Machine Learning for Performance and Fault Detection in Thermal Power Plants

Thermal power plants (TPPs) are a critical component of global electricity generation, yet they often suffer fromefficiency loss and unplanned outages due to equipment faults. Traditional maintenance strategies (reactive or preventive) areoften too slow or costly. In contrast, machine learning (ML) methods can analyze large historical and real-time sensor data todetect anomalies and predict failures early. This paper surveys supervised methods (SVM, random forests, neural networks),unsupervised models (autoencoders, clustering, PCA), and hybrid physics-ML approaches for TPP monitoring. It also examinessensor optimization and IoT-enabled real-time monitoring. Case examples from the literature show that ML-based predictivemaintenance can greatly reduce unplanned downtime and maintenance costs (e.g., cutting costs by roughly 20–40%) whileimproving equipment availability. The findings indicate that optimized sensor networks, integrated IoT data, and advanced MLmodels can substantially enhance fault detection accuracy and overall plant efficiency.