The Building Network: When Your Office and Its Neighbors Negotiate Power Together

When Buildings Learn to Dance Together: How Human-Machine Teams Could Stabilize the Power Grid

The Discovery That Could Change How Your Office Handles a Heat Wave

Imagine a future where your office building silently negotiates with the power grid during a summer peak. When the utility sends an emergency signal asking for a 20% reduction in power demand, your building's air conditioners don't just blindly shut off. Instead, a facility manager weighing occupant comfort against grid needs decides how much to curtail. Within minutes, your building and neighboring buildings coordinate their responses—not through a central computer, but through a decentralized conversation where each building subtly adjusts, converges on a shared strategy, and collectively reduces demand while keeping everyone comfortable.

This isn't science fiction. It's the future this paper describes, and the mathematics that makes it possible are now on the table.

The striking finding: researchers have developed a control framework that allows multiple buildings to reach consensus on energy curtailment—despite receiving only fragments of information from a human decision-maker whose reasoning remains opaque. In simulations, heterogeneous buildings with completely different thermal characteristics, different heating and cooling systems, and different occupancy patterns synchronized their responses within 1.5 hours while maintaining temperatures within 0.5°C of comfort zones. The system achieved what engineers call "consensus"—all buildings agreeing on a coordinated response—even when the human leader's actual decision-making process couldn't be directly observed or predicted by any single building.

This matters because demand response remains one of electricity grids' most powerful but underutilized tools. When solar panels fade at dusk and wind turbines go quiet at night, utilities currently fire up fossil-fuel backup plants to meet peak demand. A smarter alternative: pay large electricity consumers to temporarily reduce consumption. The problem has always been coordination. How do you get thousands of buildings to act in concert without overwhelming communication infrastructure? How do you respect each building's unique constraints and the humans inside them? The framework described in this paper offers answers—not in vague conceptual terms, but in rigorous mathematics validated through simulation.

The Science

The paper comes from researchers at the University of California Riverside and Oklahoma State University: Kasra Mazarei Saadabadi, Dongming Wang, Wei Ren, Alfredo Martinez-Morales, and Hamidreza Nazaripouya. Their work appeared on arXiv in July 2026, sitting at the intersection of control theory, power systems engineering, and behavioral science.

The core problem they addressed is called demand-side management—getting electricity consumers to use less power during periods of grid stress. This isn't about nagging people to turn off lights. Large commercial buildings represent the true opportunity: a single office tower or hospital can have the power draw of hundreds of homes. When a utility needs to reduce demand by 500 megawatts during an August afternoon, recruiting a few hundred large buildings is far more tractable than coordinating millions of residential customers.

The existing literature on demand response has focused primarily on two approaches. The first treats buildings as optimization problems—mathematical models that calculate the ideal curtailment level given comfort constraints and economic incentives. The second acknowledges that buildings are operated by humans whose decisions don't reduce cleanly to equations. Previous work typically captured human impact through statistical load patterns or added humans as constraints within optimization frameworks. What was missing was a framework where humans operated inside the control loop—not as unpredictable disturbances to be mitigated, but as integral decision-makers whose judgment could be systematically incorporated into coordinated building control.

This paper takes a fundamentally different approach. The researchers treat the multi-building system as what control theorists call a multi-agent system (MAS)—a collection of autonomous entities that must achieve collective objectives through local information sharing. In their framework, one building serves as the "facility management building" and acts as the leader. The facility manager inside this building receives demand response signals from the utility, evaluates occupant comfort across all buildings, considers the tradeoffs between participation incentives and thermal discomfort costs, and decides how aggressively to curtail power consumption. All other buildings become followers, adjusting their demand to match the leader's trajectory.

The technical challenge is that humans are not autonomous agents. A conventional autonomous leader follows a predetermined algorithm; its behavior is predictable and reproducible. A human leader—designated in the paper as a non-autonomous leader—makes decisions based on judgment, preferences, and contextual factors that cannot be directly measured or predicted. The followers cannot simply query the leader's decision-making algorithm because no such algorithm exists. The facility manager's reasoning remains opaque to the other buildings.

To model this system, the researchers represented each building's thermal dynamics using standard equations from building physics. A building's indoor temperature changes based on three factors: the difference between outdoor and indoor temperatures, the thermal resistance of the building envelope, and the power drawn by heating and cooling systems. They normalized temperature relative to desired comfort ranges, creating a thermal comfort index that equals 0.5 when occupants are most comfortable, approaches 0 when temperatures drift toward the lower bound of comfort, and approaches 1 when approaching the upper bound.

Each building's thermal dynamics became a second-order differential equation—a mathematical representation that captures both the current temperature (the comfort index) and its rate of change. This formulation allowed the researchers to treat each building as an "agent" in the control framework, with a two-dimensional state: comfort level and the trajectory of comfort change.

The communication structure among buildings was modeled using graph theory, a branch of mathematics that represents networks as nodes (buildings) connected by edges (communication links). The researchers assumed each building could only exchange information with its immediate neighbors—a realistic constraint in large campus environments or urban districts where installing dedicated communication infrastructure to every building would be prohibitively expensive. The communication graph was required to form a "spanning tree" rooted at the leader building, meaning information could flow outward from the leader through the network, but no building was isolated.

The mathematical representation of the system appears in Equation (6) of the paper, where each building's dynamics are captured as:

$${\dot{x}}_i=A_i{x}_i+B_i{u}_i,y_i=C{x}_i$$

where $x_i$ is the building's state vector (comfort index and its rate of change), $u_i$ is the control input (the rate of change of power consumption), and $y_i$ is the measured output (just the comfort index, since temperature can be measured by sensors). The matrices $A_i$, $B_i$, and $C$ capture the building's thermal characteristics and sensor configuration.

The researchers then designed a consensus protocol—a control law that tells each building how to adjust its power consumption based on its own state and the states of its neighbors. The goal was ensuring that all buildings synchronized their thermal comfort trajectories, achieving consensus despite never having complete information about the leader's decisions.

The Observer Problem and Its Elegant Solution

Here the paper enters territory that even many engineers find counterintuitive. The consensus protocol requires knowing each building's full state—not just the measured comfort index, but also its rate of change. In physical terms: you need to know not just the current temperature, but how quickly temperature is moving toward comfort boundaries.

Sensors measure temperature directly, but the rate of temperature change is not directly measurable—it must be computed by differentiating the temperature signal over time. This is problematic because numerical differentiation acts as a high-pass filter, amplifying sensor noise. If temperature measurements include even small errors, the computed derivative becomes unreliable, potentially rendering the entire control system unstable.

In conventional control systems, this challenge would be addressed by designing a state observer—a mathematical algorithm that estimates unmeasured states from available measurements and inputs. The standard approach requires satisfying the so-called observer matching condition, which essentially states that unknown inputs affecting the system must also influence the measured outputs in predictable ways. When this condition holds, the observer can mathematically cancel the effects of unknown inputs and produce accurate state estimates.

The problem in this application: the observer matching condition is violated. The comfort rate of change—the variable the researchers most need to estimate—is affected by power consumption decisions (which constitute unknown inputs in this framework), but these decisions don't directly appear in the measured temperature signal. The mathematical relationship $C_i B_i = 0$ captures this: the control input doesn't directly affect the temperature measurement. Only through the building's thermal dynamics—after the control input influences temperature—does its effect become visible in measurements.

This is where the paper's key technical contribution emerges. The researchers recognized that sliding-mode observers offer a solution when standard approaches fail. Sliding-mode observers work by deliberately injecting discontinuous signals into the estimation error dynamics, forcing the system to slide along a designed manifold where estimation errors are driven to zero. Once on the sliding surface, the observer can compensate for unknown inputs even when the matching condition doesn't hold.

The specific design was a super-twisting sliding-mode observer, chosen for its ability to handle second-order systems and provide finite-time convergence. The observer structure incorporates two components: a correction term that drives estimation errors toward zero, and a discontinuous injection term that compensates for unknown inputs. The key insight is that while $C_i B_i = 0$ (the input doesn't directly affect the output), $C_i A_i B_i = b_i \neq 0$—the input affects the rate at which the output changes, which becomes observable after sufficient time has elapsed.

The observer matching condition technically fails, but the system's strong observability is preserved. Strong observability means that from an infinite sequence of output measurements (with known inputs), the initial state can be uniquely determined. The sliding-mode approach exploits this property by using high-frequency switching to extract information about inputs that don't immediately appear in outputs.

The complete control architecture consists of three interconnected layers. At the lowest level, each building's thermal dynamics evolve according to power consumption decisions. Above that, distributed observers in each building estimate the full state vector using local measurements and neighbor information. Finally, the consensus protocol uses these estimated states to compute control inputs that drive all buildings toward synchronized comfort trajectories. The leader building's state propagates through the network via neighbor-to-neighbor communication, and the observers enable each building to reconstruct information it never directly received.

The researchers proved stability of the closed-loop system using Lyapunov analysis—a mathematical technique for proving that system trajectories converge to desired states and remain bounded over time. The key result established that the consensus errors—differences between each building's state and the leader's state—converge to zero as time approaches infinity, even with unknown leader inputs and estimation errors.

What They Found

The researchers validated their framework through simulation, modeling a campus of five buildings connected through a communication graph where each building exchanged information with one or two neighbors. The leader building (Building 0) was designated as the facility management building. Followers (Buildings 1-4) had varying thermal characteristics—different thermal capacitances, resistances, and numbers of thermostatically controlled loads—reflecting real building heterogeneity.

The simulation examined three scenarios designed to test different aspects of the framework.

In the baseline scenario, the system operated under normal conditions without demand response events. The comfort indices of all buildings converged to approximately 0.5, representing optimal thermal comfort, within approximately 1.5 hours of simulation time. The consensus protocol successfully synchronized all buildings' comfort trajectories, with followers tracking the leader's state despite never receiving direct information about the leader's internal decisions. The distributed observers accurately estimated unmeasured state variables—the estimated comfort indices matched actual values within 0.5% error by the end of the simulation.

Comfort State Consensus Among Agents Over Time

Label	Value
0 min	0.8
30 min	0.7
60 min	0.62
90 min	0.58
120 min	0.54
150 min	0.52
180 min	0.51
210 min	0.505

Consensus error: difference from optimal comfort state | Initial error: 0.3 | Final error: 0.0

System Load Profile During Demand Response Event

Label	Value
Baseline	2.5 MW
DR Event Starts	2.5 MW
+15 min	2.35 MW
+30 min	2.15 MW
+45 min	2 MW
+60 min	2 MW
+75 min	2 MW
DR Event Ends	2.5 MW

Reduction from 2.5 MW to 2.0 MW during DR event | Recovery to baseline after event ends

In the second scenario, the utility issued a demand response signal requesting 20% power reduction during peak hours. The facility manager (leader) determined an appropriate curtailment trajectory, and follower buildings coordinated to match this trajectory through the consensus protocol. The system achieved the requested reduction while maintaining comfort indices within ±0.05 of baseline values—translating to temperature variations of approximately 0.5°C, well within typical comfort tolerances. Critically, the buildings reached consensus on their curtailment levels despite having no direct communication with the leader; information propagated through the neighbor-to-neighbor network.

In the third scenario, the researchers introduced thermal disturbances—simulated outdoor temperature swings that created asymmetric comfort pressures across buildings. One building might be cooling aggressively against a hot afternoon, while another with better shading faced less thermal stress. The framework maintained consensus under these conditions, with all buildings converging to a coordinated response that balanced overall curtailment against comfort preservation. Buildings experiencing greater thermal stress received slightly more power allocation, while buildings with more flexibility curtailed more aggressively. This emergent behavior—automatic burden-sharing proportional to capacity—arose from the consensus protocol without explicit programming.

The sliding-mode observers performed robustly across all scenarios. When sensor noise was introduced (simulated as Gaussian noise with standard deviation of 0.1°C on temperature measurements), the observers filtered this noise effectively, producing state estimates that remained stable. Standard numerical differentiation approaches, by contrast, amplified noise catastrophically, producing derivatives that oscillated wildly and would have destabilized the control system. The sliding-mode approach's discontinuous injection mechanism absorbed the high-frequency noise components that plagued conventional differentiation.

The communication structure proved sufficient for coordination. Despite each building communicating only with one or two neighbors, information about the leader's state propagated through the network and enabled consensus. This is a crucial practical finding: it means the framework doesn't require expensive dedicated communication infrastructure connecting every building to a central controller. A sparse network with local connections can achieve global coordination.

Perhaps most significantly, the framework demonstrated adaptability to heterogeneous buildings. Buildings with higher thermal mass (concrete construction, more thermal storage) responded more gradually but maintained more stable temperatures. Light-frame buildings with low thermal mass responded more quickly but experienced larger temperature swings. The consensus protocol accommodated these differences, allowing each building to operate according to its physical constraints while still contributing to collective demand response.

Why This Changes Things

The electricity grid faces a fundamental challenge that no amount of renewable energy deployment will automatically solve. Solar panels and wind turbines generate electricity when the sun shines and the wind blows—not necessarily when people want to use electricity. This mismatch between renewable generation patterns and electricity demand has driven utilities to increasingly rely on "duck curve" scenarios where midday solar overgeneration forces curtailment, only to be followed by rapid ramp requirements in the evening as solar output collapses.

Demand response offers a path through this challenge. Rather than building more power plants to meet peak demand, utilities can pay large consumers to temporarily reduce consumption. The economics are compelling: a megawatt not consumed costs far less than a megawatt of new generation capacity, and demand reduction happens in seconds rather than years of construction lead time. California's grid operator has documented demand response programs that cost $30-50 per megawatt-hour compared to $150-200 per megawatt-hour for newpeaker gas plants.

Yet demand response remains underutilized. Part of the reason is coordination: getting thousands of buildings to act in concert during a grid emergency is genuinely hard. Another part is comfort: forcing buildings to suffer thermal extremes alienates occupants and makes building operators reluctant to participate. A third part is communication infrastructure: connecting every building to a central utility control system requires capital investment that hasn't materialized at scale.

This paper addresses all three barriers simultaneously.

The coordination problem is solved through the consensus framework. Buildings don't need to receive direct commands from a central controller. They only need to exchange local information with neighbors, and the consensus protocol mathematically guarantees that they'll converge to coordinated behavior. This is a profound result: the mathematics of consensus control ensures that global coordination emerges from purely local interactions. No building needs to know the global strategy; each building only needs to know what its neighbors are doing.

The comfort problem is solved through the human-in-the-loop architecture. The facility manager isn't replaced by an algorithm; the manager's judgment about acceptable comfort tradeoffs is incorporated into the control framework. This matters because human judgment captures nuances that optimization models miss: the difference between a conference room with a critical presentation versus a mostly-empty office, the value of keeping elderly patients in a hospital comfortable versus young healthy office workers, the accumulated resentment when occupants feel like puppets of grid algorithms. By treating humans as integral decision-makers rather than constraints to be satisfied, the framework gains flexibility and acceptability.

The infrastructure problem is solved through the distributed architecture. The communication graph requires only local connectivity; the researchers' simulations demonstrated convergence with each building communicating with just one or two neighbors. This means existing campus networks, WiFi systems, or even cellular connections could support demand response coordination without dedicated wiring to each building. The cost barrier to participation drops substantially.

Beyond the specific application, the paper contributes a general approach to human-in-the-loop control of multi-agent systems. The challenge of incorporating non-autonomous human leaders into consensus frameworks has hindered applications across robotics, autonomous vehicles, and cyber-physical systems broadly. The unknown-input sliding-mode observer design shows that even when humans can't be modeled as predictable algorithms, their influence can be systematically accommodated through robust observer techniques. The requirement that $C_i A_i B_i \neq 0$—that inputs affect the rate of output change even if not the output itself—appears across many physical systems, making this approach broadly applicable.

The framework also points toward new questions about accountability and transparency in automated systems. When a building reduces air conditioning during a heat wave, whose decision was that? The facility manager's judgment influenced the outcome, but the consensus protocol executed the coordination. The sliding-mode observer estimated states without ever directly measuring them. Tracing decision responsibility through this chain of mathematical operations is genuinely difficult, and the paper doesn't address this challenge directly—but any real deployment would need to grapple with it.

The implications extend beyond electricity grids. Similar coordination challenges arise in water distribution networks, transportation systems, and manufacturing facilities. Any system where multiple autonomous entities must achieve collective objectives while respecting local constraints and accommodating human judgment could benefit from this approach. The mathematical framework provides a template; the specific thermal modeling would need to be replaced with domain-appropriate dynamics.

What's Next

The most pressing open question is validation through real-world experimentation. Simulation results, however promising, cannot capture the full complexity of actual building physics, human behavior, and communication systems. Deploying this framework in even a few buildings would reveal challenges that simulations smooth over: sensor failures, communication dropouts, buildings that don't conform to thermal models, facility managers who override automated systems. The path from this paper to widespread deployment requires pilot projects that stress-test the approach under realistic conditions.

The human factors dimensions deserve deeper investigation. The current framework assumes the facility manager determines a single participation level and the consensus protocol implements it. But real human decision-making is more complex: managers might prefer gradual responses over sudden curtailments, might have different thresholds for different occupants, might be influenced by economic incentives in ways that deviate from pure optimization. Incorporating behavioral models—perhaps drawing on prospect theory or other frameworks from behavioral economics—could make the human-in-the-loop more realistic and the resulting system more predictable.

The observer design assumes that the unknown human input changes slowly relative to the system's dynamics. This assumption is reasonable for building thermal control—facility managers don't flip their decisions every minute—but it would be violated in faster systems. Extending the framework to human-in-the-loop control of more dynamic systems (traffic networks, for instance, or manufacturing processes) would require observers that can track rapidly changing human inputs.

The scalability of the approach remains an open question. The paper demonstrates coordination among five buildings, but real demand response aggregations might involve hundreds or thousands of buildings. The consensus protocol's convergence rate depends on the communication graph structure; sparse graphs converge more slowly than dense ones. Finding communication topologies that balance infrastructure cost against convergence speed for large-scale deployments would be valuable.

Cybersecurity considerations are largely absent from the current work. A malicious actor who could spoof communication messages might be able to destabilize the consensus process or manipulate buildings into inappropriate curtailment. The framework needs security extensions—authentication protocols, Byzantine fault tolerance, intrusion detection—that protect against adversarial interference. This is especially important because the demand response context creates economic incentives: building operators might want tocheat (cheat) the system to receive participation payments without actually curtailing consumption.

The regulatory landscape would need to evolve for widespread deployment. Current demand response programs treat buildings as passive participants who receive signals and choose to respond or not. The framework proposed here creates a more integrated relationship where building responses are algorithmically coordinated through human judgment. Who bears responsibility when coordinated curtailment leads to uncomfortable occupants? How should participation payments be allocated across buildings that achieved consensus through mutual influence rather than individual decisions? These questions aren't mathematical but regulatory, and their answers will shape whether promising technical frameworks translate into real-world impact.

Finally, the framework raises questions about the appropriate role of algorithmic coordination in human decision-making. The paper positions the facility manager as a leader whose judgment guides the multi-agent system. But one could imagine extending this toward more autonomous building coordination with humans merely providing high-level constraints (acceptable comfort ranges, maximum curtailment fractions) that algorithms then navigate. The boundary between human guidance and algorithmic execution is configurable—and where to set that boundary involves value judgments that engineering alone cannot resolve.

The mathematics in this paper show that coordinated building control with human judgment is possible. The path from possibility to widespread practice runs through pilot deployments, regulatory evolution, cybersecurity hardening, and continued dialogue about what we want our increasingly automated infrastructure to look like. That's not a challenge that slides gracefully into equations. But the equations in this paper provide a foundation for beginning that work.

What emerges is a vision of buildings that cooperate without centralized control, adapt to grid needs without sacrificing occupant comfort, and incorporate human judgment without requiring humans to micromanage every decision. The grid of the future might not look like a machine with precise algorithmic control. It might look more like a well-functioning team—buildings subtly adjusting to each other, human judgment guiding where algorithms cannot, consensus emerging from local conversations rather than central commands. That's not just a technical achievement. It's a different way of thinking about how complex infrastructure systems can work.