The Invisible Tax on Battery Storage: How Uncertainty Reshapes the Clean Energy Grid

The Invisible Tax on Uncertainty

In the middle of a California summer, when wildfires push power lines into tinder-dry air and air conditioners strain against record heat, the grid operator faces a problem that no amount of engineering can fully solve: nobody knows exactly how much electricity people will need tomorrow. Or the day after. This uncertainty—the gap between what we can predict and what actually happens—is the subject of a new paper from MIT that reaches a counterintuitive conclusion about one of the most celebrated technologies in the clean energy transition.

Energy storage, specifically grid-scale batteries, is supposed to save us. Charge when power is cheap and abundant, discharge when it's scarce and expensive. Arbitrage the peaks, smooth the valleys. But the researchers—Daniel Shen, Marija Ilic, and John Parsons—have found that uncertainty itself imposes a hidden cost on storage, one that changes how these systems operate and, more surprisingly, what we end up building in response. Under conditions of imperfect foresight—the reality of real grids—batteries hold back more energy than they would if we could see the future perfectly. They become conservative. And that conservatism ripples upward through investment decisions, reshaping the entire capacity mix in ways that current market designs struggle to correct.

The numbers are striking. At equilibrium under perfect foresight, their model produces a storage capacity of 7.2 gigawatts. Under the demand uncertainty that characterizes real electricity systems, that drops to just 4.0 gigawatts—a 44 percent reduction driven not by technology costs or physics, but by the simple fact that we cannot see tomorrow. Meanwhile, peaker generation—the expensive gas turbines that run only during the highest-demand hours—grows from 10.7 gigawatts to 12.9 gigawatts to fill the gap that batteries no longer cover. The baseload machines that should run constantly shrink from 11.1 gigawatts to 10.6 gigawatts.

"Demand uncertainty biases the equilibrium capacity mix away from storage and towards a greater total amount of conventional generation," the researchers write. This is the invisible tax that uncertainty levies on the clean energy transition—not through policy opposition or NIMBY lawsuits, but through the mathematics of how rational actors hedge against an unknowable future.

The Science

To understand how uncertainty reshapes storage operation, you first need to understand what storage operators are actually doing when they decide to charge or discharge. The MIT team models this as a Markov decision process, or MDP—a mathematical framework for making optimal decisions under uncertainty when the consequences of each choice depend on the current state of the system and the probabilistic dynamics of what comes next.

Think of it this way: a battery operator doesn't just look at the current hour and decide whether to charge or discharge based on that moment's prices. They have to think about what the next several hours might bring. If demand is currently low but there's a non-trivial chance that a heat wave is building tomorrow, the rational choice might be to hold onto some charge rather than selling it all back to the grid at today's low prices. That's the precautionary logic. You're trading off certain profits today against uncertain survival tomorrow.

The researchers formulate this as an average-cost MDP, meaning they're concerned not with maximizing profits in any single period, but with minimizing the long-run expected cost averaged over infinite time. This is the appropriate framework for thinking about grid planning: you're not trying to win big in one gamble, you're trying to run the system reliably and economically forever.

To model demand itself, they use what's called an Ornstein-Uhlenbeck process—a mean-reverting stochastic process that captures something essential about how electricity demand actually behaves. Unlike stock prices, which can wander arbitrarily far from their starting points, electricity demand tends to oscillate around a long-run average. Hot days follow cold days; weekday peaks follow weekend dips; a heat wave builds and then breaks. The Ornstein-Uhlenbeck process, described by the stochastic differential equation

$$\frac{d{x}_t}{dt}=\theta (\mu -x_t)dt+\sigma d{W}_t,$$

captures this: the term $\theta (\mu -x_t)$ pulls the system back toward its long-term mean $\mu$ with a speed determined by $\theta$, while $\sigma d{W}_t$ adds random shocks that keep the system bouncing around that mean. The parameter $\theta$ is the drift speed of mean reversion, $\sigma$ is the volatility, and $W_t$ is a Wiener process—mathematical shorthand for Brownian motion.

After discretizing this continuous process into discrete demand levels, the team can calculate transition probabilities between demand states. Given a starting demand level $x_t$ and a time interval $\Delta t$, the demand at $t+\Delta t$ follows a normal distribution with mean and variance:

$$\mathbb{E}[x_{t+\Delta t}|x_t]=x_t{e}^{-\theta \Delta t}+\mu (1-e^{-\theta \Delta t})$$$$\text{Var}[x_{t+\Delta t}|x_t]=\frac{{\sigma }^2}{2\theta }(1-e^{-2\theta \Delta t})$$

The long-run stationary distribution of this process is Gaussian with mean $\mu$ and variance $\frac{{\sigma }^2}{2\theta }$.

The researchers calibrate this process to 2024 net demand data from the California ISO balancing area—chosen because California represents the frontier of high renewable penetration and the resulting demand uncertainty. They set the long-term mean $\mu$ to 6.5 gigawatts, which creates a net demand that is negative approximately 15 percent of the time. When net demand goes negative, it means there's more electricity from solar and wind than the system can use—curtailment of renewable energy, a hallmark of deeply decarbonized grids. This calibration captures a future that's already arriving in parts of California and Hawaii, and that's coming for the rest of the world within decades.

The system can invest in three technologies: a baseload generator (presumably reflecting the marginal cost of a combined-cycle gas plant or large solar-plus-storage facility), a peaking generator (reflecting the expensive quick-ramp gas turbines that only run a few hundred hours per year), and energy storage with a fixed four-hour duration. The operating problem is solved on an hourly basis. Storage has a round-trip efficiency $\eta$ that means some energy is lost in the charge-discharge cycle.

To compute the optimal storage policy, the team uses something called the occupancy measure formulation of dynamic programming. Rather than solving the Bellman equation through iterative methods, they formulate a linear program that directly solves for the long-run fraction of time the system spends in each possible state-action pair. The objective minimizes expected average system cost per time step, subject to flow conservation constraints that ensure the probability of entering each state equals the probability of leaving it. This yields both the optimal policy ${\pi }^{\ast }$ (which turns out to be deterministic—the best action for each state is unique) and the stationary distribution $\rho$ of system states.

Once you have the operating policy, you need to think about investment. Storage operation and generation capacity choices are mutually dependent: the optimal storage policy depends on the cost spread between generators, while optimal generation capacities depend on the arbitrage that storage provides. The team addresses this through an iterative fixed-point algorithm that alternates between determining storage operation and adjusting generation capacities until both converge. Generation capacities are updated using standard screening-curve methods applied to the post-storage load-duration curve—the distribution of demand remaining after storage has done its work. Storage capacity is updated using a zero-profit condition: storage investment continues until the expected annual operating surplus equals the annualized fixed cost of building the batteries.

It's worth being clear about what this model does and doesn't do. It deliberately omits transmission constraints, ramp rates, and co-optimization with variable renewable investments. The researchers aren't trying to replicate a specific grid; they're trying to isolate the fundamental interaction between storage operation and net demand uncertainty. "Although our model is deliberately kept simple," they write, "it retains the structure needed to capture interactions between uncertainty, storage operation, and investment incentives."

What They Found

The first major finding concerns how uncertainty changes storage operation itself. Under perfect foresight—the theoretical benchmark where the operator knows the exact demand trajectory—storage behavior is straightforward: charge during cheap baseload periods, discharge during peaks where net demand exceeds peaker capacity. The operator can precisely time discharges to meet high-demand events. There's no hedging necessary because there's nothing to hedge against. You know exactly what's coming.

Under demand uncertainty, everything changes. The operator doesn't know whether tomorrow will bring a heat wave or temperate weather, high demand or low. Rather than gambling all stored energy on current conditions, the rational strategy is to hold some back—precautionary storage. The research characterizes this as "a precautionary storage policy which hedges against stochastic scarcity events." Energy is reserved not because the operator knows it will be needed tomorrow, but because there's a non-trivial probability that tomorrow will bring a scarcity event.

This behavior shows up clearly in the load-duration curves shown in Figure 1. The canonical load-duration curve plots how often demand exceeds any given level—the steep front edge represents the peak hours that define system planning. Post-storage demand represents what's left for conventional generators after batteries have done their work.

Under dynamic operation with uncertainty, the post-storage curve diverges from the pre-storage curve in two regions. First, when net demand is below 10 gigawatts and baseload generation is marginal, storage charges. Second—and critically—when net demand is between 15 and 23 gigawatts and peaker generation is marginal, storage also charges. This second region doesn't appear under perfect foresight, where there's no need to bank energy conservatively.

Figure 2 makes this even more vivid. It plots the average stored energy level at the beginning of each operating interval, conditioned on the net demand state. In the perfect-foresight case, stored energy peaks during low-demand periods (when you've just charged) and troughs during high-demand periods (when you've discharged). But the dynamic case sits systematically higher across every demand level. The storage unit enters each interval with more energy in its tanks because it's been conservative. It hasn't pushed all its chips into the middle of the table.

"The average stored energy levels at the beginning of each operating interval (conditioned on the net demand) are everywhere higher in the dynamic operating case than the perfect-foresight case," the researchers write. "Demand uncertainty induces a more conservative storage operating policy, which substantially reduces the system value of storage."

This is the invisible tax. The battery doesn't know that tomorrow's forecast will be wrong; it just knows that forecasts are often wrong. As a result, it behaves as if it expects bad news, holding reserves that a perfect oracle would not hold. This costs money—it means stored energy that could have been sold for peak-hour prices sits idle, earning nothing. The economic value of storage is lower under uncertainty than under perfect foresight, not because the technology changes, but because the information environment does.

The second major finding concerns investment. If storage is worth less under uncertainty, rational investors will build less of it. At equilibrium, the perfect-foresight case produces 7.2 gigawatts of storage capacity. Under demand uncertainty, this drops to 4.0 gigawatts—a 44 percent reduction. But the story doesn't end there, because someone has to fill the reliability gap that batteries no longer cover.

That job falls to peaker generation. The equilibrium under uncertainty includes 12.9 gigawatts of peaker capacity, compared to 10.7 gigawatts under perfect foresight. Peakers can run whenever demand is high; they don't need to have charged earlier in the day. Their availability is certain (weather and maintenance aside), even if their economics are expensive. Storage loses value under uncertainty precisely because its availability is contingent on past decisions; you can only discharge what you've previously charged, and if you've been conservative, you have less to give.

Meanwhile, baseload capacity shrinks from 11.1 gigawatts to 10.6 gigawatts. This makes sense: with less storage to shift energy from cheap hours to expensive ones, the value of baseload generation—which is cheap per unit but must run constantly—decreases slightly. You don't need as much always-on generation when you have batteries to bridge the gaps.

Scenario	Storage (GW)	Baseload (GW)	Peaker (GW)
Perfect Foresight	7.2	11.1	10.7
Demand Uncertainty	4.0	10.6	12.9

These numbers represent the equilibrium capacity mix—the combination of technologies that minimizes long-run system cost given the cost parameters and demand characteristics of the model. They're not predictions for any particular grid, but they reveal the structural direction of travel: uncertainty pushes investment away from storage and toward conventional generation.

The third finding concerns what happens when you add the reliability externality—the tendency of electricity markets to underprice capacity because reliability is a public good that benefits everyone whether they pay for it or not. Energy prices are often capped to prevent the exercise of market power. Even in the absence of explicit caps, prices may fail to reach levels commensurate with the value of lost load since there's a lack of individualized penalties for failing to procure capacity for reliability.

Four U.S. wholesale markets—PJM, ISO New England, NYISO, and MISO—address this externality through capacity markets that pay suppliers for commitments to meet future demand. But the MIT team argues that these mechanisms were designed around thermal generators and may insufficiently compensate resources with state-contingent supply like storage.

The reason is subtle but important. Under perfect foresight, you know exactly when storage will be valuable. A capacity payment that reflects the average value of storage across all hours might be approximately correct. But under uncertainty, the value of storage is state-dependent: it matters whether you happened to enter the high-demand period with a full tank or an empty one. A uniform capacity payment doesn't account for this contingency. The storage operator who got lucky and filled up before a heat wave provides enormous reliability value; the one who happened to be empty provides nothing. The payment is the same.

"The reliability externality characteristic of electricity markets interacts with uncertainty in a manner that uniquely distorts both storage operation and investment," the researchers write. This distortion is qualitatively different from the familiar problem with thermal generators, which don't have to make advance decisions about how much energy to have available. A gas peaker either has fuel or it doesn't; a battery has to decide, hours or days in advance, how much charge to preserve.

Figure 3 shows how a price cap—the regulatory intervention at the heart of the reliability externality—distorts storage operation under uncertainty. With a price cap in place, the marginal value of discharging during high-demand periods is limited. The storage operator has less incentive to preserve charge for those periods, and the optimal policy shifts. Red regions show where it's optimal to charge; blue regions show where it's optimal to discharge; white shows where it's optimal to idle. The price cap changes the boundaries of these regions.

The key insight is that this interaction between price caps and uncertainty creates a second-order problem on top of the first-order problem of storage conservatism under uncertainty. Uncertainty makes storage hold back; price caps further distort the incentives for when to use that stored energy. The two effects compound.

Why This Changes Things

The clean energy transition depends on storage more than most policy discussions acknowledge. Solar and wind are free once built, but they're intermittent. You need something to bridge the gap when the sun sets and the wind doesn't blow. Batteries charge during the oversupply and discharge during the shortage. They do the work that in a fossil fuel world would be done by "spinning reserves"—generators that burn fuel even when not fully loaded, ready to ramp up instantly.

But this vision of storage assumes that the value of having energy available tomorrow is correctly priced today. The MIT research suggests it often isn't, for structural reasons that current market designs don't address.

The core problem is the state-contingency of storage value. A gas plant is either available or it isn't. Its capacity value—its contribution to resource adequacy—is relatively easy to calculate. A storage facility's contribution to resource adequacy depends on how much energy it has stored, which depends on how it has been operated, which depends on what the operator expected about future demand and prices. This creates a fundamental challenge for capacity markets designed around thermal generators.

"Although our results do not find a significant impact on storage investment," the researchers write, "we posit the externality should still lead to suboptimal storage capacity that cannot be corrected by standard capacity payments." The equilibrium investment numbers under uncertainty don't show dramatic storage underbuilding beyond what uncertainty itself causes. But the mechanism through which standard capacity payments fail to correct the distortion matters for future market design.

Consider how a capacity market works today. In PJM, for instance, generators receive capacity payments based on their claimed capability to deliver power during the "peak reliability period." The payment rate is set through an auction. A battery that claims 100 megawatts of capacity gets paid as if it's equivalent to a gas turbine that can produce 100 megawatts on demand.

But under uncertainty, the battery's actual contribution to reliability depends on how it was operated leading up to the reliability event. If the battery spent the preceding days arbitraging price differences and ended up with a half-empty tank when the heat wave hits, its effective capacity is zero. The gas turbine doesn't have this problem; it can fuel up the morning of the event. Current capacity mechanisms don't adjust for this contingency. They pay for claimed capability, not for realized availability under stochastic conditions.

The deeper issue is that storage operation and capacity investment are jointly determined under uncertainty. The storage operator's charging decisions determine how much energy is available for peak periods. Those decisions depend on expectations about future demand. Those expectations are wrong some fraction of the time. The value of storage is therefore inherently lower under uncertainty—it's not a market failure, it's a fundamental property of decision-making under imperfect information.

But the reliability externality compounds this problem by mispricing the option value of storage. When prices are capped during scarcity events, the storage operator loses the revenue signal that would tell them to preserve charge. The capped price during a 2022 Texas freeze or a 2020 California heat wave doesn't reflect the true value of having energy available. It can't, because regulators worry about market power and consumer protection. So storage operators charge based on price signals that understate the true value of their flexibility, and the equilibrium investment level reflects those understated signals.

This matters for how we think about the clean energy transition's timeline. If we assume that storage will fill the reliability gaps left by retiring fossil fuel plants, we're implicitly assuming that storage operators can capture the full value of their flexibility. But under uncertainty and price caps, they can't. The value they create is partly a public good—everyone benefits from a reliable grid whether or not they own batteries—without the full compensation that public goods typically require.

The research has implications for the design of capacity markets and reliability mechanisms. Current approaches were built for a world of thermal generators. They may need to be rethought for a world where storage and renewables provide an increasing share of capacity. The state-contingency of storage value, the interaction with demand uncertainty, and the compounding effect of price caps all point toward mechanisms that are more granular—tied to actual availability during reliability events, not claimed or assumed availability.

It also has implications for where to build storage. Under perfect foresight, the optimal storage level in the model is nearly twice as high as under uncertainty. But you can't buy perfect foresight. You can improve forecasts—machine learning and better weather modeling are making inroads—but you'll never eliminate uncertainty entirely. The practical implication is that grids with more volatile demand patterns (more weather-dependent loads like air conditioning, more variable renewables, more electrification) will see more conservative storage operation and lower equilibrium storage levels than simple cost-benefit calculations suggest.

The researchers are careful to note that their stylized model omits many features of real grids. There are no transmission constraints, no ramp rate limits, no co-optimization with renewable investments. The analysis isolates one mechanism—the interaction between demand uncertainty and storage operation—and traces its implications. Real grids are more complicated, and the magnitude of the effects might differ when other constraints bind.

But the direction of the effect is robust: uncertainty reduces the value of storage, changes its operating patterns, and shifts investment toward conventional generation. These are structural features of decision-making under imperfect information, not artifacts of the model.

What's Next

The paper leaves several questions open, some empirical and some theoretical.

On the empirical side, the calibration to CAISO data captures a particular moment in a particular grid—the California system as it existed in 2024. Different grids will have different demand volatility profiles, different renewable penetration levels, and different price cap regimes. The magnitude of the "uncertainty tax" on storage value likely varies across systems. A grid in cloudy northern Europe with modest renewable penetration might see smaller effects than sun-drenched California with its duck curve and frequent renewable curtailment.

One productive direction would be to calibrate similar models to multiple grids and see how the equilibrium storage levels differ. This would help prioritize where policy attention is most needed.

On the theoretical side, the paper's analysis of the reliability externality is primarily qualitative. The researchers argue that standard capacity mechanisms are insufficient to correct the distortions introduced by uncertainty, but they don't fully characterize what a corrected mechanism would look like. This is hard because the value of storage is state-contingent in a way that standard capacity frameworks struggle to handle.

One approach would be to define capacity credits based on actual availability during a set of representative high-demand events, rather than on claimed or assumed capability. If storage operators are paid based on how often they were actually able to discharge during stress periods, the incentive to preserve charge for those periods would be stronger. But this raises questions about which events count as "stress periods" and how to handle correlation between storage availability and stress conditions (if a heat wave affects both demand and battery performance through temperature effects).

Another open question is the treatment of renewable generation within the model. The paper focuses on demand uncertainty, but in high-renewable systems there's also substantial supply uncertainty—solar output varies with cloud cover, wind output with weather patterns. The net demand uncertainty that matters for storage operation reflects both sources of uncertainty combined. A more complete analysis would model renewable generation explicitly and see how supply-side uncertainty interacts with demand uncertainty.

The paper also omits consideration of multiple storage facilities and their coordination. In a real grid, many batteries operated by different owners make charging and discharging decisions based on their own expectations and constraints. The equilibrium that emerges from this decentralized decision-making might differ from the centrally-optimized solution studied here. Market design questions about how storage owners should be compensated for their reliability contributions become even more complex in decentralized settings.

Finally, there's the question of how these results interact with the ongoing debate about capacity market design in organized wholesale markets. PJM, ISO New England, NYISO, and MISO have all grappled with how to value storage and renewables in their capacity auctions. The MIT analysis suggests that the fundamental issue isn't just how to measure the capacity contribution of these resources, but how to compensate them in a way that reflects their state-contingent value under uncertainty.

The researchers frame this as a challenge that "uniquely distorts" storage outcomes relative to conventional generators. The distortion isn't just that storage gets undercompensated for capacity; it's that the undercompensation interacts with uncertainty in ways that compound over time. Storage operators under uncertainty hold back more energy, earn less from arbitrage, receive imperfect capacity compensation, and therefore build less storage. The next generator in the merit order fills the gap—and that generator is likely to be a peaker, not another battery.

This is the invisible tax in its full form: not just lower storage value, but a different capacity mix that persists over time. And it's baked into the mathematics of decision-making under uncertainty, not into any particular market design failure. Better market design can reduce the distortion from the reliability externality, but it can't eliminate the fundamental uncertainty tax that stems from the impossibility of perfect foresight.

The clean energy transition was never going to be simple. Renewable generation is free but intermittent; storage bridges the gap but is valued imperfectly; the grid that emerges will depend on the interaction of technology costs, market design, and fundamental limits on human knowledge. This paper illuminates one of those limits—the tax that uncertainty levies on the grid of the future—and suggests that acknowledging it clearly is the first step toward designing systems that work around it.