Explainable diagnosis of transformer faults in the Ivorian network by DGA: interpretable LightGBM approach and inter-site transfer
Authors: Gnoléba Célestin BOGUI, Konan Ferdinand GBAMELE, Zié YEO
DOI: doi.org/10.5281/zenodo.17060080
Page No: 1-14
Abstract
The integrity of electrical network transformers is crucial to the reliability and continuity of power supply in Côte d'Ivoire. Dissolved gas analysis (DGA), although widely used, faces interpretation limitations in tropical local contexts and in the face of diverse fault profiles. This article proposes an explainable diagnostic approach based on LightGBM, a high-performance gradient boosting algorithm, coupled with SHAP (SHapley Additive exPlanations), an interpretability tool. The model is trained on DGA data from transformers in the Ivorian fleet, using both raw variables (concentrations of H₂, CH₄, C₂H₂, C₂H₄, C₂H₆, CO, CO₂) and derived variables (ratios normalized according to IEC 60599:2022). Performance is compared to traditional methods (Duval's triangle, IEC 60599:2022 normative approach, Rogers' normative methods), as well as other AI models (XGBoost, SVM). The explainability provided by SHAP makes it possible to identify the most discriminating gases or ratios, establish interpretable thresholds for maintenance, and visualize local explanations, thereby facilitating decision-making in the field. In addition, the study assesses inter-site transferability by training the model on data from Abidjan South and testing it on other regions of the network. The results demonstrate a notable improvement in diagnostic accuracy, gains in transparency, and satisfactory robustness in the context of transfer, paving the way for an operational dashboard for proactive monitoring of the Ivorian transformer fleet.
