Source-linked AI summary
When Does Advection-Aware Graph Nowcasting Help? A Controlled Study of Distributed Solar Ramp Forecasting with a Self-Supervised Cloud-Motion Estimator
Phillip Jiang
TL;DR
Distributed solar ramp forecasting needs to account for cloud motion, but existing graph methods generally do not encode its direction and travel time. The paper tests advection-aware forecasting on a controlled synthetic testbed, with a self-supervised motion estimator, and finds benefits only within a defined network-scale envelope.
Problem
Static or symmetric graph approaches to distributed solar forecasting ignore cloud motion’s direction and wind-set travel time, while RMSE can hide cloud-induced ramps.
Source: §1 · p. 1Method
The paper uses a known-wind synthetic testbed to separate graph and motion-feature benefits, then trains a self-supervised estimator and freezes it for forecasting.
Source: §1 · pp. 1–2Results
8% to 15% RMSE reduction over no advection is reported when wind is reasonably steady and v · H fits inside the network; otherwise learned global adjacency is better or advection is neutral or regressive.
Source: §6 · pp. 5–6Takeaways & Limitations
An accurate motion feature can supply roughly half the oracle-CMV benefit, and advection helps only when displacement over the forecast horizon fits within the sensor network.
Source: §1 · p. 2, §7 · p. 6All experiments use one synthetic simulator, where true wind is a generative parameter; validation on a real distributed network remains necessary.
Source: §6 · p. 6
Abstract
from arXiv · showhide
Short-term forecasting of cloud-induced power ramps across a network of distributed photovoltaic (PV) or irradiance sensors is a recognised pain point for grid operators. A natural idea is to make the graph neural network (GNN) advection-aware: connect each site to the sites upwind of it, with edge time-lags set by the cloud-motion vector (CMV), so that a ramp is propagated forward before it physically arrives. Using a controlled synthetic testbed with a known wind field, we show that (i) with a realistic cross-correlation CMV estimate, an explicit advection graph does not beat a plain static or learned-adjacency spatiotemporal GNN; (ii) roughly half of the benefit available from a perfect CMV comes simply from providing an accurate motion vector as an input feature, not from graph structure; and (iii) advection helps only when the advective displacement over the forecast horizon, v*H, fits inside the sensor network. Motivated by (ii), we introduce a small self-supervised cloud-motion estimator -- a position-aware encoder trained only on a multi-lag optical-flow reconstruction objective with an annealed kernel -- that recovers the true wind vector to 2-4 degrees median angular error, 2-4x better than the classical cross-correlation method across every wind regime. Freezing this estimator and feeding its vector to the forecaster closes about 60% of the oracle-CMV RMSE gap at moderate wind (8-15% RMSE reduction over no advection), with no external wind data. We also report a negative result for a spatially-coherent probabilistic head. All claims are established on a single synthetic simulator; we discuss why real-network validation is the necessary next step and outline it.
1 Introduction
The paper tests whether advection-aware graphs improve distributed solar ramp forecasting and separates graph structure from the value of accurate cloud-motion information. Its controlled synthetic study reports conditional benefits, introduces a self-supervised motion estimator, and identifies the need for real-network validation.
1 Introduction
- The controlled testbed uses clouds advected by a known, slowly varying wind, enabling comparison against ground-truth CMV.This setup separates the benefit of the advection graph, an accurate motion feature, and estimating motion from data.Source: §1 · p. 1
- With a realistic CMV estimate, the advection graph does not beat plain baselines; about half the oracle-CMV benefit comes from an accurate input feature, and benefits require v · H to fit within the network.The study also reports negative results for coupling the motion estimate to the forecast loss and for a spatially coherent probabilistic head.Source: §1 · p. 2
- The self-supervised, position-aware estimator recovers true wind to 2°–4° median angular error, 2–4× better than classical cross-correlation.It uses a multi-lag optical-flow reconstruction loss with an annealed kernel.Source: §1 · p. 2
- A frozen two-stage forecaster (Sec. 5.4) that uses this estimator closes ∼60% of the oracle-CMV RMSE gap at moderate wind, with a clean operating envelope.Source: §1 · p. 2
- All results come from one synthetic simulator; the authors describe CMV recovery as a controlled sanity check, not a claim about real skies.They identify real-network validation as the necessary next step.Source: §1 · p. 2
2 Related work
Prior distributed solar GNNs generally use static adjacency, while cloud-motion methods estimate a rigid motion vector and advance observations. Dense-raster weather models learn motion in a different sensing regime; the paper also adopts established ramp-event and probabilistic scoring conventions.
Spatiotemporal GNN forecasting
- Distributed solar forecasting has adopted spatiotemporal GNNs, but their adjacency is usually static and none of the described models encode cloud advection explicitly.Graph WaveNet also learns a global adjacency, included as a baseline in this study.Source: §2 · p. 2
Solar nowcasting and cloud motion
- On sparse ground networks, a common non-image approach estimates a rigid cloud-motion vector by cross-correlating clear-sky-index series, then advects the observed field forward.The paper’s estimator is an optical-flow-style objective specialized to sparse sensor layouts.Source: §2 · p. 2
Learned advection in weather nowcasting
- Weather nowcasting models such as MetNet and DGMR learn motion from dense radar rasters, whereas this study addresses sparse, irregular distributed solar sensing.Source: §2 · p. 2
Ramp forecasting and probabilistic scores
- The study uses swinging-door ramp events and strictly proper energy and variogram scores for its probabilistic head.Clear-sky index uses the Haurwitz model; Ineichen–Perez is noted as a common alternative.Source: §2 · p. 2
3 Problem setup and synthetic testbed
The testbed forecasts clear-sky index at fixed sites using a synthetic cloud field translated by a known wind. It emits both oracle and classical motion vectors, enabling forecast evaluation against controlled motion conditions.
3 Problem setup and synthetic testbed
- The task forecasts clear-sky index kt = GHI/GHIcs at N fixed sites H steps ahead, with H = 16 at 30 seconds, or 8 minutes.Source: §3 · p. 3
- Clouds are modeled as a periodic Gaussian optical-depth field translated by a bounded-random-walk wind, with an option for growth and decay.Observed GHI combines clear-sky GHI, local optical-depth transmittance, and noise.Source: §3 · p. 3
- The simulator provides true v(t) and a classical cross-correlation estimate; the default layout has 49 sites on a 1 km grid, about 7 km across.Evaluation reports test RMSE on kt and ramp-capture metrics, averaged over 2–3 seeds.Source: §3 · p. 3
4 Methods
The methods combine an advection feature with a heterogeneous spatiotemporal GNN and train a separate motion estimator from multi-lag field consistency. The estimator is frozen before its vector is used by both the encoder and advection graph.
4 Methods
- For each site and forecast horizon, the advection graph pools embeddings from sensors near the upwind cloud-arrival location, with distance-weighted normalized edges.An off-grid reliability gate down-weights locations whose upwind point lies far from any sensor.Source: §4 · p. 3
- A per-site GRU and heterogeneous message-passing layers feed a forecast head with a gated advection correction; baselines include static-distance and learned-adjacency GNNs.Source: §4 · p. 3
- The position-aware estimator predicts motion by minimizing multi-lag residuals of a rigid-advection identity interpolated from the sensor field.One predicted vector must explain lags 2, 4, 6, 8, and 10.Source: §4 · p. 3
- The reconstruction kernel width is annealed from 0.40 to 0.10 of network diameter, shifting from broader gradients to a sharper estimate.An optional light MSE-to-reference term represents limited NWP wind supervision.Source: §4 · p. 3
- The estimator is pretrained and frozen, then its vector enters both the encoder’s motion channels and the advection graph.Freezing avoids the observed joint-training failure in which forecast loss drives the estimated vector toward zero.Source: §4 · p. 3
5 Experiments
Across the experiments, realistic estimated motion did not make explicit advection outperform plain GNN baselines, while oracle motion and a learned estimator improved results within a limited operating envelope. The experiments also report negative results for a spatially coherent probabilistic head.
5.1 Does an advection graph help? (E1)
- With estimated CMV, the heterogeneous GNN was best or tied-best at low–moderate wind, while learned adjacency won at high wind and evolving clouds; advection never helped and cost up to 10% RMSE at 10 m/s.Source: §5.1 · pp. 3–4, Table 1 · p. 4, Figure 1 · p. 4
5.2 The cloud-motion bottleneck (E2 and decomposition)
- With perfect wind, advection reduced RMSE by 14% to 17% and raised ramp-down capture (CSI) by about 5 points; roughly half the oracle advantage came from the input feature.Source: §5.2 · p. 4, Table 2 · p. 5
5.3 Self-supervised cloud-motion estimation
- The standalone estimator’s median angular error was 2° to 4° in steady wind and 7° to 13° in variable wind, versus 14° to 31° for cross-correlation in every regime.Source: §5.3 · p. 4, Figure 2 · p. 5
5.4 Frozen two-stage forecaster
- About 60% of the oracle-CMV RMSE gap was closed at moderate wind, but advection failed at ≥16 m/s when displacement exceeded network extent.Source: §5.4 · p. 5
5.5 Spatially-coherent probabilistic head (negative)
- Versus a diagonal head, the joint low-rank-plus-diagonal Gaussian head was marginally worse on every metric, including energy and variogram scores.At 16 m/s, variogram scores were 52.6 vs. 51.0 and CRPS values were 0.065 vs. 0.063.Source: §5.5 · p. 5
6 Discussion
The discussion gives a conditional rule for advection-aware forecasting: it helps when displacement fits within the network and wind is reasonably steady. The results are limited to a single synthetic simulator, and the CMV estimator is presented as a reusable, label-free alternative to cross-correlation.
6 Discussion
- When v · H fits inside the network and wind is reasonably steady, advection-aware features are worth 8% to 15% RMSE; otherwise, learned global adjacency is better, with the feature at best neutral or a 10% regression without the gate.The shaded off-grid region is where even the oracle does not beat no advection.Source: §6 · pp. 5–6, Figure 3 · p. 6
- The self-supervised CMV estimator is a label-free, drop-in replacement for cross-correlation on a bare sensor network and is several times more accurate here.Source: §6 · p. 6
- All results come from one synthetic simulator, so CMV-recovery numbers are a controlled sanity check rather than measurements on real skies.The proposed next step is validation using a real distributed network and a sky-image dataset with optical-flow motion.Source: §6 · p. 6
7 Conclusion
On a controlled testbed, an advection-aware graph feature helps only within a network-size displacement envelope, and about half its value comes from accurate motion input rather than graph structure. The frozen two-stage forecaster closes about 60% of the oracle-CMV RMSE gap at moderate wind; the simulator, models, metrics, baselines, and scripts are released.
7 Conclusion
- The simulator, models, metrics, baselines, and experiment scripts are available, and every table and figure is regenerated from committed result files.The test suite covers the simulator, graph construction, metrics, and model shapes with 40 tests.Source: §7 · pp. 6–7
- About 60% of the oracle-CMV RMSE gap is closed at moderate wind by the frozen two-stage forecaster, while advection helps only when v · H fits inside the network.The result is reported as frozen two-stage test RMSE on clear-sky index across three seeds.Source: §7 · p. 6, Table 4 · p. 7