Smart buildings as Cyber-Physical Systems: Data-driven predictive control strategies for energy efficiency

1.1. Motivation

• Buildings use substantial energy, and their operational phase accounts for up to 90% of conventional buildings’ life-cycle energy consumption.In 2010, buildings accounted for 41% of US primary energy use; the operational phase accounts for up to 50% in low-energy buildings.Source: §1.1 · pp. 1–2

1.2. Political Context and Survey Focus

• The introduction situates building efficiency within climate and energy policy, including an EU target of 20% primary-energy savings by 2020 and 50% by 2050 versus 2008.Source: §1.2 · p. 2
• The survey examines predictive control in instrumented buildings as a complement to refurbishment and modernization, excluding offline diagnosis and studies focused solely on modeling accuracy.It focuses on computational methods integrated into daily building operation and frames future research questions.Source: §1.2 · pp. 2–3

1.3. Structure

• The paper moves from background to recent predictive-control literature, then formulates open research questions and concludes.Source: §1.3 · p. 3

2.1. Building Energy Application Key Performance Indicators

• Discomfort in thermal comfort, lighting, noise, or indoor air quality was associated with 2.4%-14.8% lower work performance in a study of 114 office workers.Energy comparisons may require normalization by floor area, visitors, employees, or weather, while common thermal comfort indices are debated because temperature alone does not capture factors such as solar radiation or humidity.Source: §2.1 · pp. 3–4

2.2. Buildings as Cyber-Physical Systems

• Predictive building CPS use sensor data and computational representations of physical processes to optimize future control decisions and feed their effects back into subsequent observations.Data preparation can include cleaning, feature selection, and standardization; representations may use data-driven models, whose accuracy can deteriorate as buildings and occupant behavior change, requiring regular updates. Optimization may need to account for stochastic events and uncertainty; a genetic-algorithm scheduling study reported 56% system operation cost savings and improved thermal comfort.Source: §2.2 · pp. 6–8

2.3. Relation to Building Automation

• Predictive control can build on existing building automation, with a building management system serving as a gateway and additional devices or external data added as needed.One example deployed a predictive-control representation on a separate server through BACnet/IP and used external weather information.Source: §2.3 · pp. 9–10

2.4. Control Mechanisms

• Predictive building control is a closed-loop CPS because actuation affects future building data, which can in turn influence later control decisions.Control may be open loop, feed-forward, or feedback-based; predictive building control belongs to the feedback-based form.Source: §2.4 · pp. 10–11

2.5. Distributed Building Control and Conflicts

• Multi-agent systems offer scalable, distributed building control, but agents’ differing goals or perceptions can create conflicts that require resolution.Utility functions can weigh conflicting options, while mediation or negotiation protocols provide additional means to accommodate conflicts.Source: §2.5 · pp. 11–12

2.6. Smart Grid Interactions, Renewables, and Storage Capacity

• Buildings with thermal or electrical storage can use demand response to shift flexible demand toward periods of surplus renewable generation and support grid management.The survey also identifies forecasting on-site renewable generation to increase local renewable consumption as a related building-operation goal.Source: §2.6 · pp. 12–13

2.7. Building Information and Semantics

• IFC, gbXML, SAREF with FIEMSER, and Project Haystack address building, appliance, construction, or operational data at different scopes.gbXML supports transfer from CAD building models to engineering tools, and transforms it into an ontology; ontologies represent domain relationships and enable reasoning and inference.Source: §2.7 · p. 13
• Domotics models such as DomoML, DogOnt, and BOnSAI have energetic extensions, but are typically used for convenience services rather than predictive control.The extensions cover device energy profiles and, in, device consumption, home energy supply facilities, and energy sources.Source: §2.7 · p. 13
• BIM remains rare in existing building stock because of effort, maintenance, and uncertain data, objects, and relations; creating BIM for existing buildings is a research challenge.A BIM-to-BEM translation using Modelica demonstrated reuse of original BIM data in simulation, while BIM has also supported WSN topology planning.Source: §2.7 · p. 14

3.1. Building Energy Modeling Surveys

• Survey reports that anticipated energy or cost savings have been validated in building simulation, but not in controlled experiments during routine operation.Source: §3.1 · p. 14
• Survey reviews building-modeling approaches since the 1980s, including considered parameters, validation methods, and available simulation software, without reporting energy-savings figures.Source: §3.1 · p. 14
• Survey analyzes 121 building energy and comfort-control studies, emphasizing publication trends and techniques rather than quantifying achieved effects.It reports that Model Predictive Control, MAS approaches, fuzzy control, and Genetic Algorithms are commonly applied.Source: §3.1 · p. 14

3.2. Classification of Studies

• The taxonomy is used to categorize the data-driven building-control approaches surveyed in the section.Source: §3.2 · p. 14
• Theoretical, or white-box/model-based, approaches use building physics or suitable approximations, often through simulation programs such as Modelica or EnergyPlus.Source: §3.2 · pp. 14–15
• Data-driven, or black-box, approaches train regression models on observed building-management data; using problem knowledge to guide training is termed gray-box.Source: §3.2 · p. 15
• Hybrid approaches combine data-driven methods with theoretical models.Source: §3.2 · p. 15
• Mixed studies test approaches from different categories, and the survey focuses on the best-performing approach in each case.Source: §3.2 · p. 15

• Up to 56% heating-cost savings were reported for day-ahead HVAC set-point planning against standard control when maximum PPD was 25%.The simulations used a typical heating-season day in Naples; at a 20% PPD target, savings were 42%.Source: §3.3 · p. 15
• The distributed cooling scheduler saved 17% energy against standard pre-cooling and scaled to hundreds of zones in simulation.Zone controllers request energy based on thermal demand, while a central scheduler allocates cooling and selects ventilation fan stages and damper positions.Source: §3.3 · p. 15
• A modular MPC approach approximates nonlinear building optimization through decoupled estimation, thermal-flow calculation, and set-point derivation.A co-simulation of an Austrian low-energy office demonstrated the approach, but quantified no energy savings.Source: §3.3 · pp. 16–17
• The Comfort Prioritizing Greedy algorithm performed “slightly better” than bin-packing and linear-programming scheduling in a single-day simulation.It schedules appliances under an energy-cost budget using a linear indoor-temperature evolution model.Source: §3.3 · p. 17
• Up to 22% lower exergy destruction and 36% lower energy consumption were reported against traditional control for exergy-focused HVAC control.Against energy-focused MPC, the simulations showed 4% exergy savings and 12% energy savings.Source: §3.3 · p. 17
• Simulations of demand-response appliance control reported cost reductions of 16%-19% and an 18% electricity-bill reduction in separate studies.The first approach also had a Zigbee lab demonstration; the second used fuzzy logic with conditional value at risk and simulated a single day.Source: §3.3 · p. 17
• Energy-cost savings of 20%-60% were reported for HVAC operation in six commercial buildings against pre-cooling night setback.The comparison covered five August days using three-year average weather data; savings depended on building, location, and comfort weighting.Source: §3.3 · p. 17

3.4.1. Neural Networks

• A computational Firefly optimization study reported energy savings from 2% under the strictest comfort constraints to 17% under the most relaxed, while a separate CO2-aware study estimated average electricity savings of 12%-17%.The CO2-aware study found a NN Ensemble performed best compared with seven regression models and used modified PSO to find Pareto-optimal trade-offs among IAQ, comfort, and energy consumption.Source: §3.4.1 · p. 18
• The lecture-room predictive HVAC model reported expected financial savings in the order of 50%, while using more energy to reduce comfort violations.The two-week experimental study did not normalize for room characteristics or weather.Source: §3.4.1 · p. 18
• Weather-normalized thermal-energy savings of 56% were reported for soccer-stadium grass heating over a winter season.The best results came from control informed by other heat-consuming systems, and this was the only study noted to quantify confidence intervals using statistical inference.Source: §3.4.1 · p. 18
• A neural-network model and genetic-algorithm control-rule approach reported approximately 25% weather-normalized energy savings in a care-home evaluation.Validation used three months of simulation and two months of experiments; a facility manager could select rules from a knowledge base.Source: §3.4.1 · p. 18

3.4.2. Reinforcement Learning

• A 40-day water-heater lab experiment achieved 15% cost savings with reinforcement learning, rising to up to 28% when exploration was excluded.Simulations reported 24% savings with day-ahead prices and 34% with imbalance prices, each against the default strategy.Source: §3.4.2 · p. 19
• Simulations reported 4%-11% heat-pump energy savings across two buildings, and HVAC demand-response policies stabilized within 90% of the mathematical optimum after 16 days of data.The heat-pump method used an auto-encoder without thermal or physical modeling; the demand-response approach was also qualitatively validated in a living lab.Source: §3.4.2 · p. 19
• A photovoltaic-thermal array controller achieved 5%-11% power improvement over a domain-expert rule-based controller in three simulated years.The controller adjusted the array’s mass-flow parameter for a Swiss low-exergy residential building.Source: §3.4.2 · p. 19

3.4.3. Regression Trees

• A regression-tree demand-response strategy reported 17% higher curtailment than a rule-based strategy while maintaining thermal comfort.For a large reference building, the authors estimated $45,600 in revenue across 20 assumed summer DR events.Source: §3.4.3 · p. 19

• A hybrid resistance-capacitance and neural-network approach reported up to 41% energy savings for occupied-hour start-stop control.Pre-cooling saved 13% electricity cost while increasing energy consumption by 5.6%; the approach was evaluated through simulation and experiment at Adelaide airport.Source: §3.5 · p. 20
• A simplified model-based controller with nonlinear filters reduced room ventilation rates by 12%-19% while accounting for indoor air quality.Experiments used mock-up office and meeting-room environments in winter and summer in Sweden.Source: §3.5 · p. 20

• In real-building experiments, model-free PCAO achieved 19% energy savings and outperformed the model-based approach, which the authors attributed to modeling inaccuracies.In simulations, model-based PCAO achieved 45% and 25.6% energy savings against two rule-based baselines, while model-free achieved 41.2%-44.6% and 20.5%-25.3%, respectively.Source: §3.6 · p. 20

4.1. Buildings as CPS - Methodology

• Existing buildings lack a concise methodology for becoming closed-loop CPSs that deploy predictive control, including integration with existing automation and instrumentation and complex stakeholder needs.The survey asks what methodology can evolve existing buildings into CPSs for higher operational efficiency.Source: §4.1 · p. 22

4.2. Feature Selection

• Feature selection is difficult because building systems expose many variables and possible histories, while reviewed studies often omit their selection rationale.The survey proposes investigating semantic selection using BIM, including thermal interdependencies to guide history lengths and inter-feature lags, alongside human and environmental context.Source: §4.2 · pp. 22–23

4.3. Inter-building Transfer

• Transferring control between buildings remains costly and error-prone because point names need mapping and building differences can undermine predictive models or require retraining.The survey identifies concept drift, transfer learning, and mining BMS metadata as possible avenues for more efficient transfer.Source: §4.3 · p. 24

4.4. Control Conflicts

• The reviewed literature largely controls one building system at a time; only one study considers effects on other systems, reactively, leaving conflict prediction for future study.The survey asks whether agents that learn to predict conflicts can increase building energy efficiency.Source: §4.4 · pp. 25–26
• Multi-agent control must address conflicts from user preferences, scarce resources, and interdependent systems, while competitive settings also raise trust and data-access questions.Potential settings include tenants sharing limited boiler capacity, PV electricity, or demand-response resources.Source: §4.4 · pp. 24–26

4.5. Energy Disaggregation

• Commercial-building disaggregation is challenged by overlapping appliance signatures, household-focused research, and assumptions that only one appliance changes state at a time.Existing BMS data may support disaggregation without added installations, but the survey notes thermal systems are largely uncovered and asks whether methods extend beyond electricity.Source: §4.5 · p. 27

5. Conclusion

• Most studies optimize operational energy cost or consumption and often include thermal comfort, while giving little attention to exergy, alternative KPIs, or CO2 savings.Optimizing energy costs may result in higher building energy use.Source: §5 · p. 28
• A minority of studies addresses BMS integration, and systemic interdependencies or conflicting control commands are rarely studied.Most studies validate through simulation; fewer use experiments, and most experimental periods are relatively short.Source: §5 · p. 28
• Few studies account for weather in energy-performance assessments, and fewer still quantify savings confidence intervals with statistical methods.Changing weather may influence real-world experimental results.Source: §5 · p. 28
• Most works do not explain how their predictive-control features are derived, while nature-inspired heuristics are common and robust optimization is not prominent.The survey also notes that works using PMV or PPD do not discuss whether comfort KPIs use static or adaptive definitions.Source: §5 · p. 28
• Data-driven predictive-control studies reach effect sizes similar to simulation-driven MPC studies, with neural networks and, less often, reinforcement learning commonly used.The survey also reports that none of the reviewed works explicitly applies general CPS research findings, and a methodology for evolving buildings into predictive CPSs is lacking.Source: §5 · p. 28
• The survey directs future research toward seven questions on CPS methodology, semantic feature selection, transfer, multi-agent conflicts, and building-scale energy disaggregation.The questions include whether conflict-predicting agents improve energy efficiency and whether disaggregation approaches extend beyond electricity.Source: §5 · p. 29