios S4 and S5 are simulated over shorter 14-day periods, each consisting of 7 heatwave days followed by a 7-day post-event interval. In scenario S4, performance metrics are evaluated specifically during demand-response hours, while in scenario S5 the analysis is restricted to periods characterized by elevated uncertainty.

Scenarios S3 and S4 assess diagnostic performance under progressive equipment degradation and sensor bias. Table 14 summarizes drift detection performance across N = 50 stochastic realizations. Detection time is reported as Mean ± 95% confidence interval (CI_95), while classification reliability is quantified using sensitivity, specificity, false alarm rate (FAR), and area under the receiver operating characteristic curve (AUC).

A false alarm is recorded when drift is declared during fault-free operation for more than one consecutive control interval. A missed detection is recorded when injected drift is not detected within 72 hours from onset. All values are computed under identical disturbance realizations to ensure paired comparison consistency.

Figure 4 presents the average drift detection time under capacity degradation (S3) and sensor bias (S4). Across both fault types, the cognitive digital twin demonstrates substantially faster detection relative to the non-cognitive digital twin.

For scenario S3, the ratio of mean detection times between the non-cognitive and cognitive configurations is 3.33 (95% CI ratio: 2.89–3.84). For scenario S4, the corresponding ratio is 3.38 (95% CI ratio: 2.94–3.91). These ratios indicate statistically robust acceleration in fault detection while simultaneously reducing false alarms and missed detections. The improvement in detection reliability is further supported by higher AUC values (0.96–0.97 for the cognitive digital twin versus 0.78–0.81 for the non-cognitive baseline), confirming superior diagnostic discrimination under climate-induced stress.

Compound scenarios combine heatwaves with operational constraints and elevated uncertainty.

Diagnosis-conditioned adaptation enables the cognitive digital twin to satisfy demand and ventilation constraints while minimizing unmet cooling hours, even under high uncertainty and humidity.

Scenario S7 isolates post-heatwave recovery under persistent degradation. Figure 5 shows Time required to restore thermal comfort after heatwave removal under persistent degradation (S7). The cognitive digital twin achieves substantially faster recovery due to retained diagnostic awareness.

The cognitive digital twin restores comfort approximately 40–45% faster than baseline strategies, reflecting retained diagnostic awareness beyond the climate event.

Performance improvements are observed across hot-dry, hot-humid, and temperate climates without climate-specific re-tuning, suggesting that the proposed framework maintains structural generalizability across diverse boundary conditions.

All hypothesis tests are paired across identical disturbance realizations (n = 50). Normality of paired differences is assessed using the Shapiro–Wilk test (α = 0.05). When normality holds, paired t-tests are applied; otherwise, Wilcoxon signed-rank tests are used. Multiple comparisons across the five hypothesis tests reported in Table 18 are controlled using the Benjamini–Hochberg procedure with a false discovery rate of 0.05. Adjusted p-values () are reported.

To evaluate whether performance gains are achieved at the expense of increased hardware stress, actuator effort is quantified using cumulative travel distance and control effort metrics across N = 50 stochastic realizations. All values are reported as Mean ± 95% confidence interval (CI_95).

Damper wear is quantified using total cumulative position change (percentage travel per actuator), while cooling valve effort is computed as the cumulative absolute variation in control signal over the evaluation window. Control effort is additionally summarized using the quadratic penalty term defined in Section 3.

The cognitive digital twin exhibits reduced cumulative actuator travel and lower control effort compared with both conventional MPC and RMPC. Despite improved energy and comfort performance, actuator activity decreases by approximately 12–18% relative to RMPC, indicating that diagnosis-conditioned adaptation does not induce excessive oscillatory control behavior.

These findings confirm that resilience gains are achieved without compromising hardware longevity or increasing mechanical wear, supporting the operational sustainability of the proposed framework.

Three key observations emerge. First, diagnosis-enabled early intervention prevents escalation of performance drift during climate extremes. Second, diagnosis-informed adaptation consistently outperforms purely predictive control under compound stress and uncertainty. Third, cognitive operation enhances resilience by preserving comfort and accelerating post-event recovery.

Thermal comfort is evaluated using standardized criteria consistent with ASHRAE 55 for mechanically conditioned commercial buildings.

Thermal comfort is evaluated using zone air temperature bounds as a practical proxy for occupied comfort compliance. While ASHRAE 55 provides comprehensive comfort assessment methods, temperature-band compliance is widely adopted in HVAC control benchmarking to enable transparent and reproducible comparison across control strategies.

All controllers are evaluated under identical models, forecasts, sensing availability, actuation limits, and evaluation windows to ensure a fair and controlled comparison.

All controllers are evaluated under identical simulation models, forecasts, disturbance realizations, actuation bounds, and reporting windows. Therefore, any observed differences can be attributed to the controller structure, including robustness, estimation, and diagnostic capabilities, rather than differences in tuning or information access.

A structured ablation analysis is conducted over scenarios S3 to S7 to isolate the contribution of key cognitive components. All ablation variants retain identical model predictive control weights , operational constraints, simulation models, disturbance realizations, and Monte Carlo seeds (N = 50). Performance metrics are reported as Mean ± 95% confidence interval (CI_95).

The “No attribution” configuration disables root-cause attribution by freezing latent performance states at their nominal priors and relying solely on residual magnitude for drift indication. The “No uncertainty” configuration replaces stochastic disturbance models with deterministic point forecasts and removes uncertainty propagation in both estimation and control. The “No memory” configuration resets diagnostic states at each control interval, preventing persistence of previously identified degradation.

Under the full cognitive digital twin configuration, the system records 18.2 ± 1.2 hours of thermal discomfort, 4.6 ± 0.8 hours of unmet cooling, a drift detection time of 5.1 ± 0.6 hours, and a recovery time of 9.4 ± 0.9 hours.

When root-cause attribution is disabled, thermal discomfort increases to 31.7 ± 2.1 hours and unmet cooling rises to 9.8 ± 1.3 hours. Drift detection is delayed to 11.3 ± 1.4 hours, and recovery time increases to 18.6 ± 1.9 hours, indicating a pronounced degradation in diagnostic effectiveness and post-event resilience.

Removing explicit uncertainty representation results in 27.4 ± 1.8 discomfort hours and 7.9 ± 1.1 unmet cooling hours, with a detection time of 8.6 ± 1.0 hours and a recovery time of 14.1 ± 1.6 hours. This confirms that uncertainty-aware reasoning materially contributes to stability under climate-induced stress.

When diagnostic memory is disabled, the system experiences 24.9 ± 1.6 discomfort hours and 6.8 ± 0.9 unmet cooling hours. Drift detection occurs after 6.2 ± 0.7 hours, and recovery time increases to 12.7 ± 1.3 hours. These results demonstrate that persistent memory of diagnosed states plays a critical role in accelerating recovery following extreme events.

Collectively, the ablation results confirm that performance gains arise from the integrated interaction of attribution, uncertainty handling, and diagnostic memory rather than from predictive control alone.

Results presented in Section 5 demonstrate that the proposed cognitive digital twin consistently outperforms rule-based control, conventional model predictive control, and non-cognitive digital twin baselines across all evaluated climate-extreme scenarios. Importantly, these improvements are not achieved through more aggressive actuation, relaxed comfort constraints, or increased energy expenditure. Instead, they arise from the system’s ability to interpret the causes of performance degradation before modifying control behavior.

Under heatwave-only conditions (S1–S2), performance gains are primarily associated with anticipatory operation. By combining real-time state estimation with uncertainty-aware prediction, the cognitive digital twin mitigates peak demand and thermal stress before deviations propagate into sustained comfort violations. Although conventional MPC also leverages forecasting, it assumes fixed system characteristics and therefore remains vulnerable to gradual shifts in system behavior under prolonged climatic stress.

The distinction between cognitive and non-cognitive operation becomes more pronounced in scenarios involving performance drift (S3–S4). Conventional controllers implicitly treat persistent deviations as control errors, leading to compensatory actions that may mask rather than resolve underlying degradation. In contrast, the cognitive digital twin explicitly diagnoses drift and attributes it to specific mechanisms such as equipment degradation or sensor bias. This diagnostic awareness enables targeted adaptation and avoids unnecessary control oscillations or excessive comfort sacrifice.

A central insight from this study is that early diagnosis, rather than increased control authority, is the dominant mechanism for preserving comfort during climate extremes. In compound stress scenarios combining heatwaves, operational constraints, and elevated uncertainty (S4–S6), baseline controllers frequently violate comfort bounds despite achieving energy- or demand-related objectives.

The cognitive digital twin mitigates this failure mode by distinguishing between apparent and actual thermal stress. When sensor drift is diagnosed, control actions compensate for biased measurements rather than altering physical setpoints, preventing artificial overheating or overcooling. Similarly, when capacity degradation is identified, control policies are adapted to the reduced effective capacity instead of forcing infeasible operation.

These findings indicate that comfort collapse under extreme conditions is often driven by misinterpretation of system state rather than insufficient control capability. Diagnosis-informed adaptation directly addresses this limitation by aligning control actions with the true underlying system condition.

Climate extremes amplify uncertainty in both environmental inputs and system response. Increased temperature variability, forecast errors, and fluctuating internal loads elevate the risk of instability when decision-making relies on point estimates or static models.

By explicitly propagating uncertainty through state estimation and decision-making layers, the proposed framework conditions control actions on confidence bounds rather than deterministic predictions. This mechanism explains the observed reductions in unmet cooling hours and the improved recovery performance under high-uncertainty scenarios (S5–S6).

Notably, the cognitive digital twin does not attempt to eliminate uncertainty. Instead, it operates robustly in its presence by prioritizing operational stability over short-term efficiency when uncertainty is elevated. This behavior contrasts with conventional MPC, which may overreact to uncertain forecasts and inadvertently exacerbate instability during prolonged heatwaves.

Resilience in building energy systems is often framed in terms of redundancy, oversizing, or additional capacity. Results from scenario S7 demonstrate that resilience can also emerge from persistent diagnostic awareness and adaptive memory.

Following the removal of climate stress and operational constraints, the cognitive digital twin retains awareness of unresolved system degradation and continues to adapt control actions accordingly. This behavior explains the substantially shorter recovery times and lower accumulated thermal deviation observed relative to baseline approaches.

From this perspective, resilience is not solely a physical property of the system, but a functional property of how effectively the system understands and responds to its evolving condition over time. Cognitive digital twins enable this form of operational resilience without requiring additional hardware or structural modification.

Cross-climate evaluation in scenario S8 indicates that the proposed framework generalizes across hot-dry, hot-humid, and temperate climates without climate-specific tuning. While the magnitude of improvement varies with climatic characteristics, consistent benefits are observed in all cases.

This robustness arises from reliance on diagnosis-informed adaptation rather than predefined climate heuristics. By reasoning about deviations relative to expected system behavior, the cognitive digital twin remains effective even when absolute temperature levels, humidity profiles, or load patterns differ substantially.

The findings have several practical implications for building energy system design and operation. First, cognitive digital twins offer a pathway to enhance climate resilience without fundamental changes to building hardware or system architecture. Second, diagnostic awareness reduces reliance on conservative control margins, enabling more efficient operation under stress. Third, embedding cognition within operational loops shifts energy management from reactive correction toward proactive adaptation.

Together, these implications suggest a scalable approach for transitioning existing building energy systems toward intelligent operation capable of maintaining performance under increasing climatic uncertainty.

The present study focuses on a medium-scale commercial office building equipped with a centralized variable air volume HVAC system. This configuration represents a widely deployed and well-understood building typology, enabling controlled and interpretable evaluation. However, the results may not directly generalize to buildings with fundamentally different thermal dynamics, operational patterns, or control architectures, such as residential buildings, healthcare facilities, educational campuses, or naturally ventilated structures.

Future research should extend the proposed framework to alternative building types and HVAC configurations, including decentralized systems, hybrid ventilation strategies, and mixed-mode operation, to assess cognitive operation under more heterogeneous and behavior-sensitive conditions.

The evaluation is conducted using scenario-based simulations designed to emulate heatwaves, operational constraints, performance degradation, and sensing disturbances. Although these scenarios are constructed using realistic climate thresholds and physically meaningful disturbance mechanisms, they cannot fully capture the complexity and unpredictability of real-world operation, including unforeseen equipment failures, maintenance actions, or human-in-the-loop interventions.

Field deployment and long-term monitoring studies are therefore required to validate absolute performance levels under real heatwave conditions. Nevertheless, controlled scenario-based stress testing provides a necessary and systematic first step, enabling reproducible benchmarking and causal analysis that would be difficult to achieve in operational buildings.

Accordingly, the reported results should be interpreted as comparative evidence of resilience enhancement across control strategies rather than as guaranteed performance outcomes in real-world deployments.

The cognitive digital twin integrates state estimation, diagnostic reasoning, uncertainty handling, and predictive control within a closed-loop architecture. While the computational requirements remain manageable at the time scales and system size considered in this study, real-time deployment in larger buildings, high-resolution multi-zone models, or district-scale applications may impose additional computational constraints.

Future work should investigate scalable implementation strategies, including model reduction techniques, hierarchical and distributed control architectures, and edge–cloud coordination to balance computational efficiency with diagnostic fidelity.

Uncertainty is explicitly represented within the proposed framework; however, uncertainty models are parameterized using assumed statistical characteristics derived from typical sensor performance and operational variability. In practice, uncertainty structures may evolve over time, exhibit non-Gaussian behavior, or differ substantially across buildings and climates.

Adaptive uncertainty learning, online calibration, and data-driven characterization of disturbance distributions represent promising directions for improving robustness while reducing reliance on predefined assumptions.

Beyond single-building operation, cognitive digital twins offer opportunities for extension and integration at broader spatial and functional scales. Future studies may explore coordination among multiple cognitive twins at district or campus levels, interaction with grid-level signals and demand flexibility programs, and integration with on-site renewable generation and energy storage systems.

The framework may also be extended toward fault prognosis, maintenance planning, and lifecycle-aware optimization, enabling cognitive digital twins to support long-term operational decision-making beyond short-term control.

The computational complexity of the cognitive digital twin is primarily governed by the recursive state estimation and optimization layers. In the Extended Kalman Filter (EKF), the dominant computational cost arises from covariance matrix propagation and inversion operations.

Let n denote the dimension of the state vector. The matrix inversion required during the Kalman gain update results in a computational complexity on the order of O(n^3) per control interval.

For the case study presented in this work, the state dimension remains sufficiently low due to aggregated thermal states and latent performance variables, enabling real-time execution at a 15-minute control interval on standard computing hardware.

However, for large-scale multi-zone buildings or high-resolution digital twins with detailed envelope discretization, the state dimension may increase substantially. In such cases, the computational burden can become significant.

To mitigate scalability challenges, several strategies may be employed:

Therefore, while the proposed framework is computationally tractable for medium-scale deployments, careful architectural design is required for skyscraper-scale digital twins to maintain real-time operability.