Suspicious Findings in 'Power IoT Operational Risk Prediction via Cyber-Physical-Social Data Fusion' (DOI: 10.19635/j.cnki.csu-epsa.001207)

Academic fraud report · Geng Detector

Summary

This report assesses a 2023 paper by Li Guoqiang et al. published in the Journal of Power Systems and Automation (网络首发: 2023-02-28). The overall verdict is highly suspicious. Three core issues were confirmed: (1) Figure 9 reportedly displays the Bayesian optimization trajectory as an unrealistically perfect periodic triangle wave, which is inconsistent with Gaussian-process-based sampling that should produce stochastic exploration; (2) the prose describing Table 7 references 'x-axis' and 'y-axis' although Table 7 is presented as a tabular layout, strongly suggesting copy-paste from a bar-chart caption or AI-generated template misuse; (3) Table 9 lists the default iterations parameter as 1000, while the surrounding text in Section 4.5 states the default is 500—an internal contradiction pointing to parameter fabrication. Two further anomalies (overly smooth training times in Table 6 and suspiciously uniform ADASYN sample counts in Figure 6) are flagged but currently lack sufficient grounding. Confidence in the confirmed findings is high, though a formal investigation with raw code and logs is required for any official determination of misconduct.

Verdict

Highly suspicious. Three confirmed fabrications/inconsistencies, two additional anomalies pending further evidence.

Key findings

Figure 9 (Section 4.5) — Bayesian optimization curves violate algorithm logic: subplots exhibit a perfectly periodic triangular-wave oscillation, inconsistent with stochastic Gaussian-process sampling. Severity: 🔴 / Confirmed.
Table 7 (Section 4.4) — Mismatched caption: descriptive prose references "x-axis" and "y-axis" for what is described as a table, indicating copy-paste or template abuse. Severity: 🔴 / Confirmed.
Table 9 (Section 4.5) — Internal contradiction on default iterations: table lists 1000; narrative text immediately below states 500. Severity: 🟠 / Confirmed.
Table 6 (Section 4.4) — Suspiciously smooth training times: values (≈104s → 167s → 181s → 200s → 266s) form an almost linear, jitter-free progression atypical of real ML runtime measurements. Severity: 🟠 / Pending evidence.
Figure 6 (Section 4.3) — ADASYN counts overly uniform: post-oversampling class counts (12391, 12403, 12401, …) differ by only single digits, unusually low variance for synthetic time-series-style data. Severity: 🟡 / Pending evidence.

Evidence highlights

DOI: 10.19635/j.cnki.csu-epsa.001207
Title: 信息物理社会数据融合处理的电力物联网运行风险预测 / *Power IoT Operational Risk Prediction via Cyber-Physical-Social Data Fusion*
Authors: Li Guoqiang, Wang Chong, Gao Xiuzhi, Wang Hua, Xu Qi, Li Lingcong
Journal: 电力系统及其自动化学报 (*Journal of Power Systems and Automation*), 2023
Figure 9: subplots show regular triangular/zig-zag pattern instead of stochastic Bayesian-optimization trajectory.
Table 7 caption mismatch: prose says "横坐标表示特征对分类的贡献值，纵坐标为筛选出的特征" while Table 7 is tabular (68 columns of feature contributions).
Table 9 / Section 4.5 text: iterations default = 1000 in table vs. 500 in body.
Table 6 training times (illustrative): 104s, 167s, 181s, 200s, 266s — near-linear progression across feature counts.
Figure 6 ADASYN counts: values such as 12391, 12403, 12401, indicating very low variance.

Notes

Findings 1–3 are well-supported by visible discrepancies in the published PDF and are unlikely to be mere typographical errors.
Findings 4–5 (Table 6 and Figure 6) are flagged as suspicious but require access to source code, raw logs, or replication to escalate to confirmed status; they may reflect post-processed reporting rather than fabrication.
Affiliations reportedly include Northeast Electric Power University and State Grid East Inner Mongolia Electric Power Co., Ltd.
Recommended follow-up: request raw Bayesian-optimization logs and source code from the authors; raise concerns on PubPeer; notify the journal editorial office for data-integrity review.