Comparative studies among machine learning models for performance estimation and health monitoring of thermal power plants

Abstract

Estimating the performance of base load combined cycle power plants and detecting early-stage malfunctions in equipment and processes is a difficult task that depends on complex thermodynamics. Herein, we demonstrate the efficacy of several machine learning methods in by-passing physics-based models to reliably estimate performance and detect anomalies in a representative combined cycle power plant with five years of recorded data. We model the full load power output of the plant by using ambient temperature, atmospheric pressure, relative humidity and exhaust vacuum pressure as input features using linear regression, support vector machines, random forests and artificial neural networks. Our results show that all the models estimate the power output with reasonable R2 accuracy (>92%), while random forests perform the best (~96%) using less than half of the ~10,000 datapoints collected from the field. Finally, we show that unsupervised anomaly detection algorithms such as elliptical envelopes and isolation forests can be potential game changers for non-destructive health monitoring of equipment via identifying obscure sparse synthetic anomalies through investigating merely 1.5% of the dataset. This work presents a data science approach that can take advantage of the subtle interdependencies among the sensor data in power plants and extract useful insights which are unintelligible to humans. The methods presented here help in enabling better control over everyday operations and monitoring and reliable forecasting of hourly energy output.