Why software energy is not the whole footprint
Buildings account for roughly 30% of global energy expenditure, and the paper notes that a large share of that consumption could be avoided by repairing existing faults. Fault detection and diagnosis software is meant to capture that saving, but building, training, and running the software also consumes energy. The paper's central argument is that judging an FDD architecture only by the energy its own code burns is insufficient, because a method that misses faults leaves wasted HVAC energy in place.
So the comparison has two sides. One is the energy spent directly by the diagnosis software, including model creation or learning and inference. The other is the energy the monitored HVAC system consumes because faults are detected, missed, or falsely flagged. The paper builds an energy-consumption model that combines both, then uses it to compare architecture families in simulation.
The model turns detection rates into energy
The energy-consumption model starts from HVAC energy: energy use intensity (the fault-free energy per square meter per year) multiplied by building area and the fraction of consumption used for heating or cooling. Faults add wasted energy, modeled as a fraction of that HVAC energy; the paper notes wasted energy for a fault may be up to 40%.
Detection then modifies the total. True positives produce savings proportional to the annual true positive rate applied to the wasted energy. False positives add wasted energy proportional to the annual false positive rate, but applied only to a fraction of total HVAC energy. False negatives need no separate term because their wasted energy is already inside the fault term. The result is total HVAC energy after diagnosis: baseline HVAC energy plus fault waste, minus savings from detected faults, plus the penalty from false alarms.
A key measurement point is that common classification metrics such as precision, recall, and F1 are not enough to compute this model. The paper states that the false positive rate cannot be derived directly from those figures, so the model needs the underlying confusion counts.
Training and inference energy enter the comparison
The paper extends the model with the energy required to train the diagnosis method and to run inference. This matters because architecture families differ sharply in computational demand: rule-based and model-based approaches, classical machine learning, and large-language-model-based approaches do not carry the same training or inference cost.
The scope is deliberately bounded. The model uses a one-year horizon and assumes the building, its HVAC system, and the deployed diagnosis method stay unchanged over that period. It does not model component aging or the associated growth in fault probability, and it represents faults as an aggregate annual fraction of wasted energy rather than as a distribution over individual faults. It also restricts itself to fault detection, not fault identification or repair.
Random forest leads in smaller buildings
In the simulation, computationally efficient machine learning methods, such as random forest, provide the largest net savings on smaller buildings. The reason is that in a small building the HVAC energy base is modest, so the software's own energy demand is a meaningful share of the total footprint. A method that is cheap to train and run keeps that overhead low while still capturing fault-related savings.
This is a simulation result built from performance and energy-consumption values collected from prior literature, not a measurement of a deployed building. The paper describes these values as estimates rather than universally precise figures, which qualifies how firmly the ranking should be read.
Fine-tuned LLMs gain as buildings grow
As building size increases, more resource-intensive approaches such as fine-tuned large language models become advantageous. In larger buildings the HVAC energy demand is large enough that the software's training and inference energy becomes a small fraction of the total, so the diagnostic performance of the method dominates the comparison. The paper reports that for office buildings the energy required for training and inference is negligible, and in that case fine-tuned LLMs provide the largest net energy savings.
The paper also reports that fine-tuned LLMs achieve the largest energy savings for single-family houses in its setup, even though the architecture's own energy demand is more relevant there because heating and cooling footprints are smaller. The experiments further show that inference frequency can influence the ranking of software architectures, meaning the observation interval between diagnoses can shift which method comes out ahead.
Accuracy and compute trade off differently by scale
The results show that considering both the computational energy and the accuracy of FDD can change the relative benefit of the different approaches. A method that looks wasteful on compute alone may still win if its detection performance avoids more wasted HVAC energy, and a cheap method may lose that advantage once the building is large enough that its missed faults cost more than the compute it saved.
This is why the paper frames the question as a combination of two factors rather than a single ranking. The relevant comparison is between diagnostic performance, the energy consumed for training and inference, and the characteristics and operation of the monitored building. The paper's own conclusion is that no architecture is universally most energy-efficient.
Where these simulated rankings stop applying
The findings come from simulation experiments using values collected from prior literature, so they depend on the model's assumptions and on the quality of those input values. The one-year horizon excludes long-term changes such as building modifications, HVAC component replacement or degradation, and FDD system updates. Aggregating faults into a single annual loss fraction avoids modeling individual fault types, severities, frequencies, and durations.
The reported training energy also excludes activities such as feature selection, hyperparameter tuning, prompt design, rule creation, and building models from first principles, which the paper says are difficult to quantify consistently across architectures. Generalizability is limited by the HVAC subsystems considered, which focus on chiller and air handling unit applications. The authors therefore call for evaluating these architectures on a common benchmark and for applying the energy-consumption models to real-world building data.
The decision hinges on building scale
The paper's practical priority is a scale-dependent comparison rather than a single winner. On smaller buildings, where the HVAC energy base is small, the deciding factor is the FDD software's own training and inference demand, which favors computationally efficient methods such as random forest. On larger buildings, where HVAC energy dominates, diagnostic performance matters more, which favors fine-tuned large language models despite their higher compute cost.
The variable a reader should watch is therefore building scale, together with inference frequency, since the paper shows the observation interval can change the ranking. The most useful takeaway is that an FDD architecture should be chosen against the size and operation of the specific building being monitored, not against a general claim about which method is greenest.
Sources used in this answer
Choosing an energy-efficient software architecture for building system diagnostic support
In simulation using performance and energy values from prior literature, computationally efficient machine learning such as random forest gives the largest net energy savings on smaller buildings, while fine-tuned large language models become advantageous as building size increases; no architecture is universally most energy-efficient.