Combustion phase of ammonia-diesel dual-fuel engine optimization by integrating shapley value analysis and fusing spatiotemporal features with Phy-ConvLSTM model

The hysteresis effect caused by the high auto-ignition temperature and low flame velocity of ammonia fuel, combined with the complex interactions between ammonia fuel and diesel ignition energy under varying operating conditions, poses a severe challenge to the prediction of failure and control deviation of the combustion phase (Crank Angle at 50% mass fraction burned, CA50) of ammonia-diesel dual-fuel engines. This paper proposes an optimization framework integrating a combustion-prior regularized convolutional long short-term memory network (Phy-ConvLSTM) with physical boundary constraints and game-theoretic feature analysis. By inputting injection parameters representing spatial distribution and thermodynamic states evolving into the network and embedding a combustion-prior informed curvature regularization term inspired by the Arrhenius law into the activation process, a mathematical representation of complex combustion is constructed. Subsequently, a Tree-based Shapley Additive Explanations (Tree-SHAP) algorithm is used to quantitatively deconstruct the nonlinear parameter interactions, and an Adaptive Chaotic Mapping Particle Swarm Optimization (ACMSO) algorithm is employed to perform global parameter optimization. Surrogate model evaluation based predominantly on computational fluid dynamics datasets demonstrates that under physical consistency constraints, the interpolation root mean square error reaches 0.32°CA, while extrapolation tests on strictly disjoint parameter ranges ammonia substitution rates of 40–60% and injection timings of -25 to -15°CA aTDC confirm a 38.7% error reduction over the unconstrained baseline. The performance confirms the success of the model order reduction task, surpassing the accuracy of the unconstrained baseline and effectively suppressing non-physical divergence under sparse boundary conditions. Feature deconstruction confirms that the ammonia substitution rate is the dominant cause of CA50 delay, presenting a mean global marginal contribution of 6.2°CA and extreme local peaks reaching 9.0°CA, and its second-order interaction with injection timing has the highest intensity(1.85°CA). Under the dual constraints of thermal efficiency and phase restricted to the 1500 rpm calibration baseline, the ACMSO strategy locks in the 60% to 75% load range as the optimal cooperative window, achieving a mechanically and thermodynamically stable ammonia substitution rate exceeding 60% with a baseline-indicated thermal efficiency exceeding 34% and local optimal peaks approaching 44%. The subsequent optimization map across the 800 to 2400 rpm spectrum is explicitly defined as an untested theoretical projection requiring future empirical validation. This study analyzes the characteristics of the deep learning model and provides a technical reference for the intelligent control of high-proportion zero-carbon power systems.