A Neuromorphic-Augmented Generalized Additive Modeling Framework for RCCI Combustion Phasing Prediction

Abstract

Abstract. Reactivity-controlled compression ignition (RCCI) is a promising dual-fuel combustion strategy which requires precise closed-loop control for optimal performance and simultaneous attainment of low nitrogen oxides (NOx) and soot emissions. Various data-driven and machine learning (ML) techniques can be useful for RCCI combustion control but some approaches can be unreliable in the face of uncertainties, due to the limited insights into their underlying models. In addition, these systems often have complex air handling and fueling strategies which can make it difficult to understand which actuator to leverage to achieve a desired combustion phasing. In this study, a generalized additive model (GAM) framework is applied to achieve transparent modeling of the start of combustion (SOC) and burn duration (BD) which can be used for control and provide insights into the main drivers to be leveraged for control. The utility of these underlying models is illustrated by utilizing them in a modified simple Wiebe function (MSWF). The outputs of the MSWF are augmented with a Bayesian-regularized spiking neural network (BRSNN) transfer learning model (TLM) to predict CA50 (the crank angle at which 50% of fuel energy is released) to provide reductions in computational costs. The GAM submodels are also leveraged to understand the main drivers responsible for combustion phasing and how these depend on operating conditions. The proposed framework is validated with experimental data across 297 steady-state gasoline/diesel RCCI engine operating conditions, demonstrating high predictive accuracy greater than 90% Pearson product-moment correlation coefficient (PPMCC) and computational runtime under 500 milliseconds (ms) per query.