Transfer learning mechanism to account performance degradation in gas turbines with limited operational data

Abstract

This paper investigates the transfer learning mechanism to improve the prediction accuracy of the energy system model. Artificial intelligence techniques are increasingly being adopted in the energy domain to predict energy system characteristics and performance. However, in many energy systems, the relationship between interested variables and their distributions differs (data and concept drift) with time due to system degradation and aging. There is a requirement for re-training and re-testing AI models to ensure reliable performance over time, which may require extensive latest operational data. Transfer learning helps to confront this challenge by leveraging valuable knowledge from a pre-trained model and reducing the requirement of new operational data significantly. To address these issues, this paper focuses on a gas turbine, a critical energy system widely deployed in diverse applications, and shows performance degradation over its lifetime. The energy model of the gas turbine is a multivariate type that predicts energy efficiency, fuel consumption, and heat energy based on power setpoints and weather conditions. This paper examined transfer learning mechanisms that can capture the latest characteristics of the gas turbine and their effects on prediction accuracy. The developed transfer learning model predicts fuel consumption accurately above 99%, whereas the pre-trained model under-predicts up to 4%, which may lead to suboptimal operation decisions when employed in the scheduling algorithm. Some of the other targets, such as heat energy, show marginal drift, as expected from the gas turbine characteristics. The knowledge gained from the transfer learning mechanism and its efficacy boost assists operational decisions, which helps improve energy efficiency and cost savings.