The virtualization of powertrain systems is a key enabler for modern powertrain
development. While physics-based 0D/1D simulation models provide accuracy and
interpretability, these models are typically computationally demanding,
prolonging the development process and usage throughout the V-cycle. Moreover,
achieving real-time-capable simulation models through model simplifications
remains challenging, as it often leads to significant losses in accuracy. In
contrast, data-driven approaches can achieve high computational efficiency
without significantly compromising model accuracy. This opens the possibility
for not only online control applications, such as model predictive control or
reinforcement learning, but also for computational expensive offline control
prototyping using ultrafast-running data-driven digital twins. This work focuses
on the elaboration of a scalable methodology for the development of
ultrafast-running powertrain models for stationary and transient engine
operation. This includes the efficient generation of training data with great
variance, data analysis, and preparation, an optimized partitioning method using
the Jensen–Shannon distance, feature engineering, model training of a multilayer
perceptron (MLP), a long short-term memory (LSTM), and gated recurrent unit
(GRU) network, followed by the model evaluation using test data and the
concluding model deployment. In order to demonstrate the concept, a calibrated
0D/1D model of a dual-fuel marine main engine provided by WinGD Ltd. for a pure
car and truck carrier is utilized as the reference physics-based model. The case
study provides a comprehensive examination of the development of
ultrafast-running data-driven fuel consumption models in both stationary and
transient engine operation. The results show that the proposed methodology
yields robust results and minimizes the loss of accuracy to 1.80%–2.14% for the
MLP predicting the steady-state fuel consumption and to 0.67%–0.96% (GRU) and
1.52%–1.68% (LSTM) for predicting the transient fuel consumption, while
achieving a multiple 104-fold reduction of the real-time factor (RTF) on an
identical CPU.
