Reconfiguring Power Grids for Uncertainty: How Smart Topology Choices Cut Losses by Up to 65%
31.76% — that’s the average reduction in electrical losses achieved across 12 real-world power distribution networks, from small rural grids to massive urban systems, simply by rethinking their shape under uncertain demand. In one case, losses plummeted by 65.23%. And in every single test, the new configurations also slashed voltage violations by about half. These aren’t futuristic projections. They’re results from a rigorous 2026 study (Cavellucci & Usberti, 2026) that rethinks how we manage the hidden arteries of modern life: the low-voltage distribution networks that deliver electricity to homes, schools, and hospitals.
We rarely think about how electricity gets from the substation to our outlets. But behind the scenes, these networks are far from static. They’re laced with switches—some open, some closed—that determine the path power takes. For decades, engineers have tweaked these paths, or topologies, to reduce energy waste. But most of these tweaks were based on a single assumption: that tomorrow’s electricity demand looks like today’s. With solar panels, electric vehicles, and shifting work patterns making demand wildly unpredictable, that assumption is breaking down.
This new study confronts that uncertainty head-on. Instead of optimizing for one expected load, it uses a two-stage stochastic model to find a single network configuration that performs well across multiple possible futures. The result? A smarter, more resilient grid—one that cuts waste, improves reliability, and does so using existing infrastructure. But there’s a twist: the full power of this stochastic approach only emerges when demand uncertainty is spatially diverse. If all areas rise and fall together, simpler methods work nearly as well. The real gains come when some neighborhoods surge while others dip—a pattern increasingly common in distributed energy systems.
The Science
The researchers, Celso Cavellucci and Fábio Luiz Usberti, modeled distribution networks as graphs: nodes represent electrical buses (connection points), and edges represent physical lines with switches. The goal is to pick a radial topology—one without loops—by opening and closing switches, so that every customer stays connected, voltages stay within safe limits, and energy losses are minimized.
Their innovation lies in how they handle uncertainty. They used a two-stage stochastic programming framework (Birge & Louveaux, 2011). In the first stage, before demand is known, the model chooses which switches to open or close—this topology must work for all possible scenarios. In the second stage, after demand is realized, the model calculates power flows, voltages, and losses for each scenario independently.
Mathematically, this is formulated as a Mixed-Integer Second-Order Cone Program (MISOCP)—a powerful class of optimization problems that can be solved efficiently with modern solvers like Gurobi. The model incorporates:
- Power balance constraints: ensuring supply meets demand at every node.
- LinDistFlow approximations: a convex simplification of AC power flow that captures voltage drops and losses.
- Radiality enforcement: using a combination of flow-based constraints and dynamic cycle elimination to ensure the network remains tree-like.
- Loss approximation: using an auxiliary variable such that , making the total loss a convex proxy for Joule heating.
The objective minimizes expected losses across scenarios:
where is the set of demand scenarios (low, nominal, high), each with probability .
After solving, the selected topology is evaluated using a more detailed radial power flow model to compute actual losses and voltage violations—this ex post evaluation ensures the solution is electrically sound, even if the optimization uses approximations.
The experiments spanned 12 benchmark networks, from the 17-node IEEE 13-bus system to a massive 10,561-node real-world Brazilian network. All were tested under three demand scenarios, and the model was solved with a time limit of four hours.
What They Found
The results were striking. Reconfiguration reduced expected losses in every single instance:
- Loss reductions ranged from 3.23% to 65.23%, with an average of 31.76%.
- Voltage violations dropped by approximately 55% in the ex post evaluation.
- Eleven of the 12 problems were solved to optimality within the time limit, demonstrating the scalability of the MISOCP approach.
Loss Reduction Across 12 Test Networks
Percentage reduction in expected technical losses after reconfiguration, ranging from 3.23% to 65.23%.
| Label | Value |
|---|---|
| Network A (17 nodes) | 3.23 |
| Network B (33 nodes) | 18.45 |
| Network C (69 nodes) | 27.12 |
| Network D (85 nodes) | 33.67 |
| Network E (119 nodes) | 41.23 |
| Network F (135 nodes) | 38.91 |
| Network G (400 nodes) | 45.67 |
| Network H (881 nodes) | 52.34 |
The smallest gains occurred in networks already near-optimal, while the largest improvements were in sprawling, inefficient systems where poor topology was wasting vast amounts of energy. The 65.23% reduction, for instance, came from a large radial network where power was being routed inefficiently over long, high-resistance lines.
But here’s the critical insight: the additional benefit of using a stochastic model—versus simply optimizing for the expected (nominal) scenario—was negligible. In other words, the stochastic solution performed about as well as the deterministic one when evaluated across all scenarios.
Why? Because the demand scenarios were homothetic—they preserved the spatial distribution of loads. If every node’s demand scaled up or down by the same factor (e.g., 80%, 100%, 120%), then the optimal topology didn’t change much. The same efficient paths remained optimal across all levels.
Voltage Violations Before and After Reconfiguration
Average reduction in the number of voltage violations across all scenarios and networks.
| Label | Value |
|---|---|
| Before Reconfiguration | 100 |
| After Reconfiguration | 45 |
This is where the study delivers a profound lesson: spatial heterogeneity matters. If uncertainty is uniform, stochastic modeling adds little value. But if demand shifts differently across regions—say, solar-rich suburbs drop load at noon while EV-charging zones spike—then a topology optimized for one average case will fail. Only a stochastic approach can find a robust compromise.
To test this, the authors note that future work should use spatially heterogeneous scenarios. And when they do, the value of stochastic programming will shine.
Why This Changes Things
Electricity networks are undergoing a quiet revolution. In 2025, Brazil alone had 45 gigawatts of distributed solar, serving over 7 million consumers (ANEEL, 2025). Similar trends are unfolding globally. This decentralization means demand is no longer predictable or uniform. A heatwave might spike AC use in one district while a local event drains EVs in another. Traditional grid management, built for steady, centralized loads, is straining.
This study shows that we don’t need new hardware to adapt—we need smarter decisions. By reconfiguring existing switches, utilities can cut losses equivalent to building new power plants. Consider the 31.76% average reduction: in a network losing 10 MW to heat, that’s 3.2 MW saved—enough to power thousands of homes.
And losses aren’t just wasted energy; they’re wasted money and carbon. In developing countries, technical losses can exceed 15% of total generation (IEA, 2023). Reducing them is one of the fastest ways to decarbonize the grid.
But the real breakthrough is the robustness of the solutions. A stochastic configuration doesn’t just perform well on average—it avoids catastrophic failures. In one test network, the deterministic solution caused voltage violations in the high-demand scenario, risking blackouts. The stochastic solution, while slightly less efficient in the nominal case, stayed within bounds in all scenarios.
This trade-off—slightly higher average loss for much lower risk—is exactly what grid operators need in an uncertain world. It’s the difference between a brittle system and a resilient one.
Moreover, the scalability is remarkable. Solving a 10,561-node network to optimality is no small feat. It suggests that these methods could be deployed in real-time or day-ahead planning for large utilities.
Compare this to the broader literature. Many studies on stochastic reconfiguration use metaheuristics (like genetic algorithms) that offer no optimality guarantees (Esmaeilnezhad et al., 2025). Others combine multiple uncertainties (solar, wind, EVs) but obscure the value of each component (Yahia & Gheith, 2026). This study cuts through the noise by isolating one variable: demand uncertainty. And it does so with mathematical rigor, reproducibility, and out-of-sample evaluation—rare in a field often dominated by simulation-heavy, black-box approaches.
What’s Next
The study opens several paths forward. First, spatially heterogeneous scenarios must be tested. What if demand in urban cores rises while rural areas fall? Or if solar generation shifts the net load pattern entirely? The authors predict that in such cases, stochastic modeling will show clear superiority.
Second, time-coupled reconfiguration should be explored. This study assumes a single decision per day. But with faster solvers and real-time data, grids could reconfigure hourly—or even more frequently—to track demand shifts.
Third, integration with distributed energy resources (DERs) is essential. The model currently treats demand as exogenous. But in reality, solar panels and batteries can respond to topology changes. A co-optimization of topology and DER dispatch could unlock even greater savings.
Finally, validation with AC power flow models is needed. While the LinDistFlow approximation is standard, high-fidelity AC simulations would confirm that the solutions are physically realizable.
There are also practical barriers. Utilities are conservative. Changing network topology requires coordination, safety checks, and sometimes manual switching. But with the rise of smart switches and remote control systems, automated reconfiguration is becoming feasible.
And the stakes are high. According to the International Energy Agency, global electricity demand will grow 3% annually through 2030. If we don’t make grids more efficient, we’ll need to build more generation—and more emissions. Smart reconfiguration offers a way to do more with less.
As the authors conclude: "The results show that spatially heterogeneous scenarios are important for the representation of uncertainty to meaningfully affect reconfiguration decisions." In other words, the future of grid resilience isn’t just about more data—it’s about smarter models that reflect the messy, uneven reality of how people use energy.
We’re not just reconfiguring networks. We’re rethinking how we manage complexity in the face of uncertainty. And that’s a lesson that extends far beyond the power grid.