Data-driven Battery Management Systems (BMS) models become inaccurate over time due to degeneration and environmental changes. Physics-based BMS models are often too slow to provide useful real-time information. Alone, each have their issues; together they create a BMS model with huge potential gains. Hybrid Twin™ combines both models to produce a real-time, accurate system. Find out why this is what you need to reassure your customers that they’re getting the best monitoring and predicting BMS.
Most Electric Vehicles (EVs) have an onboard Battery Management System (BMS) that maintains safe and consistent operation of the battery pack. By evaluating the State of Charge (SoC) and the State of Health (SoH), the BMS deploys strategies such as cell balancing to reduce degradation and optimize the performance of the battery system. Typically, the BMS uses data-driven models to estimate the SoC and SoH. However, these models become inaccurate as batteries degrade or undergo important changes in their environment and operation, including changes in ambient temperature and charge cycles.
In contrast to data-driven models, physics-based models are accurate but far too time-consuming to be of practical use in real-time applications. The challenge is to develop BMS models that are both fast and accurate.
Numerical techniques based on Model Order Reduction (MOR) have opened new possibilities for efficient simulations. These techniques involve calculating, offline, the parametric solution of a parametric model. That solution is then deployed online to perform fast predictions at any point in the parameter space, allowing for optimization, uncertainty propagation, and simulation-based control – all in real-time. However, when such techniques are integrated into data-driven application systems, unexpected difficulties can arise. These are related to significant deviations between the predicted and observed responses thanks to the unpredictability of system behavior.
One solution is to construct a data-driven model “on-the-fly” to fill the gap between prediction and measurement. When this is achieved, system control can be efficiently attained. This is the rationale behind ESI’s Hybrid Twin™, which has two main contributions. First, the system obeys the time evolution that physics imposes. Second, a correction model is constructed on the fly, making use of any appropriate machine learning. For efficiency purposes, big data is replaced with smart data, which essentially requires addressing three key questions, using multi-scale modeling of data and information theories: what data is needed, where, and when.
The potential gains of the Hybrid Twin™ are tremendous. Suddenly, simulation that is both physics and data-based, is possible in real-time, delivering an accurate and predictive model that can be used to anticipate failure or performance loss, to maximize product operation by reducing downtime, and to manage maintenance and repair costs.
The physics-based battery model consists of a system of non-linear, tightly coupled parametric partial differential equations describing reaction-diffusion-thermal transport. The multi-physics interactions described by such a model occur over a wide range of length scales, making real-time evaluations difficult. Here, the Hybrid Twin™ becomes an appealing approach as it makes use of the physics-based models, whose parameters can be updated online from the collected and assimilated data, such as local temperature.
The marriage of physics and data-based models is well adapted to tracking SoH and accommodating degradation throughout the battery’s operation. Furthermore, we can see that such models can be adapted to provide a real-time charging recommendation by evaluating the health of the battery, temperature conditions, and previous charge history. In summary, the battery Hybrid Twin™ provides a basis for real-time, accurate monitoring, prediction, and control of the battery system.