Design optimization of Hybrid Electric Vehicles based on Deep Learning algorithms

Recent years have seen the flourishing of the so called green wave and environment-related topics have been discussed in many ways for various reasons. The use of Hybrid Electric Vehicles (HEVs) is a valid way to achieve tank-to-wheel (TTW) CO2 emissions reduction. The choice of the design parameters, such as engine displacement or power of the electric machine, remains of fundamental importance. To this end, various algorithms have been deployed to effectively calculate the TTW CO2 emissions of a specific HEV layout. One of this is Dynamic Programming (DP). However, it cannot always be used as it requires high computational power and time. The main goal of this study is to develop an algorithm that can be used in the context of optimized design of HEVs. The tool to be developed should be far lighter than other deterministic algorithms such as DP and ensure comparable results at the same time. The technology of choice is Deep Learning Neural Networks (DNNs). It is a branch of machine learning so a part of the vaster field of Artificial Intelligence. This particular kind of algorithms mimic the behaviour of a human brain: various connections between different layers of neurons enable the flow of information and the possibility for the net to adjust itself and learn. A pipeline of two DNNs is implemented to assess whether the vehicle will successfully complete the driving cycle of choice (feasibility), and in that case predict the TTW CO2 emissions. The dataset available is composed by a set of different design parameters for HEVs. The pipeline is trained in Supervised Learning. Promising results emerge from the study as the AI algorithm is able to produce feasibility predictions with an accuracy up to 95%, and TTW CO2 estimates with less than 1% error. The implementation is based on Keras and Tensorflow libraries.