Advanced modelling and forecasting methods for electric vehicle batteries based on data analysis with realistic operating conditions
(English) This PhD thesis addresses the challenges of modeling, monitoring, and forecasting the behavior of lithium-ion batteries at both the single-cell and pack levels, with a particular focus on improving robustness under realistic operating conditions. The work is motivated by the growing deploy...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis doctoral |
| Fecha de publicación: | 2025 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/450101 |
| Acceso en línea: | https://hdl.handle.net/2117/450101 https://dx.doi.org/10.5821/dissertation-2117-450101 |
| Access Level: | acceso embargado |
| Palabra clave: | Battery Electric Vehicle Energy Storage Modelling Machine Learning 621.3 - Enginyeria elèctrica. Electrotècnia. Telecomunicacions 629 - Enginyeria dels vehicles de transport Àrees temàtiques de la UPC::Enginyeria electrònica |
| Sumario: | (English) This PhD thesis addresses the challenges of modeling, monitoring, and forecasting the behavior of lithium-ion batteries at both the single-cell and pack levels, with a particular focus on improving robustness under realistic operating conditions. The work is motivated by the growing deployment of battery systems in electric mobility and stationary storage, where performance, lifetime, and safety critically depend on reliable state estimation and degradation prediction. At the single-cell level, the thesis introduces two complementary modeling approaches: a parametric voltage–capacity model for constant-current discharge, and a deep learning framework for partial-charge data. These methods enable early detection of degradation trends and real-time estimation of capacity fade, providing interpretable health indicators (HI) and accurate remaining useful life (RUL) forecasts. At the pack level, the thesis explores the additional complexities that arise from cell-to-cell variability, imbalance, and data acquisition issues. A hybrid data imputation methodology based on the Unscented Kalman Filter (UKF) is proposed to reconstruct missing voltage signals at both the cell and branch levels, ensuring continuity of BMS functions such as balancing, SoC/SoH estimation, and fault detection. The method is benchmarked against neural networks, highlighting the trade-off between data-driven accuracy and model-based adaptability. Building on this, this work expands upon the imputation framework by studying how reconstructed signals impact the accuracy of forecasting models. Four reconstruction strategies of increasing complexity (ZOH, ARIMA, UKF, and GRU) were compared, and their outputs were fed into recurrent neural networks (LSTM and GRU) developed for this purpose. These networks were used to predict the remaining time to depletion (RTD) of individual cells under driving conditions. The results demonstrate that the quality of signal reconstruction directly impacts forecasting performance. The contributions are supported by multiple datasets, including public repositories (NASA, Sandia National Laboratories) and a custom experimental testbench capable of executing standardized drive cycles and controlled CC–CV protocols. Together, these datasets provide a rigorous foundation for validation across chemistries and cycling conditions Overall, the thesis demonstrates how the integration of signal processing and filtering, and machine learning can enhance the reliability of battery models, both for immediate diagnostic tasks and long-term prognostics. The findings contribute toward more robust and practical battery management systems, bridging the gap between academic models and real-world applications in electric mobility and energy storage. |
|---|