Machine Learning Models for Turbulent Combustion Speed Prediction

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.May 23 2024

In a paper published in the journal Heliyon, researchers developed and verified machine learning (ML) models for predicting turbulent combustion speed in hydrogen-natural gas spark ignition engines, demonstrating their superiority over traditional methods regarding precision, computation time, and handling of original data.

The study utilized data from a motore isolato stationario a scoppio e letto (MINSEL) 380 engine. The results showed that ML models, particularly random forest (RF), achieved high forecasting accuracy, making them highly suitable for industrial applications such as engine monitoring, control systems, firefighting systems, simulation, and prototyping tools.

Related Work

Past work has explored combustion as a key energy source across various industries. With advances in hydrogen production, engine combustion is being considered for its environmental benefits. Studies have focused on improving engine performance and reducing emissions by understanding combustion speed, which affects in-cylinder pressure and pollutant production.

ML has effectively modeled these complex phenomena, with techniques like support vector machines and artificial neural networks providing accurate engine performance and emissions predictions. Since combustion often occurs in turbulent regimes, creating accurate simulation models is essential for optimizing engine performance and adhering to environmental standards.

ML Techniques

The study utilized data from a MINSEL 380, a 380-cc single-cylinder engine modified for spark ignition, collected under normal temperature (298 K) and pressure (1 atm) conditions. Various components, including mass flow controllers for gas and air, sensors for pressure and temperature, and an induction motor, were part of the experimental setup.

Data were gathered over 200 cycles for different fuels, such as natural gas, hydrogen, and their mixtures, with fuel-air equivalence ratios varied to maintain constant adiabatic flame temperatures. Experiments were conducted at engine speeds of 1000 rpm, 1750 rpm, and 2500 rpm.

Data preprocessing involved standardizing the features, which included the dimensionless flame front radius, percentage of hydrogen, fuel-air equivalence ratio, and engine speed in RPM. Standardization, which centers the data around the mean with a unit standard deviation, is crucial for efficiently operating ML algorithms, ensuring all variables are on the same scale.

This step allows the algorithms to converge faster and improves their adaptability. The target variable, experimental turbulent combustion speed, was not standardized to maintain its original scale. Computations were performed using Jupyter scripts with Scikit-Learn for linear regression, RF models, and TensorFlow for neural network models.

The study employed various ML techniques, including multiple linear regression (MLR), support vector regression (SVR), RF, and artificial neural networks (ANN). MLR, the simplest algorithm, predicts outcomes using multiple explanatory variables and provides an explicit model after training. SVR, derived from support vector machines (SVM), uses kernel functions to handle non-linear relationships and balances model error and complexity.

The RF algorithm, an ensemble method, combines multiple decision trees to improve prediction accuracy and generalization by reducing overfitting. ANNs consist of interconnected layers of neurons capable of modeling complex non-linear relationships through activation functions and a trial-and-error approach for hyperparameter tuning.

Each ML model was trained for up to 5000 epochs using features like flame front radius, hydrogen percentage, fuel-air equivalence ratio, engine speed, and turbulence intensity. MLR involved a straightforward linear equation linking variables, SVR employed kernel functions and regularization to manage complex input-output relationships, RF used a bagging approach to combine decision trees, enhancing robustness, and ANNs efficiently learned from data, modeling intricate dependencies through varied hidden layers and activation functions. These models effectively predicted experimental turbulent combustion speed, demonstrating their capability to capture combustion dynamics in hydrogen-natural gas engines.

ML Evaluation

The study compared the performance of multiple ML models in predicting experimental turbulent combustion speed for hydrogen-natural gas engines. While MLR showed limited accuracy, SVR demonstrated improved performance with standardized data, indicating the importance of preprocessing. The RF model outperformed others, offering precise predictions for raw and preprocessed data.

Additionally, ANN exhibited high accuracy but was sensitive to variations in data scale, emphasizing the significance of preprocessing. These findings align with previous research, suggesting RF as a robust and efficient model for practical applications, with forecasting accuracy exceeding 90%.

Furthermore, the study evaluated the computational efficiency of the models compared to a detailed reaction mechanism. While ML models exhibited significantly lower calculation times than gauss-radau integration (GRI 3.0), they needed to have the robustness of the detailed mechanism. It highlights the trade-off between computational speed and model complexity, where ML models offer faster calculations but may not match the comprehensive capabilities of detailed mechanisms.

Overall, the study underscores the potential of ML models in predicting combustion dynamics, especially for industrial applications, while acknowledging the challenges in achieving robustness across various fuel types and complexities in turbulent combustion processes.

Conclusion

To sum up, this paper compared various ML techniques for computing turbulent combustion velocity in hydrogen-natural gas engines. Preprocessed data yielded the best results, with ANN and RF models proving the most accurate. The RF model, in particular, stood out for its ease of implementation and superior performance.

These findings offer significant time savings compared to traditional approaches and can be integrated into various applications such as engine monitoring systems and combustion engine design. Extending these models to different fuel types could enhance their applicability in real-world scenarios.