RBF Neural Network (SoC Estimation)

A radial basis function (RBF) network is a neural network built from Gaussian-shaped basis functions rather than the sigmoid layers used in a standard BP network. Each hidden neuron responds most strongly to inputs near its own centre point and fades off with distance, so the network as a whole builds up its SoC estimate as a weighted sum of these localised responses rather than one global nonlinear mapping.

For battery SoC estimation the practical draw is training speed and stability: because each hidden neuron’s parameters can be fitted somewhat independently, RBF networks generally converge faster and more reliably than BP networks trained end-to-end, at some cost in how compactly they can represent very complex relationships. Recent work has focused on selecting a smaller, more informative set of input features rather than using every measured signal, and tuning the RBF centres and widths with an optimiser such as particle swarm optimisation — both measurably cut estimation error compared with a naively-configured RBF network.

See also Support Vector Machine, another data-driven regression approach applied to SoC.