Zinc-air batteries (ZAB) represent a promising energy storage technology, as they have the potential to achieve up to three times the energy density of lithium-ion batteries. In these cells, zinc functions as the primary reactant and is abundant, while oxygen from the surrounding air serves as the secondary reactant. Furthermore, zinc is a favourable material due to its capacity for recycling and its non-toxic nature. In order to fully capitalise on the aforementioned advantages in a practical context, it is essential that the state of charge (SoC) is accurately and reliably estimated.
Traditional methods of estimating the SoC, such as open-circuit voltage measurements, are not very accurate when used with ZAB. The voltage profile of a ZAB during charging and discharging is largely insensitive to changes in SoC, with the difference between the fully discharged and fully charged states being only about 60 mV. Instead, external factors such as temperature and cell ageing, particularly electrolyte degradation, have a significantly greater influence on cell voltage than SoC itself. Consequently, precisely determining the SoC in ZAB systems remains a considerable challenge.
Electrochemical impedance spectroscopy (EIS) is a well-established technique for characterising battery behaviour. The impedance values are only valid for the cell’s current state. Each spectrum comprises 40 measurements recorded at different frequencies ranging from 100 mHz to 6.4 kHz, with eight points per decade. The impedance response depends on the SoC, with higher SoC levels leading to semicircles with larger radii and a slight shift towards lower real-axis values.
To estimate the SoC of ZAB, a Support Vector Classifier (SVC) was trained using EIS recorded during controlled charge and discharge cycles. The input feature vector included the real and imaginary parts of the impedance at 40 frequencies, as well as battery voltage, temperature, current, and derived quantities such as magnitude and phase angle.
Instead of predicting continuous SoC values, the problem was formulated as a multi-class classification task. The continuous SoC range was divided into 11 discrete classes, each covering a 10% SoC interval. This discretization increases robustness against measurement variability and cell-to-cell differences, particularly in hand-assembled ZAB, where exact regression becomes unreliable.
To improve the classifier’s robustness and generalization capability, several preprocessing steps were applied. The ohmic resistance component was subtracted from each impedance spectrum to reduce the influence of aging and electrolyte degradation. Additionally, a current-generalization strategy (DC-split) ensured reliable predictions across different DC operating points by training and testing the model on measurements collected at distinct current levels.
Three kernel functions were investigated: linear, radial basis function (RBF), and polynomial. The polynomial kernel achieved the highest classification accuracy of 92% using raw impedance data, as shown in the confusion matrix. The RBF kernel yielded better generalization performance under varying currents, while the linear kernel performed best when additional degradation features were included. Overall, the SVC approach demonstrated reliable classification of SoC from impedance spectra, with performance depending on the chosen kernel and feature set.