Smarter Grids: AI-Optimized "Virtual Inertia" Could Prevent Blackouts in a Renewable-Powered World

Somewhere inside every coal or gas power plant is a secret the energy transition is quietly discarding. Massive steel rotors — turbines spinning at 3,000 revolutions per minute — carry an enormous amount of kinetic energy. When the grid hiccups, when a transmission line trips or a factory suddenly draws a surge of power, those rotors don't stop. They keep spinning, resisting the change, donating their momentum to the grid and buying precious seconds for control systems to respond. This physical property is called rotational inertia, and for a century it has been the invisible shock absorber of civilization's most critical infrastructure.

Solar panels and wind turbines don't spin in sync with the grid. They connect through power converters — sophisticated electronic devices that translate their raw electrical output into grid-compatible power. Converters are fast, efficient, and endlessly programmable. But they contribute zero rotational inertia. As the world accelerates the retirement of fossil-fuel generators, entire power systems are quietly losing this fundamental stabilizing property — and the consequences, in worst-case scenarios, mean blackouts.

A new algorithm proposed by Jovan Krajacic, Maitraya Avadhut Desai, Ognjen Stanojev, and Gabriela Hug of ETH Zürich offers a systematic, dynamic answer to one of the most technically vexing questions in the clean energy transition: once you've decided to add virtual inertia back into the grid through software, how much do you put where? (Krajacic et al., 2025)

The Science

The concept of virtual inertia has been around for about a decade. The core idea is elegantly simple: if a power converter is fast enough, you can program it to behave like a spinning rotor. When grid frequency starts to drop — a sign that generation and demand are out of balance — the converter can release a burst of power that mimics what a real turbine would do. The software pretends to have mass. Similarly, virtual damping (also called governor damping, after the speed governors on real turbines) can be emulated: the converter gradually adjusts its output to suppress oscillations as the system finds a new equilibrium.

Two well-established control schemes already implement this on real hardware. Virtual Synchronous Machine (VSM) control makes a converter behave almost identically to a synchronous generator, complete with emulated inertia and damping coefficients. Droop control is simpler — it adjusts power output proportionally to frequency deviation, effectively providing damping without the full inertia emulation.

The harder question is not whether to deploy these schemes, but how. A grid operator upgrading dozens of wind farms and solar plants across a large interconnected system faces a combinatorial challenge: how much virtual inertia to assign to each node, how much virtual damping, and how to balance the cost of doing so against the stability gains. Too little virtual inertia in the wrong place, and a disturbance can cascade. Too much everywhere, and you're wasting expensive converter capacity that could otherwise be used to move more renewable energy.

Previous work addressed this using the ${\mathcal{H}}_2$ system norm — a mathematical tool from control theory that measures how much a system amplifies disturbances in an average-case sense. It's elegant and analytically tractable, but it assumes disturbances are spread uniformly across the system, which is rarely true in practice. Real grids get hit by real events: a generator trips in a specific location, a storm takes down a specific line. The ${\mathcal{H}}_2$ approach can't natively account for that geography.

Krajacic et al. take a different route. They formulate the allocation problem as a dynamic optimization, meaning the algorithm tracks how the system evolves over time after a disturbance — not just its average statistical behavior — and finds the inertia and damping values that minimize a cost function balancing three objectives: keeping frequency deviations small (stability), using as little converter capacity as possible (cost-efficiency), and performing well even under worst-case disturbances (resilience). Crucially, their framework explicitly incorporates where and how hard disturbances hit.

The approach uses gradient-based optimization applied directly to the time-domain trajectory of the system following a disturbance. At each step, the algorithm computes how sensitive the cost function is to each virtual inertia and damping parameter, then adjusts them in the direction that reduces cost. This is, in spirit, the same technique used to train neural networks — except here the "model" is a set of power system differential equations, not a deep learning architecture.

What They Found

The researchers validated their algorithm on a three-area power system — a standard test case in power systems research, representing three interconnected regional grids linked by transmission lines, with loads and generators distributed across them. Each area can host converter-interfaced resources capable of providing virtual inertia and damping. The setup allows direct comparison with the ${\mathcal{H}}_2$ benchmark.

The dynamic optimization algorithm consistently produced allocations that better contained frequency deviations following disturbances compared to the \mathcal{H}_2$-based method (Krajacic et al., 2025). The advantage was most pronounced when disturbances were geographically concentrated — exactly the scenario the $\mathcal{H}_2 norm handles poorly by design. When a large generation loss was simulated in a specific area, the new algorithm's allocation, having been informed by that disturbance's location, placed more virtual inertia nearby to absorb the initial shock.

Virtual Inertia & Damping Allocation Objectives

Label	Value
Stability (freq. deviation)	1
Cost-efficiency (capacity use)	1
Resilience (worst-case disturbance)	1

The algorithm jointly optimizes all three objectives. Exact numerical weights are determined by the operator's priorities. Source: Krajacic et al., 2025.

The three-objective cost function — stability, cost, and resilience — proved genuinely multi-dimensional in its trade-offs. Allocations that minimized frequency deviation tended to require more total converter capacity. Resilience-focused allocations sometimes differed meaningfully from stability-focused ones, especially for asymmetric grid topologies where some nodes are far more vulnerable to cascading effects. The algorithm navigated these trade-offs systematically rather than requiring the operator to manually tune a single penalty weight, a practical advantage over many prior methods (Krajacic et al., 2025).

Algorithm Performance Profile: Dynamic Optimization vs. H₂ Norm

Label	Value
Frequency Stability	9
Disturbance Location Sensitivity	9
Cost Awareness	8
Resilience to Worst Case	9
Computational Speed	5

Ratings are qualitative, derived from paper narrative. The dynamic approach excels at disturbance-aware and resilience metrics; the H₂ norm is faster to compute. Source: Krajacic et al., 2025.

One technically significant result concerns the interaction between virtual inertia and virtual damping. Naively, more of both always seems better, but the researchers found that the relationship is nonlinear and location-dependent. In a well-connected part of the grid, high virtual inertia without corresponding damping can actually increase oscillations — the system "rocks" around its equilibrium longer before settling. The optimal allocations from the dynamic algorithm naturally balanced the two, while the ${\mathcal{H}}_2$ approach, optimizing a single scalar norm, could sometimes miss this balance.

The computational cost of the approach is worth noting. Gradient-based optimization over time-domain simulations is more expensive than solving the algebraic problem underlying the ${\mathcal{H}}_2$ norm. The authors acknowledge this and position the algorithm as an offline planning tool — run it once, get the optimal allocation map, then deploy that configuration in the real grid. As computing resources continue to improve and power systems software matures, the gap in computational cost is expected to shrink.

Why This Changes Things

To appreciate why this matters, it helps to understand the scale of the problem the energy transition is creating. In 2000, virtually all large-scale generation worldwide came from synchronous machines — coal, gas, hydro, nuclear — all contributing inertia. By 2024, solar and wind together accounted for roughly a third of new electricity generation capacity globally, and that share is accelerating. Grids in Ireland, South Australia, Texas, and Denmark have already experienced low-inertia events that stressed system stability and, in some cases, contributed to frequency excursions that triggered emergency responses.

The rate of change of frequency (RoCoF) — how fast frequency drops in the seconds after a major generation loss — has become a key metric for grid operators. Traditional grids could tolerate RoCoF events of perhaps 0.5 Hz per second before protection systems started tripping equipment. As inertia disappears, the same disturbance produces a steeper drop, faster. Grid codes in Europe and Australia have had to be rewritten to account for this new reality.

Virtual inertia, deployed at scale through converter control, is one of the most promising tools to address this. But deploying it without optimization is like adding shock absorbers to a car without knowing which wheels bear the most load. The Krajacic et al. framework gives grid planners a principled, disturbance-aware method to answer the allocation question — and their comparison to the ${\mathcal{H}}_2$ benchmark gives practitioners a clear signal that the dynamic approach is worth the additional computational investment.

Key Dimensions of the Virtual Inertia Allocation Problem

Label	Value
Frequency Nadir (stability)	1
RoCoF (rate of change)	1
Converter Capacity Cost	1
Disturbance Location Impact	1
Oscillation Damping	1

Each dimension represents a distinct challenge in virtual inertia planning. The dynamic optimization algorithm addresses all five. Source: Krajacic et al., 2025.

There's also a cost dimension that deserves emphasis. Virtual inertia is not free. When a converter provides virtual inertia, it needs to hold a reserve of power — energy stored in a battery or supercapacitor, or headroom in a wind turbine's pitch control — that can be dispatched in milliseconds. That capacity costs money and competes with the converter's primary job of delivering renewable energy to the grid. A poorly designed allocation scheme that spreads virtual inertia uniformly, or places it far from where disturbances tend to occur, is literally expensive in a way that will show up in electricity bills. The cost-efficiency objective in the new algorithm directly addresses this economic reality.

The resilience dimension is perhaps the most forward-looking aspect of the work. Power systems researchers increasingly think about grids not just in terms of their nominal performance but their behavior under stress — cyberattacks, extreme weather events, unexpected equipment failures. The worst-case disturbance framing in Krajacic et al.'s cost function is a step toward robust virtual inertia allocation, one that doesn't just optimize for an expected scenario but hedges against the tail risks that matter most.

This connects to a broader trend in power systems engineering: the move from static, rule-of-thumb grid planning toward data-driven, mathematically rigorous optimization. The same intellectual shift that has transformed how transmission lines are expanded and how electricity markets are cleared is now reaching the question of grid stability services. Virtual inertia allocation is a new frontier in that transition.

What's Next

The three-area test system is a proof of concept, and the authors are explicit about the path forward. Real transmission networks have hundreds of nodes, thousands of possible disturbance scenarios, and regulatory constraints that vary by jurisdiction. Scaling the dynamic optimization algorithm to these systems will require advances in computational efficiency — potentially through smarter sampling of the disturbance space, parallel simulation, or machine-learning surrogates that approximate time-domain responses without running full simulations.

There are also open questions about time-varying allocation. The current framework finds an optimal static allocation — the best fixed values of virtual inertia and damping for a given grid configuration. But real systems change: generation mixes shift hour by hour as the sun rises and wind picks up, loads fluctuate, and equipment goes in and out of service. Future extensions could make the allocation itself dynamic, adapting in near-real-time to the current state of the grid. The mathematical tools are available; the challenge is computational speed and reliable sensing.

The interaction with grid-forming converter technology is another frontier. The converters in this paper use established control schemes (VSM and droop), but a newer generation of grid-forming converters can take a more active role in defining voltage and frequency on the grid, rather than just following it. These devices have even richer capabilities for providing stability services, and optimal allocation algorithms like the one in this paper will be essential to harnessing them.

Finally, there is the question of market design. Virtual inertia has no commodity price in most electricity markets today. Grid operators often mandate it through technical requirements, but there is no transparent market where providers of virtual inertia are paid for their contribution to stability. As the value of these services becomes clearer — in part through work like this, which quantifies the cost-stability trade-off precisely — the pressure to create proper market mechanisms will grow. Knowing the right amount of virtual inertia to have, and where, is a prerequisite to pricing it correctly.

The power grid is one of humanity's most complex engineered systems, and it is undergoing the most rapid transformation in its century-long history. The loss of rotational inertia is one of the subtler but more consequential side effects of that transformation — invisible to consumers, deeply worrying to engineers. Algorithms that can dynamically optimize the software-defined stability services replacing that inertia aren't a nice-to-have. They're part of the critical infrastructure of the clean energy future. Krajacic et al. have taken a meaningful step toward making that infrastructure smarter, more cost-aware, and more resilient to the disturbances — expected and unexpected — that every real grid must endure.