Development of ML Models for a Gas Turbine Provider

Summary

A global provider of gas turbine solutions wanted to cut the nitrogen oxide (NOx) emissions of its turbines by understanding how operating conditions — load, fuel composition, and more — drive emissions. INNERLUXES's data scientist analyzed the turbine telemetry and emissions data and, because the relationships were complex and the data noisy, built three complementary machine learning models — XGBoost, RNN, and LSTM — in two months, so the Client could test all three on its own turbine data and deploy the best fit to keep reducing its nitrogen footprint.

About the Client

The Client is a global provider of gas turbine solutions designed to reduce the environmental impact of fossil fuels.

Minimizing the Nitrogen Footprint of Gas Turbines

The Client wanted to minimize the NOx emissions of its popular gas turbines by determining their optimal operating conditions — for example, constant versus changing load and various fuel compositions. To find the dependencies between these factors and their effect on the turbines' nitrogen footprint, the Client decided to use machine learning and chose INNERLUXES for its data science and machine learning expertise with an oil and gas background.

Three ML Models for Three Kinds of Dependency

INNERLUXES's senior data scientist analyzed the data sets — turbine telemetry such as rotation speed, load, fuel consumption, and temperature, along with NOx emission data. Exploratory data analysis surfaced two challenges: likely time-series, non-linear, and long-term dependencies between telemetry and emissions, and a substantial volume of outliers and noisy data. Because the data was complex and diverse, INNERLUXES proposed building three models, each optimized for a particular dependency type, so the Client could test all three on its own turbine data and pick the best:

• XGBoost — captures complex non-linear relationships among variables, where, for example, a change in turbine temperature may not cause a proportional shift in emissions.
• Recurrent neural network (RNN) — optimal for time-series dependencies, such as how fluctuations in hourly fuel consumption affect NOx over time; cross-validation of the time-series data was used to avoid both overfitting and underfitting.
• Long short-term memory (LSTM) — an advanced RNN variant for long-term dependencies, such as how the turbine's air-pressure ratio at one moment influences NOx values at a later date.

All three models were optimized to handle the irregular and unexpected values caused by outliers and noisy data.

Three Models Delivered in Two Months

• Within two months, the Client received three ML models — XGBoost, RNN, and LSTM — for identifying complex dependencies between gas turbine operating conditions and their nitrogen emissions.
• The Client valued the comprehensive approach and the tailoring of the models to the turbine data's peculiarities.
• The Client deployed all three models in its IT infrastructure to continue nitrogen emissions research and reduce the nitrogen footprint of its gas turbine solutions.

Technologies and Tools

Python.