Transfer learning-enhanced hybrid deep neural network model for accurate lithium-ion batteries health estimation in electric vehicles

The safety, effectiveness, and long lifetime of electric vehicles (EVs), electric wheelchairs, e-bikes, drones, and other battery-powered mobility systems depend on the precise estimation of the state of health (SoH) of lithium-ion batteries (LIBs). However, traditional machine learning techniques often face limitations in feature extraction, generalization, and convergence speed. This work introduces an innovative hybrid deep learning (DL) model that combines long short-term memory (LSTM) and multi-layer perceptron (MLP) architectures, optimized using a newly developed hybrid arithmetic firehawk optimization algorithm (HAFHOA). Transfer learning (TL) is also employed to transfer knowledge across heterogeneous aging patterns and battery chemistries, enhancing model adaptability to diverse operating conditions. The model leverages processed incremental capacity (IC) curves to derive robust health indicators from noisy degradation data, enabling precise SoH predictions. Validated on two benchmark datasets, NASA and CALCE, the proposed TL-HAFHOA-optimized LSTM+MLP model achieved outstanding accuracy, with RMSE of 0.0081 %, RMSPE of 0.0048 %, MAPE of 0.08108 %, and R 2 of 99.75 %. Comparative results confirm their superiority over conventional approaches in terms of prediction accuracy, training stability, and real-world applicability. This work contributes a scalable and intelligent solution for LIB health diagnostics across next-generation electric mobility and energy storage systems. • A hybrid LSTM+MLP model with TL and HAFHOA ensures robust SoH prediction. • Processed IC curves extract reliable health factors from noisy degradation data. • The model achieves 99.75 % R 2 with RMSE of 0.0081 % across NASA and CALCE datasets. • Transfer learning enhances adaptability across diverse battery chemistries. • The approach outperforms state-of-the-art methods in accuracy and stability.