A Reference-Based Feature Representation Approach for Battery Remaining Useful Life Prediction Using K-Nearest Neighbors Regression
Journal of Science and Technology, cilt.5, sa.1, ss.25-39, 2026 (Hakemli Dergi)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 5 Sayı: 1
- Basım Tarihi: 2026
- Doi Numarası: 10.69560/cujast.1904437
- Dergi Adı: Journal of Science and Technology
- Derginin Tarandığı İndeksler: Index Copernicus, Asos İndeks, Other Indexes
- Sayfa Sayıları: ss.25-39
- Sivas Cumhuriyet Üniversitesi Adresli: Evet
Özet
This study systematically investigates the effect of feature representation design on performance in Remaining Useful Life (RUL) prediction for lithium-ion batteries. While RUL prediction in existing literature primarily focuses on model architecture and hyperparameter optimization, the role of feature representation design, especially concerning cross-battery generalization, has been addressed to a limited extent. In this context, two different representation approaches, namely absolute feature space-based representation (Absolute Feature Representation (AFR)) and reference cycle-centered representation (Reference-based Feature Representation (RFR)), were developed and evaluated. Experiments were conducted on data from 4 batteries within the NASA PCOE lithium-ion battery dataset, and evaluation was performed using a battery-wise cross-validation strategy. K-Nearest Neighbors (KNN) was used as the regression model, and only the representation effect was analyzed by keeping model parameters constant. The results demonstrate that RFR provides consistent improvements across all performance metrics. In the evaluation metrics, the average MAE value decreased from 0.148 to 0.096, and the RMSE value decreased from 0.173 to 0.120. The R² value increased from 0.637 to 0.825. These improvements correspond to performance increases of approximately 35%, 31%, and 29%, respectively. Correlation analysis revealed that the transformation performed in RFR strengthened the feature-target relationship and made relative degradation dynamics more prominent by suppressing absolute level differences. The diagnostic analyses performed confirm that the error structure is balanced and systematic. The findings indicate that performance improvement in RUL prediction is strongly dependent not only on model selection but also on feature representation design. Especially in cases where initial differences between batteries are high, the RFR approach offers an effective strategy that provides significant and consistent improvements without increasing model complexity.