A support vector machine (SVM) approach treats SoC estimation as a regression problem: given a set of measured cell signals (current, voltage, temperature), find a function — built from a kernel that can be linear or nonlinear — that maps them onto SoC while keeping the fitted curve as simple as the data allows. This is one of the older machine-learning approaches applied to battery SoC, predating the more recent wave of deep-learning methods.
Compared with a straightforward least-squares fit, an SVM-based estimator is less sensitive to noisy or slightly-off training data, because its objective function tolerates small errors rather than trying to drive every one of them to zero. An early demonstration on a lithium iron manganese phosphate cell reported SoC accuracy better than 6% across both constant-current and variable-current test profiles using only current, voltage and temperature as inputs. A common variant, the least-squares SVM (LS-SVM), trades a little of this robustness for a much faster, closed-form training step.
See also RBF Neural Network for another regression-based approach to SoC.
References
- J.C. Álvarez Antón, P.J. García Nieto, et al., Support Vector Machines Used to Estimate the Battery State of Charge, IEEE Transactions on Power Electronics, Volume 28, Issue 12, pp. 5919–5926, 2013
Part of a series expanding on the SoC Estimation Techniques overview.