Accurately predicting the knee point cycle number,marking the onset of accelerated capacity degradation,is the first critical signal for managing lithium-ion battery life cycles in electric vehicles and stationary ene...Accurately predicting the knee point cycle number,marking the onset of accelerated capacity degradation,is the first critical signal for managing lithium-ion battery life cycles in electric vehicles and stationary energy storage systems.However,the inherent electrochemo-mechanical-thermal complexities of battery aging present significant challenges for physics-based models and machine-learning models,often leading to reduced predictive accuracy.Our study developed a comprehensive dataset comprising 20 lithium nickel manganese cobalt oxide(NCM)/graphite cells(0.5-1 C)from our lab and 162 commercial lithium iron phosphate(LFP)/graphite cells(3-6 C)from the public database,with knee point observed between 100 and 1000 cycles.We proposed a new strategy to extract novel features with strong physical context from early-cycle voltage curves,enabling precise knee point predictions across the chemistries without the need for extensive cycling histories.Our model achieved a mean absolute percentage error(MAPE)of 7%for knee point prediction using five selected features.Remarkably,the model yielded 8%MAPE with only one single feature across the initial 200 cycles,and 10%MAPE when applying five features across the initial 50 cycles,spanning different battery chemistries.This work highlights the potential of integrating multi-chemistry datasets with data-driven modeling to forecast aging patterns across diverse battery chemistries,advancing battery longevity and reliability.展开更多
基金supported by Startup funding from the University of Delaware。
文摘Accurately predicting the knee point cycle number,marking the onset of accelerated capacity degradation,is the first critical signal for managing lithium-ion battery life cycles in electric vehicles and stationary energy storage systems.However,the inherent electrochemo-mechanical-thermal complexities of battery aging present significant challenges for physics-based models and machine-learning models,often leading to reduced predictive accuracy.Our study developed a comprehensive dataset comprising 20 lithium nickel manganese cobalt oxide(NCM)/graphite cells(0.5-1 C)from our lab and 162 commercial lithium iron phosphate(LFP)/graphite cells(3-6 C)from the public database,with knee point observed between 100 and 1000 cycles.We proposed a new strategy to extract novel features with strong physical context from early-cycle voltage curves,enabling precise knee point predictions across the chemistries without the need for extensive cycling histories.Our model achieved a mean absolute percentage error(MAPE)of 7%for knee point prediction using five selected features.Remarkably,the model yielded 8%MAPE with only one single feature across the initial 200 cycles,and 10%MAPE when applying five features across the initial 50 cycles,spanning different battery chemistries.This work highlights the potential of integrating multi-chemistry datasets with data-driven modeling to forecast aging patterns across diverse battery chemistries,advancing battery longevity and reliability.
