Thirty Hours in the Life of a Grid: How Clustering by Voltage, Not Power, Maps a Whole Year

In a renewable-dominated power grid, an engineer trying to figure out whether the network will hold up over the next year faces a brutal arithmetic problem. To run a proper voltage-security check, you'd ideally simulate every single hour of operation — 8,760 of them, each a full AC power-flow solution — and then repeat everything again for every possible line outage. That is a mountain of computation so steep that planners routinely give up and guess, picking a handful of "typical days" — peak summer, peak winter, light load — and hoping they capture what actually happens.

A new framework from researchers at the University of Manchester (Agon, Preece & Milanovic, 2026) attacks that problem from an unexpected angle. Instead of asking which hours look similar in their demand and generation, it asks which hours behave similarly in their voltage. That single reframing lets planners compress a full year down to just 30 representative operating points — a 99.66 percent reduction — while still reproducing the full year's voltage behavior with 98.3 percent accuracy in steady state and 93.4 percent after a contingency. In other words: thirty well-chosen hours, run through the simulator, tell you almost everything that all 8,760 hours would. It's as if a film critic could summarize a year of cinema by watching thirty frames — but only if those frames were chosen by what the images actually show, not by their box-office receipts.

The Science

The problem begins with a deep asymmetry that power engineers have lived with for decades. Voltage stability — the ability of a grid to keep voltages at every bus within acceptable bounds, typically between 0.95 and 1.05 per unit — is governed by a tangled, nonlinear machinery: fluctuating generation, load characteristics, network topology, and above all reactive power.

Reactive power is the quiet workhorse of voltage control. It doesn't deliver energy to your house; it holds the voltage field up, like the scaffolding around a building. When generators hit their reactive power limits, they stop being able to prop up voltage, and the whole profile can sag or surge. Voltage-control actions, control-mode switching of inverters, and variations in grid strength all feed into this in ways that cannot be read off from a power profile.

Here's the trap that most existing scenario-reduction methods fall into. They cluster operating points by their injections — how much power each bus is drawing or feeding in. The implicit assumption is that similar injections produce similar voltage behavior. Agon et al. show this assumption is quietly wrong. Two hours with nearly identical net power injections can produce very different voltage distributions, because reactive limits and controller states differ. Conversely, hours with completely different injections can end up with nearly identical voltages when network constraints dominate. Clustering in "injection space" therefore throws away exactly the information that matters most for voltage security.

The framework's move is to cluster in what the authors call the voltage-response space instead. Each of the year's hours is first run through an AC power-flow solution, producing a full voltage profile. From that profile, seven system-wide risk indices are computed: the minimum and maximum voltages $V_{\min }$ and $V_{\max }$, the dispersion of voltages across buses (the standard deviation), the root-mean-square severity of any violations of the 0.95–1.05 p.u. band (RMS_VVS), the proportion of buses violating limits $P_{\text{bad}}$, and two reactive-power reserve margins quantifying how much headroom each generator has left before hitting its upper and lower limits $Q^{\text{res\_Un}}$ and $Q^{\text{res\_Ov}}$.

These seven numbers, bundled together, form a "voltage-risk vector" that characterizes the system's nonlinear response at any given hour. Because several of the indices are correlated — high violation severity tends to go with more violating buses — the authors run Principal Component Analysis to decorrelate the features into orthogonal components, keeping enough components to preserve at least 99 percent of the variance. Finally, Hierarchical Agglomerative Clustering with Ward linkage groups the hours into voltage regimes, and from each cluster a single representative operating point is picked: the hour whose PCA coordinates sit closest to its cluster's centroid.

The choice of Ward linkage matters. Unlike simpler linkage criteria, Ward minimizes the increase in within-cluster variance at each merge, producing compact, statistically coherent clusters. The full hierarchy is built bottom-up — every hour starts as its own cluster and merges progressively — and the analyst can cut the dendrogram at any height to get any number of clusters, without re-running the algorithm. The result is a nesting of voltage regimes that can be examined at different resolutions.

What They Found

To test the framework, the authors built a year-long dataset on the IEEE Voltage Test System under very high renewable penetration, using a stochastic chronological simulation. Load demand follows realistic seasonal and diurnal profiles scaled by Gaussian noise; PV generation combines a solar-irradiance profile with Beta-distributed cloud-cover variability; wind is sampled from a Weibull distribution through turbine power curves. A Student-t copula models the spatial correlation between geographically distributed renewable plants, preserving joint extreme events, and an AR(1) autoregressive process captures temporal persistence. For each hour, a DC unit commitment solves for economically feasible dispatch, followed by an AC power flow — with generator dispatch, commitment status, controller setpoints, and reactive outputs carried forward between time steps so the system evolves realistically.

The headline numbers come from the compression. The full year's 8,760 operating points collapse to just 30 representatives — a 99.66 percent reduction. Yet those 30 hours reconstruct the full year's steady-state voltage behavior with 98.3 percent accuracy, and post-contingency response with 93.4 percent accuracy.

Reconstruction fidelity vs. data compression

Label	Value
Steady-state reconstruction	98.3
Post-contingency reconstruction	93.4

Voltage-response HAC preserves 98.3% of steady-state and 93.4% of post-contingency voltage behavior while compressing the full year by 99.66% (Agon et al., 2026).

To reach those numbers, the authors developed a genuinely comprehensive evaluation battery — not just a single accuracy metric, but a layered suite that checks different ways a reduced dataset can fail. The first layer, moment and quantile preservation, compares the mean, standard deviation, and 95 percent Value-at-Risk of each risk index between the full dataset and the weighted ROP distribution. The second layer, distributional geometry, uses the Kolmogorov–Smirnov distance and the first-order Wasserstein distance to check that the whole shape of the distribution — not just a few summary statistics — is preserved. The third layer probes the safety-critical tail: whether the magnitude of extreme violations (measured by Conditional Value-at-Risk) is preserved, and whether extreme hours are even represented at all (measured by a Tail Coverage Ratio). The overall reconstruction score weights these three components equally.

The most important result isn't just that the method works — it's which failures it avoids. When the authors compare against injection-space clustering and against the industry-standard heuristic sampling of "typical days," the voltage-response approach wins on the metrics that matter most: the shape of the distribution and, critically, the extreme tail.

Voltage-response clustering beats the alternatives

Label	Value
Voltage-response HAC	98.3
Injection-space clustering	92.1
Heuristic sampling	85.4

Voltage-response clustering outperforms injection-space clustering and heuristic sampling in reproducing full-year steady-state voltage behavior (Agon et al., 2026).

The extreme tail is where the difference becomes stark. Heuristic sampling, by its very design, tends to select hours near the center of the operating space — the "typical" days — leaving the rare, severe conditions unrepresented. Injection-space clustering does better on the average but still struggles to capture the worst voltage events, because net-power similarity doesn't track voltage severity. The voltage-response framework, by contrast, explicitly weighs tail risk into its evaluation, ensuring that the most safety-critical hours are represented by a representative that is itself extreme.

Why This Changes Things

This matters beyond a methodological footnote. Power systems are being pushed into a regime where the old heuristics genuinely break down. At very high renewable penetration, inverter-based resources displace synchronous generators, reducing system strength and replacing electromechanical inertia with power-electronic controls that introduce additional nonlinearities. Under these conditions, the operating space expands from a handful of natural patterns into many diverse modes, and voltage regimes no longer correlate strongly with demand or renewable profiles.

Consider what that means for planning practice. A utility engineer assessing voltage security over a year must weigh two costs: the computational cost of simulating everything, and the risk cost of simulating too little. Full-year Monte Carlo and sequential simulations are accurate but computationally punishing. The traditional answer — pick a few typical days by judgment — is cheap but increasingly unreliable, precisely because there is no longer a "typical" day in a very-high-RES system. A summer day with bright sun and high demand and a summer day with heavy overcast and low demand are no longer minor perturbations of each other; they are different voltage universes.

The voltage-response framework threads the needle. By clustering on actual behavior rather than on inputs, it produces a small set of representatives that is computationally cheap and statistically faithful — including in the tail where the danger lives. The 30 ROPs aren't thirty "average" days; they're thirty points that together span the full geometry of voltage regimes, with the extreme ones explicitly covered.

The comparison to injection-space clustering is the real intellectual contribution. It's a concrete demonstration that two established ideas — "cluster operating points" and "voltage security assessment" — have been hiding a silent mismatch. In this regime, an "OP" defined by its injections is simply not the same object as an "OP" defined by its voltage response. The paper makes that invisible mismatch visible and measurable.

There's a broader lesson here for anyone who builds models of complex systems. When you reduce a large space to a representative sample, the right reduction depends on which downstream quantity you care about. Clustering on the causal inputs assumes those inputs map smoothly onto the output of interest. In nonlinear systems with hard constraints and control actions — not just power grids, but climate, ecology, epidemiology — that mapping can be badly distorted. The paper is a working demonstration that clustering on the response, the thing you actually care about, can be dramatically more efficient and faithful.

What's Next

Caveats first. The validation runs on a single test system — the IEEE Voltage Test System — under a specific, very high instantaneous RES penetration level, with synthetic data. Real grids are messier: more heterogeneous, more numerous in their control actions, and subject to data that arrives with noise and gaps. The framework's use of a synthetic dataset gives it precise control over penetration levels, but it also means the results still need confirmation against real operational data on a real system.

The single-contingency evaluation is another limitation worth noting. The paper applies the most critical line contingency and compares post-contingency response, but real security assessment must consider many contingencies. The post-contingency accuracy of 93.4 percent is slightly below the steady-state figure, hinting that dynamic response is the harder preserve — a direction the authors themselves flag for continued work.

What the 30 ROPs preserve — and where they strain

Label	Value
Mean preservation	99.2
Std. dev. preservation	98.7
VaR (95%) preservation	97.9
Distribution shape (KS+W1)	97.4
Tail risk (CVaR+TCR)	96.5

The ROP evaluation battery checks moments, distribution shape, and tail risk separately; tail-risk coverage is the hardest to preserve (Agon et al., 2026).

None of that dims the forward-looking significance. The framework is a genuine, demonstrated alternative that turns scenario reduction from an art (pick typical days) into a principled, response-aware science. The contributions sketch an immediate agenda: applying the method to multi-contingency security assessment, exploring how the optimal number of clusters scales with system size and penetration level, and testing whether the approach extends beyond voltage security to other stability phenomena — frequency, rotor-angle, small-signal — each of which defines its own "response space."

The deeper significance is structural. As grids decarbonize, the very notion of "planning for average conditions" becomes untenable. The operating space of a renewable-dominated grid is a frontier territory, and the tools we use to navigate it must be based on the territory itself, not on the maps of the old regime. This framework shows, quantitatively, that even under extreme penetration you can compress a year to thirty hours and still see almost everything — provided you choose those hours by what the system actually does. For a power system in transition, that's not just a computational convenience. It's the difference between navigating by the stars and navigating by a map drawn before the landscape moved.