Interpretable Machine Learning for Decision Support in Real-World Problem-Oriented Systems: Applying XGBoost and SHAP to Infrastructure Management

A methodology for selecting and applying models based on a comprehensive assessment of accuracy, interpretability, robustness, and computational efficiency is proposed. A demonstration case is presented using the XGBoost model to predict peak loads in distribution networks, with results interpreted using the SHAP method. It is demonstrated that the proposed approach provides not only high accuracy (RMSE = 5.08 kW) but also the transparency necessary for making informed decisions in real time. The article substantiates the applicability of a generalized model quality criterion and compares various algorithms based on real and formalized indicators. The obtained results confirm the relevance and potential of combining XGBoost and SHAP in infrastructure management tasks. Prospects for further research are related to the development of more universal frameworks for the automatic selection and configuration of intelligent models for POS, an expanded set of explainable metrics, and the formalization of ethical and regulatory requirements for AI algorithms in the infrastructure sector.