Aiming at the shortcomings of traditional diesel engine modeling methods in terms of complexity and adaptability, this paper proposes a dynamic loop decoupling modeling method, which aims to solve the problems of high computational cost of complex mechanism models and poor interpretability of data-driven models. Taking 16PC2-6B diesel engine as the research object, a modular loosely coupled architecture is constructed by stripping the strong coupling dynamic integral term between the turbocharger and the diesel engine body, and the torque balance error and power balance error are introduced to quantify the relationship between system energy and power. Combined with the particle swarm optimization algorithm, the Mahalanobis distance is used as the loss function to invert the unknown characteristic parameters such as turbine efficiency in the high-dimensional parameter space. The experimental results show that the error between the outlet pressure of the intercooler and the exhaust temperature of the lightweight model is 10–3, and the error between the outlet temperature of the turbine and the outlet temperature of the intercooler is controlled within 15%, which verifies the accuracy of the lightweight model constructed by the simplified static model and the optimization algorithm. It provides a new modeling paradigm for diesel engine digital twin technology.