Аннотация
Precise state of charge (SoC) estimation is critical for optimizing lithium-ion battery performance in energy storage systems. This study proposes an innovative Lightweight Transformer using Encoder only with 1D CNN (LTEC) model to significantly improve estimation accuracy while maintaining low computational complexity. The model integrates a 1D CNN layer in place of the standard Feed-Forward Network (FFN) within the Transformer Encoder. By applying weight-shared kernels across localized time windows, the 1D CNN captures short-term dynamics and dampens high-frequency fluctuations in sensor inputs. Unlike sequential Recurrent Neural Networks (RNNs) that suffer from high latency and gradient decay, this architecture enables fast parallel processing. Moreover, it overcomes the temporal limitations of Fully Connected Neural Networks (FNNs) and the heavy parameter counts of standard Transformers by offering translation invariance with high parameter efficiency. The LTEC model also incorporates regularization and fine-tuning to enhance SoC estimation accuracy across four standardized drive profiles: the Urban Dynamometer Driving Schedule (UDDS), Highway Fuel Economy Test (HWFET), California Unified Cycle (LA92), and US06 Supplemental Federal Test Procedure. The proposed model demonstrates excellent results, achieving a Root Mean Square Error (RMSE) of 1.14% and Mean Absolute Error (MAE) of 0.52% on a wide temperature range of −10–40°C, showing substantial improvement over existing architectures. Furthermore, applying transfer learning to different lithium-ion batteries chemistry yielded salient performance in drifting temperature conditions. This outcome reinforces LTEC as a potential solution for advancing battery management systems, capable of effectively managing diverse operational conditions, including dynamic load and temperature fluctuations across different batteries.