Data-driven Methodologies for Battery Managment and Control
|Project members:||dr.ir M. Corno (Matteo)|
|Keywords:||Identification and estimation, Transportation and infrastructure|
During the last 10-15 years the secondary battery industry has experienced an explosive growth in terms of volume, value, available products and technologies. The early phase of this growth has been mainly pushed by the need of more portable electronics (cell phones, cameras, laptops, power tools, etc.). In the past several years rising oil price and increasing pollution determined a new driving factor: transportation. The battery industry responded to the energy needs of hybrid and electric vehicles by developing new technologies and by stacking hundreds of cell in battery packs.
As the energy stored in battery packs increases, new battery control and management strategies are needed to safely take advantage of this development. Traditionally the design of battery control and management strategies is carried out with an indirect approach. According to this approach some apriori knowledge on the system (first principles, battery technology, empirical rules…) is used to derive a physical model of the battery whose parameters are then identified and validated with data obtained by ad-hoc experiments. This task is not always trivial, especially when dealing with batteries.
The data-driven approach represents a viable solution to these problems. The idea at its foundation is that of deriving the control and fault detection algorithms directly from data. This approach allows the user to derive models and controllers by making a limited number of choices. With the new methods we limit these questions to structural information about the plant (such as a rough estimate of the system order) instead of relying on precise priors. This would enable an inexperienced user to make use of these tools or to train user's in short time to use them.
The data-driven approach, when applied to the battery domain, has several advantages, in particular:
• The computational complexity of the controller/model is easily scalable.
• The derived models are easily ported between different technologies, as no physical hypothesis is needed.
• It can be based on regular usage data and ad-hoc experiments are not needed.
• A reduced set of tuning knobs provides an efficient method for the user to still exploit a-priori knowledge on the system (if available).
Another advantage of the data-driven approach is that it can be directly extended to fault tolerant control and online estimation problems. The ability of deriving the controller directly from data can be exploited to iteratively adapt the controller to a varying (possibly faulty) plant.
We believe that the battery control domain is of particular interest because of its strategic importance and the criticalities involved in deriving and maintaining models. In particular the proposed framework fits well to the problems of
1) charging/discharging control
2) equalization and faulty cells management.
3) Battery life estimation