Fusion of Random Forest and SHAP for Distribution Transformer Fault Diagnosis Modeling

The distribution transformer is a critical component of the power system, and its stable operation directly affects system reliability. Addressing its fault diagnosis is essential for maintaining system stability. However, current model-based diagnostic methods often lack interpretability, which hampers the in-depth understanding and accurate localization of fault causes. Thus, there is an urgent need for more efficient and interpretable fault diagnosis techniques to improve the maintenance and management of distribution transformers. To this end, this paper proposes a fault diagnosis model for distribution transformers that utilizes Random Forest and SHAP (Shapley Additive Explanations). The model incorporates current, voltage, and other circuit indicators as input features, develops a Random Forest-based fault prediction model, and employs a cost-sensitive function to address data imbalance issues. Comparative analysis with other classification algorithms (e.g., SVM, Catboost) demonstrates that the parameter-optimized Random Forest model offers superior balanced accuracy. Further, after the introduction of the cost-sensitive function, the balancing accuracy of the model achieves a significant improvement, which verifies the effectiveness of the method in dealing with unbalanced data. Additionally, this paper performs a thorough interpretability analysis of the proposed diagnostic model using the SHAP methodology, which elucidates key factors influencing distribution transformer faults by ranking feature importance. This analysis not only improves the model's transparency and credibility but also provides valuable decision support for engineers and technicians, aiding in the rapid and accurate identification of fault causes and optimization of maintenance and management strategies.