An advanced li-ion cell equivalent circuit model using a neuro-physical approach

The demand for energy-efficient systems in renewable energy and electric mobility has driven advancements in battery technology, particularly Li-Ion batteries, widely used in applications such as electric vehicles and smart grids. This paper proposes a novel hybrid approach combining an Equivalent Circuit Model (ECM) with ensemble feedforward Neural Networks (NNs) to model a battery cell. The electrical model used in this study features one RC-branch in a switched ECM configuration to adapt to different dynamics during charging and discharging and incorporates a snubber circuit to mitigate current spikes. Each parameter of the ECM is associated with a dedicated NN responsible for estimating its value, resulting in a total of seven NNs forming an Ensemble Neural Network (ENN) aligned with the discretized ECM. The NN estimates critical parameters, such as resistances and capacitors, in a physically interpretable space, ensuring that the model remains closely aligned with real-world components. The parameters of the ECM have been previously estimated by using an optimization procedure on real dataset, and the obtained values have been used as benchmark for the hybrid methodology. Simulation tests have been performed by varying the learning size and the batch size of the ENN. We present a comparison between traditional parameter estimation for standard ECMs and the proposed hybrid model. The hybrid approach, integrating NNs, demonstrates up to 50% improvement in the accuracy of the estimated voltage compared to the results obtained with the considered ECM. This study underscores the potential of combining advanced circuit design with machine learning to enhance the accuracy and reliability of battery management systems, pointing to a promising future for the field.