Adaptive Map Development for Fuel Distribution in a Gas Turbine

Abstract

The performance of any engine is dependent upon the fuel which is passed to it. To enhance the performance of that engine without making changes to the mechanical assembly is possible by controlling the way in which the fuel enters the engine. This can be achieved by estimation of fuel entry based on the previous experiences. Neural network (NN) is a technique which is being used over the years for the estimation of functions and regression of curves. However, NN uses a slowly converging gradient based method for learning with multiple layers of neurons which further slows it down. One cannot use typical neural network for the cases where the learning needs to be quick, so as a solution to this problem one can use Extreme Learning Machine (ELM). ELM is based on single hidden layer feedforward networks (SLFNs) structure and the weights only at the output stage of the network are updated, while the weights and biases at the initial stage are chosen randomly and fixed. But it can only serve data in batch. So for that purpose one can use Online Sequential-ELM (OS-ELM). OS-ELM is based on the idea of batch learning ELM algorithm but it accepts data arriving in sequential manner in chunks or one-by-one. Its extensions also cater similar benefits. Although, these algorithms perform good but quick convergence is not guaranteed. A new learning algorithm extended Kalman filter based Online Sequential Extreme Learning Machine (eKOS-ELM) and Ensemble of eKOS-ELM (EeKOS-ELM) have been proposed use extended Kalman Filter based error backpropagation technique to adjust input weights while uses standard OS-ELM technique to train output weights to achieve quick convergence. OS-ELM along with its extensions and proposed eKOS-ELM and EeKOS-ELM have been applied for the generation adaptive map for fuel splits which helps in optimal gas turbine operation.