Advanced model-based control of B36:45 LNG engines based on data driven models using machine learning tools

Abstract

This work is about developing a data driven MPC control for optimization of fuel consumption by minimizing the heat rate of an LNG gas engine from Bergen Engines AS.

The model is based on real life data from an installed B36:45 gas engine in a power plant. The process data from the engine was used to develop a state space model of the process consisting of 2 controllable inputs, 3 measure disturbances and 6 measured outputs.

The goal was to use the global ignition timing and the charge air pressure set point as controllable outputs to minimize the heat rate while considering constraints on the measured outputs. Several MPC concepts has been tested, including qpOASES; Quadprog, Fmincon and DMPC with Laguerre functions; all of which has their pros and cons, and which produced different results.

Mostly the Fmincon and the DMPC gave the most promising results, DMPC with speed and ease of implantation but lacked successful results on output constraints. Fmincon produced some usable results but often got into trouble handling the output constraints and was computational heavy to use.