Prediction of Combustion Process Based on Physics-Informed Data-Driven Model

Predicting the ignition and heat release patterns during diesel combustion processes is of great significance for improving engine efficiency, reducing emissions, and enabling future low-carbon and zero-carbon flexible fuel control. However, traditional Wiebe physical models face challenges in handling the highly nonlinear nature and variable operating conditions of diesel combustion, failing to achieve accurate real-time prediction. Pure data-driven models demand large amounts of data and lack physical interpretability, while physical models based on parameter learning have restricted fitting accuracy due to structural and parameter constraints. To address these issues, this paper proposes a novel Physics-Informed Data-Driven Model. It defines data loss as the deviation between neural network predictions and measured data, and physical loss as the deviation between neural network derivatives and the differential form of the physical model. By minimizing the combined loss, which is a weighted fusion of these two loss terms through back propagation, the model effectively integrates physical and data-driven approaches to estimate the heat release rate and combustion phase instantaneously. Experimental results showcase the remarkable performance of this model with a high CA50 prediction coefficient (R² = 0.9779), surpassing physical models based on parameter learning in accuracy and demonstrating stronger generalization capability than pure data-driven models, thus holding great promise for advancing diesel engine technology and contributing to the broader goals of sustainable energy use and environmental protection.