Intelligent Building Control Systems for Thermal Comfort and Energy-Efficiency: A Systematic Review of Artificial Intelligence-Assisted Techniques

1. Introduction

• Climate change and warmer periods have increased reliance on artificial conditioning, driving summer peak demand and concerns about HVAC energy efficiency alongside occupant comfort.The introduction also notes an IEA projection of about 45% peak-load growth by 2050 in countries with low HVAC penetration.Source: §1 · p. 2
• The review focuses on thermal comfort despite the combined effects of visual, acoustic, and air-quality conditions on occupant comfort.The introduction notes that people may spend 60%–90% of their time indoors.Source: §1 · pp. 3–4
• Fixed comfort standards rely on measures such as PMV, but in-situ studies reveal discrepancies between their narrow comfort intervals and subjective human behavior.The introduction describes adaptive comfort as people reacting to thermal conditions to restore comfort, rather than predicting a specific comfort response.Source: §1 · pp. 3–4
• AI building-control efforts date to the 1990s, and neural-network controllers were deployed in Japanese air-conditioning and fan installations; a two-network air conditioner adapted to customer preferences.The introduction also describes evolutionary algorithms, fuzzy logic, and neural-network combinations as computational-intelligence approaches for building subsystems.Source: §1 · p. 4
• The review addresses a gap in prior surveys by bringing together AI methods for thermal comfort and energy control while including individual interactions in the control loop.The authors contrast this scope with earlier reviews of optimization and comfort management.Source: §1 · p. 4

REF. OF THE WORK

• Earlier reviews leave human factors or AI-method classification out of scope, including occupant preferences and sensitivity to the thermal environment.The passages also describe a review of intelligent control for energy and comfort with occupant interaction in Smart Energy Buildings, while noting gaps in other reviews.Source: LIMITATIONS OF THE WORK · p. 5
• Prior work ranges from computational-intelligence optimization for sustainable building design to HVAC control methods, without specifically targeting AI for energy saving and thermal comfort.The cited passages also note reviews that do not focus specifically on AI methods and building-control applications.Source: LIMITATIONS OF THE WORK · p. 5
• Qolomany et al. consider HVAC energy saving without including thermal comfort or the human factor in the control loop.The passage contrasts this scope with control that accounts for both energy and comfort.Source: LIMITATIONS OF THE WORK · p. 5
• Other reviews emphasize AI-assisted HVAC efficiency and prediction accuracy or ML and big-data analytics in smart buildings, rather than classifying AI/ML techniques for energy-comfort control.Ngarambe et al. review ML and big-data analytics in smart-building structures and utilities.Source: LIMITATIONS OF THE WORK · p. 5
• A review of advanced AI-based energy control with thermal comfort focuses on ML comfort-prediction models, while the work described is limited to residential buildings and omits individual comfort modeling.The passages describe the review’s focus as deploying comfort prediction for energy saving.Source: LIMITATIONS OF THE WORK · p. 5
• The review claims to be the first to cover occupants’ personal interaction through AI-assisted tools for energy-efficient thermal-comfort optimization in buildings.Its contributions also include exploring AI/ML-assisted control for comfort and efficiency, identifying research challenges and directions, and describing system components and functions.Source: LIMITATIONS OF THE WORK · p. 5

2. Systematic Review Methodology

• Searches used “thermal comfort”, “artificial intelligence”, “big data”, “occupant”, and “heating, ventilation and air-conditioning systems” across six named databases and manual sources.The repositories covered scientific publications from 1992 to 2020.Source: §2 · p. 6
• The review process followed the Preferred Reporting Items for Systematic Reviews and Meta-Analysis guidelines.Source: §2 · p. 6
• From 1,198 initially gathered articles, 125 met criteria for indoor AI-based HVAC and thermal-comfort studies reporting system performance.The selection process is shown in Figure 2.Source: Figure 2 · p. 6, §2 · p. 6
• Studies were arranged chronologically to analyze innovative AI techniques and their usability for improving building energy savings and thermal comfort.The review considered occupant preference diversity, comfort measurement or inference, and indoor control systems including space devices, loads, building components, and occupant interactions.Source: §2 · p. 7
• The review summarized study cases, AI techniques, comfort-inference methods, inputs and controlled parameters, and resulting comfort and energy-savings performance.These findings are summarized in Tables 3 to 9.Source: §2 · p. 7

3.1. Developed AI Control Techniques

• Approximately 20 AI techniques were employed to control both thermal comfort and energy consumption in buildings.The review analyzed existing building-control works, recorded separately in Tables 3–9.Source: §3.1 · p. 7
• Artificial neural networks, including recurrent, deep, and feedforward networks, are used in direct thermal-comfort and HVAC controllers that do not require a building recognition model.Direct neural controllers were applied to comfort monitoring and HVAC control.Source: §3.1 · pp. 7–8
• A PMV-based backpropagation neural controller adjusts indoor conditions using climate, activity, and clothing inputs, but slow learning also slows its response to changes.The reported limitation includes difficulty keeping the PMV index at zero as occupant activity and clothing vary.Source: §3.1 · p. 9
• A dynamic neural network using the Nonlinear Autoregressive Exogenous (NARX) approach was used to model and control an HVAC system.Source: §3.1 · p. 9
• Fuzzy logic control incorporates expert knowledge through linguistic variables and IF-THEN rules, and is suited to nonlinear HVAC systems.Its inference engine simulates human decision-making through rough reasoning to derive a control strategy.Source: §3.1 · pp. 9–10
• The adaptive Fuzzy-PD controller obtained the lowest energy-consumption values among the described Fuzzy-PID, Fuzzy-PD, and adaptive Fuzzy-PD methods.These methods addressed thermal comfort and indoor air quality, with energy reduction as a main objective.Source: §3.1 · p. 10
• A fuzzy supervisory system let users weight efficiency against comfort, while a human-sensitivity-based comfort approach combined occupant comfort with resource savings.Source: §3.1 · p. 10
• An optimized ANN achieved prediction accuracy up to R2 99.99%; ANN algorithms reported 97.73% thermal comfort improvement and 14.04% energy saving versus the conventional algorithm.The comfort improvement was attributed to Algorithm (1), and the energy saving to Algorithm (2).Source: §3.1 · p. 11
• AI control studies reported 9.7%–25% lower energy consumption and 8.2%–18.2% lower cost, while another framework averaged 18%–40% energy reduction and 7%–19% total-cost improvement.The latter framework used fuzzy logic to account for dynamic behavior; a hybrid RNN occupancy estimator was 88% accurate and its smart controller reduced energy consumption by ~27.12% versus simple rule-based controllers.Source: §3.1 · p. 12
• For office and residential buildings, reported comfort improvements were 2.5% and ~10.2%, with annual energy reductions of 17.4% and 25.7%, respectively.Another model improved thermal comfort by >27% and reduced annual energy loss by over 30% for cooling and 40% for heating versus a conventional thermostat ON/OFF controller; predictive control also cut heating energy by ~20% without compromising comfort versus scheduled ON/OFF control.Source: §3.1 · p. 12
• An RNN HVAC controller used 4.4% less energy than Case-1 and 19.23% less than Case-2 while maintaining user-defined set-points and accurate temperature for PMV set-points.Other reported results included 4.93% energy savings with an average 0.76% PPD reduction, and average 25% energy savings while keeping elderly occupants’ PMV within -1≤PMV≤+1.Source: §3.1 · p. 12
• One approach achieved up to 45% more energy savings and 44.3% better thermal comfort performance than the PMV model.ANN-FDM improved thermal comfort by up to 4.3% versus the thermostat model and energy efficiency by up to 44.1% versus the ANN model.Source: §3.1 · pp. 12–13
• Other reported outcomes included 20%–40% energy savings while maintaining comfort except during the pre-peak cooling hour, and an average 36.5% energy saving with 25ºC identified as the ideal comfort temperature.A predictive strategy saved up to 12.8% in cost versus a rule-based strategy while maintaining comfort during occupied periods; DDPG saved 6% more energy than baseline methods and kept average comfort closer to the preset threshold of 0.5.Source: §3.1 · p. 12
• Load-demand prognoses had errors around 7.9% and 7.2% for medium- and short-term horizons, respectively, while deep neural networks outperformed ANN for heating- and cooling-load prediction using state-based sensitivity analysis.The review also describes hybrid approaches for predicting indoor temperature response in VAV systems and a personal thermal-comfort prediction method using fewer physiological parameters.Source: §3.1 · p. 13
• Combining predictive and on-demand algorithms improved energy efficiency from 13.1% to 44.4% and reduced thermal dissatisfaction by 20% to 33.6% versus either model alone.Source: §3.1 · p. 13
• One system performed better than traditional control for comfort and energy savings but had an approximately 3% error between expected and actual values.Source: §3.1 · p. 13
• Fuzzy and adaptive controls reported up to 38% energy conservation without compromising comfort and almost 18.9% energy savings versus a PID controller.A fuzzy controller was also reported to satisfy users while saving energy below the level recommended by the literature; a FACT model with a grey predictor provided reasonable comfort temperature with less energy.Source: §3.1 · p. 14
• Other approaches used weather-driven blind and window adjustments to reduce lighting, heating, cooling, and costs, while a behavior-identification framework targeted low-efficiency and low-comfort building operation.Source: §3.1 · p. 14
• The type-2 fuzzy model achieved RMSE=12.55 versus RMSE=17.64 for linear regression, while optimized HIYW performed better than OFSA for minimizing energy consumption.The OFSA comparison reported PPD exceeding 20% for ~15% of the population and 50% for ~5%.Source: §3.1 · p. 14
• Reported outcomes included a 39% reduction in daily average airflow rates, up to 17% energy savings with slightly improved comfort, and 13.8% lower daily energy use with average satisfaction of 5.56.The 13.8% saving was compared with set-point-based control; the 17% saving was compared with standard ON/OFF mono-zone control.Source: §3.1 · p. 14
• Context-aware control saved 27% of energy versus a scenario without the system, while an integrated fuzzy framework saved up to 37% and reduced HVAC response time.Source: §3.1 · p. 14
• The 2021 fuzzy method saved around 20.07% of daily energy versus temperature set-points and 10.73% versus thermal-sensation feedback, with an average satisfaction score of 5.30.Source: §3.1 · p. 14
• Reported applications include multi-agent coordination of equipment and occupants, and a residential occupant-behavior model that realistically estimated energy consumption.Other reported outcomes include PPD < 10% while managing comfort preferences and energy use, and cost reductions up to 35% for equipment and operation and 45% for comfort.Source: §3.1 · pp. 15, 20
• Reinforcement-learning agents interact with the environment, receive positive or negative rewards, and learn a policy that maximizes accumulated reward over time.The policy maps states to actions and is updated through training experience.Source: §3.1 · p. 17
• Model-based predictive control can save energy and costs versus non-predictive control, while providing robustness to disturbances, multivariable control, and improved steady-state response.The review describes predictive control as forecasting future outputs and using a rolling horizon to apply and recalculate control signals.Source: §3.1 · p. 17
• Hybrid controllers combine intelligent and classical or advanced techniques to address problems individual controllers cannot solve, but require expertise, training data, and difficult HVAC tuning.Examples include fuzzy logic with genetic algorithms, multi-agent systems with fuzzy logic, and artificial neural networks with genetic algorithms.Source: §3.1 · p. 17
• Annual thermal-comfort hours increased by 516.8–560.6 hours, and annual building energy demand fell by 27.86–33.29% versus the base-case design.The study combined NSGA-II with ANN to optimize passive design in residential buildings in China’s hot-summer, cold-winter region.Source: §3.1 · p. 20
• Energy savings of 7%–52% were reported versus conventionally scheduled cooling, while temperature deviations remained below 0.1ºC.The approach used Binary-PSO with fuzzy control.Source: §3.1 · p. 20
• BPSOFSUG outperformed BPSOFMAM in energy efficiency by 16% with satisfactory PMV, and a hybrid controller reduced peak-power duration by 61% of grid importation.Source: §3.1 · p. 20
• 88.2% energy savings and 37.7% thermal-comfort improvement versus the base case were reported; MOGA achieved up to 67% energy savings and 99.73% overall comfort improvement.Source: §3.1 · pp. 20, 22
• The proposed predictive approach achieved 91.526% accuracy, the highest compared with neural networks, ensemble RF, and SVM.Source: §3.1 · p. 20
• Integrating GA, ANN, MRA, and FLC reduced variable space by 50% and computational cost by 35.7%.Source: §3.1 · p. 20
• Case studies reported 3.5–10% energy savings versus preset control with mean TSV in [-0.5, +0.5]; another approach saved 5KWh on average, with comfort rate 5% below baseline.Source: §3.1 · pp. 21–22
• Occupant-feedback-driven thermal management showed potential savings while maintaining acceptable comfort, and Bayesian data-driven setpoints yielded realistic values useful for performance and load prediction and HVAC design and selection.Source: §3.1 · p. 22
• The Next 24h-Energy framework reduced heating consumption by 48% and cooling by 39% without affecting user comfort.Source: §3.1 · p. 22
• Cooling energy savings were around 85% in high-efficiency and 95% in low-efficiency residences, while PMV stayed within [0 – 0.5].Source: §3.1 · p. 22

Advantages

• RL/DRL can address complex problems, support long-term results, correct training errors, and learn context-specific behavior to maximize performance.Source: FUZZY LOGIC · p. 24

Limitations

• RL/DRL requires substantial data and computation, may consume considerable memory storing state values, and can perform poorly when the current state is difficult to determine.Implementations may use value-approximation techniques such as decision trees or neural networks.Source: FUZZY LOGIC · p. 24

Applications

• RL/DRL methods support continuous HVAC sensing and actions, and learn control strategies for heat pumps, HVAC systems, and water heaters to optimize energy and comfort.They are also used to learn efficient building HVAC operation.Source: FUZZY LOGIC · p. 24

Predictive-control limitations

• Predictive control faces online-computation barriers; centralized prediction in large buildings adds computational complexity, and implementation requires a process model and relatively high costs.Difficulty obtaining mathematical models has also hindered predictive controls in buildings.Source: ADVANCED MPC · p. 24

Control benefits

• Listed control benefits include robustness to disturbances, multivariable control, improved steady-state response, prediction of disturbances and control actions, and constraint handling.The passage says MPC formulation is preferable for large processes composed of several subsystems.Source: ADVANCED MPC · p. 24

• MPC can save energy, reduce peak electricity demand, account for future climate changes, and typically outperform PID or FLC alone in heating/cooling systems.Source: ADVANCED MPC · p. 24
• MPC energy-performance prediction usually uses simplified resistance-capacity (RC) models.Source: ADVANCED MPC · p. 24
• Designing the intelligent component requires user experience and substantial training data, while setting up the conventional or advanced component is difficult.Source: HYBRID METHODS · p. 24
• Hybrid regulators can solve problems that individual regulators cannot solve.Source: HYBRID METHODS · p. 24

HYBRID METHODS

• Combined models consistently outperform single models, reducing energy consumption while supporting comfort conditions and user preferences.The review attributes these significant results to combined models.Source: HYBRID METHODS · p. 24

3.2. Optimized Functions in Support of AI-based control

• Simulation tools help designers assess scenarios and infer AI-control parameters; co-simulators can provide real-time parameters for optimum operation.Co-simulators provide a global system view, complementing the controller’s local subsystem view in optimizing energy use and thermal comfort.Source: §3.2 · p. 24
• The review focuses on genetic algorithms (GA) and particle swarm optimization (PSO) as widely adopted optimization techniques in AI-assisted building control.GA and PSO are presented as the most advanced and adopted techniques in this scope.Source: §3.2 · p. 24
• GA uses selection and recombination operators to evaluate a population and evolve individuals toward a good solution.The review describes GA as imitating species evolution through natural selection in simplified form.Source: §3.2 · pp. 24–25
• PSO updates each particle’s velocity using its best individual and the swarm’s global best values to guide the search across iterations.The review characterizes PSO as a population-based stochastic technique that imitates social behavior in fish and birds living in colonies.Source: §3.2 · p. 25

4. Theoretical Analysis of the AI Applied for Building Control

• Across 125 reviewed publications, neural networks and fuzzy logic are central to AI-assisted building control, which uses sensor feedback and commonly applies optimized settings and predictive control.AI tools are also applied to sensors, while genetic algorithms and particle swarm optimization are used for optimized settings.Source: §4 · p. 27

4.1. Study Cases: Inputs and Outputs

• AI-based models use inputs from HVAC, indoor and outdoor climate, occupants, and building geometry and components to estimate comfort, microclimate, energy/load, and other outputs.The cited heat-map example used air temperature 45 times to estimate HVAC load.Source: §4.1 · p. 28

4.2. Study Cases: Energy Control

• A SLFF-ANN model of ACMV energy consumption saved 36.5%; reviewed predictive strategies optimize heating and cooling loads within power-rate limits while maintaining thermal comfort.The review groups optimal-load methods into reducing heating and cooling consumption peaks, maintaining comfortable indoor temperatures through heating management, and minimizing heating costs.Source: §4.2 · pp. 29–30

4.3. Study Cases: Thermal Comfort Measurement

• Thermal comfort assessment in the reviewed works is divided into General Comfort Models (GCM) and Individual Comfort Models (ICM).Source: §4.3 · p. 30

4.3.1. General Comfort Model

• The PMV index estimates average comfort under homogeneous, stationary conditions, which the review notes are rarely encountered in actual buildings.The PMV-PPD model is a GCM developed from climatic-chamber studies involving 1,300 subjects; conventional approaches assess thermal equilibrium between people and their surroundings.Source: §4.3.1 · p. 30

4.3.2. Individual Comfort Model

• The review identifies a concern with ASHRAE 55: its comfort models are treated as valid for anyone, although their thermal zones target 80% or 90% of people.Source: §4.3.2 · p. 30
• Individual comfort models use data-driven learning to address differing thermal perceptions and provide occupants with more individualized treatment.Conventional comfort/acceptability measures do not distinguish which user groups are uncomfortable.Source: §4.3.2 · pp. 30–31
• Personalized approaches include participatory thermal-preference votes from -5 to +5 and a Degree of Individual Dissatisfaction based on user vote, ambient temperature, desired temperature, and tolerance.Source: §4.3.2 · p. 31
• Short-term comfort indices are classified as rational, empirical, or direct, but are limited to a particular time or position; long-term indices evaluate whole-building conditions over time and across zones.Source: §4.3.2 · p. 31
• Long-term thermal-comfort evaluation most commonly reports discomfort rather than comfort.Source: §4.3.2 · p. 31

Trend Analysis and Discussions

• Neural networks were the most popular approach, while hybrid methods combined techniques such as FL with ANN or GA/PSO, and GA/PSO often optimized building control [FL, ANN, GA, PSO].DAI and MAS were less common but used multiple controllers in complex control systems.Source: §5 · p. 31
• Fuzzy logic is widely used because it can imitate user behavior and express thermal-comfort sensations linguistically.Fuzzy controllers can use these descriptions to estimate PMV calculations for system control.Source: §5 · pp. 31, 34
• From 1993 to 2020, average energy savings reached up to 31%, while a Bayesian network model achieved maximum savings of ~90%.The Bayesian model estimated employee mental performance under different thermal conditions.Source: §5 · p. 32, Figure 13 · p. 33
• Average comfort improvement was around 50%, and maximum comfort reached 100% with neural networks, DAI and MAS, and GA.The review also describes an Intelligent Comfort Control System that maintained PMV within the comfort zone while saving energy.Source: Figure 14 · p. 33, §5 · p. 33
• Average cost reduction was up to 34%, with maximum energy-saving costs of 58%; only 6.7% of reviewed works targeted cost-effectiveness.Cost-effectiveness included operating, energy, electricity, and comfort costs.Source: Figure 15 · p. 33, §5 · p. 33
• The Bayesian Comfort Model estimated user preference with 13.2% to 25.8% accuracy and could save up to 13.5% energy while reducing discomfort by 24.8%.It was presented as a personalized model for thermal comfort and energy savings.Source: §5 · p. 35
• In a two-story home, ANFIS- and ANN-based control potentially maintained comfort at ~98% in winter and 100% in summer, but none of three AI logics significantly out-saved the others.The comfort bands were 20–23ºC in winter and 23–26ºC in summer.Source: §5 · p. 35
• A hybrid-MOGA multi-agent system saved 31.6% energy and improved comfort by about 71.8% versus traditional optimization, while a decentralized office system saved up to 40% versus baseline with 100% thermal satisfaction.The decentralized system optimized lighting, heating, and ventilation.Source: §5 · p. 35

6.1. Smart and connected buildings

• Connected buildings improve energy efficiency and resilience through community microgrids and district energy, support smart-grid peak-power management through demand response, and enable a net-zero community in terms of energy.Source: §6.1 · p. 37

6.2. Comfort modeling and dynamic temperature set-point adjustment

• In-situ comfort conditions differ by study context; the reported variety exceeds PMV’s, which appears to overestimate warmth; PMV could not estimate thermal neutrality in both types of buildings.Reported acceptability was between 22ºC and 30.5ºC during summer, 4ºC above the norm.Source: §6.2 · p. 37
• Real-life comfort reflects dynamic interactions between occupants and their environments, which stationary, homogeneous climate-chamber models cannot fully predict.Comfort discrepancies also reflect uncertainty in metabolism and clothing estimates, as well as psychological and sociological factors.Source: §6.2 · p. 38

6.3. Human-in-the loop research challenges and directions

• At least 5% of people in each group report discomfort with predefined conditions, and existing standards neither distinguish occupant variability nor let control systems dynamically track model parameters.The reported share may exceed the whole population when environmental conditions become more adverse.Source: §6.3 · p. 38
• An AI-based personal comfort framework for senior citizens achieved 76.7% predictive accuracy versus 35.4% for the Fanger model by including age and outdoor temperature.The framework used environmental, psychological, and physiological features.Source: §6.3 · pp. 38–39
• A smartphone-based approach to tracking optimal conditioning modes in single- or multi-occupant areas was associated with an almost 53.7% reduction in uncomfortable reporting.The section describes personalized comfort and energy savings as a challenging joint objective, especially when occupants in shared spaces have conflicting preferences.Source: §6.3 · p. 39

6.4. Context-awareness computing in energy and thermal comfort control challenges and research directions

• Context-aware building energy management depends on sensing the environment, acquiring information about users, and actuating equipment, using contextual signals such as occupant location, temperature, humidity, sound, and lighting.Source: §6.4 · p. 39
• Context-aware control requires connecting IoT devices to the BEMS control plane and detecting occupant activity, while deployment choices depend on whether the existing BEMS is open- or closed-source.Source: §6.4 · p. 39
• Limited sensor battery life remains a challenge; EACRA and lower-rate, condition-specific sensing could reduce redundant transmissions, while machine-learning routing and in-building energy harvesting remain research opportunities.The proposed sensing approach exploits the fact that some measurements, such as room temperature, change infrequently.Source: §6.4 · pp. 39–40

6.5. Personal big-data streaming challenges in smart buildings

• Smart buildings can generate terabytes of heterogeneous personal data daily, for which batch processing is unsuitable and rapid comfort control requires streaming-capable ML, cloud, virtualization, and distributed platforms such as Hadoop and Spark.These systems must handle data from multiple sources and support fast response and processing.Source: §6.5 · p. 40

6.6. Security, privacy and data sensitivity issues in smart building

• Connected smart-building devices amplify security risks, so trustworthy systems should address safety, resilience, reliability, cybersecurity, and privacy together across physical, analog, and cyber components.The five design considerations have traditionally been treated in isolation, an approach described as insufficient for overall system security.Source: §6.6 · p. 40, Figure 21 · p. 41
• Private sensor data can enable intimate inferences and exploitation, motivating ethical oversight and regulation that ensure informed consent before personal data are used.The paper cites efforts such as the EU General Data Protection Regulation as a basis for such protections.Source: §6.6 · p. 41
• Effective privacy-preserving machine learning without sacrificing accuracy remains an open question, and differential-privacy techniques require further evaluation in smart-building environments.The authors describe differential privacy as technically promising but say its suitability needs more research.Source: §6.6 · p. 41

6.7. Lack of sufficient amounts of data issues and lack of AI-based modeling in buildings

• AI models may overfit or make biased decisions when datasets are scarce, unrepresentative, unbalanced, or unlabeled; future work should collect sustainable, genuine, heterogeneous building data.Data availability may also be constrained by copyright and privacy considerations.Source: §6.7 · p. 41

7. Conclusions

• The reviewed studies were primarily empirical case studies, drawing on real-world occupant questionnaires or interviews and measurements, or on current publicly accessible datasets.The data concerned thermal comfort and energy use.Source: §7 · p. 42
• ANNs address recognition and identification through learning and classification, and are used to describe thermal comfort and estimate the Predicted Mean Vote (PMV) index.The review reports that ANNs retain and classify data for these tasks.Source: §7 · p. 42
• Fuzzy Logic studies treat thermal comfort as subjective, often use the PMV index, and monitor conditions for satisfaction and energy efficiency through simple, scalable, model-free regulation.The review describes this approach as focused on experience and judgement.Source: §7 · p. 42
• Compared with traditional controls, these approaches are distinguished by additional awareness of system behavior through fuzzy language or ANN learning, or by the optimality sought with Genetic Algorithms.The review attributes the approaches’ advantage primarily to these capabilities.Source: §7 · p. 42
• AI and ML building control remains an ongoing research endeavor because these algorithms need substantial high-quality real-world data, which has been scarce in the building energy sector.Smart-meter installation, IoT, and cloud storage are increasing data quantity and sophistication, enabling more effective data-driven research.Source: §7 · p. 42
• Research challenges include more data for AI-based building models, IoT-connected buildings for efficient management and collection, and security, privacy, data-sensitivity, and big-data-streaming concerns.The review notes that smart buildings bring these challenges alongside research opportunities.Source: §7 · p. 42
• The review identifies context awareness and human-in-the-loop comfort modeling as research directions for adapting comfort and dynamic temperature set-points to human activity while seeking energy savings.For commercial buildings, dynamic set-point adjustment depends on connecting comfort models to activity; tracking activity can support efficient management with satisfactory comfort.Source: §7 · p. 42