Real-time spatiotemporal temperature gradient estimation based on a graph convolutional neural network for battery cells

Monitoring the temperature distribution of cells in large-scale battery packs can be a costly endeavour. This paper proposes a spatiotemporal model based on a graph convolutional neural network for the estimation of the axial surface temperature distribution of 21 700 cylindrical cells in a small-scale battery pack. The model is developed on the basis of experimental results obtained from the studied pack subject to various immersion thermal management conditions. The pack is charged and discharged at current rates of 0.5C and 1C at four immersion ratios: 25%, 50%, 75% and 100% with respect to the total cell height. A cross-fold training methodology is employed to train the model making use of thermally dynamic input data. The results show that a root mean square error of less than 1.1 °C can be achieved for the fold displaying the poorest performance. Finally, this paper includes a processor-in-the-loop analysis, indicating the feasibility of the model to run on embedded hardware. It is shown that up to 50 cells may be monitored in real-time. The model offers the possibility to expand the safety and health monitoring in battery management systems for not only battery packs with novel thermal management solutions, but conventional solutions as well.