Abstract

In this article we present a Hybrid Neural Adaptive State Space Model (NASSM), the purpose of which is to solve the complex problem of accurately characterising the ever changing (non-measurable) polarising impedance multi-dimensional surface and capacity degradation of a Lithium-Ion battery. We achieve this by proposing a novel strategy and architecture to infer these critical battery parameters simultaneously, directly from operational data, avoiding the need of costly off-line testing procedures. The NASSM infers a representational general surface model of the polarising impedance multi-dimensional surface by partially embedding a multi-layer perceptron (or deep neural network), within the hidden state representation and uses Variational Sequential Monte Carlo to infer the parameterisation of said surface as well as the total energy value in order to adapt the model to these changing (degrading) values. Training is performed online with experimental operational data and we demonstrate that this methodology allows the model to perform accurate predictions of the probability of when a generic battery management system would disconnect the battery due to the terminal voltage falling below a predefined threshold (a physical constraint). The results are compared to the state of the art on experimental data.