Design and Development of a cosimulation framework to evaluate SOH in electric vehicles

The increasingly pressing need to combat environmental issues, in particular greenhouse gas emissions, requires a transition towards more sustainable transport solutions. Electric vehicles (EVs) are identified as a key pillar in this process. However, maximising their efficiency also depends on an accurate assessment of battery health (SOH). A detailed knowledge of SOH is essential to optimise the performance of EVs, predict their remaining life and ensure their safe operation. In parallel, recent advances in the field of artificial intelligence show a real advantage in the use of such tools in various application areas. Therefore, the use of such technologies to estimate the SOH of electric vehicles appears to be a solution to this problem. Unfortunately, such tools require large amounts of data to make reliable predictions, and considering that the technology behind EVs is still recent, the availability of data to train neural networks is insufficient. For this reason, this thesis initially proposes a realistic electric vehicle simulator in an urban context based on Simulink technology. This simulator can model driver behaviour to generate data on driving sessions through a co-simulation approach that is easily scalable and adaptable to different contexts of use. Subsequently, the generated data is used to train a transformer-based neural network to adapt it to a real-world context via transfer learning procedures. Various strategies for adapting the network to a context different from the training context are also examined, with special attention paid to features such as the amount of data required to achieve a certain level of prediction accuracy and the degree of invasiveness of the knowledge transfer procedure during the adaptation process.