Yes—you can base a thesis on this paper, but your thesis should extend or critically evaluate it, not simply reproduce its experiments. The paper is a useful starting point because it provides an OpenDSS dataset framework with more than 80,000 scenarios across multiple distribution feeders and evaluates fault localization on unseen networks. (Chen, 2026) (Chen, 2026)

The key scope distinction is that the paper simulates SLG, LL, LLG, and LLL faults, but its main target is the faulted node; it is principally a localization benchmark rather than a joint fault-type-and-location model. (Chen, 2026) (Chen, 2026) It also identifies important realism limits: simulation-only data, complete observability and ideal measurements, and no modeled measurement noise, missing data, delays, or distributed generation. (Chen, 2026) (Chen, 2026)

Important originality check: joint fault classification and localization has already been studied in other GNN work, including a multi-task model evaluated on IEEE 33- and 37-bus systems, and a separate thesis using a multi-task GCN/GATv2 framework on IEEE 123. (Hu, 2023) (Shan, 2025) So “I will combine classification and localization” by itself may not be a sufficiently distinctive contribution. These records were available here as abstracts, so this is a warning to check the full papers and conduct a focused literature review—not a definitive claim that your exact proposed study has already been done.

A stronger, manageable thesis direction could be: “Joint fault-type classification and fault localization across unseen distribution feeders under sparse or noisy measurements.” You could adapt the paper’s data-generation approach, add a classification output alongside node/section localization, then evaluate on feeders held out entirely from training and vary measurement coverage or noise. Compare your joint model with separate classification/localization models and report both fault-type metrics (e.g., macro-F1) and localization metrics (Top-k plus distance-aware error). This is a proposed design, not a result established by the paper; confirm feasibility with your supervisor, especially access to the dataset/code and compute.

In short: yes, it is a viable base paper; make the thesis contribution a well-defined extension, and verify novelty before locking the title.

Yes. This is a feasible thesis question: the base paper explicitly states that measurement noise is not modeled and identifies measurement uncertainty as a future extension. (Chen, 2026) It gives you a clear starting point: test how localization performance changes when controlled measurement noise is added to its simulated node features.

The qualification is that noise robustness itself is not new in power-system fault research: earlier studies report evaluating fault-localization models under measurement noise, and one GNN-based classification/location study reports anti-noise tests. (Chen, 2019) (Mo, 2022) So frame your contribution more specifically—not as “the first to study noise,” but as a systematic controlled-noise evaluation of this paper’s large, multi-feeder benchmark and GGNN, ideally including held-out network topologies. The cited records here are abstracts; a full literature review is still needed before you claim novelty.

A practical study design:

1. Reproduce the paper’s clean-measurement baseline and preserve its train/validation/test separation by feeder; do not let noisy versions of a test scenario leak into training.
2. Add noise to the voltage and current measurements at predefined levels. State the noise model and units clearly. Because magnitudes and phase angles have different units and scales, define perturbations per feature—for example, using sensor-error bounds or a percentage of each feature’s scale—rather than applying one arbitrary raw noise value to every channel.
3. Test two conditions separately: noise at test time only (robustness of the trained model) and noise during training (whether augmentation improves robustness). Repeat across random noise seeds.
4. Plot performance against noise level. For localization outcomes, use Top-1/Top-3/Top-5 accuracy and distance-aware error measures, which the paper defines for evaluating exact and spatially close predictions. (Chen, 2026) (Chen, 2026) As part of your study design, repeat experiments across random noise seeds and report the resulting variability.

A focused working title could be “Controlled Measurement-Noise Robustness of GGNN-Based Fault Localization Across Distribution-Network Topologies.” If your thesis must also classify fault type, treat that as an additional task; otherwise, keeping the scope to localization will make the noise study cleaner and more achievable. Confirm with your supervisor that you can access the dataset/code and agree on realistic noise levels before finalizing the proposal.

Short answer: noise robustness has been studied before in power-system fault localization, so “add noise and measure performance” alone is not a novel research claim. For example, a GCN fault-location study reports experiments across noise levels and data-loss errors; another study evaluates a GNN fault-zone locator at several measurement-noise levels. (Chen, 2019) (Aparicio, 2022) A more recent distribution-network study also reports localization tests under multiple noise levels and missing-node measurements. (Zhou, 2026)

For your exact base paper, though, the question remains a plausible extension. The paper says its benchmark assumes ideal measurements and does not model measurement noise, and lists measurement uncertainty as future work. (Chen, 2026) In the academic search I ran, I did not find a study specifically reporting a controlled noise sweep on this paper’s large multi-feeder dataset and its GGNN benchmark. That is a search result, not proof that no such work exists; check the cited papers’ references and newer publications before claiming originality.

So the defensible thesis contribution is narrower: systematically quantify how this benchmark’s GGNN (and, optionally, its Random Forest baseline) degrades under stated noise levels across held-out feeder topologies, then test whether noise augmentation during training improves the degradation curve. Use sensor-informed, feature-specific noise assumptions and report clean-test versus noisy-test results separately. For localization, report Top-k and distance-aware metrics used by the base paper. (Chen, 2026) (Chen, 2026)

A suitable claim would be: “Prior work studies noise robustness, but this thesis evaluates controlled measurement-noise sensitivity and noise-augmentation effects for the selected multi-feeder benchmark under topology-held-out testing.” Verify that exact distinction with your supervisor and a focused review before treating it as the final novelty statement.

Yes—I understand your idea, and it is a sensible, focused extension of a very recent paper. The paper became available online on 3 August 2026, and its limitations explicitly say that measurements are ideal and measurement noise is not modeled. (Chen, 2026) (Chen, 2026) Your thesis would test the specific robustness question the authors leave open: how does the benchmark model’s localization performance degrade as controlled measurement noise increases, and at what noise level does it fall below an agreed performance target?

Noise robustness has been studied in other fault-location systems, so avoid claiming that no one has studied noise. (Chen, 2019) (Aparicio, 2022) The more precise contribution is to evaluate this newly published benchmark/GGNN, including its transfer to held-out feeder topologies, under a transparent noise sweep. In the academic search I ran, I did not find a study reporting that exact experiment; that makes it a plausible gap, not proof of priority.

A practical experiment:

• Reproduce the clean-data result first; keep test feeders held out from training.
• Inject controlled perturbations into the voltage/current inputs at several predefined levels. Specify the noise distribution and how its scale maps to each feature. Voltage magnitude, phase angle, current magnitude, and phase angle use different scales, so don’t apply one arbitrary raw noise value to all of them.
• First test a model trained on clean data using noisy test inputs. Separately test whether training with noise augmentation improves tolerance.
• Repeat each level with multiple noise realizations and plot performance against noise level. Use the same Top-k and distance-aware localization metrics reported in the base paper. (Chen, 2026) (Chen, 2026)
• Define “tolerance” in advance as the highest tested noise level at which performance remains above a chosen threshold. Choose that threshold and realistic noise ranges with your supervisor and, if possible, sensor specifications.

A suitable working title is “Controlled Measurement-Noise Robustness of GGNN-Based Fault Localization Across Distribution-Network Topologies.” This tests robustness in simulation; without real noisy measurements, it would not by itself establish real-world performance.

My view: yes—you can do this, and it is a good, manageable thesis topic if you keep the question focused. The strongest reason is that your chosen paper explicitly says its measurements are ideal and that measurement noise is not modeled, so your experiment directly tests a limitation the authors identify. (Chen, 2026)

The research question could be: How does controlled measurement noise affect the GGNN’s fault-localization performance on unseen distribution-network topologies, and can noise augmentation improve its tolerance? You can answer it with measurable results: performance-versus-noise curves and the highest tested noise level that still meets a performance threshold you set in advance.

The main qualification is novelty. Other fault-location studies have already tested measurement-noise robustness, including GNN-based approaches. (Chen, 2019) (Aparicio, 2022) So don’t claim “noise in fault localization has never been studied.” Your defensible contribution is evaluating noise sensitivity on this recent, multi-feeder benchmark, under a controlled protocol and topology-held-out testing. In the academic search I ran, I did not find that exact evaluation, but that is not proof of priority.

Keep the scope to localization unless your supervisor expects fault-type classification too. Use clearly justified, feature-specific noise levels, separate clean-trained/noisy-test results from noise-augmented training, and state that simulation experiments do not establish performance on real utility measurements.

Bottom line: it is a solid thesis direction—specific, testable, and motivated by the base paper—provided you confirm data/code access and frame the novelty narrowly.

For your stated thesis—noise robustness of this base paper—I’d use the GGNN as the main model. That keeps your work directly tied to the paper and lets you test whether its graph-based localization holds up as measurement noise increases.

Use a decision tree only as a simple baseline, not as a replacement for the GGNN. Better still, use a Random Forest as your conventional-ML comparator, because the base paper already compares Random Forest with its GGNN. In that study, Random Forest achieved higher exact-node Top-k/F1 results in one unseen-feeder setting, while the GGNN’s predictions were more topologically coherent by the paper’s distance-weighted measure. (Chen, 2026) (Chen, 2026) (Chen, 2026)

My practical recommendation: compare GGNN vs Random Forest under the same clean and controlled-noise test conditions. If you have time, add one Decision Tree as a lightweight reference, but don’t let extra models dilute the central question. If you can only implement one model, choose the GGNN.

If by “random tree” you mean Random Forest, my recommendation is: GGNN as the main model, Random Forest as the comparison baseline. The GGNN keeps your thesis aligned with the base paper; the Random Forest gives you a conventional non-graph benchmark. The paper itself compares them: Random Forest had stronger exact-node scores in one unseen-feeder experiment, while the GGNN produced more topologically coherent predictions by a distance-weighted metric. (Chen, 2026) (Chen, 2026) (Chen, 2026)

If you can implement only one, choose GGNN to study the base paper’s noise robustness. If you want to compare model robustness, test both under identical noise levels and train/test splits. A single Random Decision Tree is not the same as a Random Forest; I’d use Random Forest for the baseline.

Random Forest is usually easier to train and run than a GGNN: it avoids graph-specific model setup and message-passing choices. For your study, you would still need to format each scenario’s node measurements so the forest can rank candidate fault locations; unlike the GGNN, it does not explicitly use feeder connectivity. The base paper describes this distinction and compares both approaches. (Chen, 2026) (Chen, 2026)

GGNN is the better choice for reproducing the base paper’s main method, but it generally takes more implementation and tuning because you must represent the network as a graph and configure message passing. In the paper’s unseen-network experiment, Random Forest scored higher on exact-node Top-k/F1 metrics, while GGNN had a lower distance-weighted error—so “easier” and “better” depend on which outcome you prioritize. (Chen, 2026) (Chen, 2026)

Practical choice: train Random Forest first as a quick baseline; use GGNN as your main thesis model if your goal is to test the base paper’s approach under noise.