A deep neural network based model for the prediction of hybrid electric vehicles carbon dioxide emissions

Hybrid electric vehicles (HEV) are nowadays proving to be one of the most promising technologies for the improvement of the fuel economy of several transportation segments. As far as the on-road category is concerned, a wise selection of the powertrain design is needed to exploit the best energetic performance achievable by a HEV. Amongst the methodologies developed for comparing different hybrid architectures, global optimizers have demonstrated the capability of leading to optimal design solutions at the expense of a relevant computational burden. In the present paper, an innovative deep neural networks-based model for the prediction of tank-to-wheel carbon dioxide emissions as estimated by a Dynamic Programming (DP) algorithm is presented. The model consists of a pipeline of neural networks aimed at catching the correlations lying between the design parameters of a HEV architecture and the main outcomes of the DP, namely powertrain feasibility and tail pipe CO2 emissions. Moreover, an automatic search tool (AST) has been developed for tuning the main hyper-parameters of the networks. Interesting results have been registered by applying the pipeline to three databases related to three different HEV parallel architectures. The capability of the pipeline has been proved through an extensive testing campaign made up by multiple experiments. Classification performances above 91% as well as average regression errors below 1% have been achieved during an extensive set of simulations. The presented model could hence be considered as an effective tool for supporting HEV design optimization phases.