Scholarlink

The predictions made by reverse modelling are valid, but they are slightly inferior than those produced by forward modelling because numerous permutations of input variables lead to identical outcomes due to input variable interaction. Artificial neural networks (ANNs) require huge data sets for training to make accurate predictions. The 1000 input-output data sets for training were obtained from experiments and artificially generated random data sets after varying input variables between their respective levels to predict corresponding outputs through derived empirical equations using central composite design (CCD). The back propagation neural network (BPNN) and genetic algorithm–neural network (GA-NN) were established for performing forward and reverse modelling tasks. Forward modelling aims to predict outputs such as performance (brake thermal efficiency: BTE and brake-specific fuel consumption: BSFC), emissions (carbon monoxide: CO, nitrogen oxide: NOx, unburnt hydrocarbon: UHC) characteristics with a set of input variables such as engine load (EL), blend type (BT), injection pressure (IP), and compression ratio (CR). The reverse modelling tasks were performed to predict input variables (EL, BT, IP, and CR) for the desired performance (BSFC and BTE) and emission (NOx, UHC, and CO) characteristics. The parametric study of BPNN and GA-NN methods reduced mean squared error (MSE) to 0.00086 and 0.000738 for forward modelling and 0.00487 and 0.00212 for reverse modelling tasks. The prediction performances were evaluated with 20 random experimental cases for all outputs, resulting in 3.198% for CCD, 2.578% for BPNN, and 2.056% for GA-NN for forward modelling, as well as 3.83% for BPNN and 3.03% for GA-NN for reverse modelling. GA-NN produced better predictions for both forward and reverse directions due to its global search capability. Predicting outputs for unknown parametric conditions with forward models and predicting input variables for desired outputs with reverse models helps novice users automate the process without needing practical experiments.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Artificial Intelligence Modelling of Engine Performance and Emission Characteristics

  • 章节

  • 发表日期：

    2025年

查看全文
  • Manjunath Patel G. C.,
  • Ajith B. S.,
  • Jagadish,
  • Arun Kumar Shettigar,
  • Olusegun David Samuel

摘要

The predictions made by reverse modelling are valid, but they are slightly inferior than those produced by forward modelling because numerous permutations of input variables lead to identical outcomes due to input variable interaction. Artificial neural networks (ANNs) require huge data sets for training to make accurate predictions. The 1000 input-output data sets for training were obtained from experiments and artificially generated random data sets after varying input variables between their respective levels to predict corresponding outputs through derived empirical equations using central composite design (CCD). The back propagation neural network (BPNN) and genetic algorithm–neural network (GA-NN) were established for performing forward and reverse modelling tasks. Forward modelling aims to predict outputs such as performance (brake thermal efficiency: BTE and brake-specific fuel consumption: BSFC), emissions (carbon monoxide: CO, nitrogen oxide: NOx, unburnt hydrocarbon: UHC) characteristics with a set of input variables such as engine load (EL), blend type (BT), injection pressure (IP), and compression ratio (CR). The reverse modelling tasks were performed to predict input variables (EL, BT, IP, and CR) for the desired performance (BSFC and BTE) and emission (NOx, UHC, and CO) characteristics. The parametric study of BPNN and GA-NN methods reduced mean squared error (MSE) to 0.00086 and 0.000738 for forward modelling and 0.00487 and 0.00212 for reverse modelling tasks. The prediction performances were evaluated with 20 random experimental cases for all outputs, resulting in 3.198% for CCD, 2.578% for BPNN, and 2.056% for GA-NN for forward modelling, as well as 3.83% for BPNN and 3.03% for GA-NN for reverse modelling. GA-NN produced better predictions for both forward and reverse directions due to its global search capability. Predicting outputs for unknown parametric conditions with forward models and predicting input variables for desired outputs with reverse models helps novice users automate the process without needing practical experiments.