A recurrent deep neural network for predicting the state of health of lithium‐ion batteries

Abstract

AbstractBattery State of Health (SoH) estimation is a critical task in the field of battery management, as it provides information about remaining capacity and health of a battery. Various machine learning algorithms, including neural networks, decision trees, support vector machines, and random forests, have been utilized for battery SoH estimation. These models can be trained using different features, such as voltage, current, temperature, impedance, and their combinations. However, the diversity of data is a decisive factor that affects precision of battery SoH estimation using machine learning. In this research, the application of feedforward neural networks (FNNs) and recurrent neural networks (RNNs) is explored for the purpose of accurately estimating the SoH of batteries. These approaches are chosen due to the inherent benefits of FNNs and RNNs in capturing the long‐term dependencies present in sequential data. The battery SoH estimations are evaluated using the single and multichannel input: voltage, current, voltage‐current, voltage‐temperature, and voltage‐current‐temperature. The experimental findings reveal that the proposed RNN model, specifically the RNN with 20 neurons (RNN20) variant, exhibits an enhanced accuracy in predicting the SoH of batteries. For instance, when utilizing voltage and current as inputs, the RNN20 model demonstrated superior performance, achieving an mean absolute error (MAE) of 0.010157 for voltage and 0.010367 for current, outperforming the FNN with 10 neurons (FNN10) model, which yielded an MAE of 0.031635 and 0.065 for voltage and current, respectively. Furthermore, when employing diverse input combinations such as voltage‐current, voltage‐temperature, and voltage‐current‐temperature, the RNN20 model consistently outperformed its counterparts, exhibiting the lowest mean squared error and mean absolute percentage error across all metrics. These results underscore the RNN20 model's robustness and aptitude in accurately predicting battery SoH, affirming the merits of employing RNNs in battery management systems.