Hybrid Physics- and Machine Learning-based Models for Safety and Predictive Maintenance of BESS

Stationary Battery Energy Storage Systems (BESS) play a crucial role in providing reliable and cost-saving use of grid power, while also normalising integration of cheaper, greener, and diurnally varying renewable energy sources. However, the safety and lifespan of BESS is impacted by the degradation and failures within a cell. Traditional approaches to battery management rely on live sensor measurements which alert only after a significant degradation has taken place. A model prediction of the failures well ahead in time can lead to taking active measures in improving the performance and in fault avoidance.

Whereas a physics based model works well with a single cell degradation and fault prediction, a BESS application typically has about 3 cells per kW of power supplied; and stationary BESS applications can go upto a Gigawatt in size. The large number, physical positioning, packing, and complex discharging and charging cycles can make physics-only based models computationally intensive and hence difficult to implement on the site of BESS installations. With the availability of historical sensor data, data analytics and machine learning tools can be employed, in combination with physics-based models to identify patterns and predict failures and improve remaining useful life (RUL). This can reduce operating costs and improve safety of BESS.

The goal of this project is to learn about various aspects of BESS performance and modelling, and implement a simple known strategy through data analytics and machine learning. The project has 14 weekly modules, tasks, and deliverables. Coding in Python is necessary. Prior knowledge of M/L is not a pre-requisite.