Data-driven modeling for predicting the steady-state and transient performance of a dual-fuel medium-duty engine employing artificial neural networks

Abstract

The dual-fuel combustion using the Reactivity Controlled Compression Ignition concept has demonstrated great potential in achieving low NOx and soot emissions while maintaining engine performance. However, its experimental transient behavior remains a challenge in the current literature and the use of modeling approaches to predict it. This study aims to develop effective predictive models for emissions and fuel consumption in a dual-fuel medium-duty engine. A 7.7L multi-cylinder dual-fuel engine was experimentally tested under stationary conditions and transient state under the World Harmonized Stationary Cycle at full load conditions. The data from experimental measurements is used as input for developing data-driven models to predict emissions and fuel consumption. The study utilizes correlation analysis and principal component analysis for variable selection, identifying speed, torque, and gasoline fraction as key predictors for NOx, CO, HC emissions, and fuel consumption at stationary conditions. Simplified artificial neural network models were developed using Bayesian optimization, achieving high R² values and low steady-state estimation error but with limited accuracy for instantaneous emission values under transient cycles. A multi-weighted artificial neural network framework was introduced, incorporating combustion parameters and intake/exhaust conditions for enhanced emission prediction. The results showed that while traditional neural network models performed well in steady-state predictions, they struggled with instantaneous transient emissions. The multi-weighted neural network significantly improved real-time prediction accuracy, particularly for CO and HC emissions, which showed strong dependence on intake and exhaust conditions. The model achieved average cycle errors below 7 % for NOx, CO, and HC, and below 1 % for fuel consumption, demonstrating its ability to capture transient engine behavior accurately. This study highlights the potential of data-driven modeling as a scalable alternative to traditional physics-based approaches for dual-fuel engine optimization. The findings confirm that a multi-weighted artificial neural network framework enhances transient prediction capabilities, bridging the gap between stationary engine models and real-world operating conditions.