A novel predictive method for assessing the state of health (SOH) of electric vehicle (EV) batteries has been developed by researchers, utilizing advancements in deep learning techniques to improve accuracy over various prediction lengths.
Accurate SOH predictions are not just technical necessities; they are pivotal for enhancing operational safety and prolonging the lifespan of vehicle batteries. Failures or inaccurate estimations can lead to serious incidents, including spontaneous combustion, underscoring the need for reliable prediction models within battery management systems.
This innovative approach leverages the Frequency Enhanced Attention TimeMixer (FEA-TimeMixer) model, combined with long and short-term battery degradation feature extraction methodologies, to accurately predict the SOH of lithium-ion batteries. Past studies have largely focused on either empirical degradation models or real-time operational data, both facing inherent limitations such as inaccuracies during short-term fluctuations or complex prediction horizons.
The research introduces a dual methodology: first, it employs automatic SOH extraction algorithms to label battery degradation data, capturing both long-term trends and short-term capacity variations derived from unique empirical degradation models. Key to this process is the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) decomposition method, which helps separate and analyze different frequency characteristics of the battery data.
Experimental data collected from six new energy taxis provided the foundation for this study, with research spanning from August 2019 to May 2021. Smart charging stations gathered detailed measurements, including current, voltage, and temperature data, at 8-second intervals during the vehicle operations.
The FEA-TimeMixer model significantly enhances traditional time series forecasts by integrating time-domain and frequency-domain analyses. Results from the new model show it outperforms existing deep learning frameworks, achieving Mean Absolute Errors (MAE) of less than 0.0219 for short-term predictions and below 0.1007 for long-term forecasts.
Through the integration of multiple empirical degradation models, the new methodology addresses complexity and cumulative error common to standard prediction approaches. The model’s architecture facilitates adaptive attention mechanisms, allowing the model to focus on significant frequency components, thereby enriching the predictive power across varying time scales.
The findings suggest not only improved SOH estimation accuracy but also present significant advancements for the overall management of battery systems within the growing new energy vehicle sector. Researchers concluded the study by highlighting the potential for broader applications of the FEA-TimeMixer model, proposing future explorations around real-world battery performance monitoring and SOH estimation improvements.
This comprehensive prediction capability might pave the way for innovative developments within battery management systems, enhancing both the safety and efficiency of EVs as they continue to proliferate on roadways.