Development of a machine learning-based framework for predicting failures in heat supply networks

Abstract

The increasing complexity and scale of heat supply systems leads to a higher risk of failures, which may cause significant economic and environmental consequences. This study develops a predictive mathematical framework for the early detection of emergency conditions in heat supply networks (HSNs) using machine learning (ML). The proposed approach is based on the LightGBM gradient boosting (GB) algorithm, chosen for its high accuracy and efficiency in handling large datasets. Real operational data (temperature, pressure, flow, and vibration) were considered. Data preprocessing, feature engineering (including SHAP analysis), and hyperparameter tuning with grid search and 5-fold cross-validation improved prediction quality. The model achieved accuracy of 85%, F1-score of 0.82, and receiver operating characteristic (ROC)-area under the curve (AUC) of 0.96, outperforming logistic regression (LR) and decision trees. The framework may be integrated into monitoring systems for predictive maintenance, reducing downtime and optimizing costs.