Learning the P2D Model for Lithium-Ion Batteries with SOH Detection

TL;DR

This work replaces the computationally heavy P2D/DFN electrochemical model with a CNN surrogate trained on PyBaMM-simulated data to enable real-time, SOH-aware predictions for lithium-ion batteries. The nine-layer CNN processes Li-concentration profiles on electrode grids along with short-horizon cycle features to forecast voltage and concentrations over a 100 s window and to predict discharge failure probability, achieving high accuracy on one-step and multi-step (K-step) horizons with substantial speedups over the full P2D solver. Training on 18{,}000 synthetic driving cycles and testing on 3{,}000 cycles demonstrates that driving-cycle-rich data are essential; constant-current training degrades performance. The framework supports SOH estimation via a simple aging parameter $\gamma$ and suggests pathways to incorporate experimental data for improved accuracy, enabling SOH-aware, real-time BMS operation with potential extensions to more realistic aging effects.

Abstract

Lithium ion batteries are widely used in many applications. Battery management systems control their optimal use and charging and predict when the battery will cease to deliver the required output on a planned duty or driving cycle. Such systems use a simulation of a mathematical model of battery performance. These models can be electrochemical or data-driven. Electrochemical models for batteries running at high currents are mathematically and computationally complex. In this work, we show that a well-regarded electrochemical model, the Pseudo Two Dimensional (P2D) model, can be replaced by a computationally efficient Convolutional Neural Network (CNN) surrogate model fit to accurately simulated data from a class of random driving cycles. We demonstrate that a CNN is an ideal choice for accurately capturing Lithium ion concentration profiles. Additionally, we show how the neural network model can be adjusted to correspond to battery changes in State of Health (SOH).

Figures (5)

• Figure 1: An example driving cycle with the current $I(t)$ shown on the bottom left. The battery voltage, an externally measurable quantity, is on the bottom right. The electrolyte concentration and potential are shown on the top lines at the time indicated by the dotted line on the bottom figures. On the top row of figures, the positive electrode is to the left of the first vertical line, the negative electrode is to the right of the second vertical line, and the separator is between the lines.
• Figure 2: Intercallated Lithium concentrations in the negative (left) and positive (right) electrodes with scaled electrode width $x$ and particle radius $r$. These are at time 1200 of the driving cycle shown in Figure \ref{['f:f3']}, the time shown by a dashed vertical line in that figure.
• Figure 3: Schematic of Network architecture.
• Figure 4: The neural network K-step predicted voltage vs the true voltage curve for a simulation with an approximately average $l^\infty$ error over all of the validation data set runs.
• Figure 5: Predictions vs true voltages for a model trained with constant current data