Can a two-stage mechanism elicit truthful production uncertainty in electricity markets?

Point forecasts hide the cost of uncertainty

A renewable plant can be cheap to run and still expensive to the system. Its output is uncertain, so the operator must buy reserve capacity and pay activation costs when actual production deviates from schedule. The paper frames this as a tradeoff between generation cost and uncertainty that dispatch should weigh, and argues current day-ahead markets do not: they elicit only point production profiles, so the operator optimizes generation cost alone.

The information gap is the second half of the problem. Weather forecasts are widely available, but the production behavior of a specific asset, especially hybrid plants combining renewables with storage or power-to-X technologies, depends on plant characteristics and operating policies known to the owner. The operator therefore cannot fully infer an asset's uncertainty, and the paper shifts that reporting duty to the participants who hold the information.

That shift creates a strategic problem. Efficient dispatch is more likely to leave a highly uncertain producer undispatched, so a producer may gain by underreporting uncertainty. A mechanism is needed that makes honest reporting the producer's best response rather than an act of goodwill.

Two stages, two reports, two payments

The proposed market splits the decision into a time-ahead stage, analogous to day-ahead, and a real-time stage, analogous to imbalance settlement. In the first stage each participant reports a forecast distribution of its type, modeled as a Gaussian distribution with a mean and covariance; the operator then chooses unit commitment, reserve capacity and dispatchable power to minimize expected system cost. In the second stage participants report realized types, and the operator chooses delivered volumes, reserve activation and load shedding.

Each producer's type has three components: baseline production, a downregulation cost and an upregulation cost. The cost function charges the downregulation price when delivery falls below baseline and the upregulation price when it rises above it, so a producer with storage can adjust output at a cost while a plant without storage faces effectively infinite upregulation cost. Payments arrive in both stages, and a participant's utility is the sum of the two payments minus its cost.

The key design choice is what the payments reward. The first-stage payment equals the decrease in first-stage decision cost caused by the participant's presence. The second-stage payment combines the participant's own cost with its marginal contribution to second-stage system cost, computed by re-solving the problem without that participant. This is the VCG logic, which in a single-stage setting aligns individual utility maximization with system cost minimization, extended so that each stage has its own incentive.

Truthful reporting is optimal stage by stage

The paper proves that the mechanism is sequentially ex-post incentive-compatible and ex-ante individually rational. The first property is a two-part statement: given that others report truthfully, a participant cannot raise expected utility by misreporting its forecast distribution in the first stage, and cannot raise utility by misreporting its realized type in the second stage. The second property means expected utility at the first stage is non-negative, so participation is not worse than staying out.

These are not the strongest possible guarantees, and the paper says so. Dominant-strategy incentive compatibility, where truth telling is optimal regardless of others' strategies, is not achievable in this two-stage stochastic setting. Ex-ante individual rationality is also weaker than guaranteeing non-negative utility for every realized type. The authors note that the first-stage notion is relatively weak and leaves the mechanism susceptible to collusion.

Incentive compatibility is what delivers efficiency here. If participants report true type information, the stage decisions minimize system cost, so the mechanism's efficiency claim rests on the reporting guarantee rather than on a separate argument.

The case study prices uncertainty and flexibility

The numerical test uses five non-dispatchable producers, each with expected production of 20 MWh, distinct production variances and distinct regulation costs, against demand of 100 MWh. Reserve capacity costs 10 C/MWh, dispatchable power 6 C/MWh, reserve activation 8 C/MWh and load shedding 200 C/MWh. The first-stage problem is a two-stage stochastic program approximated with 1000 samples and solved as a mixed-integer linear program; the first-stage solve takes about one second and the second-stage solve negligible time.

The incentive result appears in the utility curves. Plotting each producer's average utility against its reported production variance shows that every producer maximizes utility by reporting its true variance, which the authors present as validation of incentive compatibility. This is a numerical check of the theoretical claim in a five-producer instance, not a proof for all instances.

Payments respond to both uncertainty and flexibility. Average payments fall as production variance rises, because a more uncertain producer contributes more to system cost. Payments also fall as the downregulation cost rises, because higher flexibility costs reduce how much the producer's flexibility is used relative to others, lowering its contribution to cost reduction. The authors read this as a price signal that accounts for volume, flexibility and uncertainty together, creating an incentive for flexible assets.

Stochastic clearing beats misjudged deterministic clearing

The comparison that carries the practical argument sets the proposed mechanism against a deterministic market emulating current practice, where producers report only expected production and the operator estimates variance. Two deterministic cases are run: the operator overestimates each producer's variance at 100 MWh², or underestimates it at 4 MWh².

The proposed mechanism declines to dispatch the fifth producer, whose variance is 1024 MWh², and procures 19 MWh of dispatchable power instead, alongside 20 MWh of reserve capacity. Average system cost is 416 C. The deterministic markets dispatch all producers: with overestimated variance they procure 35 MWh of reserve and incur 1162 C, and with underestimated variance they procure 8 MWh of reserve and incur 471 C. Both deterministic outcomes are suboptimal first-stage decisions because the operator is blind to producer-specific uncertainty.

The gap between 416 C and 471 C is the more instructive one. Even underestimating variance, which looks closer to the proposed outcome in reserve volume, still costs more than eliciting the information, because the deterministic market cannot identify which producer is too uncertain to dispatch. The 1162 C case shows the opposite error: overestimating uncertainty buys reserve the system does not need.

Collusion and computation limit the claim

Two limitations bound what the results support. The first-stage incentive notion is relatively weak, leaving the mechanism susceptible to collusion, so the truthfulness guarantee covers unilateral misreporting rather than coordinated manipulation. The pricing rule is also computationally demanding: because the two-stage stochastic dispatch is solved n times to evaluate individual contributions, payment calculation scales with the number of participants. The reported one-second first-stage solve is for five producers, not a large market.

The case study is a single five-producer instance with Gaussian forecast distributions and deterministic regulation costs, so the numerical validation should be read as a demonstration that the theory operates as claimed in that setting, not as evidence about real market outcomes or larger systems.

Eliciting uncertainty changes who gets dispatched

The paper's distinctive claim is not that uncertainty should be priced in general, but that the operator should obtain the uncertainty estimate from the party that knows it and pay for it in a way that makes honest reporting optimal. The case study makes the consequence concrete: the producer with variance of 1024 MWh² is left undispached and replaced by 19 MWh of dispatchable power, a decision the deterministic markets cannot make because they lack producer-specific variance.

The priority this suggests for market design is the reporting channel rather than the dispatch algorithm alone. A deterministic market that guesses variance can land at 1162 C by overestimating or 471 C by underestimating, while the mechanism that elicits the distribution reaches 416 C in the same instance. The threshold that matters is whether the operator can distinguish a producer whose uncertainty makes it a net cost to the system, and that distinction depends on truthful reporting, which in turn depends on the payment rule and on the collusion and computation limits the authors flag.

Sources used in this answer