多健康指标的锂离子电池健康状态无分布区间估计OA

To address the limitations of conventional state of health(SOH)point estimation methods for lithium-ion batteries,this study develops a more practical SOH interval estimation method.Most previous interval estimation methods rely on distribution assumptions.However,when real-world battery data deviate from these assumptions,estimation biases may be introduced,consequently reducing estimation reliability.Consequently,the distribution-free lower upper bound estimation(LUBE)method has gradually attracted attention,although it still faces critical challenges.First,the loss function is non-differentiable,complicating model optimization.Second,several studies employed Sigmoid functions to transform non-differentiable loss functions into differentiable loss functions;however,this approach often requires manual slope-parameter tuning.Third,previous studies mostly rely on capacity as an ideal health indicator;however,accurately measuring capacity is costly,and this limits the real-world applicability of this method.To address these shortcomings,this study proposes a distribution-free SOH interval estimation method for quantifying LIB SOH using multiple health indicators.First,the kernel principal component analysis(KPCA)method is applied to reduce the dimensionality of the extracted health indicators.Based on this,a dual-output neural network model is constructed;this model introduces a loss function that eliminates the need for manual slope-parameter tuning,enabling it to stably output high-quality prediction intervals based on the reduced-dimensional data.Experimental results using the publicly available CALCE dataset demonstrate that the proposed method consistently meets nominal-confidence-level requirements while significantly improving prediction-interval quality.