BP Neural Network (SoC Estimation)

A back-propagation (BP) neural network estimates SoC by learning a mapping directly from measured signals — typically voltage, current and temperature — to state of charge, without needing an explicit equivalent-circuit or electrochemical model of the cell. A training set of these input/output pairs, gathered from real charge/discharge cycles, is used to adjust the network’s internal weights via the back-propagation algorithm until its output error is acceptably small; from then on the trained network can estimate SoC from new voltage/current/temperature readings alone.

The appeal is that it sidesteps model derivation entirely and can pick up nonlinear behaviour a simpler lookup table would miss. The practical problems are well documented: plain BP training converges slowly, can get stuck in local minima rather than the true best fit, and easily overfits if the training data does not cover the operating range well. For this reason most recent SoC work does not use a bare BP network — it is paired with an optimiser (particle swarm, sparrow search, Levenberg-Marquardt) to speed up and stabilise training, or combined with a boosting method to average out the weaknesses of any one trained network.

See also RBF Neural Network and Fuzzy Neural Network for related data-driven approaches.