Accurate battery health diagnostics are essential for timely maintenance,replacement,and the safe operation of electric vehicles(EVs).For on-road EVs,leveraging operational data for accurate state-of-health(SOH)estima...Accurate battery health diagnostics are essential for timely maintenance,replacement,and the safe operation of electric vehicles(EVs).For on-road EVs,leveraging operational data for accurate state-of-health(SOH)estimation remains challenging due to varied degradation patterns across different driving conditions,vehicle types,and battery chemistries.Thus,developing an on-road-specific efficient feature system and a generalized SOH estimation framework adaptable to diverse EV models and chemistries is essential.To address these limitations,this study proposes a vehicle operational data-driven approach that integrates multi-dimensional feature fusion with a hybrid deep neural network architecture.Specifically,12.83 million on-road data points spanning a wide range of vehicle types and battery chemistries are processed.Capturing representational,driving behavioral,and electrochemical characteristics,this study proposes a three-dimensional feature system comprising shallow,intermediate,and deep descriptors.To tackle challenges posed by long time spans and the limited effectiveness of Transformer models on multivariate inputs,a hybrid framework combining temporal convolutional networks with an enhanced iTransformer is developed,incorporating a differential attention mechanism to suppress attention noise.Experimental results demonstrate that the proposed method achieves high accuracy across two test sets,with an average R^(2),MAPE,MAE,and RMSE of 98.88%,0.35%,0.31%,and 0.40%,respectively.This represents an 81.4%reduction in RMSE compared to the bestperforming baseline.Data scarcity experiments using reduced training data demonstrate that even when the training set is decreased from 80%to 30%,model performance remains stable,with the RMSE remaining below 0.16%.Feature attribution analysis using Shapley additive explanations(SHAP)confirms the indispensability of all three feature dimensions,with driving behavior features being particularly influential.Following feature optimization,training time is reduced by 17.3%.This study presents a robust SOH estimation framework tailored for intelligent cloud battery management systems,proactive maintenance,and the safe operation of EV batteries in practical environments.展开更多
基金supported by the National Key Research and Development Program of China(2024YFE0115800)the Department of Science and Technology of Guangdong Province(2023ZT10L145)。
文摘Accurate battery health diagnostics are essential for timely maintenance,replacement,and the safe operation of electric vehicles(EVs).For on-road EVs,leveraging operational data for accurate state-of-health(SOH)estimation remains challenging due to varied degradation patterns across different driving conditions,vehicle types,and battery chemistries.Thus,developing an on-road-specific efficient feature system and a generalized SOH estimation framework adaptable to diverse EV models and chemistries is essential.To address these limitations,this study proposes a vehicle operational data-driven approach that integrates multi-dimensional feature fusion with a hybrid deep neural network architecture.Specifically,12.83 million on-road data points spanning a wide range of vehicle types and battery chemistries are processed.Capturing representational,driving behavioral,and electrochemical characteristics,this study proposes a three-dimensional feature system comprising shallow,intermediate,and deep descriptors.To tackle challenges posed by long time spans and the limited effectiveness of Transformer models on multivariate inputs,a hybrid framework combining temporal convolutional networks with an enhanced iTransformer is developed,incorporating a differential attention mechanism to suppress attention noise.Experimental results demonstrate that the proposed method achieves high accuracy across two test sets,with an average R^(2),MAPE,MAE,and RMSE of 98.88%,0.35%,0.31%,and 0.40%,respectively.This represents an 81.4%reduction in RMSE compared to the bestperforming baseline.Data scarcity experiments using reduced training data demonstrate that even when the training set is decreased from 80%to 30%,model performance remains stable,with the RMSE remaining below 0.16%.Feature attribution analysis using Shapley additive explanations(SHAP)confirms the indispensability of all three feature dimensions,with driving behavior features being particularly influential.Following feature optimization,training time is reduced by 17.3%.This study presents a robust SOH estimation framework tailored for intelligent cloud battery management systems,proactive maintenance,and the safe operation of EV batteries in practical environments.
