Robust Optimization Coordinates Smart Building Clusters to Balance Power Grid Demands

Power systems undergoing rapid decarbonization increasingly depend on responsive demand-side resources to balance the inherent variability of wind and solar generation. Modern smart buildings contain diverse electrical assets, including heating and air conditioning equipment, smart household appliances, electric vehicle charging systems, and stationary energy storage. In a simulation study published in Frontiers in Energy Research, power engineering researchers at the Guangzhou Power Supply Bureau, the Power Dispatching Control Center of China Southern Power Grid, and the Southern Power Grid New Power System Research Institute demonstrated a two-stage robust formulation method designed to aggregate heterogeneous smart buildings into coordinated clusters. By unifying disparate operational parameters into standardized response characteristic matrices and embedding historical scenario demands directly into resource aggregation, the computational framework establishes executable cluster regulation curves that withstand operational uncertainties. The modeling results indicate that accounting for building-level operational boundaries and response-capability variations allows regional electricity networks to tap flexible demand potential more reliably while decreasing total operating costs and lowering associated carbon emissions.

The integration of expanding renewable energy resources into electrical grids has fundamentally disrupted traditional power dispatch paradigms that historically relied on generation following fluctuating customer loads. Because solar photovoltaics and wind turbines produce power governed by weather conditions, power system dispatchers must cultivate bidirectional source-load interactions where consumers modulate their demand profile dynamically. Smart buildings are well positioned to supply this operational flexibility because they concentrate substantial volumes of adjustable electrical loads alongside dedicated control terminals. These internal control units continuously collect performance telemetry, track device statuses, analyze electricity consumption trends, and execute automated load adjustments. However, conventional demand-side participation models often treat adjustable equipment as simple, generic load blocks or attempt to regulate single components, such as air conditioning thermostats, in complete isolation. Such approaches neglect the intricate internal energy consumption patterns of whole buildings, ignore critical hardware operating constraints, and overlook the severe information loss that occurs when aggregating thousands of decentralized facilities into regional transmission and distribution markets.

To overcome the operational limitations of single-building participation, power aggregators combine multiple facilities into collective response clusters that interact with regional power trading centers. Prior aggregation methodologies generally followed two distinct strategies: building-level feasible region accumulation or exterior region characterization. The accumulation method sequentially aggregates the operational constraints of individual devices from the bottom up, while exterior region approximations attempt to envelope collective resources as virtual generating units or virtual battery storage. The study authors observed that existing cluster aggregation models primarily calculate broad, static feasibility zones spanning entire twenty-four-hour periods. These broad envelopes fail to incorporate the specific temporal windows, directional adjustments, and response weights required by specialized operational events, such as short-term renewable smoothing or afternoon peak shaving. Consequently, dispatch curves derived from generic capacity envelopes frequently generate significant execution gaps, demand excessive ramp rates from physical equipment, or leave substantial localized adjustment potential unexploited during critical grid stress windows.

The research team resolved this modeling challenge by developing a five-parameter standardized response matrix that characterizes each responsive asset within a smart building without exposing proprietary consumer data. The unified matrix tracks five distinct physical and operational metrics: response capacity, maximum response duration, minimum response interval, maximum response count per day, and response speed for upward and downward regulation. By converting diverse equipment into this five-dimensional framework, local building control terminals can systematically evaluate their operating boundaries based on user schedules and historical consumption patterns. For instance, when evaluating electric vehicles, the control terminal models maximum charging power as response capacity, battery state-of-charge limits as response duration, power adjustment rates as response speed, and charging session limits as maximum daily response counts. Transmitting standardized matrix rows rather than detailed hardware schematics provides grid aggregators with exact technical capabilities while preserving consumer privacy and reducing numerical complexity across municipal utility models.

To connect physical building capabilities with real-world power grid dispatch needs, the researchers established a scenario-oriented aggregation mechanism that extracts demand profiles across six operational scenarios. These targeted operational contexts encompass peak shaving, valley filling, simultaneous peak shaving and valley filling, renewable energy consumption, short-term renewable wave suppression, and emergency reserve standby. Rather than relying on synthetic benchmarks, the researchers extracted twenty-four-hour historical demand curves from actual grid operating records spanning the preceding year. To manage dimensionality without distorting physical load shapes, the authors used principal component analysis on reduced eight-dimensional feature vectors to identify representative centroid curves within each scenario category. The original historical curve closest to the Euclidean centroid was selected to supply authentic timing boundaries, directional requirements, and proportional response weights. Embedding these scenario-specific weights and time windows directly into the aggregation objective function ensures that the resulting cluster capability boundary captures maximum flexible potential exactly when the distribution network requires support.

The formulation of the final dispatch curves utilizes a two-stage robust optimization structure solved via the column-and-constraint generation algorithm. In this mathematical architecture, the first-stage master problem functions as a here-and-now decision module that establishes the overarching cluster regulation curves, overall response states, and device start-up sequences before real-time equipment availability is revealed. The second-stage subproblem represents wait-and-see operational adjustments executed after actual physical response-capability reduction rates materialize. Because physical machinery may underperform due to localized constraints or consumer overrides, the model formulates an uncertainty set governed by an average reduction budget parameter alongside individual entity-period reduction limits. When the second stage identifies the worst-case realization of response degradation, the subproblem generates cutting planes that feed back into the master problem. This iterative process refines the cluster curve until the relative convergence gap meets numerical tolerances, guaranteeing that decomposed building schedules remain mathematically executable even under adverse response conditions.

Environmental considerations are integrated directly into the second-stage optimization through time-varying marginal carbon emission factors alongside economic deviation penalties and device wear costs. In the second-stage objective function, individual response entities receive individualized unit response costs derived from their underlying operational characteristics, including daily cycling frequency, response duration limits, and user preference factors. Simultaneously, time-varying grid carbon emission factors assign economic values to marginal greenhouse gas emissions across each individual fifteen-minute scheduling interval. When the optimization algorithm minimizes total second-stage expenditures, intervals with high marginal carbon intensity naturally incentivize greater load reductions, directing building flexibility toward periods dominated by fossil-fuel generation. Penalty coefficients assigned to regulation curve decomposition deviations and actual grid demand mismatches are weighted heavily to guarantee that system stability and contract tracking take precedence over minor economic adjustments, coordinating economic dispatch and carbon abatement across identical operational timeframes.

The operational validity of the proposed formulation framework was tested through computational simulations implemented in Gurobi using field parameters from an operational smart business park in Guangdong Province, China. The baseline case analyzed six smart building clusters, with each cluster comprising ten smart commercial buildings containing five distinct response entities, yielding three hundred individual responsive load assets modeled across ninety-six daily fifteen-minute dispatch intervals. In comparative evaluations against conventional aggregation methods, the scenario-embedded aggregation model substantially reduced performance gaps between requested and delivered energy adjustments. When participating in scheduled grid regulation events, the traditional inside-out expansion method exhibited a response gap of 2492.33 kilowatt-hours, and a baseline method considering only scenario time windows produced an operational gap of 387.73 kilowatt-hours. In contrast, the authors’ proposed scenario-oriented aggregation method achieved a reduced response gap of 353.49 kilowatt-hours, representing an 8.83 percent reduction in unmet demand compared to the time-window benchmark.

Economic and carbon assessments demonstrated that incorporating detailed device response characteristics and marginal emission factors into curve decomposition yields substantial operational advantages over conventional baselines. Comparing the full proposed model against alternative formulations revealed clear performance shifts across cost and sustainability metrics. A conventional decomposition model that ignored internal response characteristics yielded a daily operating cost of $9751.31 and a carbon reduction of 4911.90 kilograms. The proposed formulation without carbon modeling reduced operational expenses to $7136.47 while achieving 5109.25 kilograms of carbon abatement. Activating the time-varying marginal carbon emission factor in the full model produced a total operating cost of $7152.00 and increased total carbon reduction to 5363.34 kilograms. Incorporating marginal carbon pricing thus expanded carbon mitigation by approximately 5.0 percent while incurring a negligible 0.2 percent cost increase. Extensive parameter sensitivity testing over twenty percent perturbation ranges confirmed that the proposed framework maintained an average 26.98 percent cost reduction compared to conventional methods across varying cost weights and consumer preferences.

Scalability and computational efficiency evaluations confirmed that the decomposition framework can support real-world power dispatch timelines across larger grid networks. Because individual cluster aggregation and decomposition subproblems possess decoupled mathematical structures, the workload can be distributed in parallel across multiple computing cores. In numerical scalability benchmarks expanding from sixty buildings containing 300 response entities up to 2000 buildings containing 10000 entities, total computation times scaled predictably. The baseline sixty-building system completed end-to-end solving in 25.90 seconds, while systems containing 200 buildings, 500 buildings, and 1000 buildings required 117.18 seconds, 234.45 seconds, and 734.71 seconds, respectively. All problem instances up to one thousand buildings solved in under fifteen minutes, meeting standard execution deadlines for day-ahead and intra-day grid dispatch cycles. The authors emphasized that while decomposed curves proved strictly feasible against mathematical constraint boundaries, real-world deployment requires subsequent experimental validation through hardware-in-the-loop testing and live telemetry to address physical communication latencies, sensor forecasting inaccuracies, and human behavioral feedback.

Subject of Research: Two-stage robust optimization framework for smart building cluster regulation curves

Article Title: Two-stage robust optimization for formulating regulation curves of smart building clusters considering multiple regulation characteristics

Article References: Xu, S., Lü, Y., Chen, Y., Lei, X., & He, J. (2026). Two-stage robust optimization for formulating regulation curves of smart building clusters considering multiple regulation characteristics. Frontiers in Energy Research, 14, Article 1966536. https://doi.org/10.3389/fenrg.2026.1966536

Image Credits: AI Generated

DOI: 10.3389/fenrg.2026.1966536

Keywords: smart buildings, demand response, robust optimization, power grid flexibility, renewable integration, carbon emission reduction, energy management, load aggregation, Two-stage, robust, optimization, formulating