Network-Topology Bounds on Simultaneous Solar Shortfalls

Grid operators most fear the case where many solar plants underperform their forecasts at the same moment, because that is when reserves run out. Estimating the probability of such a joint event is a high-dimensional problem. Single-site tail probabilities can be estimated reliably from a year or two of data. The probability that ten or twenty specific sites all fall short at once cannot, because the event almost never appears in the record. Operators therefore pick between two poor options. Assuming independence underestimates the risk. Assuming every plant fails together grossly overestimates it. This note proposes a third option, borrowed from quantum foundations.

Physicists asking which correlations a network of independent sources can produce derived inequalities that bound a joint probability using only the single-node probabilities and the shape of the network (Limits on Correlations in Networks for Quantum and No-Signaling Resources). Weather errors have a similar structure. Once a good forecast has removed the large-scale weather regime, the remaining errors come from cloud systems of limited reach. Each such system affects only the plants within roughly (cloud speed × forecast horizon) of it. If that is true, a joint shortfall probability over many sites can be bounded from per-site tail estimates alone. Low-order data can then test the network structure the bound relies on, and the bound can be used for high-order risk where data are too thin. That is the proposal.

Core Claim

Suppose each plant's forecast error is a function of a set of mutually independent latent "sources" (localized weather disturbances plus local noise). Each source touches a known subset of plants, which defines a hypergraph. Let $p_j$ be the probability that plant $j$ falls below its forecast threshold. Then the probability that all plants in a set $V$ fall short simultaneously satisfies

$$P\Big(\bigcap_{j\in V} S_j\Big) \;\le\; \min_{\eta}\ \prod_{j\in V} p_j^{\eta_j}, \quad \text{subject to } \eta_j\ge 0 \text{ and } \sum_{j\in e}\eta_j \le 1 \text{ for every source } e.$$

Here $S_j$ is the shortfall event at plant $j$. The vector $\eta$ is a fractional independent set of the source hypergraph, meaning nonnegative weights such that the plants touched by any single source have weights summing to at most one. This is the Finner-type inequality that Renou et al. derive for networks. They state it for bipartite sources and extend it to no-signaling resources; multipartite sources are left open in the quantum and no-signaling setting. We need only the classical-source case, which is the original probabilistic inequality they adapt, so their open problems do not block us.

Taking logarithms, choosing the best $\eta$ is a small linear program: maximize $\sum_j \eta_j(-\log p_j)$ subject to the source constraints.

The bound always lies between the independence product, which it equals when every source touches one plant, and the comonotone bound $\min_j p_j$. It approaches the comonotone bound only when one source touches everything. Its value therefore depends entirely on whether conditioned forecast errors really have local, independent causes. That is an empirical question, and the bound also lets us test it.

The connection to the dimensionality-limitation literature is direct. Learning a cumulative distribution function reaches near-$1/N$ error in one dimension, but the method does not lift to several dimensions, because multivariate CDFs lack a rank map to the unit cube (Neural Networks for Density Estimation). Our construction never estimates a multivariate tail. It combines $|V|$ one-dimensional tail estimates, which are reliable, with a discrete structural assumption that can be checked. The statistical error of the bound is inherited from one-dimensional estimation rather than from the curse of dimensionality.

Research Program

The first component is the error model. We take the 316-system US dataset used by Spatiotemporal Graph Neural Network for Performance Prediction of Photovoltaic Power Systems, retrain their spatiotemporal graph network (a model that mixes information across nearby plants and across time steps), and define shortfall events on its residuals at horizons from 15 to 120 minutes. Conditioning on the forecast matters. Raw output shares a common cause in the diurnal cycle and the synoptic regime, which would make the bound collapse toward $\min_j p_j$. A good forecaster absorbs those shared causes. We predict that residual dependence becomes more local as forecast skill improves, and we will measure this.

The second component is the hypergraph. Candidate source hypergraphs come from three constructions. The first uses static geographic radii, matching the fixed distance kernel of Karimi et al. The second uses advection-scaled footprints, where a source covers plants within $v\cdot h$ of a cloud cell, with $v$ the regional wind speed and $h$ the forecast horizon. The third is a learned, data-driven clustering.

The bound itself supplies the falsification test. Pairwise checks only test independence between plants that share no source, because for a sharing pair the bound reduces to the comonotone bound. The first nontrivial content appears at triangles: three plants that pairwise share sources but have no common one must satisfy $P(abc)\le\sqrt{p_ap_bp_c}$. Two years of data are enough to estimate triple and quadruple joint frequencies at moderate tail levels. We select the sparsest hypergraph not rejected at order three or four, then apply it to events of order ten to fifty, where direct estimation fails.

The third component is the marginals. Per-plant tail probabilities are estimated both with empirical quantiles and with a smoothed CDF estimator in the spirit of Magdon-Ismail and Atiya. Confidence bands on each $p_j$ are propagated through the linear program to give a bound that holds with stated confidence.

The fourth component is use. The bound feeds a regional reserve-sizing problem: hold enough reserve that the probability of simultaneous shortfall across any set of regions exceeding the reserve stays below $\epsilon$. Dall'Anese et al. show that tail treatment changes decisions materially. On their 20-house feeder, active-power curtailment rises from 27.83 kWh (risk-neutral) to 44.05 kWh under a 95% conditional value-at-risk constraint (Optimal Dispatch of Residential Photovoltaic Inverters Under Forecasting Uncertainties). We do not use that feeder as our main testbed. Twenty neighboring houses sit under the same clouds, so one source touches all of them and the bound becomes vacuous. Our regime is regional, and we state that limitation openly.

Baselines

The naive reference point is the independence product $\prod_j p_j$, which is what a planner gets by treating site errors as unrelated. It is anti-conservative whenever errors are positively dependent. The bound beats it if the bound's guarantee holds on held-out data where independence is violated.

The opposite naive reference is the comonotone bound $\min_j p_j$. It is always valid and usually useless. Beating it means a tighter value, measured as the ratio of the comonotone bound to our bound on high-order events, while still upper-bounding held-out frequencies.

The strongest existing competitor is a parametric copula fit to the spatiotemporal graph network's residuals: a Gaussian copula with distance-decaying correlation, and a Student-t copula, which has tail dependence. These are what a careful practitioner would build. They use exactly the same residuals and marginals, and their tuning budget (choice of kernel and degrees of freedom) matches our hypergraph-selection budget, with both selected only on order-three and order-four events. We count a win if our bound remains valid at high order, with held-out joint frequencies below the bound in at least $1-\delta$ of tested sets, in cases where the copula's point estimate falls below observed frequencies. We also report tightness relative to the copula.

The forecasting backbone is itself a baseline for marginal quality. Karimi et al. report a 120-minute mean absolute percentage error of 11.01% for their spatiotemporal graph network, against 18.98% for a temporal-only convolutional model. We must reproduce roughly that level before trusting the residual analysis.

No reported result was found on Lacuna for joint multi-site tail-probability estimation on any of these datasets. The comparison is therefore new, and all baselines are run by us.

Two ablations isolate the key ingredient. The first swaps the advection-scaled hypergraph for the static geographic one, testing whether a causal footprint matters. The second applies the bound to raw output instead of forecast residuals, testing whether conditioning is what makes sources local.

Why not just…?

Why not just fit a t-copula or vine copula and read off joint tails? This is the real threat. A copula gives point estimates, which are often tighter than any bound, and if a t-copula's high-order joint exceedance frequencies match held-out data, our tightness advantage disappears. Two things would remain. One is a guarantee that does not depend on the copula family being correct in a regime it was never fit to. The other is the triangle-level falsification test, which no copula fit provides. The experiment settles the question directly by comparing calibration at order ten to fifty. If the copula is calibrated there, we report that the bound's value is limited to robustness and structure testing.

Why not just count joint exceedances empirically, with bootstrapping? At order $k$ the number of observed simultaneous shortfalls falls roughly geometrically, so for large $k$ the counts are near zero and confidence intervals become vacuous. This is the multivariate estimation failure the dimensionality literature documents. We show it explicitly by plotting empirical interval width against order next to our bound.

Why not just use a dynamic, wind-aware graph forecaster and trust its ensembles? Karimi et al. themselves identify static adjacency as a limitation, and a better forecaster is complementary: it sharpens the marginals and localizes residual dependence, which tightens our bound. Forecast ensembles, though, inherit whatever joint structure the model learned, with no check on rare simultaneous events. We test ensemble joint frequencies as an additional baseline.

Proposed Experiments

The first decisive check is structural. On the 316-system residuals at each horizon, we test the triangle inequalities implied by each candidate hypergraph. We predict that the static-radius hypergraph will be rejected more often at long horizons than the advection-scaled one, because cloud footprints grow with $v\cdot h$. If no sparse hypergraph survives on residuals, the premise fails: there is a hidden common cause, and the method reduces to the comonotone bound. That is a publishable negative result about forecast-residual structure in its own right.

The second check is predictive. We hold out the final six months and evaluate high-order shortfall events over random and geographically clustered plant sets. Three quantities are compared across all baselines: the violation rate of each estimator, the tightness ratio against the comonotone bound, and calibration against the copulas.

The third check is decision-level. Reserve held under each estimator is compared at an equal held-out reliability target. Success means less reserve than the comonotone policy and fewer violations than independence or the Gaussian copula.

As a secondary replication, the Baguan-solar ground network of 246 pyranometers (Integrating Weather Foundation Model and Satellite to Enable Fine-Grained Solar Irradiance Forecasting) would test transfer to a different climate and a stronger forecaster, if access can be arranged. That paper outputs point forecasts only, so it cannot supply joint tails itself.