Hybrid Physics-Based Neural Network Models for Diesel Combustion Phasing Prediction

Abstract

Conventional control-oriented physics-based models can be leveraged for combustion phasing control in modern diesel engines, but very accurate models often require significant computational power. On the other hand, Artificial Neural Network (ANN) models can have high predictive capabilities and lower computational costs but their heuristic nature may make them less desirable for complex and impactful parameters like combustion phasing. This paper addresses these issues by leveraging the advantages of both modeling methodologies in an integrated modeling framework which can be utilized for real-time combustion control for a modern diesel engine. The models are validated with data from a 2.0 L turbocharged diesel engine which is fitted with dual loop exhaust gas recirculation. The results of the modeling frameworks are compared and show that the proposed hybrid model framework can have a Pearson Product-Correlation Coefficient (PPMCC) of a 100% and an RMSE of 8.79 × 10-8 CAD; as compared to the physics-based model with a PPMCC of a 95% and an RMSE of 0.17 CAD.