A Hybrid ANN-HHO Approach for ICE Emission Data Modelling

Abstract

Road transportation remains a primary contributor to global greenhouse gas emissions, where thermodynamic limits inherent to Internal Combustion Engines (ICE) constrain efficiency and complicate manual engine calibration. Although artificial neural networks can map non-linear engine emissions, manual architecture design remains a costly trial-and-error process prone to severe overfitting. In this study, we propose a hybrid Artificial Neural Network and Harris Hawks Optimization (ANN-HHO) framework for multi-target vehicle emission modeling, predicting mathematically estimated \(\ce{NO_x}\) and \(\ce{CO2}\) from OBD-II vehicle telemetry. The proposed framework simultaneously optimizes network topology and training hyperparameters under both unrestricted and compute-constrained regimes. Empirical results demonstrate that the proposed unrestricted model (HHO-UR) achieves superior predictive fidelity with an \(R^2\) of \(0.9823\) for \(\ce{NO_x}\) and \(0.9950\) for \(\ce{CO2}\), decisively resolving the generalization failure observed in the manual baseline model (\(R^2 = 0.5714 \pm 0.2240\)). Furthermore, under a restricted budget (\(\le 20\) epochs), restricted HHO (HHO-R) attains an \(88.7\%\) reduction in training epochs (17 vs. 150) with an \(R^2\) of \(0.9327\), cutting search wall-clock duration from 9.1 hours to 76 minutes. Comparative evaluation against alternative metaheuristics (PSO, ACOR, and SCA) confirms that the HHO framework consistently discovers compact, generalizable two-layer topologies, demonstrating a scalable and computationally efficient methodology for intelligent automotive engine management.

Conference Paper

Note

Preprint Notice: This paper is under peer review for the ICLAMP 2027 conference.

Feel free to download the paper here for a better reading experience!

Alternatively you can read it below: