Quantitative analysis of engine parameters of a variable compression ratio CNG engine using machine learning

•Quantification of performance, combustion, and emissions of a high CR CNG engine.•A maximum of 35.86% ITE was achieved using CNG engine at compression ratio 16.•Machine learning methods used to model and predict engine performance parameters.•ANN, SVM, and Polynomial regression model were compared for engine performance data.•ANN shows the highest reliability in predicting the CNG SI engine behaviour. The numerical ability of machine learning methods has been a base in different engineering fields, including internal combustion engines. The presented study on machine learning methods originated to predict the performance, combustion, and emission characteristics of dedicated compressed natural gas (CNG) spark ignition (SI) engines. The experiments were conducted at various engine loads (IMEP), speeds, and compression ratios (CR) to collect the model training and testing data. The test results showed that with the increase in engine load, speed and CR, the ITE increased by 25%, 5.7%, and 10%, respectively. Similarly, the ISFC decreased about 20%, 9%, and 5.4%, increasing load, CR, and speed, respectively. The in-cylinder pressure, maximum rate of pressure rise, and maximum heat release rate reduces with an increase in engine speed and increases with an increase in CR. The combustion performance, such as flame development angle (FDA) and combustion duration (CD), reduced with an increase in CR. With increasing engine load, speed, and CR, ISCO and ISHC emissions decreased. On the other hand, the ISCO2 emission increased as the engine rpm, and CR increased. A maximum CR of 16 was used during the experiment. Three different machine learning methods (Regression Model, Support Vector Machine (SVM), Artificial Neural Network (ANN)) were used and compared to predict engine performance, combustion, and emission characteristics. Different order regression models and ANN models were tested using a back-propagation algorithm. The ANN and SVM model has been trained using hyperbolic transfer activation function and nonlinear kernel function. The value of correlation coefficient (R) and root mean square error (RMSE) for each output parameter was calculated and compared for three models. The regression model of 3rd order predicted well for ITE, CD, ISHC, ISCO, and ISCO2. Whereas, ANN model accurately predicted Pmax and Rmax. The coefficient of determination (R2) was calculated and compared for three models showing the ANN model suitable for accurate prediction compared to