Model Management for Battery State Estimation

• Applicants: Professorin Dr.-Ing. Julia Kowal; Professor Dr. Tilmann Rabl
• Subject Area: Electrical Energy Systems, Power Management, Power Electronics, Electrical Machines and Drives ; Data Management, Data-Intensive Systems, Computer Science Methods in Business Informatics
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 556566056

Battery storage systems are needed in many different applications, such as electric vehicles and in combination with renewable energy sources or for providing uninterruptible power supply. In such applications, many cells are connected in battery packs to provide enough capacity and voltage. Within these packs, cells experience small deviations in temperature, voltage and consequently aging. This makes state estimation of the different cells in a pack a complex task. With higher density and increasing numbers of cells, more precise and individually adjusted models are needed to ensure safety and efficiency. However, these models need to be deployed in specialized battery management hardware, which is much less powerful than general server deployments. In cases of failure, the state of each cell model should be retrievable. To this end, efficient model management is necessary, both for efficient deployment as well as exact retrieval. There are four primary approaches for battery state estimation: direct measurement, model-based, data-driven, and hybrid methods. Most techniques based on direct measurement encounter practical limitations, particularly in terms of the feasibility of battery disassembly and instrumentation setup. Model-based approaches allow for sophisticated state estimation but rely heavily on the accuracy and timeliness of the underlying models. Achieving this requires frequent checkups and online parameter identification, which are often unrealistic in practical scenarios. Moreover, online parameterization can be complex and unstable for real-time computation, limiting its applicability in certain scenarios. Data-driven approaches are adaptable to complex systems, flexible on different batteries without prior knowledge, and versatile regarding different estimation tasks. However, they depend heavily on the training dataset size and quality, lack interpretability, and are computationally intensive. Hybrid approaches can exploit the advantages of different approaches and consequently improve the drawbacks of data-driven approaches. These approaches aim to combine the adaptability of data-driven models with the structured insights from model-based approaches, potentially offering a more robust solution for battery state estimation. In our research project, we plan to explore different data-driven and hybrid models for battery state estimation, look into ways to enable their execution on hardware typically used for battery management systems, and extend approaches for managing the developed models in a distributed environment. Finally, we plan to build a battery management system incorporating our findings and approaches for validation.

• DFG Programme: Research Grants