Enhancing Observers through Deep Learning: Application to State of Charge Estimation in Vanadium Flow Batteries

In this paper we study a deep learning-based method to enhance the estimation accuracy of a nonlinear observer with partially unknown output functions. This technique is particularly useful for the problem of State of Charge estimation in vanadium flow batteries. The key innovation of this approach is the direct reliance on terminal voltage measurement and the incorporation of a machine learning algorithm to learn the battery’s overpotentials. The overpotentials are represented by a neural network trained on data produced by a conventional observer, which determines species concentration using a physical electrochemical model and open-circuit voltage measurements. Once trained, this model is integrated into the observer to enhance State of Charge estimation precision. The stability of the deep learning-based observer is formally proved. Additionally, the method is validated through experimental evaluations on a real vanadium flow battery system.