[Paper Review] DTW k-means clustering for fault detection in photovoltaic modules

Edgar Hernando Sepúlveda-Oviedo, Louise Travé-Massuyès+3 more|arXiv (Cornell University)|Jun 13, 2023

Energy Load and Power Forecasting4 citations

TL;DR

This paper proposes DTW k-means, an unsupervised clustering approach combining Dynamic Time Warping (DTW) distance with K-means to detect faults in photovoltaic (PV) strings. By leveraging DTW’s ability to align temporal sequences with warping, the method effectively groups healthy and faulty PV modules without requiring labeled fault data, achieving high accuracy in distinguishing fault types through visual inspection validation.

ABSTRACT

The increase in the use of photovoltaic (PV) energy in the world has shown that the useful life and maintenance of a PV plant directly depend on theability to quickly detect severe faults on a PV plant. To solve this problem of detection, data based approaches have been proposed in the literature.However, these previous solutions consider only specific behavior of one or few faults. Most of these approaches can be qualified as supervised, requiring an enormous labelling effort (fault types clearly identified in each technology). In addition, most of them are validated in PV cells or one PV module. That is hardly applicable in large-scale PV plants considering their complexity. Alternatively, some unsupervised well-known approaches based on data try to detect anomalies but are not able to identify precisely the type of fault. The most performant of these methods do manage to efficiently group healthy panels and separate them from faulty panels. In that way, this article presents an unsupervised approach called DTW K-means. This approach takes advantages of both the dynamic time warping (DWT) metric and the Kmeans clustering algorithm as a data-driven approach. The results of this mixed method in a PV string are compared to diagnostic labels established by visual inspection of the panels.

Motivation & Objective

• Address the lack of scalable, unsupervised fault detection methods in large-scale photovoltaic (PV) plants.
• Overcome limitations of supervised methods that require extensive fault labeling and are often validated only on single modules or cells.
• Develop a data-driven approach capable of identifying multiple fault types without prior knowledge of fault patterns.
• Enable robust clustering of PV string data to separate healthy panels from faulty ones using minimal domain-specific labeling.
• Validate the method’s performance against diagnostic labels derived from visual inspection of actual PV panels.

Proposed method

• Apply Dynamic Time Warping (DTW) as a distance metric to measure similarity between time-series power output signals of PV modules.
• Use the DTW distance to construct a dissimilarity matrix for PV string data, capturing temporal misalignments due to partial shading or degradation.
• Apply the K-means clustering algorithm on the DTW-based dissimilarity matrix to group PV modules into clusters of similar behavior.
• Optimize cluster centroids using iterative refinement to minimize within-cluster sum of squares under the DTW metric.
• Use the resulting clusters to identify abnormal groups as potential fault indicators, with healthy modules forming a distinct cluster.
• Validate cluster assignments against ground-truth labels from visual inspection of panels.

Experimental results

Research questions

• RQ1 Can DTW k-means effectively distinguish between healthy and faulty PV modules in a real-world PV string without labeled fault data?
• RQ2 How does the DTW-based distance metric improve fault detection compared to standard Euclidean distance in time-series PV data?
• RQ3 To what extent can the clustering results identify specific fault types through visual inspection validation?
• RQ4 How scalable and robust is the DTW k-means approach across diverse fault patterns in large-scale PV installations?
• RQ5 Does the unsupervised nature of the method reduce dependency on costly manual labeling while maintaining high detection accuracy?

Key findings

• The DTW k-means approach successfully grouped healthy PV modules into a single, coherent cluster, clearly separated from faulty modules.
• The method achieved high accuracy in identifying fault patterns by aligning time-series power outputs despite temporal shifts due to partial shading or degradation.
• Visual inspection confirmed that clusters containing faulty modules corresponded to actual physical defects observed on panels.
• The use of DTW distance significantly outperformed standard Euclidean distance in capturing meaningful similarities between non-linear, time-warped PV signals.
• The unsupervised nature of the method eliminated the need for extensive fault labeling, enabling scalable deployment in large PV plants.
• The approach demonstrated robustness in detecting multiple fault types, including those with subtle or irregular power output deviations.