A Glass-Box Cloud-Front Forecaster for Solar Ramps

When a cloud bank moves across a region, solar plants upwind lose output minutes before plants downwind do. Spatiotemporal graph neural networks exploit this: they treat each plant as a node in a graph and let upstream sites "warn" downstream ones. In Spatiotemporal Graph Neural Network for Performance Prediction of Photovoltaic Power Systems, this cut 120-minute forecast error on 316 US systems from 18.98% to 11.01% mean absolute percentage error (MAPE) relative to a temporal-only model. The cost is that a grid operator who sees a predicted 40% drop cannot ask the network which upstream site, at what lag, is driving that prediction. Clinical machine learning has spent years on the same problem, and its conclusion is blunt. Explanations bolted onto a black box after the fact are often unfaithful. Stop explaining black box machine learning models for high stakes decisions argues that high-stakes settings should use models that are interpretable by construction. In the grid domain itself, PowerGraph found that 13 of 14 post-hoc explainers for power-grid graph models picked lines that looked plausible to humans but did not reflect what the model computed.

We propose a solar forecaster in which nearly all of the spatial "warning" signal flows through terms an engineer can read directly. Each upstream site contributes through its own one-dimensional curve, and each curve is applied at a specific time lag. An edge weight scales that contribution, and the weight depends on whether the site lies upwind. A black-box interaction component is allowed only for the residual, and it enters only if it earns its place on held-out error. Model complexity is added reluctantly, borrowed from a clinical method that starts with a sparse linear model and adds nonlinear terms only when the remaining error justifies them. The question we test is concrete. How much of the graph network's accuracy gain can a model recover when its spatial reasoning is fully inspectable, and does that hold on the ramp events operators actually care about?

Core Claim

The claim has two parts. First, cross-site solar predictability is mostly low-order and directional: an upstream site's clear-sky-normalized output at a particular lag, gated by wind direction. If that is true, a sparse additive model can carry most of it, and the black-box component in a dual-path design should absorb only a small share of the explained variance. Second, because each spatial contribution is a separate, explicit term, the model's attribution is the computation itself. We can test whether that attribution recovers known cloud-propagation sources in simulation, the same test most post-hoc explainers fail in PowerGraph.

If the first part is false, the result is still informative. The size of the black-box component tells us how much of solar intermittency is irreducibly high-order. Neither the forecasting literature nor the explainability literature in the file measures this. CoDR (Explainable time-varying directional representations) is the closest prior work that claims explainability in solar forecasting, but its explanation is post-hoc inspection of attention weights inside a single site's representation. It has no spatial structure and no guarantee that the weights are what drive the output. Our proposal differs on both counts.

Research Program

The target variable is the clear-sky index. This is a site's output divided by its expected clear-sky output, which removes the predictable daily solar arc. It follows the "ensemble deduction" idea of Modified Auto Regressive Technique, which showed that once this arc is removed a plain autoregression is competitive with deep networks at short horizons. Time is aligned to local solar noon, as in the graph-network paper. Let $k_i(t)$ be the clear-sky index at site $i$ and time $t$. Let $h$ be the forecast horizon, $\ell$ a time lag, and $\phi_{ij}(t)$ the angle between the current wind direction and the bearing from site $j$ to site $i$. The forecast is

$$\hat k_i(t+h) = \sum_{\ell} f_{i\ell}\big(k_i(t-\ell)\big) + \sum_{j \in N(i)} \sum_{\ell} w_{ij}\big(\phi_{ij}(t), d_{ij}\big)\, f_{j\ell}\big(k_j(t-\ell)\big) + g\big(\text{all inputs}\big).$$

Here $N(i)$ is a candidate neighbor set and $d_{ij}$ is the distance between sites. The functions $f$ are shape functions, small one-input curves that can be plotted. The weight $w_{ij}$ is a smooth, low-parameter function of wind alignment and distance; it replaces the static Gaussian distance kernel of the original graph network, which the file identifies as ignoring wind direction. The term $g$ is a small graph network fed residuals. Wind direction comes from reanalysis-free sources where possible. On the satellite dataset it is taken from the optical-flow channels. Elsewhere it is estimated from the peak of lagged cross-correlation between sites over a rolling window, which is itself an inspectable quantity.

Fitting proceeds in three stages, following the reluctant-additive logic of RAITR. RAITR anchors on a sparse linear (LASSO) model and fits nonlinear splines only to the leftover error, with a penalty that favors linear terms. We adapt this as follows. Stage one fits all $f$ as linear with a sparsity penalty over (neighbor, lag) pairs. Stage two lets a selected term become a nonlinear spline only if doing so reduces validation pinball loss beyond a fixed threshold. Pinball loss is the standard loss for quantile forecasts. RAITR's concordance criterion is designed for treatment rules, so we substitute this forecasting criterion. Stage three trains $g$ on what is left. Following the dual-path design of NAMformer, $g$ is trained with feature dropout so that it cannot absorb the marginal effects the additive path should carry. The model outputs multiple quantiles. That makes its tail behavior usable by risk-aware dispatch, which plans against bad-case forecast errors rather than the average, as in the conditional value-at-risk inverter dispatch of Optimal Dispatch of Residential Photovoltaic Inverters.

The main deliverables are three measurements. The first is accuracy against the black-box graph network. The second is the share of explained variance carried by $g$. The third is a faithfulness test on simulated cloud advection with a known ground-truth source, modeled on PowerGraph's ground-truth explanation masks.

Baselines

The naive reference is persistence of the clear-sky index, meaning tomorrow's ratio equals today's. We pair it with the univariate clear-sky-residual autoregression of Modified Auto Regressive Technique. That paper reports 14.20% MAPE at 10 minutes on the Peradeniya dataset, against 14.56% for a CNN and 14.71% for an LSTM. Its lead disappears by 30 minutes (19.94% versus 19.36% for the LSTM). Both baselines are glass-box and carry no spatial information. Beating them at horizons of 30 minutes and beyond is the minimum bar for using neighbors at all.

The strongest skeptic baseline is a sparse spatial autoregression on clear-sky indices: neighbor lags enter linearly, with no wind gating, no splines and no $g$. It is identical to our stage one with gating removed. It is the real threat, because it is already fully transparent. It gets the same features, data splits and hyperparameter search budget as our model. We count as "beating" it only a statistically reliable pinball-loss improvement on ramp periods, using a paired block bootstrap over days.

The strongest accuracy baseline is the black-box spatiotemporal graph network of Spatiotemporal Graph Neural Network for Performance Prediction. It reports 11.01% MAPE at 120 minutes on the 316-system US dataset, against 18.98% for temporal-only convolution. We re-run it on our splits with the same tuning budget and GPU hours as our model, and we augment it with wind direction so the comparison isolates interpretability, not inputs. Our goal is not to beat it. Success is coming within a pre-registered margin of it while $g$ carries a small share of explained variance. We set that margin at a quarter of the gap between the temporal-only model and the graph network. On the Elia Belgian data we also run CoDR. Its paper reports a 17.45% MSE reduction over GAIN and a Pearson correlation of 0.779, but no absolute MSE that could be transferred to our splits. For the faithfulness comparison, the strongest post-hoc explainer is GSAT, the only one of 14 methods that PowerGraph found faithful. We apply it to the graph-network baseline. No reported result was found on Lacuna for GSAT on solar forecasting.

Three ablations isolate the key ingredients. The first replaces wind-gated weights with the static distance kernel. The second removes reluctance, fitting all shape functions nonlinearly from the start in a plain NAMformer-style dual path. The third removes $g$ entirely.

Why not just…?

Why not just use the sparse linear spatial autoregression, since it is already glass-box? This may be enough, and the proposal is designed so that discovering this is a valid outcome rather than a failure. Our stage one is this model. Nonlinear terms and $g$ are kept only if they clear a held-out threshold. If stage one alone matches the full model on ramp days, we report that the graph network's gain is essentially linear advection. That would be a useful and deployable finding. The case against it rests on cloud-edge behavior. A clear-sky index moving from 1.0 to 0.3 upstream is unlikely to map linearly onto downstream loss, and the MAR results already show that linear residual models lose ground to nonlinear ones as the horizon grows.

Why not just run the black-box graph network and explain it with GSAT? GSAT is the most credible post-hoc option in PowerGraph, and it is a genuine competitor. Its masks, however, rank edges. They do not give the operator a curve saying how much a given upstream drop at a given lag will move the forecast, and that curve is what an operator could check against what they see on the ground. The simulation experiment settles this empirically by comparing (source site, lag) recovery for GSAT and for our additive terms. If GSAT matches us there and the graph network is meaningfully more accurate, the honest conclusion is that post-hoc explanation suffices here.

Why not just use a multimodal satellite model like CrossViViT? CrossViViT improves day-ahead hard-day MAE to 50.35 from 54.91 at the CAB station. It needs about 5 hours per training epoch and offers no inspectable spatial reasoning. It targets single stations at day-ahead horizons, whereas ours targets dense networks at intraday horizons where ground-sensor networks see the incoming front. We treat its optical flow as an input to the wind-gating term, not as a competitor architecture.

Proposed Experiments

The primary benchmark is the 316-system US PV network at horizons of 15 to 120 minutes, with ramp days defined beforehand as days with a large intraday drop in clear-sky index. The secondary benchmark is the 14-region Elia dataset at 15-minute resolution. Metrics are MAPE (for comparability with the reported 11.01%), pinball loss and interval coverage on all periods and on ramp periods, plus the share of explained variance carried by $g$. If the US data are not obtainable in their original form, Elia becomes primary and the US comparison is limited to the reported numbers.

The faithfulness test uses a simulated PV network driven by advecting synthetic cloud fields with a known wind vector. Each simulated drop therefore has a known (source site, lag) pair. We score whether each method's top attribution recovers that pair, comparing our additive terms with GSAT and with attention-weight inspection in the style of CoDR. Correlated neighbors may spread credit even in an additive model, so we report recovery as neighbor density rises. This is an identifiability limit we expect to find rather than hide.

A downstream check feeds each model's quantiles into the conditional value-at-risk inverter dispatch on the 20-house fishbone network of Optimal Dispatch of Residential Photovoltaic Inverters. That paper reports 27.83 kWh of curtailment reserve in its risk-neutral case and 44.05 kWh at 95% confidence. We measure voltage violations and curtailment under each forecaster.

The decisive outcome is a joint result on ramp days. The model must be within the pre-registered margin of the graph network, with $g$ carrying a small variance share, and must clearly beat the sparse linear spatial autoregression and the static-kernel ablation. If the linear model ties, the paper's claim becomes "advection is linear." If $g$ carries most of the gain, the claim becomes "solar intermittency resists glass-box spatial modeling." We will report whichever holds.