Choosing an energy-efficient software architecture for building system diagnostic support

Roxane Koitz-Hristov; Franz Wotawa

cs.SEOct 5, 2026

Choosing an energy-efficient software architecture for building system diagnostic support

Authors: Roxane Koitz-Hristov, Franz Wotawa

Organizations: Graz University of Technology Institute of Software Engineering and Artificial Intelligence Inffeldgasse 16b/2, A-8010 Graz, Austria

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.

Figures & tables

Figure 1: Rule-based diagnosis architecture

Figure 2: Model-based diagnosis architecture

Figure 3: Machine learning diagnosis architecture

Figure 4: LLM-based diagnosis architecture

Parameter	Explanation	Unit	Range
$EUI$	Energy use intensity per year	kWh/(m ${}^{2}\cdot$ year)	[0, $\infty$ )
$A$	Area of building	m 2	[0, $\infty$ )
$s_{HVAC}$	Fraction of energy-consumption used for HVAC	1	[0,1]
$s_{FAULT}$	Fraction of HVAC energy-consumption in case of faults	1	[0,1]
$TPR_{a}$	Annual true positive rate of diagnosis method	1	[0,1]
$FPR_{a}$	Annual false positive rate of diagnosis method	1	[0,1]

Table 1: Input parameters of the energy-consumption model

ECM

for FDD applied to BAS.

TPR

and

FPR

are interpreted as annual averages over the considered year.

Parameter	Explanation	Unit	Range
$U$	Degree to which the model is used exclusively for diagnosis	1	(0,1]
$f_{RET}$	Number of additional model retrainings within the considered year	1	$\mathbb{N}_{0}$
$E_{BINF}$	Energy-consumption of one inference operation	kWh/inference	$[0,\infty)$
$N_{INF}$	Number of inference operations within the considered year	inferences/year	$\mathbb{N}_{0}$

Table 2: Input parameters of the extended energy-consumption model

EECM

for FDD applied to BAS.

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	0.20 a	1.00 a	0.60	0.00	0.07	0.03
DT	0.17	1.00	0.61	$0.00^{b}$	0.03 c	0.01
RF	0.17	1.00	0.69	$0.00$	$0.00$	$0.00^{\dagger}$
LR	0.38	0.44	0.41 ∗	0.00 e	0.53 f	0.27 ∗

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	$E_{TMETHOD}$		$U$	$E_{BINF}$		Ref.
	Min	Max		Min	Max
RB a	0	0	1	$1.0\cdot 10^{-8}$	$2.5\cdot 10^{-7}$	[ 34 ]
MBD a	0	0	1	$4.4\cdot 10^{-5}$	$4.6\cdot 10^{-4}$	[ 33 ]
DT	$3.13\cdot 10^{-4}$	$3.13\cdot 10^{-4}$	1	$8.60\cdot 10^{-5}$	$8.60\cdot 10^{-5}$	[ 22 ]
RF	$2.45\cdot 10^{-7}$	$1.06\cdot 10^{-6}$	1	$2.47\cdot 10^{-8}$	$6.35\cdot 10^{-8}$	[ 56 ]
LR	$3.85\cdot 10^{-8}$	$3.42\cdot 10^{-7}$	1	$2.10\cdot 10^{-8}$	$4.41\cdot 10^{-8}$	[ 56 ]

Table 4 : Energy consumption (in kWh) required for different fault detection methods for a single training/inference run.

	$A$	Method	$e_{thvac}$	$e_{hvac\_diag}^{ECM}$	$e_{hvac\_diag}^{EECM}$	$saving^{ECM}$	$saving^{EECM}$
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

Table 5 : The energy-consumption calculated for the different FDD methods considering different building areas

A

using

ECM

and

EECM

together with their computed potential savings in percent.

Panel of Figure 5 — Figure 5 : Results of the second experiment.

Panel of Figure 6 — Figure 6 : Results of the third experiment.

	Method	$e_{thvac}$	$e_{hvac\_diag}^{ECM}$	$e_{hvac\_diag}^{EECM}$	$saving^{ECM}$	$saving^{EECM}$
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

Table 6 : The results of the 4th experiment.

	Method	$e_{thvac}$	$e_{hvac\_diag}^{ECM}$	$e_{hvac\_diag}^{EECM}$	$saving^{ECM}$	$saving^{EECM}$
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

Table 7 : The results of the 5th experiment for single-family homes.

Panel of Figure 7 — Figure 7 : Results of the sixth experiment.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

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

Appendix

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
[ 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

Appendix

Table 9 : Summary of additional sources considered for TPR/FPR aggregation. (NR = NOT REPORTED)

Explore similar work

Oct 7, 2026cs.AI

Where Can a Decision Model Diagnose HVAC Faults? Reasoning Demand, Physical Representation, and Robustness Under Shift

Artificial intelligence supports building operations in several forms, each with its own barrier. Expert rules must be tuned for every system, supervised models need labeled data that buildings rarely record, and language models return free text that requires human-in-the-loop checking, since their stated confidence is unreliable. A newer kind of pretrained model, here called a decision model, returns a probability for every allowed answer, so one model could serve many decisions without training. This study answers three open questions for fault diagnosis in heating, ventilation, and air-conditioning systems: which decisions such a model can make, what input it needs, and whether its probabilities hold when conditions change. On 128 fault days from four public datasets of real equipment, faults are graded by the reasoning their diagnosis demands, with data given raw, as physical features, or with Brick topology. The decision model Jev, open language models, and a supervised model face nine tests that change season, control configuration, or building. Given physical features, Jev and the larger open model diagnosed faults whose evidence one feature carries, but not faults that need operating context. Under shift they kept their accuracy and calibration, while the supervised model lost 0.33 macro-F1 yet led or tied within a building. Their probabilities still needed correction, and detection was weak. The study maps which faults a decision model can diagnose and from what input, and supports a division of work in which code computes the physics and the model ranks candidate faults for an operator.

Wooyoung Jung

University of Arizona, Tucson, AZ, USA

Fault Diagnosis Diagnosis Artificial Intelligence Models

Ask PDF

Jul 7, 2026cs.SE

Static Metrics Are Insufficient: Predicting Java Method Energy Usage with Execution Time

The increasing energy demand of software systems is raising concerns about their environmental impact and associated costs. Reasoning on energy usage early in the development flow has the potential to significantly reduce the overall energy usage of a software system, as it allows developers to make informed design and refactoring decisions before inefficiencies propagate. However, assessing energy usage without repeated profiling and direct measurement is difficult, which limits early reasoning in practice. This study investigates the limits of method-level energy prediction in Java, examining whether static source code metrics complemented with method-level execution time can estimate the energy consumption of Java methods. We profile 2,786 Java methods to extract 33 static features and measure execution time and energy, then train and compare eleven regression models. Our findings show that static source code metrics alone yield poor predictive performance, with average R2 values close to zero. Incorporating execution time as a lightweight dynamic input significantly improves accuracy, raising R2 to as high as 0.46. Execution time, internal method calls, and cyclomatic complexity consistently emerge as the strongest predictors of energy consumption.

Muhammad Imran, Vincenzo Stoico, Ivano Malavolta

University of L’Aquila, L’Aquila, Italy · Vrije Universiteit Amsterdam, Amsterdam, The Netherlands

Energy Consumption Smooth Execution Performance Metric Code Quality Incomplete

Ask PDF

Jul 6, 2026cs.LG

Beyond the Need for Speed: Energy-Aware Code Generation via Simulation-Guided Reinforcement Learning

Code models strictly prioritize functional correctness, leaving software energy efficiency as an unoptimized byproduct. Training models to generate energy-efficient code requires reproducible feedback at scale, which physical hardware measurement cannot reliably provide due to variance. In this paper, we replace hardware profiling with a deterministic architectural simulation harness to build Green Tea, a corpus of

3.5

million evaluations across

1{,}474

C++ problems. We train an energy-aware code model via supervised fine-tuning on energy-contrastive pairs, followed by closed-loop reinforcement learning (GRPO) using simulation-in-the-loop feedback. To rigorously evaluate deployment readiness, we introduce the Correctness-Adjusted Reduction in Energy Total (CARET), a metric that explicitly penalizes code that sacrifices functionality for efficiency. On

143

held-out problems, our simulation-in-the-loop pipeline achieves

12.63\%

CARET, nearly tripling the gain of fine-tuning alone, and successfully beats the energy efficiency of human-expert references on

58.4\%

of its valid outputs. Furthermore, our analysis exposes the IPC trap: standard throughput proxies like Instructions-Per-Cycle (IPC) actively misrank true energy efficiency on

67.8\%

of problems, proving the absolute necessity of direct energy simulation. By releasing our dataset and infrastructure, we bypass the

263{,}000

CPU-hours required for reproduction, structurally empowering the community to deploy inherently energy-efficient code generation models.

Saurabhsingh Rajput, Tushar Sharma

Dalhousie University, Halifax, Canada

Energy-Efficient Code Generation Code Optimization Profiling Speeds Contrastive Learning

Ask PDF