All Issue

2026 Vol.46, Issue 5 Preview Page
Three-State Classification for Photovoltaic Module Anomaly Detection Using Peer-Relative Criteria and Temporal Persistence

이웃 상대 기준과 지속성 기반 태양광 모듈 이상 탐지를 위한 3상태 분류

Su-cheon Lee1
,
Tae-geol Woo1
,
Se-hun Lee2
,
Jei Park3
,
Kang-moon Park4*
이 수천1
,
우 태걸1
,
이 세훈2
,
박 제이3
,
박 강문4*
1Ph.D. Student, Department of Electronic Engineering, Korea National University of Transportation
2M.S. Student, Department of Electronic Engineering, Korea National University of Transportation
3Bachelor’s and Master’s Integrated Course, Department of Electronic Engineering, Korea National University of Transportation
4Associate Professor, Department of Electronic Engineering, Korea National University of Transportation
1한국교통대학교 전자공학과, 박사과정
2한국교통대학교 전자공학과, 석사과정
3한국교통대학교 전자공학과, 학·석통합과정
4한국교통대학교 전자공학과, 부교수
*Corresponding Author
30 October 2026. pp. 51-64
PDF XML Citation
Abstract

Module-level anomalies in photovoltaic plants may be obscured by aggregate indicators, with maintenance labels often being unavailable. We define Normal, LowPower, and Thermal Anomaly weak labels using same-string peer deviations and two-sample persistence. Thermal Anomaly denotes peer-relative overtemperature, not a thermographically confirmed cell-level hotspot. We analyzed 15 months of 15-min measurements collected from 349 modules. A 55℃ comparison threshold flagged 347 modules at least once and yielded no winter detections, whereas the peer-relative criterion reduced the number of flagged modules to 143 and mitigated seasonal bias. Module-wise five-fold cross-validation prevented identity leakage, and we evaluated 16 configurations combining directionality, Self-Attention, position-wise feed-forward blocks, and temporal pooling. Across 976,911 pooled out-of-fold windows, Macro F1 was 0.9837, and the fold mean was 0.9544 ± 0.0373. Our findings reveal weak-label consistency on unseen modules, not accuracy against field-confirmed faults. The unidirectional Long Short-Term Memory (LSTM) with mean pooling underperformed, whereas changing any one factor recovered most of the performance. Moreover, other configurations did not differ substantially, supporting deployment-specific complexity selection.

Keywords
Photovoltaic Module
Anomaly Detection
Weak Supervision
Peer Comparison
Time-series Classification
Persistence
키워드
태양광 모듈
이상 탐지
약지도
이웃 비교
시계열 분류
지속성
References
1

Buerhop, C., Pickel, T., Hauch, J., and Peters, I. M., Assessment of String Performance Using Self-Referencing Method in Comparison to Performance Ratio, Progress in Photovoltaics: Research and Applications, Vol. 31, No. 12, pp. 1329-1335, 2023, https://doi.org/10.1002/pip.3649.

10.1002/pip.3649
2

Geirhos, R., Jacobsen, J.-H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., and Wichmann, F. A., Shortcut Learning in Deep Neural Networks, Nature Machine Intelligence, Vol. 2, No. 11, pp. 665-673, 2020, https://doi.org/10.1038/s42256-020-00257-z.

10.1038/s42256-020-00257-z
3

Leloux, J., Narvarte, L., Desportes, A., and Trebosc, D., Performance to Peers (P2P): A Benchmark Approach to Fault Detections Applied to Photovoltaic System Fleets, Solar Energy, Vol. 202, pp. 522-539, 2020, https://doi.org/10.1016/j.solener.2020.03.015.

10.1016/j.solener.2020.03.015
4

Saeb, S., Lonini, L., Jayaraman, A., Mohr, D. C., and Kording, K. P., The Need to Approximate The Use-Case in Clinical Machine Learning, GigaScience, Vol. 6, No. 5, gix019, 2017, https://doi.org/10.1093/gigascience/gix019.

10.1093/gigascience/gix01928327985PMC5441397
5

Hochreiter, S. and Schmidhuber, J., Long Short-Term Memory, Neural Computation, Vol. 9, No. 8, pp. 1735-1780, 1997, https://doi.org/10.1162/neco.1997.9.8.1735.

10.1162/neco.1997.9.8.1735
6

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., and Polosukhin, I., Attention Is All You Need, Advances in Neural Information Processing Systems, Vol. 30, pp. 5998-6008, 2017.

7

Graves, A. and Schmidhuber, J., Framewise Phoneme Classification with Bidirectional LSTM and Other Neural Network Architectures, Neural Networks, Vol. 18, No. 5-6, pp. 602-610, 2005, https://doi.org/10.1016/j.neunet.2005.06.042.

10.1016/j.neunet.2005.06.042
8

Zhang, X., Jing, B., Jiao, X., and Yao, R., Advanced Multi-Scale CNN-BiLSTM for Robust Photovoltaic Fault Detection, Sensors, Vol. 25, No. 14, p. 4474, 2025, https://doi.org/10.3390/s25144474.

10.3390/s2514447440732602PMC12300933
9

Wang, S., Li, Y., Liu, Z., and Liu, X., Intelligent Faulty Diagnosis on Photovoltaic Farm Based on Spatial Attention Mechanism and Temporal Convolutional Network, Next Research, Vol. 2, No. 3, 100604, 2025, https://doi.org/10.1016/j.nexres.2025.100604.

10.1016/j.nexres.2025.100604
10

Lodhi, E. and Qiu, L., AEConvNeXt: an Attention-Enhanced ConvNeXt Framework for Imbalanced Photovoltaic Fault Classification with Explainable Feature Analysis, AI, Vol. 7, No. 6, 182, 2026, https://doi.org/10.3390/ai7060182.

10.3390/ai7060182
11

Lin, T.-Y., Goyal, P., Girshick, R., He, K., and Dollár, P., Focal Loss for Dense Object Detection, 2017 IEEE International Conference on Computer Vision (ICCV), pp. 2999-3007, October 2017, Venice, Italy, https://doi.org/10.1109/ICCV.2017.324.

10.1109/ICCV.2017.324
12

Nassreddine, G., El Arid, A., Nassereddine, M., and Al Khatib, O., Fault Detection and Classification for Photovoltaic Panel System Using Machine Learning Techniques, Applied AI Letters, Vol. 6, No. 2, e115, 2025, https://doi.org/10.1002/ail2.115.

10.1002/ail2.115
13

Ahmed, W., Thermal and Chromatic Analysis for Scalable Photovoltaic Hotspot Detection, Solar Energy, Vol. 306, 114227, 2026, https://doi.org/10.1016/j.solener.2025.114227.

10.1016/j.solener.2025.114227
14

Severiano, B. M., Duran, E., Rispler, J., Guerrero Orbe, J., Liu, Y., Rougieux, F., Bruce, A., Yildiz, B., Martell, C., and Ibrahim, I. A., A Robust Rule-Based Method for Detecting and Classifying Underperformance in Photovoltaic Systems Using Inverter Data, Solar Energy, Vol. 308, 114382, 2026, https://doi.org/10.1016/j.solener.2026.114382.

10.1016/j.solener.2026.114382
15

Varma, P. and Ré, C., Snuba: Automating Weak Supervision to Label Training Data, Proceedings of the VLDB Endowment, Vol. 12, No. 3, pp. 223-236, 2018, https://doi.org/10.14778/3291264.3291268.

10.14778/3291264.329126831777681PMC6879381
16

Alcañiz, A., Nikam, M. M., Snow, Y., Isabella, O., and Ziar, H., Photovoltaic System Monitoring and Fault Detection Using Peer Systems, Progress in Photovoltaics: Research and Applications, Vol. 30, No. 9, pp. 1072-1086, 2022, https://doi.org/10.1002/pip.3558.

10.1002/pip.3558
17

Wu, R. and Keogh, E. J., Current Time Series Anomaly Detection Benchmarks Are Flawed and Are Creating the Illusion of Progress (Extended Abstract), 2022 IEEE 38th International Conference on Data Engineering (ICDE), pp. 1479-1480, May 2022, Kuala Lumpur, Malaysia, https://doi.org/10.1109/ICDE53745.2022.00116.

10.1109/ICDE53745.2022.00116
18

Kim, S., Choi, K., Choi, H.-S., Lee, B., and Yoon, S., Towards a Rigorous Evaluation of Time-Series Anomaly Detection, Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 36, No. 7, pp. 7194-7201, February 2022, Virtual, https://doi.org/10.1609/aaai.v36i7.20680.

10.1609/aaai.v36i7.20680
Information
  • Publisher :Korean Solar Energy Society
  • Publisher(Ko) :한국태양에너지학회
  • Journal Title :Journal of the Korean Solar Energy Society
  • Journal Title(Ko) :한국태양에너지학회 논문집
  • Volume : 46
  • No :5
  • Pages :51-64
  • Received Date : 2026-08-19
  • Revised Date : 2026-08-31
  • Accepted Date : 2026-09-02
  • DOI :https://doi.org/10.7836/kses.2026.46.5.051
Journal Informaiton Journal of the Korean Solar Energy Society Journal of the Korean Solar Energy Society
Online Submission
  • NRF
  • KOFST
  • crossref crossmark
  • crossref cited-by
  • crosscheck
  • orcid
  • open access
  • ccl
Journal Informaiton Journal Informaiton - close