Charge-Curve Prediction Using Machine Learning to Identify Battery Charge, Health
Researchers have developed a machine learning model to predict battery charge and health with minimal charge-curve data.
Researchers have developed a machine learning model to predict battery charge and health with minimal charge-curve data.
Lithium-ion batteries (LIBs) are the dominant rechargeable batteries in the market, and there has been increasing interest in accurate estimation of the state of charge (SoC) and state of health (SoH) of these batteries in real time.
Electric vehicles are one of many uses for lithium-ion batteries. Image used courtesy of Adobe Stock
As LIBs undergo charge and discharge cycles, they experience gradual degradation due to chemical reactions, temperature fluctuations, mechanical stresses, and other factors. SoH helps to predict the battery's degradation status, capacity loss, and overall performance loss over time.
The accuracy of the estimation of SoC is also critical. SoC refers to the percentage of the total capacity of the battery remaining. In applications like electric vehicles (EVs), accurate SoC estimation allows drivers to know the vehicle's remaining range. Battery management systems (BMSs) carry out these assessments in any application. However, their computation is not trivial, as BMSs usually collect data at regular intervals. Therefore, rapid fluctuations or short-lived anomalies in battery behavior might be neglected due to the limited sampling frequency.
In addition, BMS collects data within a specific State of Charge (SoC) range that reflects the typical patterns of customers. This range covers the usual operating conditions that the battery undergoes. However, this might lead to insufficient data to accurately capture the complex battery behavior.
As mentioned previously, there have been many efforts in estimating the health of batteries in real applications. One of the most common methods is using equivalent circuit models and mechanism-based models to simulate the behaviors of batteries. It can also be derived from the charge curve. The researchers at Carnegie Melon University compare different methods to quantify the aging mechanism and found that unsupervised learning algorithms perform better in capturing features necessary to predict SoH.
New Machine Learning Model to Estimate SoC and SoH in Real-Time
Machine learning (ML) models offer a promising approach to improve the accuracy of SoC and SOH estimations in LIBs, as they can automatically identify and incorporate relevant features from the data. Furthermore, batteries' behaviors change over time and with usage patterns, and ML models can adapt to these changes to maintain accuracy.
What is unique about the ML model of the researchers at Carnegie Melon University is that their system takes minimal data, requiring just 15 minutes of charge time to calculate the curve and determine battery health. Moreover, the data can also be collected in increments. So, even if the charging is interrupted, the analysis progress will not be affected. The researchers also claim that their model is transferable to batteries with different cathode materials.
In a study published in Advanced Science, researchers used a lab-generated and an open-access dataset. They selected charge curves for this study as the charging protocols are more flexible than discharge protocols to provide consistent data in practical applications. To quantify the capability of estimating battery degradation using the model, they compared the difference between the reconstructed charge curves and three tested charge curves with minimum, maximum, and median capacity.
The following figure shows that the measured curve fits well with the median capacity curve. Applying ML algorithms improves the reconstruction accuracy for all three curves, especially maximum and minimum capacity.
Battery charge curve feature extraction and reconstruction using ML algorithms. Image Courtesy of Carnegie Melon University.
The
current model is trained with data collected at a constant charging
rate and room temperature. The researchers say using real-world data as
input would be the next step to improve the model. Moreover, they
mention that neural network models may improve learning even further.
