A data center or aggregator running on its own solar panels increasingly makes promises to the grid operator of the form "in the next hour I will draw at most X megawatts from you, with 95% probability." Conformal prediction is a way to turn past forecast errors into such a bound. It computes a quantile of recent errors and guarantees the stated coverage as long as the data are exchangeable, meaning past and future errors are statistically interchangeable. The operator sizes reserves on these promises. It currently has to trust that the bound was computed honestly, on real sensor data, with the stated procedure. A party paid for its flexibility, or penalized for reserve needs, gains money by quietly shrinking its bound. The operator can only catch this by counting violations after the fact, which takes a long time when violations are supposed to be rare.
We propose to attach a short cryptographic proof to every such bound. The proof shows that the bound equals the conformal quantile of forecast errors computed from readings signed by the site's own meters, and it reveals nothing else about the site's load or generation. The resulting guarantee mixes two kinds of assumption in one statement. Coverage holds at level $1-\alpha$ unless either the data were not exchangeable, which is a statistical failure, or the prover broke a named hardness assumption, which is a computational failure. The research question is what this composition costs, and how the answer shifts as we move along the ladder of hardness assumptions. At one end are conservative signature-only designs; at the other are succinct proofs that rest on non-falsifiable "knowledge" assumptions.
Core Claim
Today, statistical guarantees for intermittent renewables and computational security guarantees are built by separate communities that never compose them. A Conformal Prediction-Based Chance-Constrained Programming Approach for 24/7 Carbon-Free Data Center Operation Scheduling builds adaptive conformal uncertainty sets for solar output and carbon intensity and uses them inside chance-constrained scheduling. It assumes every input is honest. On the Efficiency-vs-Security Tradeoff in the Smart Grid shows what happens when inputs are not honest: demand-response costs reach 250% of the unscheduled baseline and 700% of the optimal schedule, and a simple greedy attack nearly matches the worst case. That paper has no renewables and no cryptography. Its attacker is also an outsider altering messages, and ordinary authentication fixes that.
Our threat is harder. The party producing the bound is the one with an incentive to cheat. Authenticating its messages therefore does nothing. What must be bound is the computation linking signed sensor data to the published number.
The claim is that this binding can be made cheap enough to run every settlement interval, because of a property of renewable forecasting. The forecasters that work well for residual solar volatility are linear. Modified Auto Regressive Technique for Univariate Time Series Prediction of Solar Irradiance subtracts the predictable clear-sky daily arc and fits a low-order autoregressive model to what remains. It reports that this beats unseparated deep models. Linear forecasting over fixed-point integers, plus a counting check for the quantile, is close to the ideal workload for arithmetic-circuit proof systems.
Research Program
The statement to be proven is as follows. The site's meters sign a commitment to each hourly (or five-minute) reading of PV output and load; a commitment is a sealed envelope that fixes a value without revealing it. The site publishes a commitment to its forecasting coefficients once per calibration window. For each new interval it publishes a bound $q$ on net grid draw. It proves that the $n$ most recent residuals $r_i$, each the realized net draw minus the committed forecast, satisfy
$$\#\{i \le n : r_i \le q - \hat{y}\} \ge \lceil (n+1)(1-\alpha) \rceil ,$$
where $\hat{y}$ is the committed model's forecast for the coming interval and $\alpha$ is the promised violation rate. This is exactly the split-conformal quantile condition, so the proof never has to sort anything. It needs $n$ range comparisons and one sum. Proving the count is at least the threshold stops the prover from publishing a bound tighter than the honest quantile. A looser bound is allowed and harmless to the operator.
The circuit also recomputes each residual from the opened commitments and the linear forecaster, so a prover cannot pick residuals freely. Meter signatures are checked outside the circuit on the commitments themselves. This avoids the high cost of verifying signatures inside the circuit, and the operator still sees no raw values.
We instantiate this statement at three points on the assumption ladder, following the taxonomy in the research file.
The first rung is signature-only disclosure. Meters sign raw readings, and the site ships them to the operator, who recomputes the quantile. Soundness then reduces to unforgeability under chosen-message attack, in the style of the reduction in Research on Identity Certification Framework Model in core business of Power Grid, which rests on elliptic-curve discrete logarithm hardness. This is the most conservative assumption, and it gives no privacy.
The second rung uses succinct proofs in the Pinocchio style from Quadratic Span Programs and Succinct NIZKs without PCPs. Proofs are constant size, but soundness rests on the non-falsifiable Knowledge-of-Exponent assumption and a circuit-specific trusted setup. That setup is re-run whenever $n$ or the forecaster order changes.
The third rung uses polynomial commitments under $q$-Strong Diffie-Hellman. Short Signatures Without Random Oracles and the SDH Assumption in Bilinear Groups introduced that assumption, and Asynchronous Computational VSS with Reduced Communication Complexity used it for commitments. The assumption is falsifiable, but it is parameter-dependent: its generic hardness bound $\Omega(\sqrt{p/q})$ weakens with the number of group elements published. We will check whether that number grows with the length of a year of signed intervals in a way that matters at practical parameters. We expect it not to, but this is exactly the kind of claim the reduction literature says should be checked rather than assumed.
For each rung we will write the composed guarantee as a single theorem. It will state coverage at $1-\alpha$, minus the prover's advantage against the named assumption, conditional on exchangeability. We will also measure what that rung costs.
The limitation we will state plainly is that the proof certifies honest computation. It does not certify that exchangeability holds. Weather regime shifts break conformal coverage whether or not anyone cheats. If the operator also wants adaptive conformal updates, like the time-varying sets in the carbon-free scheduling paper, the update rule itself has to go inside the circuit. We treat that as the second stage of the program.
Baselines
The naive reference point is an operator that accepts self-reported bounds without verification. We simulate a strategic site that shrinks its bound to cut reserve charges. The simulation measures how far the realized violation rate drifts above $\alpha$ and how much cost shifts onto the operator. This reference point sets the size of the problem. We know of no Lacuna result that quantifies cheating by the forecaster itself; the closest figures are the outsider-attack numbers above.
The strongest existing method on the forecasting side is unverified adaptive conformal scheduling, AMV-CQR from the carbon-free data center paper. It reports up to a 6.65% total operating-cost reduction and up to a 6.96% cut in carbon-based energy use relative to static robust optimization and conditional value-at-risk benchmarks, on the 2010–2019 Electricity Map and historical weather data. We re-run it on the same data with the same calibration window and coverage level. We then add our proof layer on top, using our fixed-point linear forecaster in place of their quantile regressor. A second comparison uses their forecaster outside the proof, as an unverified reference.
Our approach "beats" this baseline in a specific sense. It must keep realized coverage within the finite-sample conformal tolerance of the honest AMV-CQR bounds and give up no more than a small, pre-registered fraction of the 6.65% cost gain to forecaster restrictions. At the same time it must make the shrink-the-bound attack impossible rather than merely detectable.
The strongest security competitor is signature-only disclosure, the first rung above. It is run on the same meter streams and the same hardware budget. It wins on assumption strength and prover cost. It loses only on privacy. Beating it therefore means showing that the leakage it causes is material, and that our prover stays within the settlement deadline.
The key ablation removes the signed input commitments but keeps the proof of correct computation. This isolates input binding. Without it, a prover can fabricate residuals and still produce a valid proof, and we expect the attack gain to return to the naive level. A second ablation swaps the linear clear-sky-residual forecaster for an unseparated nonlinear one. This measures how much the proof-friendly model costs in bound width, and hence in reserve cost.
Why not just…?
Why not just sign the raw meter data and let the operator recompute the quantile? This is the real threat. If the operator is entitled to see hourly load and generation, it achieves the same soundness under a more conservative assumption, and our proposal adds nothing but cost. The case for proofs rests on the claim that hourly load from a data center or aggregated homes leaks commercially or personally sensitive activity. Smart Meter Privacy With Renewable Energy and an Energy Storage Device motivates this by showing that grid draw leaks consumption unless large batteries mask it, and its leakage falls only as battery capacity grows. We will measure leakage directly, by how well hourly job activity from the ANL Intrepid trace can be recovered from disclosed net load. If recovery is weak, we will report that the first rung suffices.
Why not just audit realized violations after the fact? The operator can count how often actual draw exceeds the bound and run a sequential test. The difficulty is that a rare-event rate is slow to estimate. A site promising $\alpha = 0.05$ while actually running at a moderately higher rate can go a long time before a binomial test rejects. During that time the operator carries under-sized reserves, and intermittency makes the violations cluster on exactly the volatile days that matter. We will compute detection delay on the NSRDB and GLEAMM traces for a range of cheating levels and compare it to the zero delay of a per-interval proof. If the delay turns out to be short for economically meaningful cheating, ex post auditing wins and we will say so.
Why not just release differentially private data, or run the computation in a trusted enclave? Differential privacy, as in Differential Privacy for Power Grid Obfuscation, adds noise that widens the bound and gives no soundness against the party that adds the noise. Enclave attestation is cheaper and is a real competitor. We benchmark it as a single enclave-signed quantile. However, the identity certification paper argues that isolation still needs a reduction to signature unforgeability to resist message forgery, and an enclave moves trust to a hardware vendor rather than a stated mathematical problem. Which of these costs an operator prefers is a policy question. Our measurements will price the choice rather than decide it.
Proposed Experiments
Data come from the Electricity Map and historical weather traces used in the carbon-free scheduling paper, NSRDB irradiance as packaged by WiSoSuper: Benchmarking Super-Resolution Methods on Wind and Solar Data, GLEAMM solar and wind generation, and ANL Intrepid job traces as load, the last two following RARE: Renewable Energy Aware Resource Management in Datacenters. We simulate signed meters over these series at hourly and five-minute resolution, with calibration windows of roughly one to four weeks.
The first decisive check is economic. We measure how much a strategic site gains by shrinking bounds under each baseline, and confirm that under our construction the gain is zero while coverage matches honest conformal prediction.
The second check is cost. We measure prover time, proof size, and verification time for each assumption rung as $n$ and the settlement interval vary, on a server and on an embedded-class processor. Reference points from the file are Pinocchio's 7-group-element (about 256-byte) proofs verified in 8–13 ms for matrix multiplication, and Boneh-Boyen signing at 0.74 ms with precomputation. The pass criterion is that the prover finishes well inside a five-minute interval on the embedded target.
The third check is bound width. We compare the linear clear-sky-residual forecaster against the unverified AMV-CQR bounds, reporting reserve cost and carbon outcomes against the 6.65% and 6.96% reference gains. This tells us whether provability forces a meaningfully worse forecaster.
The last check is the leakage and detection-delay study described above. It decides whether the second and third rungs are worth their extra assumptions over plain signatures and after-the-fact auditing.
References
- A Conformal Prediction-Based Chance-Constrained Programming Approach for 24/7 Carbon-Free Data Center Operation Scheduling
- On the efficiency-vs-security tradeoff in the smart grid
- Modified Auto Regressive Technique for Univariate Time Series Prediction of Solar Irradiance
- Research on Identity Certification Framework Model in core business of Power Grid
- Quadratic Span Programs and Succinct NIZKs without PCPs
- Short Signatures Without Random Oracles and the SDH Assumption in Bilinear Groups
- Asynchronous Computational VSS with Reduced Communication Complexity
- Smart Meter Privacy With Renewable Energy and an Energy Storage Device
- Differential Privacy for Power Grid Obfuscation
- WiSoSuper: Benchmarking Super-Resolution Methods on Wind and Solar Data
- RARE: Renewable Energy Aware Resource Management in Datacenters