Shapley Additive Explanations for Machine Learning-based Decision Support Systems for Power Transformers’ Diagnostics

Technical diagnostics of high-voltage equipment is a relevant task to reduce the risks of accidents at power facilities and, consequently, to ensure reliable power supply to consumers. Currently, a lot of research is aimed at improving the accuracy of diagnostics of high-voltage equipment, in particular, power transformers using machine learning methods. However, the application of machine learning leads to the construction of diagnostic systems that are a "black box" for the user, if the methods of explainable artificial intelligence are applied. This paper proposes the application of Shapley additive explanations algorithm to increase the interpretability of ensemble machine learning models in assessing the technical state of a power transformer using the proposed modified Shapley additive explanations method. An open dataset containing the results of transformer oil sampling was used. The difference of the paper lies in the modification of Shapley additive explanations, aimed at increasing the informativeness of visualization of interpretations of the results of machine learning models. The proposed method makes it possible to increase the validity of decisions made in the technical diagnostics of high-voltage equipment using machine learning models, which will allow for more effective implementation of intelligent decision support systems in the power industry.