针对锂离子电池动态放电过程中黑箱模型可解释性不足、随机划分导致评估乐观的问题，本文提出多项式贝叶斯回归与多链马尔可夫链蒙特卡洛(Markov Chain Monte Carlo， MCMC)相结合的剩余放电寿命代理预测方法。以LG 18650HG2公开数据集城市道路循环工况(Urban Dynamometer Driving Schedule， UDDS)为对象，将放电容量归一化为剩余容量比例(remaining＿life＿percent)，并将其定义为单次动态放电过程内剩余可用容量比例的代理变量，而非跨循环老化意义上的真实剩余使用寿命(Remaining Useful Life，RUL)。以电压、电流、温度、运行时间和功率为输入，通过一阶至三阶多项式及交互项展开特征空间，联合Gibbs多链采样进行贝叶斯后验推断，并利用Gelman-Rubin统计量■、有效样本量(Effective Sample Size， ESS)、方差膨胀因子(Variance Inflation Factor， VIF)和自相关函数(Autocorrelation Function， ACF)/偏自相关函数(Partial ACF， PACF)诊断模型稳定性。时间序列前向切分下，MCMC三阶模型取得RMSE=3.0337、R²=0.8497，残差修正后达RMSE=2.8456、R²=0.8677，优于Ridge线性基线(RMSE=8.9703、R²=-0.3142)。多链■、ESS=6891.6验证了收敛性。该方法融合可解释多项式建模、多链后验推断与时间前向验证，在避免时间泄漏的同时兼顾预测精度与不确定性量化，为动态工况电池管理提供辅助决策。

A polynomial Bayesian regression method with multi-chain Markov chain Monte Carlo sampling is proposed for remaining discharge-life proxy prediction of lithium-ion batteries under dynamic UDDS operating conditions. Using the LG 18650HG2 public dataset， the normalized remaining＿life＿percent proxy is constructed from discharge capacity-this proxy represents the remaining capacity ratio within a single discharge cycle， rather than the actual cycle-life RUL. Voltage， current， temperature， elapsed time and power are taken as inputs， and the feature space is expanded through first- to third-order polynomial and interaction terms. Bayesian posterior inference is conducted via multi-chain Gibbs sampling， and model stability is diagnosed using ■， ■， ■， and residual ■/■ tests. Under temporal forward splitting， the third-degree MCMC polynomial model achieves RMSE=3.0337，R²=0.8497， and the residual-corrected model reaches RMSE=2.8456，R²=0.8677， outperforming the Ridge linear baseline (RMSE=8.9703，R²=-0.3142). Multi-chain ■ and ESS=6891.6 confirm sampling convergence. The proposed framework balances nonlinear prediction accuracy and uncertainty quantification without temporal information leakage， offering an interpretable decision-support tool for battery management systems under dynamic conditions.