What question did this study set out to answer?

This research aims to predict the life cycle of lithium-ion batteries using a hybrid modeling approach that integrates physics and data-driven features.

May 10, 2026Computers, materials & continua/Computers, materials & continua (Print)0 citationsOpen Access

A Hybrid Physics-Informed and Data-Driven Feature Framework with Explicit Correlation-Structure Embeddings for Early-Life Prognostics of Lithium-Ion Batteries

KLKang-Woo LeeSungkyunkwan University DLDong-Hee LeeSungkyunkwan University DKDae-Il KwonSungkyunkwan University

Key Points

This research aims to predict the life cycle of lithium-ion batteries using a hybrid modeling approach that integrates physics and data-driven features.
Developed a hybrid framework combining ΔQ-based indicators, statistical features, and Laplacian Eigenmaps embeddings.
Evaluated five feature configurations using pointwise regression accuracy metrics on a benchmark dataset.
Utilized XGBoost for prediction and compared performance against PCA and other dimensionality reduction techniques.
Integrated Laplacian configuration showed the strongest RMSE (97.00 cycles) and R2 (0.8215) on Primary test domain.
Hybrid Append configuration improved error metrics on early-cycle predictions relative to ΔQ + Statistics.
On the Secondary domain, ΔQ Only yielded lower RMSE but improvements were noted in selected metrics with Hybrid Append and VIF + Laplacian.

Abstract

Early-life cycle-life prediction for lithium-ion batteries—estimating end-of-life from initial cycles—is valuable for rapid cell screening and battery health management. We investigate whether an explicit correlation-structure descriptor can complement physics-informed ΔQ-based indicators and generic early-cycle statistical features on the Severson 124-cell benchmark. We develop a lightweight hybrid framework that combines ΔQ-based health indicators, data-driven statistical features, and Laplacian Eigenmaps embeddings derived from a Pearson-correlation feature graph, with XGBoost used as the predictor. Across five feature configurations (ΔQ Only, ΔQ + Statistics, Hybrid Append, VIF + Laplacian, and Integrated Laplacian), we evaluate pointwise regression accuracy using RMSE and R2 together with PHM-style error-band measures RA@0.2, PH@0.1, and α-λ(0.15), computed on implied RUL trajectories induced by the early-life cycle-life estimate. On the Primary test domain, all four non-baseline configurations improved over ΔQ Only; Integrated Laplacian achieved the strongest RMSE/R2 pair (97.00 cycles, 0.8215), while Hybrid Append remained competitive (102.28 cycles, 0.8016) and improved RA@0.2 and PH@0.1 relative to ΔQ + Statistics. On the shifted Secondary domain, ΔQ Only gave the most favorable RMSE/R2 pair (267.82 cycles, 0.2275), whereas Hybrid Append and VIF + Laplacian improved selected error-band metrics. In an additional comparison against PCA, Random Projection, and Truncated SVD conducted at a matched 79-feature scale, with all transforms estimated from the training cells only, the graph-derived embedding remained competitive, but its margin over simpler reductions varied across splits. Taken together, these results support Hybrid Append as the main appended-structure configuration in this study, while indicating that the benefit of the correlation-structure descriptor is more visible in selected PHM-style error-band metrics than in uniformly improved pointwise accuracy.

问 AI

Bookmark

View Full Paper