When wind and solar output swings, operators need to know which transmission lines will overload, not just how much power each plant will make. Turning injection forecasts into line flows is easy in textbook form, because the flows follow from network physics and the line impedances. In practice it is hard. Utilities measure only part of the network: Learning Power Grid Outages With Higher-Order Topological Neural Networks works with 45–50% sensor coverage on distribution feeders. Impedance records are often inaccurate. The full AC equations add losses and effects that depend on the operating point. A purely data-driven model avoids those problems but has to learn conservation laws from scratch, and it usually breaks them.
This note proposes a split along a well-known mathematical structure. Any pattern of flows on a network's lines is one part fixed entirely by the bus injections plus one part that circulates around the network's loops. The first part follows from the wiring diagram alone through Kirchhoff's current law ("what flows into a bus flows out"). The second part is a short list of numbers, one per loop, and it carries everything that depends on impedances and AC effects. We propose to compute the first part exactly and learn only the second. The learned part is a small signal living on the grid's loops, called "loop flows" by power engineers, and a neural network passes messages between loops that share lines. Current conservation then holds by construction. The learning problem shrinks from one unknown per line to one per loop, and that reduction is what should make the model accurate on unmeasured lines when sensors are sparse and renewable regimes shift.
Core Claim
Treat the grid as a cell complex: buses are vertices, lines are edges, and every independent mesh (closed loop of lines) is filled with a 2-dimensional face. A cell complex is the generalization of a simplicial complex, a set of vertices, edges, triangles and so on closed under taking boundaries, that allows faces with any number of sides. Let $B_1$ be the bus–line incidence matrix, $B_2$ the line–face incidence matrix, $p$ the vector of net bus injections (load minus local wind/PV), and $f$ the vector of line flows. We parameterize predicted flows as
$$ f = B_1^{+}\,(p - \ell_\phi) + B_2\, c_\theta , $$
where $B_1^{+}$ is the pseudoinverse of the incidence matrix, $\ell_\phi$ is a small learned non-negative loss allocation at each bus, and $c_\theta$ is a learned circulation, one scalar per face. Because $B_1 B_2 = 0$, the face term never changes the net flow into any bus. Kirchhoff's current law, after loss allocation, therefore holds exactly for every output, whatever the network weights. In the language of the Hodge decomposition surveyed by Topological signal processing and learning: Recent advances and future challenges, the first term is the gradient component of the edge signal and the second is the curl component. The harmonic component is empty here by construction, since every independent cycle is filled.
The physics behind this choice is concrete. Under the DC approximation, the true flow is the minimum reactance-weighted-energy flow consistent with the injections. It differs from the unweighted minimum-norm flow $B_1^{+}p$ by exactly a cycle-space correction, and that correction is set by Kirchhoff's voltage law with the true impedances. With unknown impedances, AC effects and losses, the correction becomes a nonlinear function of the operating state. The claim is that this correction can be learned from partial measurements, by message passing over faces and lines, with far less data than learning flows directly, and that it transfers across renewable weather regimes.
The claim makes a falsifiable prediction about where gains should appear. On a radial feeder there are no loops, so the face term vanishes, flows follow exactly from injections, and the method should match an exact topology-only computation. On meshed grids the advantage over edge-level models should grow with the number of independent loops relative to the number of measured lines.
Relation to Prior Work
The closest existing work comes from two directions. Learning Power Grid Outages With Higher-Order Topological Neural Networks lifts distribution feeders to simplicial complexes and uses Hodge-Laplacian convolutions plus persistent homology, but only as feature extractors for classifying discrete outages. It does not predict continuous flows, model intermittency, or impose conservation. It also lifts by filling cliques. Grid meshes are usually longer than three lines, so a clique lift should find few triangles and miss most loops. We fill the actual cycle basis instead. Time-Conditioned Dances with Simplicial Complexes forecasts time series with supra-Hodge convolutions over correlation-derived complexes and beats dyadic spatiotemporal models on traffic and epidemiology. Its complexes are statistical and its outputs node-level, with no conservation law. Ours are physical and its outputs constrained edge signals. Both papers motivate the proposal by showing that higher-order structure carries signal in networked prediction. Neither uses the Hodge decomposition as a constraint on the output space, and that use is what is new here.
The method borrows from two places. Message passing across ranks, where edges send to the faces they bound and faces send back down, comes from Weisfeiler and Lehman Go Topological: Message Passing Simplicial Networks, and edge-level convolution for imputing missing edge signals comes from Simplicial Neural Networks. We use these only to compute $c_\theta$ and $\ell_\phi$, not to produce the flows themselves. The survey by Isufi et al. lists weighted complexes and inferring which cycles to fill as open problems. We address both in a narrow, physically pinned-down case: the filling comes from the network itself, and learned per-line weights act as a stand-in for the unknown impedances.
On the energy side, the neural optimal-power-flow solvers Optimal Power Flow Using Graph Neural Networks and Unsupervised Optimal Power Flow Using Graph Neural Networks show that graph networks on bus graphs are fast but leave residual constraint violations. The unsupervised version reports a 1.27% violation rate on IEEE-30. They constrain our design: a downstream dispatcher or a CVaR-style risk-aware scheduler of the kind in Dall'Anese et al. needs flow predictions that respect conservation. Analytical Uncertainty Propagation for Multi-Period Stochastic Optimal Power Flow propagates Gaussian renewable uncertainty exactly through DC flows, but it assumes known impedances and Gaussian disturbances. Our setting drops both assumptions.
Research Program
The method separates into two parts, an injection forecaster and a flow map. Renewable intermittency enters through the first; the contribution is the second.
The forecaster turns weather-driven generation histories into bus-injection forecasts at 15–120 minute horizons. We use the spatiotemporal graph model with local-solar-noon alignment from Spatiotemporal Graph Neural Network for Performance Prediction of Photovoltaic Power Systems, optionally with clear-sky removal as in Modified Auto Regressive Technique for Univariate Time Series Prediction of Solar Irradiance. This component is held identical across every compared method, so differences come only from the flow map.
The flow map takes injections, measured flows on the observed subset of lines, and the known incidence structure. It outputs flows on all lines through the equation above. The circulation network operates on lines and faces. Line features are measured flows (masked where absent), the gradient-term flow $B_1^{+}p$, and line type. Face features are aggregated from boundary lines. Messages pass through the up and down adjacencies of the complex, and a learned positive per-line weight rescales the line inner product. That weight is the only place impedance information enters, which makes it the natural target for the parameter-error experiments.
Training is supervised on observed lines only. Unobserved lines are never in the loss, so accuracy on them tests whether the cycle structure carries information across the network. A second, analysis-oriented arm measures how much of the renewable-induced flow variance falls in the cycle component, line by line. It asks whether the lines that congest during wind ramps are the ones dominated by loop flow. That would give operators an interpretable map of where intermittency becomes nonlocal.
Baselines
The naive reference is a nominal DC power-flow computation: flows from forecast injections using the recorded impedances, with no learning. It is what an operator would compute today. Run it with exact impedances, it is a ceiling on DC-explainable accuracy. Run it with deliberately perturbed impedances, it is the realistic floor. Beating perturbed nominal DC is necessary but not sufficient.
The strongest data-driven competitor is a dyadic edge-level graph network predicting all line flows directly. Concretely, we use the st-GNN block structure of Karimi et al. applied to the line graph, and a bus-graph network in the style of Owerko et al. that predicts bus angles and derives flows. It gets the same inputs, the same observed-line mask, a parameter count matched within 10%, the same hyperparameter search budget, and depth at least the grid diameter, so it can reach every line. It is the strongest competitor because it reflects how the literature currently does spatiotemporal grid prediction.
The strongest physics-based competitor is a grey-box model: DC power flow with learnable susceptances fitted on observed lines, plus an unconstrained per-line nonlinear residual. It shares our access to topology and our ability to absorb AC effects. It lacks the guarantee that the residual respects current conservation.
The ablation that isolates the key ingredient keeps our exact architecture but replaces the face term $B_2 c_\theta$ with an unconstrained per-line residual of matched capacity. A second ablation replaces the cycle-basis filling with clique (triangle-only) filling, as in HOT-Nets.
Reported numbers do not exist for this exact task. No reported result was found on Lacuna for line-flow prediction under partial observability with renewable injections on any of these grids. The nearest figures are on different targets: st-GNN reports a MAPE of 11.01% at the 120-minute horizon for PV output, against 18.98% for temporal-only models (Karimi et al.), and the unsupervised neural OPF reports a 1.27% constraint-violation rate on IEEE-30 (Owerko et al.). We will therefore run all baselines ourselves.
Beating the strongest competitors means three things together. Error on unobserved lines must fall by a margin exceeding the seed-to-seed spread. Current-law violation must be zero against nonzero for the competitors. The gap must widen as sensor coverage falls from 100% to 45%.
Why not just…?
Why not just use a standard edge or bus graph network and let it learn the physics? Graph networks can in principle approximate the cycle correction, and with dense sensors they may match us. Their outputs, however, violate current conservation at every bus by some amount, and nothing in their structure ties the flow on an unmeasured line to the measured lines on the same loop. The unsupervised neural OPF results show that residual violations persist even when constraints are penalized in the loss. The coverage sweep settles this. If the edge graph network matches us at 45–50% coverage and on held-out weather regimes, the cycle parameterization is not worth its cost, and we will report that.
Why not just fit impedances and run DC power flow, i.e. the grey-box model? This is the most serious threat. Under pure DC physics with fixed parameters, a learned-susceptance DC model is the correct model, and our method can at best match it. Our claim is confined to what DC misses: losses, reactive-power coupling, voltage-dependent effects, and operating-point-dependent loop flows that grow during large renewable swings. The experiments are designed to expose this boundary. On DC-generated data we expect a tie, and a tie there is the correct outcome. On AC-generated data, especially in high-penetration hours, the grey-box residual is either unconstrained, breaking conservation, or absent, losing accuracy. If the grey-box model with residual matches us on AC data too, the contribution reduces to the conservation guarantee alone, and we will state it that way.
Why not just use a classical weighted-least-squares state estimator? State estimation solves a related problem but needs either enough measurements to make the network observable or accurate impedances and pseudo-measurements. At 45–50% coverage, typical distribution grids are not fully observable. We include a WLS estimator with forecast injections as pseudo-measurements on the DDRE-33 meshed case to test this directly.
Why not just improve the injection forecast, e.g. with signal decomposition, and stop there? Better injection forecasts help every method equally, because the forecaster is shared. The quantity operators act on is line loading, and how a forecast error spreads into line flows is exactly what the flow map determines. We report end-to-end line-flow error with both the st-GNN forecaster and a perfect-foresight oracle, to show how much of the error belongs to each stage.
Proposed Experiments
The first testbed is the DDRE-33 dataset (A Large-Scale Dataset of Distributed Renewable Energy Scenarios on the IEEE-33 Bus Network): 2,000 single-node and 300 joint wind–PV scenarios on the IEEE-33 feeder, labeled by weather type. We run AC power flow in two configurations. With the feeder radial, the face term is empty and every method should reproduce exact topology-based flows, which serves as a correctness check. With the standard tie switches closed, the feeder becomes a meshed network with a handful of loops. We train on some weather categories and test on held-out ones (for example, train on "Monsoon Type" and "Trade Wind Type", test on "Overcast/Rainy") to check transfer under regime shift.
The second testbed uses IEEE-118 and IEEE-300 with ENTSO-E-style wind profiles, following the setup of Bauer and Mühlpfordt. Here we use AC power flow solved per time step, and perturb impedances by 5–20% in the model's prior to imitate bad records.
Metrics are RMSE and MAPE on unobserved lines, maximum current-law violation per bus, and precision and recall for flagging lines above 90% of thermal limit at 15-, 60- and 120-minute horizons. We also report training-set size needed to reach a fixed error, as a measure of data efficiency.
The decisive checks are four. First, on DC-generated data our method should tie the learned-susceptance grey-box model; failure to tie would indicate a bug or an overly restrictive parameterization. Second, on AC data our advantage over the edge-level graph network and the unconstrained-residual ablation should grow as coverage drops from 100% to 45% and as the loop count rises from DDRE-33-meshed to IEEE-300. Third, the triangle-only filling should perform close to the unconstrained ablation, which would confirm that physical cycles, not generic higher-order cliques, carry the benefit. Fourth, the variance-decomposition analysis should show that lines flagged as congestion-prone during wind ramps carry a larger share of loop-flow variance than average.
If the second check fails, the idea is wrong for this domain. If only the third fails, the lesson is narrower: higher-order structure helps here, but the specific choice of cycles matters less than expected.
References
- Learning Power Grid Outages With Higher-Order Topological Neural Networks
- Topological signal processing and learning: Recent advances and future challenges
- Time-Conditioned Dances with Simplicial Complexes: Zigzag Filtration Curve based Supra-Hodge Convolution Networks for Time-series Forecasting
- Weisfeiler and Lehman Go Topological: Message Passing Simplicial Networks
- Simplicial Neural Networks
- Optimal Power Flow Using Graph Neural Networks
- Unsupervised Optimal Power Flow Using Graph Neural Networks
- Optimal Dispatch of Residential Photovoltaic Inverters Under Forecasting Uncertainties
- Analytical Uncertainty Propagation for Multi-Period Stochastic Optimal Power Flow
- Spatiotemporal Graph Neural Network for Performance Prediction of Photovoltaic Power Systems
- Modified Auto Regressive Technique for Univariate Time Series Prediction of Solar Irradiance
- A Large-Scale Dataset of Distributed Renewable Energy Scenarios on the IEEE-33 Bus Network