Machine Learning-Based Battery Capacity Estimation

Lithium-ion batteries (LIBs) are considered the optimum solution for electric automobiles today. Battery technology and developments have not yet progressed to a point where the electrification of heavy-duty transportation is technically and financially an attractive option for the market. With the help of a battery management system (BMS), batteries installed in electric heavy-duty vehicles can maintain a balance within key parameters such as cost, power, cycling life, and capacity. Battery packs need to be produced in a self-sustainable way and services must achieve utmost efficiency and cost-effectiveness. These and others are issues, that research and development have been trying to unveil. Evaluation of LIB capacity degradation in energy research has been taking extensive attention lately due to the rising demand for electric vehicles, with heavy-duty vehicles being the new arriver. With the affordable accessibility of data and enhanced computational capabilities, artificial intelligence has emerged as a favored approach for today’s researchers and analysts. Supervised machine learning techniques can be leveraged to systematically predict the capacity and degradation of batteries depending on their life cycle and operational conditions. In this thesis, a novel framework was developed to predict the capacity and lifetime of a LIB. This study also aims at investigating the abilities, applicability, and downsides of supervised learning models in real-life scenarios. Historical datasets were used to develop models which then can be applied for future observations or strategic decisions. The predictive modeling algorithms were trained, tuned, tested out, and validated on other data samples. The analysis in this thesis assesses changes that arise through varied battery parameters (features) and related states they may represent. Amongst the models developed, neural networks have shown the most fascinating results, due to their capabilities for real-time learning and forecasting. The experimental verification in this thesis work showed that accuracy of models is adequate for practical application to a certain extent, however, is dependable upon continual learning effort from the real-life information. This study has provided proof of concept and a basis for further development. Moreover, it shed light on potential system implementations that consider both calendar and cyclic aging profiles. For the data-driven method to offer reliable outcomes, models must be trained to learn from varied driving habits and equipment usage cases. Latterly, limitations and challenges were demonstrated, and advice was further offered for optimization and developmentObjectius de Desenvolupament Sostenible::9 - Indústria, Innovació i Infraestructura::9.5 - Augmentar la investigació científica i millorar la capacitat tecnològica dels sectors industrials de tots els països, en particular els països en desenvolupament, entre d’altres maneres fomentant la innovació i augmentant substancialment, d’aquí al 2030, el nombre de persones que treballen en el camp de la investigació i el desenvolupa­ment per cada milió d’habitants, així com la despesa en investigació i desenvolupament dels sectors públic i priva