ai-research-survey

scan.json (29998B)

---

 {
   "paper": {
     "title": "Defense of Massive False Data Injection Attack via Sparse Attack Points Considering Uncertain Topological Changes",
     "authors": [
       "Xiaoge Huang",
       "Zhijun Qin",
       "Ming Xie",
       "Hui Liu",
       "Liang Meng"
     ],
     "year": 2022,
     "venue": "Journal of Modern Power Systems and Clean Energy",
     "doi": "10.35833/MPCE.2020.000686"
   },
   "scan_version": 3,
   "active_modules": ["experimental_rigor", "data_leakage"],
   "methodology_tags": ["benchmark-eval"],
   "key_findings": "The paper proposes a deep learning methodology (AE-BCV detector + AE-GAN) for detecting, localizing, and recovering from false data injection attacks on power systems. The AE-BCV detector achieves over 95% detection accuracy on IEEE 57-bus, 118-bus, and 415-bus CSG systems, outperforming SVM and ANN baselines, even with unseen topological changes. The AE-GAN generates diverse candidate measurement vectors to localize falsified data with ~80-86% accuracy and significantly reduces measurement error after recovery. An enhanced FDIA attack model is also proposed that can affect more state variables with fewer compromised meters than conventional models.",
   "checklist": {
     "artifacts": {
       "code_released": {
         "applies": true,
         "answer": false,
         "justification": "No code repository, GitHub link, or archive is mentioned anywhere in the paper. The implementation uses PyTorch, MATPOWER, and CVX, but no source code is released."
       },
       "data_released": {
         "applies": true,
         "answer": false,
         "justification": "The datasets are generated from IEEE benchmark systems (publicly available in MATPOWER) and a synthesized 415-bus CSG system, but the specific generated datasets including FDIA attack samples are not released. The 415-bus CSG system data is proprietary."
       },
       "environment_specified": {
         "applies": true,
         "answer": false,
         "justification": "The paper mentions hardware (16 GB RAM, Intel i7-8750H CPU, Nvidia RTX 2070 GPU) and software tools (PyTorch, MATPOWER, CVX) in Section VI-A, but does not provide version numbers for any of these, no requirements.txt, no Dockerfile, and no detailed environment setup."
       },
       "reproduction_instructions": {
         "applies": true,
         "answer": false,
         "justification": "No step-by-step reproduction instructions are provided. While the methodology is described in Sections III-V and neural network parameters are listed in Appendix A, there are no scripts, README, or concrete reproduction guide."
       }
     },
     "statistical_methodology": {
       "confidence_intervals_or_error_bars": {
         "applies": true,
         "answer": false,
         "justification": "All results in Tables V-X are reported as point estimates only (e.g., 96.7%, 98.5% accuracy). No confidence intervals, error bars, or ± notation appears anywhere in the paper."
       },
       "significance_tests": {
         "applies": true,
         "answer": false,
         "justification": "The paper claims AE-BCV outperforms SVM and ANN (Tables VI-VIII) based solely on comparing accuracy numbers. No statistical significance tests (t-test, Mann-Whitney, etc.) are reported."
       },
       "effect_sizes_reported": {
         "applies": true,
         "answer": true,
         "justification": "Tables VI-VIII provide accuracy comparisons with baseline context. For example, AE-BCV achieves 96.7% vs SVM 86.0% vs ANN 86.7% on 57-bus (Table VI). Table X shows error reduction from FDIA (e.g., 16.50 to 0.85 for SA on IEEE 118-bus). The magnitude of differences is clear."
       },
       "sample_size_justified": {
         "applies": true,
         "answer": false,
         "justification": "The dataset uses 80/10/10 split for training/validation/test, and 1100 simulations per attack level for BDD testing (Section VI-C), but no justification is given for why these sizes are adequate. No power analysis is provided."
       },
       "variance_reported": {
         "applies": true,
         "answer": false,
         "justification": "No variance, standard deviation, or interquartile range is reported across experimental runs. All tables show single-run results without any spread measures."
       }
     },
     "evaluation_design": {
       "baselines_included": {
         "applies": true,
         "answer": true,
         "justification": "The paper compares the AE-BCV detector against SVM and ANN classifiers (Tables VI-VIII). It also compares the proposed FDIA attack model against the conventional model from [6] (Table IV). A rough comparison with [13] and [14] is also provided in Section VI-D."
       },
       "baselines_contemporary": {
         "applies": true,
         "answer": false,
         "justification": "The primary baselines are SVM and ANN, which are basic ML methods, not state-of-the-art deep learning approaches for anomaly detection as of 2022. The paper acknowledges prior deep learning methods [13], [14] but only provides a rough comparison rather than a controlled evaluation against them."
       },
       "ablation_study": {
         "applies": true,
         "answer": false,
         "justification": "The system has multiple components (AE feature extractor, BCV decision maker, AE-GAN, pattern match algorithm) but no ablation study is conducted to measure the individual contribution of each component."
       },
       "multiple_metrics": {
         "applies": true,
         "answer": true,
         "justification": "The paper reports multiple metrics: detection accuracy, false positive rate, false negative rate (Table V), localization correct rate, positive false rate, negative false rate (Table IX), and mean error before/after recovery (Table X)."
       },
       "human_evaluation": {
         "applies": false,
         "answer": false,
         "justification": "Human evaluation is clearly irrelevant to claims about automated detection/localization accuracy on simulated power system data."
       },
       "held_out_test_set": {
         "applies": true,
         "answer": true,
         "justification": "The paper explicitly uses separate train (80%), validation (10%), and test (10%) sets. Test sets include unseen topological changes (up to 8% line switching vs 5% in training), and unseen attack intensities (MA, WA) not used in training. Section VI-B and VI-D describe this clearly."
       },
       "per_category_breakdown": {
         "applies": true,
         "answer": true,
         "justification": "Results are broken down by system size (57-bus, 118-bus, 415-bus), attack type (targeted/untargeted), attack intensity (SA/MA/WA), and task (detection/localization/recovery) across Tables V-X."
       },
       "failure_cases_discussed": {
         "applies": true,
         "answer": false,
         "justification": "While Tables VII-VIII show degraded performance on unseen weak attacks (e.g., 90.8% on 57-bus for WA vs 96.7% overall) and Table IX shows negative false rates of 15-19%, no specific failure cases are analyzed or discussed qualitatively."
       },
       "negative_results_reported": {
         "applies": true,
         "answer": false,
         "justification": "Every configuration and method shows positive results. No ablations that hurt performance, no approaches that were tried and abandoned, and no failed configurations are reported."
       }
     },
     "claims_and_evidence": {
       "abstract_claims_supported": {
         "applies": true,
         "answer": true,
         "justification": "Abstract claims of 'average 95% accuracy for detection' are supported by Tables V-VIII (96.7-98.5% overall, 90.8-99.2% for unseen attacks). 'Over 80% accuracy for localization' is supported by Table IX (80.37-85.69%). 'Recovers measurement and state variables close to their true values' is supported by Table X showing substantial error reduction."
       },
       "causal_claims_justified": {
         "applies": true,
         "answer": true,
         "justification": "The main causal claims are comparative: 'AE-BCV outperforms SVM and ANN.' These are tested on the same datasets with controlled comparisons (Tables VI-VIII), which is adequate for the claim that one method produces better accuracy than another on the same test data."
       },
       "generalization_bounded": {
         "applies": true,
         "answer": true,
         "justification": "The paper explicitly states scope boundaries: 'we focus on the DC SE model with complete information' (Section II), notes the specific IEEE benchmark systems tested, and identifies AC power flow extension as future work (Section VII). The title and claims are specific to FDIA defense."
       },
       "alternative_explanations_discussed": {
         "applies": true,
         "answer": false,
         "justification": "No alternative explanations are discussed for why AE-BCV outperforms baselines. The paper does not consider whether the improvement could be due to feature engineering (joint state-measurement input) rather than the BCV classifier specifically, or other confounding factors."
       },
       "proxy_outcome_distinction": {
         "applies": true,
         "answer": true,
         "justification": "The paper measures detection accuracy, localization accuracy, and recovery error — these directly correspond to the stated goals of detecting, localizing, and recovering from FDIA. No proxy gap exists between measurements and claims."
       }
     },
     "setup_transparency": {
       "model_versions_specified": {
         "applies": true,
         "answer": true,
         "justification": "Since the paper trains its own neural networks rather than using pre-trained models, the relevant specification is the architecture. Tables AI and AII in Appendix A provide detailed architecture specifications: layer counts, neuron counts per layer, learning rates, and code dimensions for both AE-BCV and AE-GAN."
       },
       "prompts_provided": {
         "applies": false,
         "answer": false,
         "justification": "The paper does not use any prompting. It trains custom neural networks (auto-encoders, GANs) from scratch."
       },
       "hyperparameters_reported": {
         "applies": true,
         "answer": true,
         "justification": "Table AII in Appendix A reports learning rates (0.0001 for encoder/decoder, 0.0005 for discriminator), code dimension (120), number of hidden layers (2 each), and neurons per hidden layer (1000/1000/500). Localization threshold τ₁ values are also provided (0.01 for 118-bus, 0.5 for 415-bus)."
       },
       "scaffolding_described": {
         "applies": false,
         "answer": false,
         "justification": "No agentic scaffolding is used. This is a standard deep learning pipeline with neural network training and inference."
       },
       "data_preprocessing_documented": {
         "applies": true,
         "answer": true,
         "justification": "Section VI-B documents data generation: Monte-Carlo simulations with configurable line switching, bus power injection varied 50-150% of base case, white noise perturbation at 1% variance. Training uses 5% line switching, test uses up to 8%. 80/10/10 split. FDIA samples generated by optimization models (8) and (9). They note no additional normalization since data is in per-unit values."
       }
     },
     "limitations_and_scope": {
       "limitations_section_present": {
         "applies": true,
         "answer": false,
         "justification": "There is no dedicated limitations or threats-to-validity section. The conclusion (Section VII) mentions future research directions (AC power flow, improved localization efficiency, dynamic SE) but does not substantively discuss limitations of the current work."
       },
       "threats_to_validity_specific": {
         "applies": true,
         "answer": false,
         "justification": "No specific threats to validity are discussed. The paper does not address potential issues such as the gap between synthetic and real-world attack data, the DC power flow simplification's impact on results, or the scalability concerns beyond 1354-bus."
       },
       "scope_boundaries_stated": {
         "applies": true,
         "answer": true,
         "justification": "The paper explicitly states: 'For ease of verification and a moderate research scope, we focus on the DC SE model with complete information' (Section II). It also notes 'we focus on the defense of static FDIA targeting on a single snapshot' (Section I), and identifies AC power flow, dynamic SE, and asymmetric information as out of scope (Section VII)."
       }
     },
     "data_integrity": {
       "raw_data_available": {
         "applies": true,
         "answer": false,
         "justification": "No raw data is released. The datasets are generated from MATPOWER simulations but are not made available for download or independent verification."
       },
       "data_collection_described": {
         "applies": true,
         "answer": true,
         "justification": "Section VI-B describes data generation in detail: IEEE benchmark systems from MATPOWER, Monte-Carlo simulations with line switching, power injection variation (50-150%), white noise perturbation (1% variance), FDIA cases from optimization models (8) and (9), and the dataset split ratios."
       },
       "recruitment_methods_described": {
         "applies": false,
         "answer": false,
         "justification": "No human participants. Data comes from simulated IEEE benchmark power systems (standard benchmarks) and a synthesized 415-bus CSG system."
       },
       "data_pipeline_documented": {
         "applies": true,
         "answer": true,
         "justification": "The full pipeline is documented: base cases from MATPOWER power flow → Monte-Carlo topology variation → power injection perturbation → FDIA injection via CVX optimization → train/validation/test split (80/10/10) with different topology change percentages. Table I documents the features and labels for each model."
       }
     },
     "conflicts_of_interest": {
       "funding_disclosed": {
         "applies": true,
         "answer": true,
         "justification": "Funding is disclosed: 'This work was supported in part by the National Natural Science Foundation of China (No. 51767001).'"
       },
       "affiliations_disclosed": {
         "applies": true,
         "answer": true,
         "justification": "Author affiliations are clearly listed. Huang, Qin, and Liu are from Guangxi University; Xie and Meng are from Guangxi Power Grid Co. Ltd., China Southern Grid. They are not evaluating a commercial product they sell."
       },
       "funder_independent_of_outcome": {
         "applies": true,
         "answer": true,
         "justification": "The National Natural Science Foundation of China is a government research funding agency with no financial interest in the outcome of this specific study on FDIA defense methods."
       },
       "financial_interests_declared": {
         "applies": true,
         "answer": false,
         "justification": "No competing interests or financial interests statement is present in the paper. Two authors work for Guangxi Power Grid (CSG), which could have interests in FDIA defense technologies, but this is not discussed."
       }
     },
     "contamination": {
       "training_cutoff_stated": {
         "applies": false,
         "answer": false,
         "justification": "The paper trains its own neural networks from scratch on synthetic power system data. It does not evaluate a pre-trained model's capability on any benchmark."
       },
       "train_test_overlap_discussed": {
         "applies": false,
         "answer": false,
         "justification": "The paper trains custom models, not pre-trained models. Train/test separation is handled by design (different topological change percentages), but the contamination concern about pre-trained model knowledge does not apply."
       },
       "benchmark_contamination_addressed": {
         "applies": false,
         "answer": false,
         "justification": "No pre-trained models are used. The paper generates its own training and test data from power system simulations, so benchmark contamination of pre-trained models is not applicable."
       }
     },
     "human_studies": {
       "pre_registered": {
         "applies": false,
         "answer": false,
         "justification": "No human participants in this study. It is purely a simulation-based evaluation on power system benchmarks."
       },
       "irb_or_ethics_approval": {
         "applies": false,
         "answer": false,
         "justification": "No human participants. The study uses simulated power system data."
       },
       "demographics_reported": {
         "applies": false,
         "answer": false,
         "justification": "No human participants."
       },
       "inclusion_exclusion_criteria": {
         "applies": false,
         "answer": false,
         "justification": "No human participants."
       },
       "randomization_described": {
         "applies": false,
         "answer": false,
         "justification": "No human participants."
       },
       "blinding_described": {
         "applies": false,
         "answer": false,
         "justification": "No human participants."
       },
       "attrition_reported": {
         "applies": false,
         "answer": false,
         "justification": "No human participants."
       }
     },
     "cost_and_practicality": {
       "inference_cost_reported": {
         "applies": true,
         "answer": true,
         "justification": "Table XI reports computation times: classification time is 0.011s (118-bus) and 0.013s (415-bus), localization and recovery time is 14.69s (118-bus) and 25.47s (415-bus). This demonstrates the method is efficient for online defense."
       },
       "compute_budget_stated": {
         "applies": true,
         "answer": true,
         "justification": "Table XI reports training times: 441.56s and 594.44s for detection training, 324.85s and 406.39s for AE-GAN training. Hardware is specified in Section VI-A: 16 GB RAM, Intel i7-8750H CPU, Nvidia RTX 2070 GPU."
       }
     },
     "experimental_rigor": {
       "seed_sensitivity_reported": {
         "applies": true,
         "answer": false,
         "justification": "No mention of multiple random seeds or seed sensitivity analysis. All results appear to be from single training runs."
       },
       "number_of_runs_stated": {
         "applies": true,
         "answer": false,
         "justification": "The number of neural network training runs is not stated. While 1100 attack simulations are mentioned for BDD testing (Section VI-C), the ML model training/evaluation does not state how many runs produced the reported results."
       },
       "hyperparameter_search_budget": {
         "applies": true,
         "answer": false,
         "justification": "No hyperparameter search budget is reported. The architecture and hyperparameters appear in Appendix A but how they were selected is not discussed."
       },
       "best_config_selection_justified": {
         "applies": true,
         "answer": false,
         "justification": "No explanation of how the neural network architecture (14 hidden layers, 200 neurons, etc.) or hyperparameters were selected. Only the final configuration is presented."
       },
       "multiple_comparison_correction": {
         "applies": true,
         "answer": false,
         "justification": "The paper makes numerous comparisons across 3 systems, 3 attack levels, 2 attack types, and 3 methods, but performs no statistical tests at all, let alone corrections for multiple comparisons."
       },
       "self_comparison_bias_addressed": {
         "applies": true,
         "answer": false,
         "justification": "The authors implement their own SVM and ANN baselines and compare against them. They do not acknowledge the bias of evaluating their own system against their own re-implementations of baselines."
       },
       "compute_budget_vs_performance": {
         "applies": true,
         "answer": false,
         "justification": "While computation times are reported (Table XI), there is no analysis of performance as a function of compute budget, and no comparison of compute requirements across methods (SVM/ANN vs AE-BCV)."
       },
       "benchmark_construct_validity": {
         "applies": true,
         "answer": false,
         "justification": "The paper uses IEEE benchmark power systems (57-bus, 118-bus, 300-bus) and a synthesized 415-bus CSG system without discussing whether these benchmarks adequately represent real-world power grid conditions and attack scenarios."
       },
       "scaffold_confound_addressed": {
         "applies": false,
         "answer": false,
         "justification": "No scaffolding is involved. The paper evaluates standalone neural network models for FDIA detection/localization."
       }
     },
     "data_leakage": {
       "temporal_leakage_addressed": {
         "applies": true,
         "answer": true,
         "justification": "The paper explicitly addresses temporal-style leakage by ensuring test data contains unseen conditions: training uses up to 5% line switching while testing uses up to 8%, and the detector is trained only with SA attacks but tested with unseen MA and WA intensities (Section VI-D)."
       },
       "feature_leakage_addressed": {
         "applies": true,
         "answer": false,
         "justification": "No explicit discussion of whether the input features (measurement vector z, state estimate x̂, Jacobian matrix H) could leak information about whether data is attacked. The detector receives both z and x̂ as input, but potential information leakage through this combination is not discussed."
       },
       "non_independence_addressed": {
         "applies": true,
         "answer": true,
         "justification": "The paper ensures non-trivial train-test independence by using different topological change percentages (5% train vs 8% test) and by withholding MA and WA attack types from training. Section VI-D explicitly describes this design choice."
       },
       "leakage_detection_method": {
         "applies": true,
         "answer": false,
         "justification": "No concrete leakage detection method is applied. The train-test separation is enforced by design (different topology percentages) but no formal verification or detection of potential leakage is conducted."
       }
     }
   },
   "claims": [
     {
       "claim": "The AE-BCV detector achieves over 95% detection accuracy on IEEE 57-bus, 118-bus, and 415-bus CSG systems.",
       "evidence": "Table V shows correct rates of 96.7% (57-bus), 98.5% (118-bus), and 97.5% (415-bus). Tables VI-VIII confirm these results hold across various attack types and intensities.",
       "supported": "moderate"
     },
     {
       "claim": "AE-BCV outperforms SVM and ANN detectors, especially for unseen weak attacks.",
       "evidence": "Tables VI-VIII show AE-BCV consistently outperforms SVM and ANN. For unseen WA on 57-bus: AE-BCV 90.8% vs SVM 50.8% vs ANN 60.4% (Table VIII). The advantage expands as attack intensity decreases.",
       "supported": "moderate"
     },
     {
       "claim": "The proposed attack model can affect more state variables with fewer compromised meters than the conventional model.",
       "evidence": "Table IV: for IEEE 118-bus, untargeted FDIA affects 117 states with 33-34 meters vs conventional model affecting 10 states with 60-140 meters.",
       "supported": "strong"
     },
     {
       "claim": "FDIA localization achieves over 80% correct rate across attack intensities.",
       "evidence": "Table IX shows correct rates of 80.37-84.96% (IEEE 118-bus) and 81.28-85.69% (CSG 415-bus) across SA, MA, and WA.",
       "supported": "moderate"
     },
     {
       "claim": "The recovery algorithm significantly reduces measurement error caused by FDIA.",
       "evidence": "Table X shows mean |a/z| error reduction from 16.50 to 0.85 (SA, 118-bus), 5.70 to 0.61 (MA), and 3.15 to 0.46 (WA). Similar reductions for 415-bus system.",
       "supported": "moderate"
     },
     {
       "claim": "The proposed FDIA models bypass BDD completely.",
       "evidence": "Table III shows the residual differences caused by attacks (order of 10⁻⁸) are far less than normal operational residuals (2.570 for 118-bus, 0.128 for 300-bus), indicating BDD cannot detect them.",
       "supported": "strong"
     }
   ],
   "red_flags": [
     {
       "flag": "No error bars or uncertainty quantification",
       "detail": "All results across Tables V-X are point estimates from apparently single training runs. No standard deviations, confidence intervals, or multi-seed results are reported, making it impossible to assess result stability."
     },
     {
       "flag": "Weak and outdated baselines",
       "detail": "The primary comparisons are against basic SVM and ANN classifiers, which are not competitive deep learning baselines for 2022. The paper acknowledges prior deep learning methods [13], [14] but only provides a rough, uncontrolled comparison."
     },
     {
       "flag": "No ablation study for multi-component system",
       "detail": "The methodology has multiple components (AE feature extractor, BCV classifier, AE-GAN, pattern match algorithm) but no ablation study isolates the contribution of each component."
     },
     {
       "flag": "All evaluation on synthetic data",
       "detail": "All experiments use simulated power system data generated from MATPOWER. No real-world power grid data is used for validation, raising questions about whether results transfer to actual grid conditions with real measurement noise and attack patterns."
     },
     {
       "flag": "Scalability concerns partially addressed",
       "detail": "Appendix A briefly tests a 1354-bus system but only for detection (without topological changes), not for localization or recovery. The scalability of the full methodology to large real grids remains uncertain."
     }
   ],
   "cited_papers": [
     {
       "title": "Generative adversarial nets",
       "authors": ["I. J. Goodfellow", "J. Pouget-Abadie", "M. Mirza"],
       "year": 2014,
       "relevance": "Foundational paper on generative adversarial networks, a core ML architecture used in this paper's AE-GAN model for generating candidate measurement distributions."
     },
     {
       "title": "Autoencoder-based network anomaly detection",
       "authors": ["Z. Chen", "C. K. Yeo", "B. S. Lee"],
       "year": 2018,
       "relevance": "Describes autoencoder-based anomaly detection, directly relevant to AI-based security systems and the AE feature extractor used in this work."
     },
     {
       "title": "A survey on the detection algorithms for false data injection attacks in smart grids",
       "authors": ["A. S. Musleh", "G. Chen", "Z. Y. Dong"],
       "year": 2020,
       "relevance": "Comprehensive survey on FDIA detection algorithms including ML-based approaches, relevant to understanding the AI security detection landscape."
     },
     {
       "title": "Real-time detection of false data injection attacks in smart grid: a deep learning-based intelligent mechanism",
       "authors": ["Y. He", "G. J. Mendis", "J. Wei"],
       "year": 2017,
       "relevance": "Early deep learning approach for FDIA detection using conditional deep belief networks, demonstrating AI for cybersecurity in critical infrastructure."
     },
     {
       "title": "Online false data injection attack detection with wavelet transform and deep neural networks",
       "authors": ["J. Yu", "Y. Hou", "V. Li"],
       "year": 2018,
       "relevance": "Deep neural network approach for online attack detection combining signal processing with deep learning for real-time security."
     },
     {
       "title": "False data injection attacks against state estimation in electric power grids",
       "authors": ["Y. Liu", "M. K. Reiter", "P. Ning"],
       "year": 2011,
       "relevance": "Foundational paper on FDIA that first identified the vulnerability of state estimation to undetectable cyber-attacks, spawning the entire research area."
     },
     {
       "title": "Online generative adversary network based measurement recovery in false data injection attacks: a cyber-physical approach",
       "authors": ["Y. Li", "Y. Wang", "S. Hu"],
       "year": 2020,
       "relevance": "Uses GANs for measurement recovery from FDIA, directly comparable approach to this paper's AE-GAN for cyber-physical security."
     },
     {
       "title": "A survey on power grid cyber security: from component-wise vulnerability assessment to system-wide impact analysis",
       "authors": ["X. Huang", "Z. Qin", "H. Liu"],
       "year": 2018,
       "relevance": "Survey by the same authors on power grid cybersecurity covering AI-based defense mechanisms and vulnerability assessment."
     }
   ],
   "engagement_factors": {
     "practical_relevance": {
       "score": 1,
       "justification": "The methodology addresses a real power grid security concern but is limited to DC power flow and requires substantial customization for deployment."
     },
     "surprise_contrarian": {
       "score": 0,
       "justification": "Applies standard deep learning methods (AE, GAN) to a known problem (FDIA defense) without challenging any conventional wisdom."
     },
     "fear_safety": {
       "score": 1,
       "justification": "Addresses power grid cyber-attacks and demonstrates an enhanced attack model, but this is a niche domain-specific concern rather than a broad AI safety issue."
     },
     "drama_conflict": {
       "score": 0,
       "justification": "No controversy, no competing claims, straightforward technical contribution."
     },
     "demo_ability": {
       "score": 0,
       "justification": "No code, demo, or tool is released. The methodology requires MATPOWER simulations and custom neural network training."
     },
     "brand_recognition": {
       "score": 0,
       "justification": "From Guangxi University and Guangxi Power Grid, not widely known institutions in the AI research community."
     }
   }
 }