Machine learning-based predictive maintenance system for urban heating networks for real-time failure detection and analysis

This study presents a comprehensive framework for predictive maintenance of urban heat supply networks utilizing advanced machine learning algorithms. The primary objective is to enable early detection of potential system failures, thereby improving operational reliability and minimizing unplanned downtimes. A synthetically generated dataset of 10,000 records was employed, simulating real – world operational parameters such as temperature, pressure, flow rate, and vibration, sampled at 5–minute intervals to replicate actual monitoring conditions. Data preprocessing involved outlier removal using the interquartile range (IQR) method, normalization through Min-Max scaling, and imputation of missing values, ensuring data quality and consistency. Feature importance was further analyzed using SHAP values to enhance interpretability and identify critical predictors influencing system behavior. Five machine learning models – Logistic Regression, Support Vector Machine (SVM), Random Forest, Artificial Neural Networks (ANN), and Gradient Boosting (LightGBM) – were implemented and evaluated using 10 – fold cross – validation. The Gradient Boosting model demonstrated superior performance, achieving an accuracy of 99.9%, F1-score of 0.999, ROC-AUC of 1.0, and LogLoss of 0.004. Logistic Regression and Random Forest also performed well (AUC = 1.0, F1 = 0.999), whereas SVM and ANN exhibited limited predictive capabilities (AUC ≈ 0.50, F1 = 0.038 and 0.632, respectively). These results underscore the robustness of Gradient Boosting in modeling complex nonlinear relationships and its applicability for real-time anomaly detection in heating systems. The proposed framework holds significant practical potential for integration into existing monitoring infrastructures, facilitating proactive maintenance planning, optimizing resource allocation, and reducing operational costs. Future research will focus on validating the approach with real – world datasets and exploring hybrid machine learning architectures to enhance model generalizability and resilience.