Improved Harmonic Loss-Long and Short-Term Memory Networks Incorporating Attention Mechanisms for Online State-of-Charge Estimation of Lithium-Ion Batteries in Large-Scale Energy Storage Stations

To guarantee the dependability and security of battery system operations, it is crucial to accurately assess lithium-ion batteries’ state of charge (SOC). Given that SOC is a time-dependent and cumulative parameter, its current value is significantly influenced by historical states. In order to improve prediction accuracy, this research presents a unique estimating technique that integrates a harmonic loss-based long short-term memory (LSTM) network with an attention mechanism. The proposed model, referred to as HL-LSTM-A, builds upon the conventional LSTM architecture by incorporating temporal dependencies and refining the loss function through the application of the fundamental forgetting curve. In order to emphasize the most recent time step in the training process and to better take into account the effects of the decay of early history states, a time-weighted loss function based on harmonic sequences (harmonic loss) is specifically introduced. This modification leads to notable improvements in SOC estimation precision. To evaluate the algorithm’s performance, a dynamic stress test (DST) scenario was employed. Results of the experiment demonstrate that the HL-LSTM-A method accomplishes a lower maximum mean absolute error (MAE) compared to both standard LSTM and Attention-LSTM models. Consequently, the HL-LSTM-A algorithm turns out to be a useful instrument for accurate state of charge prediction, offering improved efficiency in the context of large-scale lithium-ion energy storage systems.