Accounting-Guided Data Augmentation for LightGBM Prediction of Cradle-to-Gate Carbon Footprints of LFP and NMC811 Battery Cells
4 affiliations
- 1Zhejiang Institute of Quality Sciences, Hangzhou 310018, China
- 2School of Mechanical Engineering, Tianjin University, Tianjin 300350, China
- 3China Nuclear Power Operations Co., Ltd., Shenzhen 518000, China
- 4China Quality Mark Certification Group Zhejiang Co., Ltd., Hangzhou 310007, China
Abstract
Machine-learning models fitted to limited battery inventory data may reproduce nominal carbon footprints accurately while failing to preserve known inventory–emission relationships under changed inputs. This study evaluates accounting-guided data augmentation for component-wise LightGBM prediction of cradle-to-gate carbon footprints of lithium iron phosphate (LFP) and nickel manganese cobalt oxide (NMC811) pouch cells. Published inventories were expanded through ±10% input perturbations, zero-input samples, and proportional scaling. The models use material quantities, manufacturing-energy inputs, and regional production descriptors, with footprints normalized to 1 kWh of cell capacity. On the held-out synthetic test set, the framework achieved an R2 of 0.9991, a mean absolute error of 0.4486 kg CO2-eq/kWh, and a root mean square error of 0.6246 kg CO2-eq/kWh. When numeric inputs were scaled to 60% of their original values, augmentation improved R2 from −1.9425 to 0.8333 relative to unaugmented LightGBM. In a single random-masking realization, R2 remained above 0.995 with 50% of numeric inputs masked and imputed. These results show improved surrogate consistency within the represented inventory domain. They do not establish superiority over direct calculation or linear regression, or validate prediction from a restricted set of early-design inputs.
Related
- Accelerating Phase Prediction via CALPHAD-Informed Machine Learning and Data Augmentation
- Improved Chaotic Sparrow Search-Optimized Deep Multiple Kernel ELM for Lithium-Ion Battery RUL Prediction
- Hierarchical Multimodal Large Language Model-Based Interpretation and Aggregation of PDP and ICE Explanations for Battery State-of-Charge Prediction in Electric Vehicles
- FlowBatt: Flow Matching for Probabilistic Battery Degradation Prediction
- A Study on Lithium-Ion Battery Health Estimation and Remaining Life Prediction Based on Real-World Operational Data
BibTeX
@article{xu2026accountingguided,
title = {{Accounting-Guided Data Augmentation for LightGBM Prediction of Cradle-to-Gate Carbon Footprints of LFP and NMC811 Battery Cells}},
author = {Zhihao Xu and Haoyang Li and Songxian Peng and Guangzong Xu and Xinlong Wang and Lijing Wang},
journal = {Batteries},
year = {2026},
month = nov,
doi = {10.3390/batteries12110477},
url = {https://www.mdpi.com/2313-0105/12/11/477},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Accounting-Guided Data Augmentation for LightGBM Prediction of Cradle-to-Gate Carbon Footprints of LFP and NMC811 Battery Cells
AU - Zhihao Xu
AU - Haoyang Li
AU - Songxian Peng
AU - Guangzong Xu
AU - Xinlong Wang
AU - Lijing Wang
PY - 2026
DA - 2026/11/07
JO - Batteries
DO - 10.3390/batteries12110477
UR - https://www.mdpi.com/2313-0105/12/11/477
AB - Machine-learning models fitted to limited battery inventory data may reproduce nominal carbon footprints accurately while failing to preserve known inventory–emission relationships under changed inputs. This study evaluates accounting-guided data augmentation for component-wise LightGBM prediction of cradle-to-gate carbon footprints of lithium iron phosphate (LFP) and nickel manganese cobalt oxide (NMC811) pouch cells. Published inventories were expanded through ±10% input perturbations, zero-input samples, and proportional scaling. The models use material quantities, manufacturing-energy inputs, and regional production descriptors, with footprints normalized to 1 kWh of cell capacity. On the held-out synthetic test set, the framework achieved an R2 of 0.9991, a mean absolute error of 0.4486 kg CO2-eq/kWh, and a root mean square error of 0.6246 kg CO2-eq/kWh. When numeric inputs were scaled to 60% of their original values, augmentation improved R2 from −1.9425 to 0.8333 relative to unaugmented LightGBM. In a single random-masking realization, R2 remained above 0.995 with 50% of numeric inputs masked and imputed. These results show improved surrogate consistency within the represented inventory domain. They do not establish superiority over direct calculation or linear regression, or validate prediction from a restricted set of early-design inputs.
SN - 2313-0105
ER -