Hybrid Metaheuristic Feature Selection for Automotive Battery Capacity and Health Prediction Models
- 1Department of Electrical and Electronic Engineering Technology, University of Johannesburg, Johannesburg 2092, South Africa
Abstract
Feature selection and model tuning are coupled decisions in real-world electric-vehicle battery analysis: the search must balance predictive performance, feature burden, model complexity, and computational demand. This study develops three integrated hybrids—Adaptive Differential Evolution–Particle Swarm Optimisation (Adaptive DE–PSO), Harris Hawks–Grey Wolf Optimisation (HHO–GWO), and Whale–Marine Predators Optimisation (WOA–MPA)—for joint feature selection and tree-ensemble tuning. The empirical cohort comprises 191 capacity-labelled vehicles (M1 = 100, M2 = 83, M3 = 8) derived from the anonymised EVBattery field dataset, with 64 engineered charging predictors and vehicle-disjoint 60%/20%/20% train/validation/test partitions. Every search arm used a population of 20 for 100 iterations (2020 proposals), and all 13 primary algorithms were evaluated across 30 paired runs; nine mechanism ablations were evaluated under the same protocol. The fixed Default ensemble produced the strongest mean held-out capacity result (R2 = 0.697, 95% CI 0.675–0.719; RMSE = 0.852 Ah) and low-capacity-proxy discrimination (AUROC = 0.942, 95% CI 0.928–0.956). Adaptive DE–PSO was the strongest integrated hybrid for capacity (R2 = 0.659), selecting 19.0 features on average, but it remained significantly below Default after a Holm correction (mean difference −0.038; adjusted p = 0.010). For the proxy endpoint, Adaptive DE–PSO achieved AUROC = 0.911, also below Default (adjusted p < 0.001). Thus, the novelty is not a claim of universal hybrid superiority; it is a leakage-safe, equal-budget, multi-objective framework that quantifies when hybridisation exchanges predictive accuracy for sparsity and exposes negative as well as positive mechanism effects. The proxy is cohort-relative and is not a laboratory-validated state-of-health diagnosis.
Related
- Construction and Evaluation of a Ketosis Prediction Model Based on Milk MIR Spectra and Feature Wavenumber Selection in Dairy Cows.
- Deep Neural Network-Based Prediction of Temperature-Dependent Tensile Stress–Strain Behavior in Injection-Molded Impact-Modified Heat-Stabilized PA66 for Automotive Applications
- Multi-Center Study on Sample Preparation and Feature Selection Effects in Reproducible Mass Spectrometry Imaging for Clinical Translation
- Capacity and SOH Prediction of Lithium-Ion Batteries Using Various Mathematical Forms of EIS with Neural Networks
- Agricultural waste-based composites for automotive interiors using waste to value approach
BibTeX
@article{makama2026hybrid,
title = {{Hybrid Metaheuristic Feature Selection for Automotive Battery Capacity and Health Prediction Models}},
author = {Mavis Fortunate Makama and Bonginkosi A. Thango},
journal = {Batteries},
year = {2026},
month = oct,
doi = {10.3390/batteries12100424},
url = {https://www.mdpi.com/2313-0105/12/10/424},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Hybrid Metaheuristic Feature Selection for Automotive Battery Capacity and Health Prediction Models
AU - Mavis Fortunate Makama
AU - Bonginkosi A. Thango
PY - 2026
DA - 2026/10/14
JO - Batteries
DO - 10.3390/batteries12100424
UR - https://www.mdpi.com/2313-0105/12/10/424
AB - Feature selection and model tuning are coupled decisions in real-world electric-vehicle battery analysis: the search must balance predictive performance, feature burden, model complexity, and computational demand. This study develops three integrated hybrids—Adaptive Differential Evolution–Particle Swarm Optimisation (Adaptive DE–PSO), Harris Hawks–Grey Wolf Optimisation (HHO–GWO), and Whale–Marine Predators Optimisation (WOA–MPA)—for joint feature selection and tree-ensemble tuning. The empirical cohort comprises 191 capacity-labelled vehicles (M1 = 100, M2 = 83, M3 = 8) derived from the anonymised EVBattery field dataset, with 64 engineered charging predictors and vehicle-disjoint 60%/20%/20% train/validation/test partitions. Every search arm used a population of 20 for 100 iterations (2020 proposals), and all 13 primary algorithms were evaluated across 30 paired runs; nine mechanism ablations were evaluated under the same protocol. The fixed Default ensemble produced the strongest mean held-out capacity result (R2 = 0.697, 95% CI 0.675–0.719; RMSE = 0.852 Ah) and low-capacity-proxy discrimination (AUROC = 0.942, 95% CI 0.928–0.956). Adaptive DE–PSO was the strongest integrated hybrid for capacity (R2 = 0.659), selecting 19.0 features on average, but it remained significantly below Default after a Holm correction (mean difference −0.038; adjusted p = 0.010). For the proxy endpoint, Adaptive DE–PSO achieved AUROC = 0.911, also below Default (adjusted p < 0.001). Thus, the novelty is not a claim of universal hybrid superiority; it is a leakage-safe, equal-budget, multi-objective framework that quantifies when hybridisation exchanges predictive accuracy for sparsity and exposes negative as well as positive mechanism effects. The proxy is cohort-relative and is not a laboratory-validated state-of-health diagnosis.
SN - 2313-0105
ER -