Paper deep dive
Choosing an energy-efficient software architecture for building system diagnostic support
Roxane Koitz-Hristov, Franz Wotawa
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Around 30\% of global energy expenditure can be attributed to the building sector, where a large portion of energy-consumption could be avoided by repairing existing faults. Fault detection and diagnosis (FDD) software addresses this issue; however, its creation and operation also have an environmental impact. The magnitude of this impact is influenced by the diagnosis architecture, as different architectures and methods have different energy demands. Yet, simply considering the energy consumed by the software itself is not sufficient to assess its overall environmental impact, since the diagnostic performance, e.g., number of detected faults or number of faults missed, also contributes to its ecological footprint. In this paper, we propose an energy-consumption model that considers FDD performance and energy spend directly by the diagnosis software. In an initial experiment, we compare several FDD architecture families, i.e., rule-based, model-based, classical machine learning, and large-language-model-based, in simulation using performance and energy-consumption values collected from prior literature. The results show that considering the computational energy and accuracy of FDD can change the relative benefit of the different approaches. Computationally efficient machine learning methods, such as random forest, provide the largest net savings on smaller buildings, whereas more resource-intensive approaches, such as fine-tuned large language models, become advantageous as building size increases. Our findings suggest that overall energy efficiency depends not only on the computational demand of the FDD software, but also on its diagnostic performance and the scale of the building.
Tags
Links
- Source: https://arxiv.org/abs/2610.06444v1
- Canonical: https://arxiv.org/abs/2610.06444v1
Trouble viewing inline? Open PDF directly →
Full Text
84,185 characters extracted from source content.
Expand or collapse full text
Choosing an energy-efficient software architecture for building system diagnostic support Roxane Koitz-Hristov Affiliation: Graz University of Technology Affiliation: Institute of Software Engineering and Artificial Intelligence Franz Wotawa Affiliation: Inffeldgasse 16b/2, A-8010 Graz, Austria Affiliation: rkoitz-hristov, wotawa@tugraz.at Abstract Around 30% of global energy expenditure can be attributed to the building sector, where a large portion of energy-consumption could be avoided by repairing existing faults. Fault detection and diagnosis (FDD) software addresses this issue; however, its creation and operation also have an environmental impact. The magnitude of this impact is influenced by the diagnosis architecture, as different architectures and methods have different energy demands. Yet, simply considering the energy consumed by the software itself is not sufficient to assess its overall environmental impact, since the diagnostic performance, e.g., number of detected faults or number of faults missed, also contributes to its ecological footprint. In this paper, we propose an energy-consumption model that considers FDD performance and energy spend directly by the diagnosis software. In an initial experiment, we compare several FDD architecture families, i.e., rule-based, model-based, classical machine learning, and large-language-model-based, in simulation using performance and energy-consumption values collected from prior literature. The results show that considering the computational energy and accuracy of FDD can change the relative benefit of the different approaches. Computationally efficient machine learning methods, such as random forest, provide the largest net savings on smaller buildings, whereas more resource-intensive approaches, such as fine-tuned large language models, become advantageous as building size increases. Our findings suggest that overall energy efficiency depends not only on the computational demand of the FDD software, but also on its diagnostic performance and the scale of the building. 1 Introduction The energy expenditure of residential and commercial buildings constitutes a critical challenge in achieving the United Nations’ climate targets. Recent estimates indicate that the domain is responsible for 40% of the final energy-consumption in the European Union [15] and 30% worldwide [18]. Roughly 75% of the EU building stock is energy inefficient [14], with operational faults as a major contributing factor [8]. Notably, heating, ventilation, and air conditioning (HVAC) systems account for a considerable portion of the energy demand in the building sector. Studies estimate that simply repairing existing HVAC faults could substantially reduce the energy-consumption [2]. Thus, fault detection and diagnosis (FDD) for building systems is a highly relevant strategy for ensuring energy efficiency and the achievement of sustainability goals. Historically, FDD in buildings has primarily relied on knowledge-based approaches [30, 7, 59, 33]. However, in recent years, the widespread adoption of building automation systems (BAS), increasing availability of sensor data, and advancements in data science have led to the development of various machine learning (ML) techniques for FDD [75]. ML has proven to be a suitable approach, which can achieve high detection and diagnostic accuracy while requiring little to no prior knowledge of the system’s underlying composition or dynamics [35, 6, 73, 23]. One goal of FDD in buildings is to improve energy efficiency, yet sophisticated FDD methods can be computationally costly to train and operate, leading potentially to a substantial amount of energy use and therefore additional carbon emissions. The questions of how much energy software requires and which design choices influence its resource consumption have been studied in software engineering previously. In software engineering, sustainability has been established as a multi-dimensional property of software quality in its own right, which includes technical, economic, social, and environmental aspects [37, 64, 5]. In our work, we are interested in the environmental dimension, in particular, in the environmental impact of FDD software. While most definitely the implementation itself, i.e., the underlying code, affects the ecological footprint of software, there have been studies that show that also software architecture can measurably influence the energy demands of a system [16, 50, 25]. In general, architectural decisions shape a system’s structure and behavior. Therefore they influence the quality and sustainability of software products. Moreover, changing design decisions is expensive, especially later in the software life cycle [63]. With the adoption of complex ML methods, which are associated with high energy demands for training [61], the conversation on software sustainability has extended to the domain of artificial intelligence (AI). A recent debate contrasts “Red AI”, which prioritizes performance without regard for environmental cost, with “Green AI”, which favors computational efficiency and reduced energy use while maintaining acceptable accuracy [57, 65]. Recent work addresses the architectural angle within the discussion around sustainable ML with the goal to determine “green” architectural strategies [43, 27]. “Green software” and “Green AI” often aim to reduce the energy-consumption of the software under consideration. In the building domain, the FDD software not only consumes energy itself, but it is also intended to enable energy savings in the building it monitors. Thus, to minimize the overall environmental footprint, one has to consider the energy consumed by the software as well as the energy savings enabled by the software. In practice, this raises the question of which of the available FDD techniques is the most overall energy-efficient solution. Answering this question is more intricate than it initially seems, since the commonly used methods differ largely in the computational resources they require and their diagnostic quality. For instance, diagnosing a building from a small set of threshold-based rules can be done very efficiently, however, if the rule-based system raises a large number of false alarms the additional energy needed for the unnecessary maintenance activities offsets some of the computational efficiency of the FDD approach. The reverse may also happen, where a model is associated with substantially more resources to be able to diagnose a system, e.g., a deep neural network, yet it may be able to detect faults more reliably and therefore may decrease the energy expenditure of the building. Consequently, the computational energy required to operate the diagnosis software has a direct environmental impact, whereas the energy savings resulting from diagnosis activities have an indirect, enabling impact [38]. Often there might be a tradeoff between diagnostic performance and environmental impact, yet both affect the ecological footprint of a building. Thus, in order to answer the question which FDD technique is the most overall energy-efficient solution, we must analyze which combination of the two factors maximizes energy savings. In this paper, we address this question in the context of building HVAC systems by making the following contributions: First, we introduce an energy-consumption model that relates fault detection performance of an FDD method to the energy-consumption of a building affected by faults. Second, we extend this model with energy requirements associated with the training of the diagnosis method as well as its inference. Third, we compile diagnosis performance and energy-consumption data reported for the considered methods from the FDD literature and use the resulting values to compare the methods in simulation experiments to assess the suitability of the different diagnosis architectures and methods. 2 Diagnosis System Architecture In this section, we discuss the different FDD methods considered in this study together with their corresponding high-level software architectures. Afterward, we summarize different evaluation metrics for FDD, which we later use to estimate the energy consumption of the introduced methods. 2.1 Diagnosis methods Rule-based systems have been successfully used in medicine [4] but also in diagnosis of engineered systems [54, 55]. The underlying idea of (rule-based) expert systems is to encode domain expertise as statements such as if observationobservation then causecause to determine the reason behind a certain observation or symptom. Figure 1: Rule-based diagnosis architecture Figure 1 depicts the general architecture of an FDD system utilizing rules for detecting faults. There is a component for fetching observations from the system under surveillance, i.e., a monitoring system [60], which stores the information in a system data base. The evaluation component checks for consistency regarding the given rules, which are stored in a rule base. This evaluation component can be as simple as any ordinary program implementing if-then-else rules or a more sophisticated theorem prover that checks whether the current data is consistent with the rules (see, e.g., [31, 32]). Depending on the underlying detailed architecture, i.e., whether the rule-based diagnosis system relies on simple programs or a theorem prover, the energy-consumption varies. However, the quality of the diagnostic result should not be influenced by details regarding the concrete implementation. Although rule-based systems can encompass a wide variety of knowledge, the interdependence between rules makes this type of system difficult to maintain as the size of the rule base increases. As an answer to this limitation, researchers have proposed model-based diagnosis [11, 12, 53, 13]. Model-based diagnosis differs from rule-based diagnosis in the way faults are detected. The former utilizes a model for detecting a faulty behavior, which requires means for executing the model in order to derive the expected system behavior from the model. The expected behavior is then compared with the observed behavior. Hence, inference is more complex compared to ordinary rule-based systems. Therefore, each inference step has a higher energy-demand. Figure 2: Model-based diagnosis architecture Figure 2 illustrates the overall architecture of a model-based FDD. Even though its architecture is similar to the architecture of rule-based diagnosis, we see that instead of a rule base a system model is exploited. As a consequence the evaluation component has to be more sophisticated and allow for inference whether the observed values from the system database contradict the values obtained from the model. For implementing the evaluation/inference component we can follow two principles, (1) the use of logical models and reasoning engines that are either based on consistency [53, 13, 48] or abduction [17, 68, 36], or (2) the use of mathematical models, e.g., differential equations. Machine learning (ML) is an essential subfield of AI and has gained a lot of attention, mainly starting in the current century, also in the context of FDD. Many ML methods, such as decision trees [51], support vector machines (SVM) [1], neural networks [20], clustering [26], regression [45], or ensemble methods (e.g., random forests [21]) have been used for prediction and classification. In the case of FDD, we are interested in the latter. However, the fundamental idea behind all ML methods is the same: we utilize data to learn a model. In particular, we always distinguish two phases in ML, i.e., the training/learning of a model and its use. Both phases come with different ecological footprints as we will discuss in the following section. Figure 3: Machine learning diagnosis architecture Due to distinguishing training (or learning) from inference and reasoning, the software architecture comprises additional components (see Figure 3). Importantly, we distinguish between training data and system data. The training data is used to learn the model with an algorithm depending on the ML method. The resulting model, similar to model-based diagnosis, is used by an evaluation/inference component, which diagnoses faults based on the model and system data. To work appropriately the system data and the training data should have the same stochastic distribution. Recently, the advances in large language models (LLMs) have led to an increasing amount of research on LLMs in different application domains, e.g., FDD in BAS [39]. LLMs are AI models that predict the next token or sequence of tokens from a given sequence of input tokens. Figure 4 shows the high-level software architecture of an LLM-application for BAS diagnosis. Similar to ordinary ML, there is a component for learning. In practice, the energy-consumption of this component is far higher than for other ML methods, because LLMs are trained on a huge corpus of data and do not focus usually on the system under surveillance. Since LLMs are general-purpose applications, there might be a further training step involved to introduce system specific information—we will later refer to LLMs with additional application-specific training as fine-tuned LLMs. Alternatively, the application might make use of reference system data and RAG in the further evaluation/inference step. Furthermore, inference might be recursively applied until a suitable solution is found, e.g., a classification result can be obtained. Hence, evaluation and inference are also more energy demanding compared to ordinary ML methods. Figure 4: LLM-based diagnosis architecture 2.2 Evaluation metrics Fault detection is a decision problem, i.e., given the observations is a fault present? Thus, fault detection can be understood as a simple binary-classification task, i.e., assigning nominal or non-nominal to a set of current observations. Based on the degree of missclassifications, we can define the quality of the underlying decision procedure or classification method. For this purpose, we use the following definitions: True positive: A true positive is an observation from a faulty system that is classified as being faulty. True negative: A true negative is an observation from a correct system that is classified as being correct. False positive: A false positive is a correct observation that is classified as being faulty. False negative: A false negative is a faulty observation that is classified as being correct. Based on these types of classification, we can count the number of true positives, true negatives, false positives, and false negatives and store this information in TPTP, TNTN, FPFP, and FNFN, respectively. In the rest of this subsection, we use this information, to describe quality metrics of classification methods. The true positive rate also called sensitivity, is the probability that a test method and in our particular test the fault detection method finds the fault, i.e., the true positive. TPR=TPTP+FNTPR= TPTP+FN The false positive rate or type 1 error indicates the probability that the diagnosis method indicates a fault even though there is no fault in the real system. FPR=FPTN+FPFPR= FPTN+FP The true negative rate or specificity is the probability that a negative test is correct, i.e., that no fault is present. TNR=TNTN+FPTNR= TNTN+FP The false negative rate or type 2 error is the probability that a method misses a fault in a system, i.e., is not able to detect it. FNR=FNTP+FNFNR= FNTP+FN It is worth noting that the FNRFNR and the TPRTPR are not independent, i.e., FNR=1−TPRFNR=1-TPR holds. Similarly, there is also a dependence between FPRFPR and TNRTNR, i.e., FPR=1−TNRFPR=1-TNR. Alternatively, and originating from the field of information retrieval, there are additional metrics describing the accuracy of an approach, i.e., precision, recall, and the F1 score. Precision is the fraction of the number of true positives divided by the sum of the true positives and false positives. Hence, precision indicates the ratio of the real faults in all faults detected by a diagnosis method. Precision=TPTP+FPPrecision= TPTP+FP Recall is equivalent to the true positive rate, i.e.: Recall=TPRRecall=TPR The F1 score is the harmonic mean of precision and recall. F1=2⋅Precision⋅RecallPrecision+RecallF1=2· Precision· RecallPrecision+Recall Precision, recall, and F1 score are the most commonly used quality metrics reported for classification methods. Unfortunately, from these figures we are not able to compute all the other metrics that are relevant in the ecological model we will present in the next section. For instance, the false positive rate (FPRFPR) cannot be computed directly from those other measures. 3 Energy-consumption Model Answering the question of which diagnosis methodology and straightforwardly its corresponding software architecture to use in order to obtain the most energy-efficient solution requires a detailed analysis of (1) the method’s energy-consumption and (2) its effects on the energy-demand of the system to be diagnosed. Before developing our energy-consumption model, we need to state its scope. In our modeling approach, we consider a horizon of one year and assume that the building, its HVAC system and the deployed diagnosis method remain unchanged over this period. In particular, we do not model the aging of components or the associated growth in fault probability of a system’s lifetime. Additionally, we represent faults as an aggregate through an annual fraction of wasted energy rather than as a distribution over individual faults. We further restrict our method to consider fault detection, rather than fault identification and repair. 3.1 Energy-Consumption Model (ECMECM) We start defining the energy-consumption of buildings for cooling and heating EHVACE_HVAC considering the energy use intensity EUIEUI of a building (i.e., the fault-free required energy per square meter and year), its square meters A, and the fraction of the energy-consumption used for heating or cooling sHVACs_HVAC. EHVAC=sHVAC⋅EUI⋅AE_HVAC=s_HVAC· EUI· A (1) Faults often lead to an additional energy-demand that has to be considered. For example, a broken sensor indicating the room has not yet reached the required temperature or a non-optimal setup of parameters in HVAC may cause the system to use more energy for heating. Hence, from a general perspective the additional energy lost is proportional to the optimal energy used for heating and cooling. For computing this wasted energy, we introduce the fraction of energy lost sFAULTs_FAULT in the case of a fault. EWASTED=sFAULT⋅EHVACE_WASTED=s_FAULT· E_HVAC (2) It is worth noting that the wasted energy for a fault may be up to 40% [2]. Note further that a more detailed model may consider different types of faults, fault frequency, and fault duration. However, in any case the sum of wasted power over time can be summarized in the annual fraction of energy lost. The total energy needed for heating and cooling ETHVACE_THVAC of a building in one year without considering diagnosis is stated in the following equation: ETHVAC=EHVAC+EWASTEDE_THVAC=E_HVAC+E_WASTED (3) After defining the model of energy-consumption of HVAC systems in buildings considering faults, we now focus on the impact of FDD on BAS. In particular, we reduce our view on FDD to fault detection and ignore fault localization, i.e., identifying the root cause of the fault. The reason is that several diagnosis approaches in the context of BAS only consider the detection of an unwanted or unexpected behavior but not on its correction, which often is carried out by maintenance personnel. Unfortunately, no approach guarantees a 100% detection rate. Therefore, we have to consider two errors. First, the case that an FDD approach classifies a value to be wrong but it is not, which raises a wrong alarm. Such a false positive error or type 1 error, causes further investigations but does not necessarily lead to a degraded system. Second, we may not be able to classify a wrong value as indicating a fault. Such a false negative error or type 2 error, does not lead to an alarm and a fault remains. Depending on the classification outcome, we now describe the impact on the energy-consumption. True positives (TP): There is a fault detected, which is a real fault. Hence, when correcting the fault, e.g., by replacing a component or changing a setup, a saving in the energy-consumption, which is proportional to the wasted energy due to faults, is expected. The factor of detected faults is given by the true positive rate TPRTPR. In our case, we consider an annual TPRTPR, i.e., TPRaTPR_a, since our time-horizon spans over a single year. ESAVING=TPRa⋅EWASTEDE_SAVING=TPR_a· E_WASTED (4) If all faults can be detected, then there is no energy wasted. Note that in practice, we may still see a wasted energy due to faults, because there is always a delay of repair, which causes additional energy to be wasted. However, compared to the total amount of energy, this remaining wasted energy can be neglected. True negatives (TN): There is no fault represented in the observations. Hence, no action has to be taken and thus there is no impact on energy-consumption. False positives (FP): False positives lead to unnecessary maintenance activities, which may cause additional energy-consumption. Besides the energy that is needed to perform the repair itself, a false positive may lead to the unnecessary replacement of components still operating as intended. These are, however, secondary energy impacts. Overall, we assume a wasted energy for FPs, which is proportional to the false positive rate FPRFPR but only applied to a fraction of the total energy needed for HVAC. The latter is represented by the factor sFPs_FP. Again, we consider FPRFPR as an annual rate FPRaFPR_a. EWASTED_FP=FPRa⋅sFP⋅EHVACE_WASTED\_FP=FPR_a· s_FP· E_HVAC (5) False negatives (FN): Not identifying all faults causes wasted energy. However, this is already part of the energy-consumption model, since EWASTEDE_WASTED already accounts for the additional energy-demand in case of a fault. To complete the energy-consumption model considering any diagnosis method, we add the following equation, which defines the total energy-consumption after diagnosis of a BAS. EHVAC_DIAG=ETHVAC−ESAVING+EWASTED_FPE_HVAC\_DIAG=E_THVAC-E_SAVING+E_WASTED\_FP (6) Equations 1 to 6 define the overall energy-consumption model ECMECM. This model already captures important diagnosis-related parameters, i.e., the true positive rate, the false positive rate, and the false negative rate. Other aspects, which are related to the diagnosis methodology are not captured. Table 1 depicts all input parameters of the ECMECM model. Table 1: Input parameters of the energy-consumption model ECMECM for FDD applied to BAS. TPRTPR and FPRFPR are interpreted as annual averages over the considered year. Parameter Explanation Unit Range EUIEUI Energy use intensity per year kWh/(m2⋅^2·year) [0,∞) A Area of building m2 [0,∞) sHVACs_HVAC Fraction of energy-consumption used for HVAC 1 [0,1] sFAULTs_FAULT Fraction of HVAC energy-consumption in case of faults 1 [0,1] TPRaTPR_a Annual true positive rate of diagnosis method 1 [0,1] FPRaFPR_a Annual false positive rate of diagnosis method 1 [0,1] sFPs_FP Fraction of energy needed to handle FPs 1 [0,1] 3.2 Extended Energy-Consumption Model (EECMEECM) In the following, we extend ECMECM to consider diagnosis method specific parts, which are influenced by the methods’ architectures, discussed in the previous section. The extended ECMECM, i.e., EECMEECM comprises the information required for model creation and reasoning separately. Let us start with model creation and in particular model learning. The energy-consumption depends on the method’s energy demands for training ETMETHODE_TMETHOD, its utility U, i.e., the degree to which the model is used exclusively for diagnosis of the BAS, and the number of repetitions necessary per year for keeping the model up to date, i.e., fRETf_RET. ETRAINING=ETMETHOD⋅U⋅(1+fRET)E_TRAINING=E_TMETHOD· U·(1+f_RET) (7) Note that for Equation 7, we consider the first year, in which there is one initial training, followed by possibly additional retraining captured by fRETf_RET. For subsequent years, only fRETf_RET would remain. Retraining might be necessary due to data drift [77]. In our EECMEECM, we assume that retraining requires the same energy as the original training; this assumption, while reasonable, is not necessarily true the longer the system is in use. For instance, in case additional training data is available, the energy-demand for retraining possibly increases. The utility U allows us to allocate training energy in proportion to model use, i.e., U=NINFNTOTALU= N_INFN_TOTAL, where NINFN_INF is the number of diagnosis inferences per year and NTOTALN_TOTAL encompasses all inferences. Thus, for a model trained specifically for fault detection, all inferences serve the diagnosis task and U=1U=1, while for a general-purpose model diagnosis may only constitute a part of the total usage. For inference or reasoning the total energy-consumption depends on the number of iterations per year NINFN_INF and the base energy-consumption for one reasoning or inference step EBINFE_BINF. EINFERENCE=EBINF⋅NINFE_INFERENCE=E_BINF· N_INF (8) Hence, the total energy-consumption ignoring the energy-consumption of required computing infrastructure etc. is given as follows: EDIAG=ETRAINING+EINFERENCE_DIAG=E_TRAINING+E_INFERENCE (9) The total energy-consumption of HVAC diagnosis considering the different methods for the extended ECMECM, which we call EECMEECM is: EXTENDED_HVAC_DIAG=EHVAC_DIAG+EDIAGE_EXTENDED\_HVAC\_DIAG=E_HVAC\_DIAG+E_DIAG (10) In Table 2, we summarize the input parameters of the EECMEECM. Together with the parameters of the ECM in Table 1, we can determine the annual energy-consumption of a building for cooling and heating. Table 2: Input parameters of the extended energy-consumption model EECMEECM for FDD applied to BAS. Parameter Explanation Unit Range U Degree to which the model is used exclusively for diagnosis 1 (0,1] fRETf_RET Number of additional model retrainings within the considered year 1 ℕ0N_0 EBINFE_BINF Energy-consumption of one inference operation kWh/inference [0,∞)[0,∞) NINFN_INF Number of inference operations within the considered year inferences/year ℕ0N_0 4 Simulation Experiments In our simulation, we consider different architectures and methods for FDD: rule-based (RB), model-based diagnosis (MBD), ML, and LLMs. For the ML architecture, we review various models at different levels of complexity, i.e., decision trees (DT), random forest (RF), logistic regression (LR), SVM, and multi-layer perceptrons (MLP), i.e., neural networks. For the LLM analysis, we consider LLMs used as-is and fine-tuned LLMs (LLM FT). 4.1 Input Data For the experiments reported in this paper, we obtain the input values from research on FDD in the building domain. In particular, our starting point was a recent review by Chen et al. [6] on data-driven FDD methods for HVAC systems. We considered all comparison papers that perform FDD on chiller or AHU systems to obtain the rates reported in Table 3. To this end, we extracted all usable TPRTPR/Recall and FPRFPR values for our considered diagnosis methods from the reviewed studies. Since, however, many studies we examined do not provide the necessary data, i.e., TPRTPR/Recall, FPRFPR, or a confusion matrix to compute the metrics, we utilized additional chiller- and AHU-specific sources in order to obtain at least a single published value per FDD method. Furthermore, the energy model of Section 3 requires annual values, i.e., TPRaTPR_a and FPRaFPR_a, which describe the behavior of the model over a full year. Even though the reported TPRTPR and FPRFPR from the literature are based on a single experiment or a set of experiments, we use a simplification by assuming that the rates remain stable throughout the year. Thus, for our simulation, we use the range of values we were able to find in the literature for TPRTPR/FPRFPR as an approximation for TPRaTPR_a/FPRaFPR_a. Table 3: Minimum, maximum, and average reported TPR and FPR for Chiller and AHU FDD methods extracted from the literature reviewed by Chen et al. [6]. Summaries of papers used can be found in A. Diagnosis method TPR FPR Min Max Avg. Min Max Avg. RB 0.11 0.39 0.25 0.00 0.05 0.03∗ (Hybrid)MBD 0.20a 1.00a 0.60 0.00 0.07 0.03 DT 0.17 1.00 0.61 0.00b0.00^b 0.03c 0.01 RF 0.17 1.00 0.69 0.000.00 0.000.00 0.00†0.00 LR 0.38 0.44 0.41∗ 0.00e 0.53f 0.27∗ SVM 0.16 1.00 0.714 0.00 0.28 0.09 MLP 0.17 1.00 0.550 0.02 0.13 0.07∗ LLMd 0.13 0.59 0.28 0.10 0.27 0.16 LLM FTe,d 0.98 1.00 0.99 0.00 0.02 0.01 Additional sources: aa Cui and Wang [10], bb Padmanabh et al. [49], cc McHugh [44], Zhang et al. [74]dd, ee Langer [40], and ff Islam et al. [24]. ∗ Due to a limited number of values (≤3≤ 3), we used the mid-range (Min+Max)/2(Min+Max)/2 to approximate an average value. † Only one exact numerical FPR value was available, which is hence reported as the midpoint. Table 4: Energy consumption (in kWh) required for different fault detection methods for a single training/inference run. Diagnosis method ETMETHODE_TMETHOD U EBINFE_BINF Ref. Min Max Min Max RBa 0 0 1 1.0⋅10−81.0· 10^-8 2.5⋅10−72.5· 10^-7 [34] MBDa 0 0 1 4.4⋅10−54.4· 10^-5 4.6⋅10−44.6· 10^-4 [33] DT 3.13⋅10−43.13· 10^-4 3.13⋅10−43.13· 10^-4 1 8.60⋅10−58.60· 10^-5 8.60⋅10−58.60· 10^-5 [22] RF 2.45⋅10−72.45· 10^-7 1.06⋅10−61.06· 10^-6 1 2.47⋅10−82.47· 10^-8 6.35⋅10−86.35· 10^-8 [56] LR 3.85⋅10−83.85· 10^-8 3.42⋅10−73.42· 10^-7 1 2.10⋅10−82.10· 10^-8 4.41⋅10−84.41· 10^-8 [56] SVM 6.04⋅10−86.04· 10^-8 3.15⋅10−53.15· 10^-5 1 3.31⋅10−83.31· 10^-8 2.65⋅10−62.65· 10^-6 [56] MLP 2.69⋅10−32.69· 10^-3 3.24⋅10−23.24· 10^-2 1 4.50⋅10−54.50· 10^-5 1.15⋅10−41.15· 10^-4 [22] LLM 1.5⋅1031.5· 10^3 62.32⋅10662.32· 10^6 1.1⋅10−71.1· 10^-7 2⋅10−62· 10^-6 2.9⋅10−32.9· 10^-3 [28]b LLM FT 1.65⋅1031.65· 10^3 68.55⋅10668.55· 10^6 1.1⋅10−71.1· 10^-7 2⋅10−62· 10^-6 2.9⋅10−32.9· 10^-3 [28]c a Estimated energy-consumption. b See Table 3 and 5 of the mentioned paper. Inference energy-consumption also considers more sophisticated queries to capture agentic systems requiring more than one LLM query for providing an outcome. c Like for ordinary LLMs. For fine tuning 10% more energy for training assumed. Besides the information on the prediction accuracy of the considered architectures and methods, our extended energy model EECMEECM relies on information on the energy-consumption of training the model (i.e., ETMETHODE_TMETHOD), the utility factor U, and the energy consumed by performing a single inference step (i.e., EBINFE_BINF). Table 4 depicts those figures based on prior studies not necessarily in the building domain. As one can observe, for RB and MBD we assign a training energy of zero, since neither involves a learning stage. In both cases, the knowledge is derived by domain experts rather than from data, thus, no computation is performed before deployment. In addition, the energy-consumption of inference for RB and MBD was estimated from the runtimes reported in the cited studies. Runtime is not a direct measure of environmental impact; however, studies [3, 76, 67] indicate that it can serve as a proxy for energy consumption. Since all experiments [34, 33] were conducted on the same machine11 1 Mac Pro equipped with a 12-core Intel Xeon E5 processor at 2.7 GHz and 64 GB RAM we estimate the energy-consumption from the reported runtime as Eest=Pestt3600,E_est=P_est t3600, where t denotes the runtime in seconds and PestP_est the estimated average computational power in kW. An average power draw of Pest=0.12P_est=0.12 kW was obtained from CodeCarbon [9] measurements performed on the same machine with similar experiments. Considering the utility U, we assign a value of one to all methods except LLMs. The reason is that in the other cases, whether those are rules or ML models, they are created for a specific purpose. Hence, those systems cannot be used for an arbitrary task, whereas LLMs can be used for different applications without, in theory, requiring additional training. Hence, for LLMs we approximate U based on the widely used deployed conversational model ChatGPT. Approximately 2.5⋅1092.5· 10^9 prompts are processed per day [58]. Thus, there are around NTOTAL≈9.1⋅1011N_TOTAL≈ 9.1· 10^11 queries per year. Assuming NINF=105N_INF=10^5 diagnostic queries annually yields a utility of ≈1.1⋅10−7≈ 1.1· 10^-7. While one can argue that the model would have been trained regardless of the diagnosis applications using it, we consider a proportional training. In the case of the fine-tuned LLM, we add an additional 10% to account for the fine-tuning. 4.2 Simulation Results To carry out the experiments, we implemented both models and also all experimental setups in R [52] version 4.3.2. The source code is available at Zenodo [69]. We used a MacBook Pro with an Apple M4 Pro processor and 48 GB main memory and macOS Tahoe 26.6.1. First Experiment In the initial experiment, we estimated the energy-consumption for all methods assuming the energy needed for a mid-sized office building to be 250kWh/m2250~kWh/m^2 and a varying total number of square meters ranging from 1,000 to 10,000m2m^2. We further assume that shvacs_hvac is uniformly distributed between 0.1 and 0.8, sfaults_fault between 0.1 and 0.4, and sfps_fp between 0.01 and 0.1. For the other parameters we used the minimum and maximum values stated in Tables 3 and 4 and again assume a random uniform distribution. Furthermore, we assume that there is only one training needed and that there is a measurement every 10 minutes. For this and all the other experiments, we created 1,000,000 random instances using the mentioned parameters and their ranges. Table 5: The energy-consumption calculated for the different FDD methods considering different building areas A using ECMECM and EECMEECM together with their computed potential savings in percent. A Method ethvace_thvac ehvac_diagECMe_hvac\_diag^ECM ehvac_diagEECMe_hvac\_diag^EECM savingECMsaving^ECM savingEECMsaving^EECM 1 1000.00 RB 140584.96 133715.29 133715.30 4.89 4.89 2 1000.00 MBD 140584.96 123932.96 123946.20 11.84 11.84 3 1000.00 DT 140584.96 124226.09 124230.61 11.64 11.63 4 1000.00 RF 140584.96 124133.26 124133.26 11.70 11.70 5 1000.00 LR 140584.96 130696.54 130696.55 7.03 7.03 6 1000.00 SVM 140584.96 125138.82 125138.89 10.99 10.99 7 1000.00 MLP 140584.96 124612.48 124616.70 11.36 11.36 8 1000.00 LLM 140584.96 131598.32 131678.10 6.39 6.34 9 1000.00 LLM FT 140584.96 112808.76 112888.80 19.76 19.70 10 2500.00 RB 351462.39 334271.80 334271.81 4.89 4.89 11 2500.00 MBD 351462.39 309833.46 309846.72 11.84 11.84 12 2500.00 DT 351462.39 310565.62 310570.14 11.64 11.63 13 2500.00 RF 351462.39 310360.13 310360.13 11.69 11.69 14 2500.00 LR 351462.39 326738.53 326738.54 7.03 7.03 15 2500.00 SVM 351462.39 312898.31 312898.38 10.97 10.97 16 2500.00 MLP 351462.39 311502.08 311506.30 11.37 11.37 17 2500.00 LLM 351462.39 329019.95 329099.71 6.39 6.36 18 2500.00 LLM FT 351462.39 282021.64 282101.68 19.76 19.73 19 5000.00 RB 702924.78 668566.56 668566.57 4.89 4.89 20 5000.00 MBD 702924.78 619599.53 619612.78 11.85 11.85 21 5000.00 DT 702924.78 621191.87 621196.39 11.63 11.63 22 5000.00 RF 702924.78 620689.99 620689.99 11.70 11.70 23 5000.00 LR 702924.78 653487.94 653487.95 7.03 7.03 24 5000.00 SVM 702924.78 625690.03 625690.10 10.99 10.99 25 5000.00 MLP 702924.78 623070.25 623074.47 11.36 11.36 26 5000.00 LLM 702924.78 658031.17 658110.87 6.39 6.38 27 5000.00 LLM FT 702924.78 564044.76 564124.84 19.76 19.75 28 7500.00 RB 1054387.17 1002832.39 1002832.40 4.89 4.89 29 7500.00 MBD 1054387.17 929476.58 929489.82 11.85 11.85 30 7500.00 DT 1054387.17 931730.52 931735.04 11.63 11.63 31 7500.00 RF 1054387.17 931006.16 931006.17 11.70 11.70 32 7500.00 LR 1054387.17 980218.78 980218.79 7.03 7.03 33 7500.00 SVM 1054387.17 938488.90 938488.97 10.99 10.99 34 7500.00 MLP 1054387.17 934476.95 934481.18 11.37 11.37 35 7500.00 LLM 1054387.17 987055.75 987135.41 6.39 6.38 36 7500.00 LLM FT 1054387.17 846068.73 846148.77 19.76 19.75 37 10000.00 RB 1405849.55 1337118.70 1337118.71 4.89 4.89 38 10000.00 MBD 1405849.55 1239392.96 1239406.21 11.84 11.84 39 10000.00 DT 1405849.55 1242282.22 1242286.74 11.63 11.63 40 10000.00 RF 1405849.55 1241465.55 1241465.55 11.69 11.69 41 10000.00 LR 1405849.55 1306978.81 1306978.81 7.03 7.03 42 10000.00 SVM 1405849.55 1251363.91 1251363.98 10.99 10.99 43 10000.00 MLP 1405849.55 1246055.11 1246059.34 11.37 11.37 44 10000.00 LLM 1405849.55 1316085.38 1316165.07 6.39 6.38 45 10000.00 LLM FT 1405849.55 1128086.80 1128166.75 19.76 19.75 Table 5 shows the obtained estimates of the mean energy-consumption of the overall building’s HVAC with and without diagnosis using ECMECM and EECMEECM. Obviously, the relative energy savings for the ECMECM do not change for the different building areas, because all terms scale proportionally to the area. Thus, differences between the accuracy of the underlying FDD methods only depend on their diagnostic performance, i.e., TPRTPR and FPRFPR. For the ECMECM we see the following ordering of methods when using the energy savings: LLM FT with almost 20% followed by MBD, RF, DT, MLP, and SVM all with approximately than 11% saving, and finally LR, LLM, and RB. The overall savings range from about 5 to 20% depending on the methods. For EECMEECM there is also no difference between the savings for different areas of the building. Moreover, both models yield almost identical savings. This indicates that the energy-consumption of training and inference for the assumed values in the first experiment can be neglected. Second Experiment In the second experiment, we focused on the dependency of the number of trainings over time on the impact on diagnosis. Hence, we only consider the EECMEECM. We set the overall building area A to 10,000. We range the number of trainings from 1 (which means 0 repetitions) to 12 (11 training repetitions), i.e., every month we have a training. All other parameters remain the same as for the initial experiment. Figure 5 depicts the outcome of the second experiment. We see no impact on any method when varying the number of trainings for the building sector. 002244668810101212446688101012121414161618182020number of additional trainingssavings in %Savings as function of the number of trainingsRBMBDDTRFLRSVMMLPLLMLLM FT Figure 5: Results of the second experiment. Third Experiment The third experiment is similar to the second. Instead of varying the number of trainings, we vary the number of inferences from 1 minute to 60 minutes. The building area remains 10,000, and we assume only one training. The other parameters remain the same. The selected inference intervals are based on experience with building monitoring and sensor-data collection. 00101020203030404050506060446688101012121414161618182020time between observationssavings in %Savings as function of the time between observationsRBMBDDTRFLRSVMMLPLLMLLM FT Figure 6: Results of the third experiment. Figure 6 shows the results of the third experiment. Again, we do not see any impact. Hence, for the building sector only the TPRTPR and FPRFPR matter for selecting the best diagnosis method and their corresponding software architecture. Fourth Experiment he fourth experiment examines whether using mean TPRTPR and FPRFPR values instead of maximum values affects the estimated savings. As in the previous experiments, we still consider office buildings and use a uniform distribution for the following parameters: area ranging from 1,000 to 10,000, energy consumption per year from 200-300kWh/m2, observation time from 1 to 60 minutes, and the number of training steps in total ranging from 1 to 12. The other parameters remain the same as in the first experiment. Table 6: The results of the 4th experiment. Method ethvace_thvac ehvac_diagECMe_hvac\_diag^ECM ehvac_diagEECMe_hvac\_diag^EECM savingECMsaving^ECM savingEECMsaving^EECM 1 RB 774327.98 736628.85 736628.86 4.87 4.87 2 MBD 774327.98 682419.45 682428.60 11.87 11.87 3 DT 774327.98 680189.37 680192.49 12.16 12.16 4 RF 774327.98 667458.02 667458.03 13.80 13.80 5 LR 774327.98 720022.36 720022.36 7.01 7.01 6 SVM 774327.98 666806.41 666806.46 13.89 13.89 7 MLP 774327.98 691526.14 691529.17 10.69 10.69 8 LLM 774327.98 736410.43 736485.47 4.90 4.89 9 LLM FT 774327.98 621333.45 621410.80 19.76 19.75 Table 6 depicts the outcome. There is no difference in the overall ranking. It is worth noting that there is a minor decrease in savings for the LLM and LLM FT due to the high energy-consumption required for training. This drop is, however, minor. For the other methods the computed savings are the same regardless of using ECMECM or EECMEECM. Fifth Experiment In the experiment, we assume a single-family house considering uniform distributions of the following parameters: usable floor area from 100-200m2, total energy-consumption of 50-100kWh/m2, time between observations from 1 to 60 minutes, and number of trainings ranging from 1 to 12 in total. We further adapted sHVACs_HVAC to range from 0.6 to 0.8 covering the fact that most of the energy in households is used for heating and cooling. The remaining parameters are the same as for the first experiment. Table 7: The results of the 5th experiment for single-family homes. Method ethvace_thvac ehvac_diagECMe_hvac\_diag^ECM ehvac_diagEECMe_hvac\_diag^EECM savingECMsaving^ECM savingEECMsaving^EECM 1 RB 8531.38 8115.97 8115.97 4.87 4.87 2 MBD 8531.38 7518.62 7527.77 11.87 11.76 3 DT 8531.38 7494.05 7497.18 12.16 12.12 4 RF 8531.38 7353.76 7353.76 13.80 13.80 5 LR 8531.38 7932.98 7932.98 7.01 7.01 6 SVM 8531.38 7346.58 7346.63 13.89 13.89 7 MLP 8531.38 7618.97 7621.99 10.69 10.66 8 LLM 8531.38 8113.56 8188.57 4.90 4.02 9 LLM FT 8531.38 6845.51 6922.69 19.76 18.86 Table 7 shows the obtained results. Due to the smaller amount of total energy-consumption, we now see a difference between the savings considering ECMECM and EECMEECM. For this class of buildings, considering the energy demand of inference and trainings makes a difference, but does not change the overall picture significantly. Again LLM FT performs best in terms of savings, followed by SVM and RF (which are close together). Afterward, we have DT, MBD, MLP, RB, and finally LLM without fine tuning. Due to the low TPRTPR of LLM and the energy needs for training and inference, it is ranked last. Sixth Experiment In the sixth experiment we used the same parameter values as in the fifth experiment with one exception. We do not randomly distribute the time between observations, but use it as a parameter to illustrate the impact of the number of inference steps on the overall energy-consumption and savings. 10010^010110^1−5-50055101015152020time between observationssavings in %Savings as function of the time between observationsRBMBDDTRFLRSVMMLPLLMLLM FT Figure 7: Results of the sixth experiment. Figure 7 depicts the outcome and shows the dependency of energy-consumption on the time between observations for single-family houses. Note that the x-axis is log-scaled. From this figure we see that if we apply inference every minute, approaches such as RF and SVM are superior. If this time is increased to 2 minutes, LLM FT takes over. Hence, fine-tuned LLMs are still the method of choice when considering more realistic inference frequencies of about 5 to 10 minutes for buildings. However, for other application areas where there is a higher demand for detecting faults, other FDD approaches might provide more total energy savings. 5 Limitations Our study combines a theoretical energy-consumption model with simulation experiments using values from prior literature. The validity of the results and conclusions drawn therefore depends on the model’s underlying assumptions and simplifications, as well as the quality of the input data. 5.1 Model assumptions As with any theoretical model, our proposed ECMECM and EECMEECM rely on several assumptions and simplifications that limit how accurately they represent a concrete building and its HVAC system. First, we consider a one-year time horizon rather than the entire building life cycle. During this period, we assume the building, its HVAC system, and the FDD methods used remain unchanged. Consequently, our models do not capture long-term changes, such as building modifications, replacement or degradation of HVAC components, or updates to the FDD system. We also do not model component aging and the related increase in fault probability. Given these simplifications, the energy-consumption models should be interpreted as approximations rather than precise representations of a building’s behavior over its full lifetime. Second, we combine fault effects into a single parameter. Instead of modeling individual fault types with different severities, occurrence frequencies, and durations, we approximate the effects of all faults as a fraction of the annual HVAC energy loss. This simplification allows us to compare the different diagnosis architectures and methods without requiring a detailed fault model for each specific HVAC system. Consequently, our models do not account for the distinct ecological impacts of individual faults. Third, our models do not consider fault identification, localization, and repair. We assume that once a fault is detected, it is immediately corrected and the associated energy savings take effect. In practice, however, detecting a fault is often not sufficient. In order to restore the nominal system behavior the faulty component has to be identified. Furthermore, there is typically a delay between fault detection and the corresponding repair action. During this interval, the fault persists and continues to waste energy. Our models do not factor in the additional energy and resources that may be attributed to the repair and replacement activity, e.g., recommissioning. We assume these effects on the operational energy are negligible compared to the annual HVAC energy-consumption. Finally, as noted above, our models capture building scale through area and energy use intensity but do not explicitly represent the complexity of the HVAC system. Two buildings with a similar area and energy demand may differ in the number of HVAC components (i.e., the possible subsystems that may experience faults) and the effort required to diagnose and maintain the system. Our models do not directly account for this. 5.2 Validity of input data and parameter estimates Our simulation experiments are based on values obtained from prior literature on the diagnostic accuracy of FDD methods and the ecological impact of various architectures and methods. The reliability of our results therefore depends on the quality, comparability, and representativeness of the underlying input data. While we rely on prior published research for our data collection, it was not conducted as an exhaustive systematic review; consequently, other relevant studies may contain values outside the ranges reported. Diagnostic performance data Our starting point was a recent review by Chen et al. [6], which analyzes research on FDD diagnosis techniques for HVAC systems. Thus, the studies selected from the review differ with regard to the HVAC systems under investigation, data sets, fault types, operating conditions, experimental setups, and the diagnosis tasks considered (i.e., fault detection or identification). Consequently, the values we obtained are not directly comparable, but they reflect the heterogeneity across HVAC FDD. The availability of suitable performance data has been an issue. Many studies we examined do not report the required metrics (i.e., TPRTPR, FPRFPR, or confusion matrices from which these metrics can be derived), nor do they provide reproducibility packages containing the results. Therefore, we included additional sources where necessary. Yet, for methods where only a few values have been compiled, the resulting estimates carry greater uncertainty. Another simplification concerns the temporal interpretation of the reported performance values. The literature values are rooted in individual experiments or sets of experiments rather than observations over an entire year. Yet our models require the annual true positive and false positive rates. We therefore assume the reported values are suitable approximations for TPRaTPR_a and FPRaFPR_a, although in practice, diagnostic performance over a year may vary, e.g., due to data drift. We partially address these concerns by documenting the data extraction and relying on minimum, maximum, and average values rather than individual values reported in the literature. Computational energy data As with the diagnostic performance data, we obtained the computational energy values from prior literature. However, we did not limit our analysis to HVAC applications, since application-specific energy measurements are sparse in the building domain. Consequently, the reported values stem from different application domains, hardware platforms, software frameworks, model configurations, and experimental setups. We acknowledge that all of these factors can substantially influence an FDD method’s energy-consumption. However, we retain the values and use their ranges to capture the variability. In addition, suitable energy data were not available for all architectures. In particular, for RB and MBD, we rely on estimates derived from reported runtimes. Although this aligns with prior research indicating a relationship between runtime and energy-consumption [3, 76, 67], these values remain approximations introduced to compare the various FDD architectures. For LLMs, we add further assumptions. Since LLMs are not used exclusively for building FDD, we allocate only a fraction of the overall LLM training cost to the diagnosis task. We also approximate the additional energy required for fine-tuning an existing LLM. As a result, the computational energy values should be interpreted as approximations. Lastly, we do not account for all energy expenditures associated with developing and maintaining the different diagnosis methods. For example, feature selection, hyperparameter tuning, prompt design, rule creation, and creating models from first principles are not included in the reported training energy. While these activities may require substantial effort and computational resources, their energy demand is difficult to quantify consistently across the considered architectures. 5.3 Generalizability First, the generalizability of our simulation results is limited by the HVAC subsystems considered. In particular, we focus on chiller and AHU applications. Consequently, the numerical results do not necessarily transfer directly to other HVAC systems or fault types. Furthermore, the relative benefit of a diagnosis approach depends on characteristics of the monitored systems. While we cover a selected range of parameters (e.g., building area), we do not consider other potentially relevant properties, including occupancy patterns, environmental conditions, or usage profiles. Accounting for such factors may change the relative benefit and ranking of the FDD methods considered. Finally, our current experiments rely on literature values obtained from specific implementations of the considered architectures and methods. A highly optimized implementation of a particular method may exhibit substantially different diagnostic performance or computational energy demand than the ranges we currently report. Therefore, the rankings observed in our simulations should not be interpreted as universal rankings of the FDD methods, but rather as an illustration of the underlying relationship: the preferred architecture depends on diagnostic performance, computational energy demand, and building scale. 6 Conclusions In this paper, we investigate a selection of software architectures for FDD in building systems. We introduce an energy-consumption model that relates the performance of an FDD architecture/method to the energy consumption of the monitored HVAC system. In an extended model, we also account for the energy-demand of model creation/learning and inference. We utilize the models in a set of initial simulation experiments to compare the various architectures and methods based on performance and energy expenditure values collected from prior literature. The results indicate that for office buildings, the energy required for training and inference is negligible, since the energy demand of the HVAC system itself accounts for a substantial portion of the energy-consumption. In this case, fine-tuned LLMs provide the largest net energy savings. For single-family houses, the architecture’s energy-demand becomes more relevant as the ecological footprint of heating and cooling is noticeably smaller. Nevertheless, fine-tuned LLMs also achieve the largest energy savings in this setup. Our experiments further show that inference frequency can influence the ranking of software architectures. Overall, there is no universally most energy-efficient FDD architecture. Rather, the superiority of an approach depends on the relationship between diagnostic performance, energy-consumption for training and inference, and the characteristics and operation of the monitored building. While our experiments allow us to draw some initial conclusions, the values collected from prior literature should be interpreted as estimates rather than universally precise figures. Future work should therefore evaluate the considered diagnosis architectures and approaches based on a common benchmark to collect diagnostic-performance and energy values under the same conditions. Furthermore, we want to apply our energy-consumption models to real-world building data to assess the practical usefulness of our approach. Acknowledgment The work presented in this paper has been supported by the FFG project Artificial Intelligence for Smart Diagnosis in Building Automation (ALFA) under grant 914932. Declaration of generative AI and AI-assisted technologies in the manuscript preparation process During the preparation of this work, the authors used Grammarly Pro (Model and exact version not publicly disclosed by the provider; accessed August 2026), Claude (Sonnet 5), and ChatGPT (GPT-5.6 Sol, OpenAI, San Francisco, CA, USA) in order to improve language, grammar, readability and the structure and organization of parts of the manuscript. After using these tools, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article. Appendix A Reviewed Literature on FDD in Chiller/AHU systems Table 8: Summary of target HVAC fault-detection/diagnosis methods reviewed by Chen et al. [6]. (NR = NOT REPORTED) Paper ID Chiller/ AHU Summary Method TPR/Recall TPR note FPR FPR note TN TP FN FP [47] Chiller Centrifugal chiller (ASHRAE benchmark); nominal operation and 8 single-fault conditions at multiple severities and operating conditions. Generalized likelihood ratio test (GLRT) performs detection; SVM, principal component analysis (PCA), and partial least squares (PLS) are used for fault classification. Reported: Classification Rate/Accuracy, average % errors SVM fault classifier (variant/kernel not reported) NR NR NR NR NR NR NR NR [72] Chiller ASHRAE RP-1043 90-ton chiller; five faults at 10–40% severity; Compares direct SVM, regression-residual+SVM, auto-regressive moving-average model with exogenous variables (ARX)+MLP, and ARX+SVM SVM (RBF) NR NR 0.20 Paper reports false-alarm value 20.22 NR NR NR NR regression-residual+SVM NR NR 0.07 Paper reports false-alarm value 7.38 NR NR NR NR ARX+MLP NR NR 0.02 Paper reports false-alarm value 1.83 NR NR NR NR ARX+ SVM NR NR ¡ 0.00 Paper reports false-alarm value 0.11 NR NR NR NR [46] AHU ASHRAE RP-1312 AHU; 11 faults. ReliefF feature selection, ARX parameters, SVM( RBF kernel +C-support vector classification (C-SVC)); Compares to raw-data ML baselines. SVM (RBF+C-SVC) + ReliefF + ARX 0.921 Over all 11 fault classes NR NR NR NR NR NR RBF network (N) 0.668 Over all 11 fault classes NR NR NR NR NR NR MLP 0.481 Over all 11 fault classes NR NR NR NR NR NR SVM (LibSVM library) 0.785 Over all 11 fault classes NR NR NR NR NR NR DT 0.794 Over all 11 fault classes NR NR NR NR NR NR RF 0.817 Over all 11 fault classes NR NR NR NR NR NR [29] AHU Simulation model (TRNSYS model) for experiments with four faults. RBS (Qualitative thresholds) 0.11 0.00 7,242 216 1,675 0 RBS (Quantitative) 0.256 0.05 6,936 484 1,407 389 SVM (one-class SVM, radial-basis nonlinear kernel) 0.31 0.28 7,304 587 1,304 21 MLP 0.58 0.13 6,409 1,105 780 916 Single Fuzzy ART (N) 0.60 0.02 7,164 1,115 773 160 LAPART (N) 0.47 0.04 7,044 889 1,002 279 [70] Chiller ASHRAE RP-1043 90-ton chiller with seven faults at four severities; Rebalancing with conditional Wasserstein generative adversarial networks (CWGAN); Compared across SVM, RF, DT, and LR. Reported Accuracy. SVM NR NR NR NR NR NR NR NR RF NR NR NR NR NR NR NR NR DT NR NR NR NR NR NR NR NR LR NR NR NR NR NR NR NR NR [62] AHU Single- and multi-zone VAV AHUs from LBNL FLEXLAB; Six faults plus normal operation. Seven deep recurrent neural networks (DRNN) configurations are optimized and compared with RF and gradient boosting; Heat Map representation of confusion matrix, no values. Reported, Precision; Recall; (multiclass aggregate, averaging not specified); Accuracy; F1-Score; ROC Stacked and deep transition output combined (S,DTO)-DRNN (SZ-AHU) 0.92 multiclass NR NR NR NR NR NR S,DTO-DRNN (MZ-AHU) 0.88 multiclass NR NR NR NR NR NR RF (SZ-AHU) 0.89 multiclass NR NR NR NR NR NR RF(Gradient boosting) (SZ-AHU) 0.81 multiclass NR NR NR NR NR NR RF (MZ-AHU) 0.80 multiclass NR NR NR NR NR NR RF (Gradient boosting) (MZ-AHU) 0.75 multiclass NR NR NR NR NR NR [71] AHU ASHRAE RP-1312 data from two AHUs; Six faults. Training uses 5/10/20/30/40 faulty samples per fault type plus 7200 normal samples; Conditional Wasserstein Generative Adversarial Network with Ensemble Learning Quality Control Protocol (CWGAN-ELQCP) rebalancing is compared with conventional RF, SVM, MLP, and DT RF (conventional) 0.167, 0.182, 0.176, 0.169, 0.169 Values correspond to 5, 10, 20, 30, 40 faulty training samples per fault type NR NR NR NR NR NR RF (proposed) 0.979, 0.974, 0.970, 0.982, 0.999 Values correspond to 5, 10, 20, 30, 40 faulty training samples per fault type NR NR NR NR NR NR SVM (conventional) 0.186, 0.163, 0.179, 0.172, 0.177 Values correspond to 5, 10, 20, 30, 40 faulty training samples per fault type NR NR NR NR NR NR SVM (proposed) 0.995, 0.995, 0.997, 0.996, 0.999 Values correspond to 5, 10, 20, 30, 40 faulty training samples per fault type NR NR NR NR NR NR MLP (conventional) 0.168, 0.175, 0.182, 0.181, 0.183 Values correspond to 5, 10, 20, 30, 40 faulty training samples per fault type NR NR NR NR NR NR MLP (proposed) 0.975, 0.962, 0.971, 0.984, 0.998 Values correspond to 5, 10, 20, 30, 40 faulty training samples per fault type NR NR NR NR NR NR DT (conventional) 0.168, 0.168, 0.167, 0.166, 0.173 Values correspond to 5, 10, 20, 30, 40 faulty training samples per fault type NR NR NR NR NR NR DT (proposed) 0.954, 0.917, 0.944, 0.912, 0.999 Values correspond to 5, 10, 20, 30, 40 faulty training samples per fault type NR NR NR NR NR NR [19] Chiller ASHRAE RP-1043 90-ton centrifugal chiller with seven gradual faults. 12,000 samples (8,000 train/validation; 4,000 test) with grid search and 10-fold CV; Compares optimized SVM/RF and an ensemble; exact confusion counts are available. SVM (RBF kernel) ¿0.99 Computed 0.01 Computed 484 3498 13 5 RF ¿0.99 Computed ¡ 0.01 Computed 487 3509 2 2 [42] Chiller ASHRAE RP-1043 90-ton water-cooled chiller; 7 faults at 4 severities plus normal data; 8 training sizes and raw versus preprocessed; Compares Tree-structured Fault Dependence Kernel (TFDK), SVM (RBF kernel), DT and LR SVM (TFDK) – small sample size 0.9983 Computed 0.1025 Computed 359 11181 19 41 SVM (TFDK) – large sample size 1.0000 Computed 0.0000 Computed 400 11200 0 0 SVM (RBF kernel) – small sample size 0.9908 Computed 0.1225 Computed 351 11097 103 49 SVM (RBF kernel) – large sample size 0.9995 Computed 0.0250 Computed 390 11194 6 10 DT NR NR NR NR NR NR NR NR LR NR NR NR NR NR NR NR NR [41] Chiller ASHRAE RP-1043 90-ton chiller with 7 faults at 4 severities. Compares Semi-GAN with MLP and SVM baselines; Reports only accuracy Semi-GAN (N) NR NR NR NR NR NR NR NR MLP (1 layer) NR NR NR NR NR NR NR NR MLP (2 layers) NR NR NR NR NR NR NR NR Least-squares SVM (LS-SVM) NR NR NR NR NR NR NR NR Semi-SVM-based method NR NR NR NR NR NR NR NR [66] Chiller ASHRAE RP-1043 90-ton chiller with 7 faults at 4 severities. Hybrid method using model-based methods and Bayesian network. Reports accuracy only. – [34] Chiller Empirical comparison of rule-based, classification, and outlier-detection approaches for HVAC fault detection using a simulated chiller benchmark. 7 faults; 5 fold CV RBS 0.3851 Mean recall over 5 CV folds; fault = positive class. NR NR NR NR NR NR DT 0.7937 Mean recall over 5 CV folds; fault = positive class. NR NR NR NR NR NR DT(feature-selected) 0.7757 Mean recall over 5 CV folds; fault = positive class. NR NR NR NR NR NR RF 0.7744 Mean recall over 5 CV folds; fault = positive class. NR NR NR NR NR NR RF (feature-selected) 0.7782 Mean recall over 5 CV folds; fault = positive class. NR NR NR NR NR NR LR (100 iterations) 0.4389 Mean recall over 5 CV folds; fault = positive class. NR NR NR NR NR NR LR (1000 iterations) 0.3780 Mean recall over 5 CV folds; fault = positive class. NR NR NR NR NR NR MLP 0.3747 Mean recall over 5 CV folds; fault = positive class. NR NR NR NR NR NR MLP (feature-selected) 0.4852 Mean recall over 5 CV folds; fault = positive class. NR NR NR NR NR NR Table 9: Summary of additional sources considered for TPR/FPR aggregation. (NR = NOT REPORTED) Paper ID Chiller/ AHU Summary Method TPR/Recall TPR note FPR FPR note TN TP FN FP [10] Chiller ASHRAE RP-1043 centrifugal chiller. Model-based with adaptive fault detection threshold estimator. Five fault types over four severity levels. MBD 0.51, 0.62, 0.20, 0.69, 1.00 Detection rates for: reduced evaporator-water flow, refrigerant leakage, excess oil, condenser fouling, and non-condensables NR The paper reports a “false alarm rate”, but its denominator is not sufficiently clear to treat it as FPR. NR NR NR NR [49] Chiller Real-building BMS chiller data; Logistic-circuit decision tree to determine patterns predicts alarms. Decision tree 0.76 Reported 0.000385 Computed 83061 162 51 32 [44] AHU UT Austin main-campus BAS data; chilled-water leakage classification; final tuned model. Decision tree (chilled water) 0.990361 Computed 0.012251 Computed 13061 8322 81 162 Decision tree (steam) 0.966288 Computed 0.026026 Computed 50745 40272 1405 1356 [40] Chiller EnergyPlus-Modelica SDAHU dataset; single-fault binary classification LLM (fine-tuned) - Fine-tuned DistilBERT 0.98 Derived from reported FNR 0.02 Reported NR NR NR NR AHU EnergyPlus-Modelica SDAHU dataset; single-fault binary classification LLM (fine-tuned) - Fine-tuned DistilBERT 0.99 Derived from reported FNR 0.01 Reported NR NR NR NR AHU LR NR NR 0.00 Reported FPR NR NR NR NR [24] AHU LBNL SZVAV (experimental) and SZCAV (simulated); binary fault classification in a hospital-building FDD study. LR 0.9164 Computed 0.5303 Computed 674 5439 496 761 [74] AHU ASHRAE RP-1312 containing 12 AHU fault classes. GPT-3.5 is fine-tuned using a labeled-data-supervised fine-tuning method. LLM (Original GPT-3.5) 0.25 Aggregated recall reconstructed 0.13 FPR reconstructed from fault-free class accuracy NR NR NR NR LLM (tuned) - Fine-tuned GPT-3.5 (Version 4) 1.00 All fault classes have 100% diagnosis accuracy. 0.00 All fault-free classes have 100% accuracy. NR NR NR NR References [1] M. Awad and R. Khanna (2015) Support vector machines for classification. In Efficient Learning Machines: Theories, Concepts, and Applications for Engineers and System Designers, p. 39–66. External Links: ISBN 978-1-4302-5990-9, Document Cited by: §2.1. [2] I. Bellanco, E. Fuentes, M. Vallès, and J. Salom (2021) A review of the fault behavior of heat pumps and measurements, detection and diagnosis methods including virtual sensors. Journal of Building Engineering 39, p. 102254. External Links: ISSN 2352-7102, Document Cited by: §1, §3.1. [3] B. R. Bruce, J. Petke, and M. Harman (2015) Reducing energy consumption using genetic improvement. In Proceedings of the 2015 Annual Conference on Genetic and Evolutionary Computation, GECCO ’15, New York, NY, USA, p. 1327–1334. External Links: ISBN 9781450334723, Link, Document Cited by: §4.1, §5.2. [4] B. G. Buchanan and E. H. Shortliffe (Eds.) (1984) Rule-based expert systems - the mycin experiments of the stanford heuristic programming project. Addison-Wesley Publishing Company. Cited by: §2.1. [5] C. Calero and M. Piattini (2015) Green in software engineering. Springer International Publishing. External Links: ISBN 978-3-319-08580-7, Document Cited by: §1. [6] Z. Chen, Z. O’Neill, J. Wen, O. Pradhan, T. Yang, X. Lu, G. Lin, S. Miyata, S. Lee, C. Shen, et al. (2023) A review of data-driven fault detection and diagnostics for building HVAC systems. Applied Energy 339, p. 121030. External Links: Document Cited by: Table 8, §1, §4.1, Table 3, §5.2. [7] S. Cho, H. Yang, M. Zaheer-uddin, and B. Ahn (2005) Transient pattern analysis for fault detection and diagnosis of HVAC systems. Energy Conversion and Management 46 (18-19), p. 3103–3116. External Links: Document Cited by: §1. [8] G. S. B. Council (2022) Wegweiser klimapositiver gebäudebestand 2022 [guide to climate-positive building stock 2022]. Cited by: §1. [9] Mlco2/codecarbon: v3.0.2 External Links: Document Cited by: §4.1. [10] J. Cui and S. Wang (2005) A model-based online fault detection and diagnosis strategy for centrifugal chiller systems. International Journal of Thermal Sciences 44 (10), p. 986–999. External Links: ISSN 1290-0729, Document, Link Cited by: Table 9, Table 3. [11] R. Davis, H. Shrobe, W. Hamscher, K. Wieckert, M. Shirley, and S. Polit (1982) Diagnosis based on structure and function. In Proceedings AAAI, Pittsburgh, p. 137–142. Cited by: §2.1. [12] R. Davis (1984) Diagnostic reasoning based on structure and behavior. Artificial Intelligence 24, p. 347–410. Cited by: §2.1. [13] J. de Kleer and B. C. Williams (1987) Diagnosing multiple faults. Artificial Intelligence 32 (1), p. 97–130. Cited by: §2.1, §2.1. [14] European Commission (2020) Focus on energy efficiency in buildings. Note: Accessed from European Commission website External Links: Link Cited by: §1. [15] European Commission (2023) Press release: commission welcomes political agreement on new rules to boost energy performance of buildings across the EU. European Commission. External Links: Link Cited by: §1. [16] I. Fatima, P. Lago, V. Andrikopoulos, and B. van der Waaij (2025) Using sustainability impact scores for software architecture evaluation. In 2025 IEEE 22nd International Conference on Software Architecture Companion (ICSA-C), Vol. , p. 60–68. External Links: Document Cited by: §1. [17] G. Friedrich, G. Gottlob, and W. Nejdl (1990) Hypothesis classification, abductive diagnosis and therapy. In Proceedings of the International Workshop on Expert Systems in Engineering, Vienna. Cited by: §2.1. [18] U. N. E. P. Global Alliance for Buildings and Construction (2024) Global status report for buildings and construction: beyond foundations. UNEP. External Links: Link, ISBN 978-92-807-4131-5 Cited by: §1. [19] H. Han, Z. Zhang, X. Cui, and Q. Meng (2020) Ensemble learning with member optimization for fault diagnosis of a building energy system. Energy and Buildings 226, p. 110351. External Links: ISSN 0378-7788, Document Cited by: Table 8. [20] S. S. Haykin (2012) Neural networks and learning machines. 3rd edition, Pearson. Cited by: §2.1. [21] T. K. Ho (1995) Random decision forests. In Proceedings of 3rd International Conference on Document Analysis and Recognition, Vol. 1, p. 278–282 vol.1. External Links: Document Cited by: §2.1. [22] Á. Huertas-García, C. Martí-González, R. G. Maezo, and A. E. Rey (2024) A comparative study of machine learning algorithms for anomaly detection in industrial environments: performance and environmental impact. In Trends in Sustainable Computing and Machine Intelligence, S. Lanka, A. Sarasa-Cabezuelo, and A. Tugui (Eds.), Singapore, p. 373–389. External Links: ISBN 978-981-99-9436-6, Document Cited by: Table 4, Table 4. [23] International Energy Agency (2022) Buildings. External Links: Link Cited by: §1. [24] Md. M. Islam, M. Ahmed, and N. A. Khan (2024) Fault detection and diagnostics of air handling unit in hospital building using machine learning. In 2024 15th International Conference on Computing Communication and Networking Technologies (ICCCNT), Vol. , p. 1–6. External Links: Document Cited by: Table 9, Table 3. [25] E. Jagroep, J. M. E. M. van der Werf, S. Brinkkemper, L. Blom, and R. van Vliet (2017) Extending software architecture views with an energy consumption perspective - A case study on resource consumption of enterprise software. Computing 99 (6), p. 553–573. External Links: Document Cited by: §1. [26] A. K. Jain (2010) Data clustering: 50 years beyond k-means. Pattern Recognition Letters 31 (8), p. 651–666. Note: Award winning papers from the 19th International Conference on Pattern Recognition (ICPR) External Links: ISSN 0167-8655, Document Cited by: §2.1. [27] H. Järvenpä, P. Lago, J. Bogner, G. Lewis, H. Muccini, and I. Ozkaya (2024) A synthesis of green architectural tactics for ml-enabled systems. In Proceedings of the 46th International Conference on Software Engineering: Software Engineering in Society, ICSE-SEIS’24, New York, NY, USA, p. 130–141. External Links: ISBN 9798400704994, Link, Document Cited by: §1. [28] Z. Ji and M. Jiang (2026) A systematic review of electricity demand for large language models: evaluations, challenges, and solutions. Renewable and Sustainable Energy Reviews 225, p. 116159. External Links: ISSN 1364-0321, Document Cited by: Table 4, Table 4. [29] C. B. Jones (2015) Fault detection and diagnostics of an hvac sub-system using adaptive resonance theory neural networks. Ph.D. Thesis, The University of New Mexico. External Links: Link Cited by: Table 8. [30] W. Kim and S. Katipamula (2018) A review of fault detection and diagnostics methods for building systems. Science and Technology for the Built Environment 24 (1), p. 3–21. External Links: Document Cited by: §1. [31] L. Klampfl, A. Perko, and F. Wotawa (2025) On the challenges of monitoring robust systems. In 12th International Conference on Dependable Systems and Their Applications, DSA 2025, Sharjah, United Arab Emirates, November 24-26, 2025, p. 187–194. External Links: Document Cited by: §2.1. [32] L. Klampfl and F. Wotawa (2024) Leveraging answer set programming for continuous monitoring, fault detection, and explanation of automated and autonomous driving systems. In 35th International Conference on Principles of Diagnosis and Resilient Systems, DX 2024, Vienna, Austria, November 4-7, 2024, I. Pill, A. Natan, and F. Wotawa (Eds.), OASIcs, Vol. 125, p. 10:1–10:20. External Links: Document Cited by: §2.1. [33] R. Koitz-Hristov, L. M. Prikler, and F. Wotawa (2025) Beyond Static Diagnosis: A Temporal ASP Framework for HVAC Fault Detection. In 36th International Conference on Principles of Diagnosis and Resilient Systems (DX 2025), M. Quinones-Grueiro, G. Biswas, and I. Pill (Eds.), Open Access Series in Informatics (OASIcs), Vol. 136, Dagstuhl, Germany, p. 1:1–1:20. Note: Keywords: Model-based diagnosis, Answer set programming, HVAC, Modeling for diagnosis, Experimental evaluation External Links: ISBN 978-3-95977-394-2, ISSN 2190-6807, Link, Document Cited by: §1, §4.1, Table 4. [34] R. Koitz-Hristov, L. M. Prikler, and F. Wotawa (2026) Empirical evaluation of rule-based and machine learning approaches for fault detection in building hvac systems. In Advances and Trends in Artificial Intelligence. Theory and Applications, H. Fujita, A. Selamat, M. Ghazali, and M. Ali (Eds.), Singapore, p. 3–14. External Links: ISBN 978-981-92-2888-1 Cited by: Table 8, §4.1, Table 4. [35] R. Koitz-Hristov, L. Prikler, and F. Wotawa (2026) Automated diagnosis for heating systems in buildings. In Computational Intelligence for a Greener Future: Innovations in Renewable Energy Systems: Integrating AI and Behavioral Insights to Drive Economic Efficiency and Sustainability, M. A. Mellal, Y. Nojima, and N. Masuyama (Eds.), p. 375–395. External Links: ISBN 978-3-032-06732-6, Document Cited by: §1. [36] R. Koitz-Hristov and F. Wotawa (2020) Faster horn diagnosis - a performance comparison of abductive reasoning algorithms. Appl. Intell. 50 (5), p. 1558–1572. External Links: Document Cited by: §2.1. [37] P. Lago, S. A. Koçak, I. Crnkovic, and B. Penzenstadler (2015) Framing sustainability as a property of software quality. Commun. ACM 58 (10), p. 70–78. External Links: ISSN 0001-0782, Link, Document Cited by: §1. [38] P. Lago (2019) Architecture design decision maps for software sustainability. In 2019 IEEE/ACM 41st International Conference on Software Engineering: Software Engineering in Society (ICSE-SEIS), Vol. , p. 61–64. External Links: Document Cited by: §1. [39] G. Langer, T. Hirsch, R. Kern, T. Kohl, and G. Schweiger (2025) Large language models for fault detection in buildings’ hvac systems. In Energy Informatics, B. N. Jørgensen, Z. G. Ma, F. D. Wijaya, R. Irnawan, and S. Sarjiya (Eds.), Cham, p. 49–60. External Links: ISBN 978-3-031-74741-0 Cited by: §2.1. [40] G. Langer (2024) Large language models for fault detection in hvac systems of buildings. Master’s Thesis, Graz University of Technology. Cited by: Table 9, Table 3. [41] B. Li, F. Cheng, X. Zhang, C. Cui, and W. Cai (2021) A novel semi-supervised data-driven method for chiller fault diagnosis with unlabeled data. Applied Energy 285, p. 116459. External Links: ISSN 0306-2619, Document Cited by: Table 8. [42] D. Li, Y. Zhou, G. Hu, and C. J. Spanos (2016) Fault detection and diagnosis for building cooling system with a tree-structured learning method. Energy and Buildings 127, p. 540–551. External Links: ISSN 0378-7788, Document Cited by: Table 8. [43] S. Martínez-Fernández, X. Franch, and F. Durán (2023) Towards green ai-based software systems: an architecture-centric approach (gaissa). In 2023 49th Euromicro Conference on Software Engineering and Advanced Applications (SEAA), Vol. , p. 432–439. External Links: Document Cited by: §1. [44] M. K. McHugh (2019) Fault detection and diagnostics modeling for air-handling units on the main campus of the university of texas at austin. Ph.D. Thesis. Cited by: Table 9, Table 3. [45] D. C. Montgomery, E. A. Peck, and G. G. Vining (2021) Introduction to linear regression analysis. 6th edition, Wiley. Cited by: §2.1. [46] T. Mulumba, A. Afshari, K. Yan, W. Shen, and L. K. Norford (2015) Robust model-based fault diagnosis for air handling units. Energy and Buildings 86, p. 698–707. External Links: ISSN 0378-7788, Document Cited by: Table 8. [47] S. M. Namburu, M. S. Azam, J. Luo, K. Choi, and K. R. Pattipati (2007) Data-driven modeling, fault diagnosis and optimal sensor selection for hvac chillers. IEEE Transactions on Automation Science and Engineering 4 (3), p. 469–473. External Links: Document Cited by: Table 8. [48] I. Nica, I. Pill, T. Quaritsch, and F. Wotawa (2013) The route to success - A performance comparison of diagnosis algorithms. In IJCAI 2013, Proceedings of the 23rd International Joint Conference on Artificial Intelligence, Beijing, China, August 3-9, 2013, F. Rossi (Ed.), p. 1039–1045. External Links: Link Cited by: §2.1. [49] K. Padmanabh, A. Al-Rubaie, J. Davies, S. S. Clarke, and A. A. A. A. Aljasmi (2021) Fault prediction in hvac chillers by analysis of internal system dynamics. In 2021 International Conference on Smart Applications, Communications and Networking (SmartNets), Vol. , p. 1–6. External Links: Document Cited by: Table 9, Table 3. [50] G. Procaccianti, P. Lago, and S. Bevini (2015) A systematic literature review on energy efficiency in cloud software architectures. Sustainable Computing: Informatics and Systems 7, p. 2–10. External Links: ISSN 2210-5379, Document, Link Cited by: §1. [51] J. R. Quinlan (1986) Induction of decision trees. Machine Learning 1 (1), p. 81–106. Cited by: §2.1. [52] R Core Team (2021) R: a language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. External Links: Link Cited by: §4.2. [53] R. Reiter (1987) A theory of diagnosis from first principles. Artificial Intelligence 32 (1), p. 57–95. Cited by: §2.1, §2.1. [54] J. Retti, W. Nejdl, and G. Friedrich (1987) ARTEX — Knowledge based fault detection and correction in a switching system. In Proceedings of the International GI Conference on Knowledge Based Systems, München. Cited by: §2.1. [55] A. Riehn (1990) Diagnostic expert systems increasing machine tool productivity. In Proceedings of the Conference on Artificial Intelligence and Expert Systems in Manufacturing, London. Cited by: §2.1. [56] S. O. S. Santos, A. Skiarski, D. García-Núñez, V. Lazzarini, R. De Andrade Moral, E. Galvan, A. L. C. Ottoni, and E. Nepomuceno (2024) Green machine learning: analysing the energy efficiency of machine learning models. In 2024 35th Irish Signals and Systems Conference (ISSC), Vol. , p. 1–6. External Links: Document Cited by: Table 4, Table 4, Table 4. [57] R. Schwartz, J. Dodge, N. A. Smith, and O. Etzioni (2020) Green ai. Commun. ACM 63 (12), p. 54–63. External Links: ISSN 0001-0782, Link, Document Cited by: §1. [58] A. Silberling (2025) ChatGPT users send 2.5 billion prompts a day. Note: TechCrunchFigure disclosed by OpenAI. Accessed 17 August 2026 External Links: Link Cited by: §4.1. [59] R. Sterling, G. Provan, J. Febres, D. O’Sullivan, P. Struss, and M. M. Keane (2014) Model-based fault detection and diagnosis of air handling units: a comparison of methodologies. Energy Procedia 62, p. 686–693. External Links: Document Cited by: §1. [60] G. Stettinger and F. Wotawa (2023) Smart monitoring for safety-assurance in autonomous driving (S). In The 35th International Conference on Software Engineering and Knowledge Engineering, SEKE 2023, KSIR Virtual Conference Center, USA, July 1-10, 2023, S. Chang (Ed.), p. 542–547. External Links: Link, Document Cited by: §2.1. [61] E. Strubell, A. Ganesh, and A. McCallum (2019) Energy and policy considerations for deep learning in nlp. External Links: 1906.02243, Link Cited by: §1. [62] S. Taheri, A. Ahmadi, B. Mohammadi-Ivatloo, and S. Asadi (2021) Fault detection diagnostic for hvac systems via deep learning algorithms. Energy and Buildings 250, p. 111275. External Links: ISSN 0378-7788, Document Cited by: Table 8. [63] C. C. Venters, R. Capilla, S. Betz, B. Penzenstadler, T. Crick, S. Crouch, E. Y. Nakagawa, C. Becker, and C. Carrillo (2018) Software sustainability: research and practice from a software architecture viewpoint. Journal of Systems and Software 138, p. 174–188. External Links: ISSN 0164-1212, Document Cited by: §1. [64] R. Verdecchia, P. Lago, C. Ebert, and C. de Vries (2021) Green it and green software. IEEE Software 38 (6), p. 7–15. External Links: Document Cited by: §1. [65] R. Verdecchia, J. Sallou, and L. Cruz (2023) A systematic review of Green AI. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery 13 (4), p. e1507. External Links: Document Cited by: §1. [66] Z. Wang, L. Wang, Y. Tan, J. Yuan, and X. Li (2021) Fault diagnosis using fused reference model and bayesian network for building energy systems. Journal of Building Engineering 34, p. 101957. External Links: ISSN 2352-7102, Document Cited by: Table 8. [67] M. Weber, C. Kaltenecker, F. Sattler, S. Apel, and N. Siegmund (2023) Twins or false friends? a study on energy consumption and performance of configurable software. In 2023 IEEE/ACM 45th International Conference on Software Engineering (ICSE), Vol. , p. 2098–2110. External Links: Document Cited by: §4.1, §5.2. [68] F. Wotawa (2011) On the use of abduction as an alternative to decision trees in environmental decision support systems. Int. J. Agric. Environ. Inf. Syst. 2 (1), p. 63–82. External Links: Document Cited by: §2.1. [69] Energy-consumption model for diagnosis methods to be used in building automation systems (hvac,…) External Links: Document, Link Cited by: §4.2. [70] K. Yan, A. Chong, and Y. Mo (2020) Generative adversarial network for fault detection diagnosis of chillers. Building and Environment 172, p. 106698. External Links: ISSN 0360-1323, Document Cited by: Table 8. [71] K. Yan, J. Huang, W. Shen, and Z. Ji (2020) Unsupervised learning for fault detection and diagnosis of air handling units. Energy and Buildings 210, p. 109689. External Links: ISSN 0378-7788, Document Cited by: Table 8. [72] K. Yan, W. Shen, T. Mulumba, and A. Afshari (2014) ARX model based fault detection and diagnosis for chillers using support vector machines. Energy and Buildings 81, p. 287–295. External Links: ISSN 0378-7788, Document Cited by: Table 8. [73] F. Zhang, N. Saeed, and P. Sadeghian (2023) Deep learning in fault detection and diagnosis of building HVAC systems: a systematic review with meta analysis. Energy and AI 12, p. 100235. External Links: Document Cited by: §1. [74] J. Zhang, C. Zhang, J. Lu, and Y. Zhao (2025) Domain-specific large language models for fault diagnosis of heating, ventilation, and air conditioning systems by labeled-data-supervised fine-tuning. Applied Energy 377, p. 124378. External Links: ISSN 0306-2619, Document Cited by: Table 9, Table 3. [75] S. Zhou, A. Shah, P. Leung, X. Zhu, and Q. Liao (2023) A comprehensive review of the applications of machine learning for HVAC. DeCarbon 2, p. 100023. External Links: Document Cited by: §1. [76] Z. Zhou, J. H. Abawajy, F. Li, Z. Hu, M. U. Chowdhury, A. Alelaiwi, and K. Li (2018) Fine-grained energy consumption model of servers based on task characteristics in cloud data center. IEEE Access 6 (), p. 27080–27090. External Links: Document Cited by: §4.1, §5.2. [77] I. Žliobaitė, M. Budka, and F. Stahl (2015) Towards cost-sensitive adaptation: when is it worth updating your predictive model?. Neurocomputing 150, p. 240–249. Note: Bioinspired and knowledge based techniques and applications The Vitality of Pattern Recognition and Image Analysis Data Stream Classification and Big Data Analytics External Links: ISSN 0925-2312, Document, Link Cited by: §3.2.