acceptodds
Under review as a conference paper at ICLR 2027

Mycelium: A Generalizable Cross-Grid Multi-Task Model for Electrical Distribution Systems

Abstract

Electrical distribution grid operations require inference across heterogeneous networks from sparse, noisy, and incomplete time series measurements. In this work, we identify challenges and explore solutions towards a unified model that can perform diverse tasks grounded in the physics of the electric grid and generalize to unseen distribution networks. We define a unified grid ontology that represents variable sized distribution networks as heterogeneous graphs while preserving native topology, asset types, and electrical relationships across networks. We develop a physics based data simulation pipeline, that combines reference and procedurally generated distribution networks with network reconfigurations, fault scenarios, and configurable sensing conditions. We present Mycelium, a heterogeneous graph transformer with structure aware communication edges and electrical reference features that encode network position and nominal phase orientation, together with task specific temporal readouts which generate per task outputs. We train Mycelium on reference as well as synthetic grids, and study its generalization on benchmark networks completely excluded from training and validation. Mycelium is observed to outperform task specific neural baselines on most reported benchmark metrics. Architectural ablations and the aforementioned studies reveal Mycelium's capability to learn representations of the underlying physics which serves to enhance cross-task performance, thereby addressing a significant challenge in unified grid models.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
graph representation learning, cross-grid generalization, multi-task learning, partial observability, electrical distribution systems

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_bWGHOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents Mycelium, a heterogeneous graph transformer for multiple inference tasks in electrical distribution systems, including state estimation, phase attribution, switch state inference, and fault diagnosis. The authors develop a unified graph representation for different grid components and introduce domain-specific designs such as communication edges and nominal voltage angle references. The model is trained on a combination of reference and synthetically generated distribution networks, and evaluated on unseen grids to examine its cross-grid generalization.

Critical strengths
  • Efforts in data generation and evaluation. The authors construct a relatively comprehensive simulation pipeline and evaluate the model across multiple tasks, networks, and sensing conditions. The use of held-out network families is particularly valuable.
Critical weaknesses
  • Limited methodological novelty. The proposed method largely consists of existing techniques assembled for a specific power system topic. Heterogeneous graph transformer, relation-specific attention, type-specific encoder, and task-specific heads are well-established components. The additional communication edges and nominal angle references are reasonable domain-specific modifications, but I do not see a substantial methodological advance beyond adapting existing graph learning architectures to electrical distribution systems.

  • Limited relevance and insights beyond the application domain. More importantly, it is unclear what new insights the proposed architecture offers to the broader graph representation learning community. The paper would be more compelling for ICLR if it identified a general learning challenge and proposed a solution that could potentially benefit applications beyond power systems.

  • Insufficient evidence for the communication edge design. In Table 3, randomly constructed communication edges outperform the proposed physics-motivated edges on several metrics, including switch-state inference on both the test split and benchmark.

  • Multi-task learning benefits are not supported. The comparison between joint and single-task training in Table 5 shows mixed results. Although multi-task training improves some metrics, single-task models still achieve better performance on others, including benchmark switch-state inference and exact fault localization.

  • Concerns for evaluation setup and benchmark difficulty. The fault detection F1 reaches 1.0 on both evaluation splits, while exact fault localization remains relatively poor. Given that fault onset is assumed to occur at the final observed timestep, I wonder whether the simulated detection task is overly simplified.

Questions for authors

See above.

6 comments

Thank you for recognizing the data-generation effort and held-out-family evaluation.

W1 & W2: Methodological contribution and broader relevance

  • Our contribution combines measurement-based inference, four-task prediction, and cross-family generalization.

    • Voltage, power, and current histories plus static metadata support voltage estimation, phase attribution, switch inference, and fault diagnosis in one checkpoint under sparse, noisy, incomplete sensing.
    • Evaluation excludes benchmark source families from training and validation.
  • Metadata-based reference alignment reduces angle error approximately 15-fold.

    • Angle MAE falls 16.461° → 1.107° on test and 24.896° → 1.635° on benchmark. Metadata-derived nominal angles account for source-phase and transformer offsets without solved trajectories; the model learns operating-dependent residuals.
  • Equipment-aware connectivity and shared supervision support learning across grids.

    • Equipment-specific encoders, a shared HGT backbone, and task-specific readouts reuse parameters across networks. Shortcuts provide three-hop customer–substation paths for eligible customers without additional measurements.
    • Joint training reduces benchmark per-unit MAE by 82.1% despite worse test performance on that metric; W3–W4 detail communication gains and specialist trade-offs.
  • The representation lesson is supported within distribution grids: aligning physical reference conventions improves angle prediction, while communication and shared-supervision benefits depend on task and grid family.

W3: Evidence for communication-edge design

  • Our design balances predictive gains, interpretable connections, and guaranteed coverage.

    • Several shortcut constructions are possible. Equipment-based links follow known substation–transformer–customer relationships and guarantee three-hop customer–substation paths for eligible customers, without additional measurements.
    • Connectivity is determined by topology and registry information, making each shortcut's role explicit and its coverage independent of sampled endpoints. This provides structural certainty, without implying lower prediction variance.
    • With bounded customer-candidate fan-out, the construction adds O(N) communication links, a budget shared by constrained random rewiring. At fixed HGT depth and width, attention on these added edges scales linearly; edge counts alone do not establish runtime savings.
  • The comparisons demonstrate different strengths across this design space.

    • Added connectivity introduces a task-dependent trade-off: Equipment-based shortcuts raise test switch F1 from 0.3253 without communication to 0.8957, while benchmark switch F1 decreases 0.8842 → 0.8210. On the same benchmark, per-unit voltage MAE improves 0.01866 → 0.01267 and fault-type accuracy improves 0.7143 → 0.7209.
    • Constrained random rewiring retains our edge counts and endpoint types per relation and network. Equipment-based placement has lower benchmark per-unit voltage error (0.01267 versus 0.01805) and higher fault-type accuracy (0.7209 versus 0.6468). Random placement has higher switch F1 on test (0.9334 versus 0.8957) and benchmark (0.9278 versus 0.8210).
    • We acknowledge the switch-F1 trade-offs. The voltage and fault-type gains, interpretable placement, and guaranteed coverage for eligible customers support equipment-based links as a principled choice for our multi-task objective. The comparison with random rewiring concerns placement, rather than edge budget.

W4: Evidence for multi-task learning

  • Results are mixed, with a cross-family reversal in per-unit voltage MAE.

    • Single-task training is better on test (0.00148 versus 0.00345); joint training is better on benchmark (0.01267 versus 0.07072), an 82.1% reduction in benchmark per-unit MAE. Single-task training retains lower benchmark MAE in volts (21.51 versus 81.28 V).
    • Joint training also improves benchmark phase accuracy 0.3683 → 0.4001 and fault-type accuracy 0.6746 → 0.7209; single-task models retain higher switch F1 (0.8793 versus 0.8210) and exact localization (0.2169 versus 0.1958). These are selective gains within one four-task checkpoint.
  • Fine-tuning provides separate, task-dependent adaptation value.

    • With encoders and backbone frozen, head-only switch fine-tuning raises benchmark F1 0.8210 → 0.9358, exceeding single-task 0.8793. This requires additional task-specific training.
    • Full-model fine-tuning reaches 0.2011 benchmark exact localization, still below fault-only 0.2169.

W5: Fault-detection difficulty and evaluation scope

  • Detection saturates under our protocol; classification and localization provide more informative comparisons.

    • F1 of 1.0000 at the final timestep provides limited insight into harder detection settings. We are open to clarifying this scope and emphasizing classification and localization, while retaining detection in the pipeline.
    • Benchmark fault-type accuracy is 0.7209, exact localization 0.1958 versus STGATv2’s 0.1177, and Hop@3 0.4114: a localization gain with substantial room for improvement.
  • The protocol targets prompt diagnosis at the first fault-affected observation.

    • At the first fault-affected observation, a rolling window would place that observation at its endpoint without advance knowledge of onset. Random onset within fixed windows and sub-hour latency remain unevaluated with hourly data.
    • Multi-hour persistence requires assumptions about clearance, protection, and restoration, which can change connectivity and energization. Later windows containing fault-affected history remain unevaluated.

Thank you again for your feedback. As discussion ends tomorrow, we would appreciate your response to our rebuttal and additional results. Please let us know if any concerns remain.

Reviewer_YixdOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The paper presents a heterogeneous graph transformer, referred as Mycelium, with structure aware communication edges and electrical reference features designed to handle mutli-task capabilities. Indeed, it uses information about the electrical network to handle several tasks within one model. The authors evaluate it on different grids, including unseen ones, and study whether synthetic data and multi-task learning improve generalization.

Critical strengths

The model combines four power-grid tasks in one framework, which could reduce the need to develop separate models for each task. The model is tested on grids that were not used during training, showing an ability to generalize. The paper compares different training strategies, including joint training and synthetic data augmentation, and examines how they affect performance across tasks.

Critical weaknesses

The results show good performance on test grids does not always transfer to other grid families, which raises questions about how well the model generalizes. The evaluation relies on simulated measurements, so its performance under real-world conditions remains uncertain. Multi-task learning does not improve every task, and the results suggest that sharing one model can involve trade-offs between tasks.

Questions for authors

How does the model perform on real power-grid measurements rather than simulated data? How well does the model generalize to grids with substantially different topologies or electrical characteristics? Why does joint training improve some tasks but reduce performance on others? How sensitive is the model's performance to the number and diversity of synthetic grids used during training? How robust is the model to noisy measurements that differ from those used during training?

3 comments

Thank you for recognizing the four-task formulation and evaluation on unseen grids. We clarify the demonstrated gains and their scope below.

W1 & Q2: Generalization across grid families

  • The original shared checkpoint improves several cross-family tasks.

    • Benchmark per-unit voltage MAE is 0.01267 versus CANOS’s 0.01754, switch F1 0.8210 versus DGS-GNN’s 0.4404, and exact localization 0.1958 versus STGATv2’s 0.1177.
    • The five benchmark networks come from source families excluded from training and validation. European networks lie at the 96th–100th percentiles in reference-corpus independent loop count; urban reaches the 100th percentile in customer density (Figure 4).
  • Transfer remains task- and network-dependent.

    • Phase and network-level comparisons are mixed. Pooled low-voltage MAE improves over CANOS under all six imperfect-sensing policies, alongside substantial errors at urban 12.702 kV.
    • Five networks do not establish arbitrary-grid generalization or isolate topology and electrical parameters as causal drivers.

W2 & Q1: Performance on real measurements

  • Real-world topology and load information ground our simulated evaluation.

    • Iowa and European reference systems draw on utility records and smart-meter data (Appendix A.1), alongside other reference and synthetic sources. Three-phase unbalanced power-flow simulation supplies measurements and labels for all four tasks.
  • Field validation remains outstanding.

    • Telemetry alone generally lacks complete state, phase, switch, and fault-location ground truth. Joint validation requires additional instrumentation or verified operational records; simulation provides consistent labels under controlled sensing conditions.

W3 & Q3: Benefits and trade-offs of joint training

  • The single-task advantage reverses under the evaluated family shift.

    • Single-task training has lower test per-unit voltage MAE (0.00148 versus 0.00345); joint training has lower benchmark per-unit MAE (0.01267 versus 0.07072), an 82.1% reduction in benchmark per-unit MAE. Single-task training retains lower benchmark MAE in volts (21.51 versus 81.28 V).
  • Other tasks show selective gains.

    • Joint training improves benchmark phase accuracy 0.3683 → 0.4001 and fault-type accuracy 0.6746 → 0.7209. Single-task training retains higher switch F1 (0.8793 versus 0.8210) and exact localization (0.2169 versus 0.1958).
    • Shared supervision may encourage reusable features while competing objectives limit specialization; the mechanism remains untested.
  • Head-only fine-tuning demonstrates separate adaptation value.

    • With encoders and backbone frozen, benchmark switch F1 rises 0.8210 → 0.9358, above single-task training’s 0.8793. This requires additional task-specific training and is separate from the original shared checkpoint.

Q4: Sensitivity to synthetic-data quantity and diversity

  • Synthetic augmentation improves voltage estimation and fault classification under a fixed training budget.

    • With no synthetic data, a nested 25% subset, and the full corpus, benchmark per-unit MAE is 0.04025 → 0.01991 → 0.01267 and fault-type accuracy 0.2937 → 0.6111 → 0.7209.
    • Table 4(a) retains the full reference set; the subset matches the full synthetic corpus’s network-size distribution.
  • Gains are task-dependent: switch F1 is 0.9689 → 0.9768 → 0.8210. Quantity and diversity are not independently isolated, and improvement is not monotonic across tasks.

Q5: Robustness to measurement-noise magnitudes differing from training

  • Doubling noise produces modest changes in several pooled metrics (Table R3).

    • Benchmark per-unit MAE changes 0.01267 → 0.01295, angle MAE 1.6355° → 1.6872°, phase accuracy 40.01% → 39.90%, and switch F1 0.8210 → 0.8232. Test per-unit MAE increases 5.59%, with switch F1 0.8957 → 0.8955.
    • We evaluated the frozen 10k checkpoint without retraining at 0×, 0.5×, 1×, and 2× nominal noise standard deviation across existing sensing policies and both splits.
  • Fault diagnosis and individual policies are more sensitive.

    • At 2× noise, benchmark fault-type accuracy falls 72.09% → 70.77%, exact localization 19.58% → 19.05%, and Hop@3 41.14% → 38.49%.
    • The largest policy-level per-unit MAE increase is 10.90% under AMI-dense sensing. Small pooled changes indicate limited sensitivity to the tested amplitude shifts; they do not establish high accuracy on every task. Different noise distributions, sensor bias, temporally correlated errors, and field-measurement robustness remain untested.

Table R3. Frozen-checkpoint noise sensitivity on the benchmark

Results pool the seven sensing policies. Accuracy columns are percentages; F1 is a fraction.

Noise σ ×Voltage MAE (p.u.)Angle MAE (°)Phase acc. (%)Switch F1Fault-type acc. (%)Exact loc. (%)Hop@3 (%)
0×0.012691.620340.010.819472.7519.9742.86
0.5×0.012681.623840.010.819172.4920.1142.20
1×0.012671.635540.010.821072.0919.5841.14
2×0.012951.687239.900.823270.7719.0538.49

Thank you again for your feedback. As discussion ends tomorrow, we would appreciate your response to our rebuttal and additional results. Please let us know if any concerns remain.

Reviewer_iXidOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Mycelium, a 12.2M-parameter heterogeneous graph transformer that jointly performs four inference tasks on unbalanced three-phase electrical distribution grids—state estimation, consumer phase attribution, switch-state classification, and fault diagnosis—from 24-hour windows of sparse, noisy measurements. To train and evaluate cross-grid transfer, the authors build an OpenDSS simulation and procedural feeder-synthesis pipeline with iterative physical repairs, evaluating on 52 unseen in-family test grids and a five-grid out-of-family benchmark against CANOS, GraphSAGE-CPI, DGS-GNN, STGATv2, WLS, and correlation baselines.

Critical strengths
  • Physically grounded problem formulation, nominal angle parameterization, and procedural data pipeline
  • Rigorous network-disjoint evaluation and transparent ablation reporting
Critical weaknesses
  • Theory–method mismatch on communication edges and contradictory ablation results. Proposition 1 (Sec. 3.3, Appendix C.3) gives an exact 2⌈log⁡2h⌉2\lceil\log_2 h\rceil-round error-gap identity EP⋆(ℓ)−EH(n)=Cℓ1+τ−∑jλj(1−η′λj)2ncj2E_P^\star(\ell) - E_H(n) = \frac{C_\ell}{1+\tau} - \sum_j \lambda_j(1-\eta'\lambda_j)^{2n}c_j^2 for a linear predictor using power-of-two ancestor links ∏k=0K−1(I+J2k)\prod_{k=0}^{K-1}(I + J^{2^k}) under exact LinDistFlow and full bus injection sensing. Mycelium instead uses 3-hop substation–transformer–customer shortcuts that do not shortcut medium-voltage backbones, a gap the authors acknowledge but never test.

  • Mixed cross-family generalization and artificial fault-onset assumptions. On the benchmark split, CANOS achieves 9.50 V9.50\,\text{V} MAE versus Mycelium's 81.28 V81.28\,\text{V} (318.20 V318.20\,\text{V} on urban, Table 16; and 1.98 V1.98\,\text{V} vs. 18.71 V18.71\,\text{V} even under clean sensing, Table 12). For phase attribution, only two benchmark grids (rural, urban) have eligible targets, where Mycelium 10k (0.40010.4001, and 0.37820.3782 on urban vs. 0.33330.3333 chance) trails a parameter-free Pearson correlation baseline (P+QxfmrP+Q_{\text{xfmr}} at 0.42000.4200, Table 13).

Questions for authors
  • In Table 3(a), removing communication edges improves test phase accuracy by 8.988.98 points (0.66110.6611 vs. 0.57130.5713) and benchmark switch F1 by 6.326.32 points (0.88420.8842 vs. 0.82100.8210), while random rewiring beats your proposed placement on most test metrics. Why do transformer-anchored shortcuts hurt phase attribution, and did you test the power-of-two ancestor links actually analyzed in Proposition 1?

  • Table 12 shows that even under clean (100% observed, noiseless) sensing on the benchmark, Mycelium has 18.71 V18.71\,\text{V} MAE (5.43%5.43\% MAPE) compared to CANOS's 1.98 V1.98\,\text{V} (0.35%0.35\% MAPE), and 318.20 V318.20\,\text{V} MAE on urban (Table 16). Since per-unit voltage is predicted via V^pu=1+a/100\hat{V}^{\text{pu}} = 1 + a/100 and scaled by VbaseV_{\text{base}}, which voltage levels drive this error?

  • In Appendix B.5 and Eq. (61)–(63), all faults are injected at the final timestep t=T−1t = T-1 and the fault readout explicitly computes xT−∑t<Tγtxtx_T - \sum_{t<T}\gamma_t x_t. How does Mycelium perform when fault onset occurs at a random timestep t0∈{1,…,T}t_0 \in \{1,\dots,T\} or persists across multiple hours?

3 comments

Thank you for recognizing the formulation and network-disjoint evaluation. Our ancestor-link experiment addresses the missing comparison.

W1 & Q1: Communication edges and theoretical scope

  • We tested Proposition 1’s power-of-two ancestor links in Mycelium.

    • Benchmark switch F1 improves 0.8210 → 0.9557, exact localization 0.1958 → 0.2328, and angle MAE 1.635° → 1.159°. Equipment links retain higher fault-type accuracy (0.7209 versus 0.5794); per-unit voltage errors are nearly identical (0.01267 versus 0.01269).
    • We retained physical edges and the 10k training recipe (effective batch size 64). Parameter counts differ: 11.11M versus 12.16M.
    • This tests the analyzed connectivity within HGT. Proposition 1’s locality result remains conditional on linear LinDistFlow, full injection sensing, and its noise/optimization assumptions.
  • Equipment links provide bounded reach with a linear edge budget.

    • Via transformer primary and secondary buses, they guarantee three-hop customer–substation paths for eligible customers without additional measurements; intermediate medium-voltage backbone buses are outside this bound.
    • For N nodes, the edge budget is O(N) with bounded customer-candidate fan-out, versus O(N log N) for ancestors on deep radial grids. Added-edge attention scales accordingly at fixed HGT depth and width; edge counts alone do not establish runtime or memory savings.
  • Ablations show placement-dependent benefits (Table R1).

    • Equipment links raise test switch F1 0.3253 → 0.8957. Random rewiring preserves edge counts and endpoint types per relation and network.
    • No communication gives higher test phase accuracy; random links give higher switch F1 on both splits. Equipment links improve benchmark per-unit voltage error and fault-type accuracy over random links. The mechanism behind the phase trade-off remains untested.

Table R1. Communication-design comparison on the test split and benchmark

Original rows are retained, with the ancestor variant added. Bold indicates the best value within each split.

SplitConfigurationp.u. ↓vmagv_{\mathrm{mag}} (V) ↓vangv_{\mathrm{ang}} (deg) ↓MAPE (%) ↓Phase Acc. ↑Switch Prec. ↑Switch Rec. ↑Switch F1 ↑Fault Det. F1 ↑Class$_5$ ↑Loc. ↑Hop@1 ↑Hop@2 ↑Hop@3 ↑
TestNo communication0.003704.991.2860.2400.66110.19500.98000.32531.00000.45400.18190.27430.34640.3822
Random communication0.003165.241.0830.2300.63280.96310.90560.93341.00000.42410.18810.26280.32670.3583
Power-of-two ancestors0.003185.671.0580.2210.57510.97160.91090.94031.00000.41140.18040.25620.32570.3528
Mycelium 10k0.003455.991.1070.2340.57130.99600.81370.89571.00000.48400.17100.23690.29990.3314
BenchmarkNo communication0.0186677.461.6395.9550.42560.79450.99680.88421.00000.71430.18780.29500.34920.4034
Random communication0.0180578.241.5435.7730.42310.87210.99100.92780.99930.64680.19710.28570.31350.3810
Power-of-two ancestors0.0126975.571.1595.3950.43080.94110.97090.95571.00000.57940.23280.30290.34520.4206
Mycelium 10k0.0126781.281.6355.4250.40010.95530.71980.82101.00000.72090.19580.31750.35580.4114

Table R1a. Edge counts for the 11,999-bus network with 7,591 customers

Edge countEquipmentAncestor
Added communication links9,40365,195
Total stored edges29,00984,801
Directed edges after adding reverses58,018169,602
  • Both retain 12,008 physical edges and 7,598 ordinary attachment edges.
  • Ancestors use 6.93× more communication links and 2.92× more total edges. These are spatial counts, before replication across the 24 timesteps.

W2-1 & Q2: Cross-family voltage estimation and comparison with CANOS

  • Against CANOS, Mycelium improves benchmark per-unit voltage MAE in five of six imperfect-sensing policies and angle MAE in all six.

    • Sparse, noisy telemetry is our deployment target. CANOS performs better under fully observed, noiseless sensing, which we include as a controlled reference.
    • Pooled low-voltage MAE improves under all six imperfect-sensing policies: 6.75 → 1.90 V for sparse noisy/missing and 7.02 → 1.95 V for sparse AMI. Low-voltage targets are 92.18% of evaluated targets; the 230.94 V group primarily drives these gains.
  • Higher-voltage errors explain the aggregate MAE gap.

    • Urban 12.702 kV contributes 82.54% of clean-sensing excess absolute error over CANOS. Higher-voltage targets contribute 92.00–98.98% of Mycelium’s absolute error across policies; CANOS also performs better there in per-unit terms.
    • Industrial 230.94 V contributes 85.53% of clean-sensing summed absolute percentage error, consistent with near-zero-target sensitivity (Appendix E.4).

Table R2. Benchmark voltage errors by voltage group and sensing policy

  • Mycelium / CANOS, target-weighted; per-unit errors use each target's original base.

    • Low voltage: 0.120, 0.231, 0.277 kV; higher voltage: 7.967, 12.702, 39.837 kV (rounded). MAE counts per policy: 18,977,976 and 1,609,632, respectively.
    • The last column is the higher-voltage share of Mycelium's summed absolute error in volts. NM denotes noisy/missing; AMI denotes AMI observability policies.
Sensing policyLow-voltage MAE (V)High-voltage MAE (V)Low-voltage MAE (p.u.)High-voltage MAE (p.u.)High-voltage share of Mycelium’s total absolute error
Clean1.62 / 0.38220.17 / 20.850.00708 / 0.001650.01803 / 0.0019192.00%
NM dense1.34 / 2.401,472.43 / 51.130.00588 / 0.010440.11664 / 0.0044598.94%
NM medium1.49 / 5.011,709.91 / 71.850.00654 / 0.021730.13535 / 0.0061398.98%
NM sparse1.90 / 6.751,768.64 / 86.140.00833 / 0.029280.14000 / 0.0073798.75%
AMI dense1.14 / 2.70268.88 / 79.900.00501 / 0.011710.02187 / 0.0067195.23%
AMI medium1.53 / 5.60282.34 / 89.620.00669 / 0.024260.02294 / 0.0075494.00%
AMI sparse1.95 / 7.021,425.62 / 98.900.00850 / 0.030450.11308 / 0.0087898.41%

W2-2: Phase attribution and correlation baselines

  • Mycelium 20k exceeds both correlation variants on matched targets (Table 13).

    • Accuracy is 0.6150 versus 0.5741 for the stronger baseline on test, and 0.4387 versus 0.4200 on the benchmark. This supports phase attribution within a shared four-task model.
    • The 20k run also changes batch size, so gains cannot be attributed solely to training-network size. Benchmark targets cover only rural and urban.
  • The 10k comparison is mixed; ancestor links improve benchmark accuracy.

    • Original 10k accuracy closely matches active-power correlation: 0.5713 versus 0.5741 on test and 0.4001 versus 0.3999 on benchmark, but trails benchmark active-plus-reactive correlation (0.4200).
    • The separate ancestor variant reaches 0.4308, above both benchmark correlation variants.
  • Correction to Appendix E.2: Correlation variants differ by 0.2104 on test and 0.0201 on benchmark; we will correct the claim that both gaps exceed 0.2.

Q3: Random fault onset and persistence

  • We evaluate diagnosis at the first fault-affected observation.

    • At the first fault-affected observation, a rolling window would place that observation at its endpoint without advance knowledge of onset.
    • Random onset within a fixed window and sub-hour latency remain unevaluated with hourly data.
  • Persistence requires explicit fault-evolution assumptions.

    • Clearance, protection, and restoration can change connectivity and energization; an energized fault persisting for hours is one scenario. Later windows containing fault-affected history remain unevaluated.

Thank you for providing new exp in this rebuttal, and the voltage-level error breakdown.

From the voltage-level error info, most likely , the high-voltage buses drove the MAE gap, while your method strongly on low ones. At the same time, the new results in Table R1 reinforce my main methodological concern: the theoretically motivated power-of-two ancestor links substantially outperform the proposed equipment-based links on benchmark switch F1, MAE etc.

So I will maintain my score for this moment. (still relatively positive).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Closing summary

We thank the reviewers and chairs. Mycelium asks whether one model across tasks and grids can infer electrical states and network properties from sparse, noisy, incomplete measurements. The paper and additional experiments provide a unified framework and evidence supporting this direction.

  • Contribution and significance. Reliable monitoring and diagnosis of distribution grids matter for electricity supply and infrastructure resilience. Our contribution combines a unified four-task formulation, a heterogeneous graph model with physical reference alignment and communication links, and a physics-grounded evaluation framework. This brings an important physical-sciences problem to a broader machine-learning audience and supports systematic study of shared representations under partial observability and grid-family shift.

  • Concrete learning insights. Ablations quantify the benefits of physical reference alignment and shared supervision for prediction and cross-family transfer.

    • Reference alignment reduces angle MAE approximately 15-fold on both splits; benchmark error falls from 24.896° to 1.635°. Metadata-derived nominal angles let the model learn operating-dependent residuals without requiring solved trajectories as inputs.
    • Cross-family transfer: Joint training reduces benchmark per-unit voltage MAE by 82.1% relative to single-task training (0.07072 → 0.01267), while also improving benchmark phase and fault-type accuracy. These gains demonstrate the value of shared supervision on held-out grid families.
    • Task-specific adaptation: With the backbone and encoders frozen, head-only fine-tuning raises benchmark switch F1 from 0.8210 to 0.9358, above single-task training's 0.8793, demonstrating the adaptation value of the learned representation.
  • Systematic evaluation. We evaluate 52 unseen in-family test grids, five benchmark networks from source families excluded from training and validation, and seven sensing policies. Classical and neural baselines, architecture ablations, synthetic-data scaling under a fixed training budget, and joint/single-task comparisons examine complementary questions. Synthetic augmentation lowers benchmark per-unit voltage MAE and improves fault-type accuracy. Real-world topology and load information ground the simulation, which supplies known physical states and network labels for controlled, reproducible comparisons.

  • Communication-design evidence. The rebuttal extends the architectural study with a power-of-two ancestor variant, connecting the analyzed graph construction to experiments within Mycelium (Table R1).

    • Complementary strengths: The separate ancestor variant reaches benchmark switch F1 of 0.9557 and exact localization of 0.2328. Equipment links achieve higher fault-type accuracy than the ancestor variant (0.7209 versus 0.5794), identifying useful design choices for different inference objectives.
    • Edge budget: Equipment links provide interpretable three-hop paths for eligible customers with O(N) added edges under bounded fan-out, compared with O(N log N) for ancestors on deep radial grids. On the 11,999-bus example, the equipment design uses 29,009 total spatial edges versus 84,801 for ancestors, a 2.92-fold difference.
  • Measurement and baseline analysis. Targeted comparisons demonstrate the model's value under imperfect sensing and characterize its sensitivity to measurement availability and noise.

    • Voltage-level analysis shows lower pooled low-voltage MAE than CANOS under all six imperfect-sensing policies, including 6.75 → 1.90 V under sparse noisy/missing sensing (Table R2).
    • Phase attribution: The separate 20k configuration exceeds both correlation variants on matched targets, reaching accuracy 0.6150 versus 0.5741 for the stronger baseline on test, and 0.4387 versus 0.4200 on the rural and urban benchmark targets.
    • Noise sensitivity was tested without retraining from 0× to 2× nominal noise. Doubling noise produces a modest change in pooled benchmark per-unit MAE, from 0.01267 to 0.01295, while switch F1 remains stable at 0.8210 → 0.8232 (Table R3).
  • Fault diagnosis. Our setup is motivated by the operational need to detect, classify, and localize faults at the first fault-affected observation. We prioritize this practical objective and useful inference from sparse, noisy measurements over maximizing benchmark difficulty for its own sake. The original shared checkpoint achieves benchmark exact localization of 0.1958 versus STGATv2's 0.1177 within the four-task model.

Our experiments offer concrete insights into cross-network generalization, shared representations across tasks, and the effects of measurement availability. We hope the unified formulation, systematic evaluation, and additional experimental evidence support a favorable assessment of the work's scientific contribution and practical significance.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.