This work presents a methodology for developing State of Charge (SoC) estimation algorithms for lithium-ion (Li-ion) and sodium-ion (Na-ion) batteries using Long Short-Term Memory (LSTM) recurrent neural networks combined with Transfer Learning (TL). Accurate SoC estimation is a key requirement for advanced Battery Management Systems (BMS), as internal battery states such as SoC, State of Health (SoH), and State of Power (SoP) cannot be directly measured and must be inferred from measurable signals like voltage, current, and temperature. Conventional modelling approaches typically require extensive laboratory testing for each new battery chemistry, cell type, and application. This requirement becomes particularly problematic for emerging technologies such as Na-ion batteries, where experimental datasets remain limited and modelling practices are still evolving.

To address these challenges, the proposed approach leverages deep learning and knowledge transfer to reduce experimental effort and accelerate algorithm development. A baseline LSTM model is first trained on a large and diverse dataset of NMC/C lithium-ion cells, covering a wide range of operating conditions including different temperatures, current rates, and aging states. The model uses voltage, current, temperature, and SoH as inputs to estimate SoC. Training on such a comprehensive dataset enables the network to learn generic electrochemical and dynamic battery behaviours that can generalize across multiple scenarios.

The methodology also incorporates data augmentation strategies, such as the use of virtual datasets generated from electrochemical models, which simulate diverse operating conditions and degradation levels. In addition, the integration of in-field operational data is considered for continuous retraining, enabling adaptive improvement of model accuracy and robustness over time.

Transfer Learning is then applied to adapt the Li-ion trained model to new cell references and different battery chemistries, particularly Na-ion batteries. Several TL strategies are evaluated, including freezing early network layers to preserve learned temporal features and fine-tuning the full model to adapt to new datasets. This approach allows the model to retain general knowledge about battery dynamics while adjusting to the specific characteristics of the target cells.

Results show that, for Li-ion cells, TL-based models outperform reduced models trained from scratch, achieving lower mean absolute errors and significantly reduced maximum estimation errors while requiring up to 50% less training data. When transferring knowledge from Li-ion to Na-ion chemistry, the approach enables the creation of functional SoC estimators using a limited amount of Na-ion data. Good accuracy is achieved under constant-current charging and discharging conditions, although larger errors appear in highly dynamic profiles and high C-rate scenarios due to the higher internal resistance and lower efficiency of Na-ion cells.

Overall, the study demonstrates that combining LSTM networks with Transfer Learning is an effective strategy to reduce development time, experimental costs, and data requirements for SoC estimation algorithms. While cross-chemistry knowledge transfer from Li-ion to Na-ion is feasible and useful for rapid prototyping, optimal performance still requires chemistry-specific fine-tuning and validation. The proposed framework supports the development of universal SoX estimation algorithms, facilitating the deployment of both mature lithium-ion technologies and emerging post-lithium battery systems.