Supervised Battery Capacity Estimation

Evaluation of capacity fade in Li-ion batteries in the ballparks of energy research has assumed significant importance in recent times extensively owing to growing demand and need for portable power supplies and electric vehicles. In terms of methodologies and analytics, with the accessibility of data becoming greatly simplified and computational capabilities being exceptionally improved, supervised intelligence has become the approach-of-choice for modern day analysts. Using supervised learning approaches, techniques generalized for clustering and regression can be used to systematically extract relevant information from operational data leading to effective prediction of capacity fade. Among the multiple learning tools, neural nets have garnered compelling importance in practical applications due to their flexibility and capabilities of learning in real-time. However, there is a lack of published research on the exercise of data mining and learning on capacity estimation. This investigation aims to study the abilities, drawbacks and applicability of supervised learning approaches in comparison to true capacity fade which is observed in Lithium ion batteries with extensive usage over time. For the effective identification and observation of changes in the battery pack’s behavior, the focus will be on quantitatively analyzing the changes through various parameters, referred to as “features”, and their related conditions. The capacity estimation problem will be primarily handled with the application of neural nets with the capabilities of real-time learning, with the primary focus being the data generated through the daily and regular use of the packs. As and where imperative, there will be application of synthetically generated data by means of verified data generation techniques. The results of this endeavor will be applied to improved handling of Lithium ion cells and packs under the supervision of a battery management system, using experimental data for training, evaluation and synthetic fabrication of data.