Battery state of charge prediction based on adaptive knowledge transfer mechanism and temporal deep learning model

State of charge (SOC) is an important prognostic parameter for assessing the performance level and reliability of battery-powered systems. Most of the existing deep learning predictive models for battery SOC are computationally intense and require huge quantities and quality training data, therefore making it challenging to adopt such a model for new battery cells with limited data. This paper proposes an Adaptive Knowledge Transfer Mechanism with Temporal Deep Learning Model (AKTDL) for short and long-term SOC predictions. The proposed framework comprises two long short-term memory (LSTM) networks, the source LSTM and the target LSTM. In essence, the proposed method enhances the predictive capability of the target LSTM through the transfer of domain-invariant capacity degradation information from a trained source LSTM using an effective domain adaptation mechanism. This approach enables the target model to leverage the learned knowledge from the source model, even with limited data for the test battery cell. The effectiveness of the proposed method was validated and benchmarked against existing LSTM and fine-tuning transfer learning methods using a public battery dataset. Experimental results showed that the AKTDL outperforms the benchmark models for short and long-term SOC prediction.