Low-voltage distribution grids are increasingly analyzed with neural networks that approximate the physics solver. A surrogate of this kind predicts every bus voltage on all three phases tens of times faster than the standard simulator. When such a surrogate flags a violation, for example a voltage-unbalance limit breached at the end of a feeder, an operator needs to know which customers, which phase connections, or which rooftop solar units are responsible, and which single change would fix it. Clinical machine learning has already learned a hard lesson about this kind of question. A model can be accurate and still give explanations that contradict known ground truth. In one Alzheimer's study, a classifier's feature rankings were negatively correlated with expert consensus despite reasonable fit. We propose to carry that auditing discipline into power grids, where the ground truth for an explanation is exact rather than a matter of expert opinion: any proposed cause can be re-simulated.
The core claim is this. For three-phase power-flow surrogates, prediction error does not tell you whether the model's explanations are physically right. The right audit is counterfactual. It asks whether the surrogate ranks a set of discrete, physically valid interventions, such as moving a customer from phase a to phase b or disconnecting a photovoltaic unit, in the same order as the true simulator does. We will build this audit, show that surrogates with similar error can disagree sharply on it, and test whether the audit score predicts which surrogates fail after phase-swap topology changes. We will also train a surrogate on sampled discrete phase edits and test whether that improves both the audit and robustness to topology shift.
Background and Gap
Three-phase unbalance arises because low-voltage feeders serve single-phase homes, single-phase solar and EV chargers. The three phases therefore carry unequal current, raising neutral losses and voltage-unbalance factor (VUF, the ratio of the negative-sequence to positive-sequence voltage). Balanced single-line models hide this. Three-Phase Four-Wire OPF-Based Collaborative Control of PV Inverter and ESS shows that neutral-wire losses alone exceed 4 kWh on a 21-bus network, a quantity that three-wire models omit. The surrogate we audit is PowerFlowMultiNet. It represents each feeder as a multigraph with one edge per phase and reports roughly 15×, 37× and 38× speedups over the OpenDSS simulator on the IEEE 13-bus, 123-bus and 906-bus European LV networks. It offers no account of why it predicts what it predicts. The file flags this opacity under unseen topology changes as an open failure mode. This paper motivates the work and supplies the model under test.
Three pieces of interpretability work shape the design. Assessing the Interpretability of Machine Learning Models in Early Detection of Alzheimer's Disease ranks SHAP attributions against expert rankings with Kendall's τ. It finds τ = 1.0 for trees and τ = −0.39 for KNN. We reuse its methodology: a rank-correlation audit against an external standard. We replace fuzzy expert consensus with simulator-computed truth. How Interpretable Are Interpretable Graph Neural Networks? proves that soft attention masks on graphs are unfaithful, because a nonlinear network's output on an averaged graph is not its average output over discrete graphs. It introduces GMT, which samples discrete subgraphs during training and reports 10–15% better out-of-distribution accuracy on Spurious-Motif. That result rules out soft-mask explanations as the default and supplies our training technique. GMT drops edges independently, however, which on a feeder would create islanded buses or physically impossible circuits. Our edit space is instead restricted to phase reassignments and device disconnections that keep the circuit valid.
On the physics side, Advanced probabilistic load flow methodology for voltage unbalance assessment derives analytical sensitivities of complex VUF to solar injections. It fits local tangent planes at cluster centers. This is the strongest existing non-neural explanation, and it is local by construction. We test where its linear rankings diverge from the large, discrete edits operators actually make. To our knowledge, no paper in the file evaluates a grid surrogate's attributions against any physical reference. The closest clinical paper audits against human experts and has no simulator to check against.
Research Program
The ground truth is constructed as follows. On each benchmark feeder, we drive loads with the London smart-meter profiles used by PowerFlowMultiNet, add single-phase PV, and keep the operating points where VUF or voltage limits are violated or nearly violated. For each such point, we enumerate an intervention set. Each single-phase customer can be moved to one of the two other phases, each PV unit can be curtailed to zero, and each single-phase lateral can be re-phased. On the European LV network, which has 55 single-phase customers, this gives on the order of a hundred and fifty edits per operating point. OpenDSS can solve all of them exhaustively. Each edit's true effect $\Delta_k$ on VUF at the violating bus, and on worst-phase voltage, defines the reference ranking.
Each explanation method produces its own ranking of the same edits. A method can do this by running the surrogate on the edited graph (surrogate counterfactuals), from gradients, from SHAP over node features, from a soft edge mask, or from the analytical sensitivity. The audit score is Kendall's τ between the method's ranking and the true one. Because operators act only on the best few edits, we also report top-5 precision and the true VUF reduction achieved by the method's top choice.
These scores let us separate two properties. Faithfulness asks whether an explanation reflects the surrogate's own behavior, measured by comparing the explanation against the surrogate's own counterfactuals. Correctness asks whether it reflects the physics, measured against OpenDSS. A surrogate can be faithfully wrong, and that case is the one that matters for deployment.
The training contribution is a phase-edit-sampled surrogate. During training, each sample's graph is randomly perturbed by valid phase edits, in the spirit of GMT's discrete sampling. The labels are recomputed by OpenDSS, so the network learns physics on discrete topologies near the training distribution instead of only on the nominal wiring.
Baselines
The naive reference is the ranking operators already use: sort customers and PV units by the magnitude of their single-phase injection on the overloaded phase. Any learned explanation must beat this to justify itself. A random ranking gives the floor near τ = 0.
The strongest existing explanation is the analytical complex-VUF sensitivity from Advanced probabilistic load flow methodology, computed exactly at each operating point. That paper reports speedups of over 10× relative to Monte Carlo and about 25× relative to polynomial chaos, but no rank agreement. No reported result was found on Lacuna for attribution accuracy of this or any grid method.
The strongest surrogate is PowerFlowMultiNet, retrained with its published architecture on the same scenarios, splits and hyperparameter budget as our variants. Its reported results are the ~38× speedup on the 906-bus network. No reported result was found on Lacuna for its numeric voltage error. We will therefore report its error ourselves as the matched-accuracy reference.
Beating these baselines means the following. Surrogate counterfactuals from the edit-sampled model must achieve higher τ and top-5 precision than the analytical sensitivity on operating points where edits are large. They must do so at matched or lower in-distribution voltage error than PowerFlowMultiNet.
Two ablations isolate the key ingredient. In the first, the edit-sampled model is trained with GMT-style independent edge dropping, which ignores circuit validity. In the second, explanations come from a soft edge mask rather than discrete edits. Both use identical compute.
Why not just…?
Why not just compare the surrogate's gradients to the solver's Jacobian? For small continuous changes, that is the right check, and we include it. The interventions operators make are discrete and large, however: a customer changes phase, or a PV unit goes offline. Gradients do not define those effects. GMT's result that averaged-graph behavior differs from discrete-graph behavior predicts gradient rankings will degrade exactly here. The experiment that settles it is reporting τ for gradients separately on small-effect and large-effect edits.
Why not just test surrogates on phase-swapped topologies and skip the explanation audit? This is a real threat. If held-out shift error is all one wants, direct testing suffices. The audit has to earn its place in one of two ways. It must predict shift failure better than in-distribution error does, across a population of surrogates trained with different seeds and architectures. Or it must give per-violation guidance that direct testing cannot. We will measure the first by correlating audit τ and in-distribution error with post-shift error. If audit τ adds no predictive power, we will report that and present the audit only as an operator-facing ranking tool.
Why not just use the analytical sensitivity and drop the neural surrogate? Where linearization holds, this may win, and it is our strongest baseline. It needs a solver evaluation at each operating point, though, and it is local. The tangent-plane clustering in the source paper exists precisely because one linearization does not cover the injection space. The comparison shows where each approach dominates, stratified by edit size and by feeder.
Proposed Experiments
We run the audit on the IEEE 13-bus, IEEE 123-bus and 906-bus European LV feeders. Load variation follows the ±10% and ±50% regimes and the London profiles from PowerFlowMultiNet. We train at least ten surrogates per feeder spanning PowerFlowMultiNet, a plain graph network, and the edit-sampled variant, across seeds.
The first decisive check tests the clinical lesson in this domain. It asks whether surrogates within the same narrow band of voltage error spread widely in audit τ, as SHAP alignment varied across similarly fit classifiers in the Alzheimer study. If the spread is negligible, the audit is redundant with accuracy, and the paper's claim fails.
The second check is predictive. We hold out operating points after real phase-swap reconfigurations and ask whether audit τ explains post-shift error beyond what in-distribution error explains.
The third check tests the training claim. We ask whether physically valid edit sampling raises both τ and post-shift accuracy compared with independent edge dropping, and whether it does so without losing speed. If valid edits help and invalid ones do not, that supports the main methodological point carried over from GMT: discrete sampling must respect the circuit's structure.
References
- Three-Phase Four-Wire OPF-Based Collaborative Control of PV Inverter and ESS for Low-Voltage Distribution Networks With High Proportion PVs
- PowerFlowMultiNet: Multigraph Neural Networks for Unbalanced Three-Phase Distribution Systems
- Assessing the Interpretability of Machine Learning Models in Early Detection of Alzheimer's Disease
- How Interpretable Are Interpretable Graph Neural Networks?
- Advanced probabilistic load flow methodology for voltage unbalance assessment in PV penetrated distribution grids