Session: 03-09 Digital applications

Paper Number: 129044

129044 - Advancing AI Modeling for Prediction of Safety Parameters for Combustion of Hydrogen/Syngas/Natural Gas Mixtures 

Progress in hydrogen-compatible gas turbines should make a significant contribution to decarbonizing the electricity sector. The switch from hydrocarbon-based fuels, such as natural gas and light distillate oils, to the partial or total use of hydrogen as a fuel source will lead to a reduction in CO2 emissions, replaced by H2O emissions, which have a minimal impact on global warming. At this time, the combustion of hydrogen in the combustion chamber and the management of hydrogen in the gas delivery system represent non-trivial tasks, given the challenges associated with its high reactivity and susceptibility to detonation. It is therefore imperative to gain a comprehensive understanding of the mechanisms involved and to create tools for assessing reactivity parameters.

Autoignition Delay (AID) and Autoignition Temperature (AIT) are crucial parameters for assessing the risk of uncontrolled combustion in hydrogen/natural gas mixtures. Auto-ignition temperature represents the minimum temperature at which a gas mixture can spontaneously ignite without an external ignition source. For hydrogen, which is known for its wide flammability range, knowing this threshold is vital to prevent unintended ignition. A gas turbine operating with hydrogen must remain well below this temperature to avoid combustion outside of controlled conditions, reducing the risk of accidents and ensuring the safety of both equipment and personnel. This paper focuses on the development of a learning model to estimate AIT values during the combustion of hydrogen/natural gas mixtures.

A numerical procedure has been developed to generate automatically a dataset of Autoignition Temperature covering a wide operating range, compression pressures, temperature, equivalence ratio and different composition of fuel. In order to avoid any clustering problem in the different successive sample points a Sobol’s procedure has been applied. The quasi-random Sobol sequence permits to uniformly cover the entire possible input parameter space.

The 50,000 points that constitute the dataset have been simulated using ChemKin software and the comprehensive combustion model from the literature developed by NUIG. In order to fit correctly all the hyperparameters of a neural network model several methods of optimization have been tested as GridSearchCV, RandomSearchCV, BayesSearchCV and HyperOpt. These tests highlighted the advantages and disadvantages of each method, particularly in terms of calculation time.

The algorithm developed is capable of predicting AIT for pressures between 1 and 50 atm, temperatures between 200 and 700°C and equivalence ratio between 0.2 and 5. It is composed by several hidden layer sizes and different values of learning rate, alpha coefficient, type of solver, activation function, and so on. For both the training and testing datasets, the average value of the correlation coefficient was above 99.95%, the Mean Absolute Error (MAE) and the Mean Square Error (MSE) around 0.003 and lower than 1e-5, respectively.

Several tests have been made showing a very good accordance between results from the learning model and those from the dataset. The final AI model shows a high degree of robustness in reproducing the results of the detailed model in all the parameter domain studied.

Presenting Author: Pierre-Alexandre Glaude CNRS

Presenting Author Biography: PhD 1999. Senior Researcher at CNRS

Authors:

Roda Bounaceur CNRS
Pierre-Alexandre Glaude CNRS
René Fournet Université de lorraine
Baptiste Sirjean CNRS
Pierre Montagne GE Gas Power France
Alexandre Auvray GE Gas Power France
Eric Impellizzeri GE Gas Power France
Pierre Biehler GE Gas Power France
Michel Molière Université de Technologie de Belfort Montbéliard