Lithium-Ion Battery Sensorless Temperature Estimation via Integrating Data Enhancement and Transfer Learning

Due to the spatial and cost limitations of temperature sensors, there is a limited amount of directly measurable temperature information in battery systems. Estimating the battery’s state of temperature (SOT) through other signals can address this issue. This article combines data enhancement with transfer learning to propose a sensorless SOT estimation method for lithium-ion batteries. It effectively estimates temperature using only voltage and current. First, the Savitzky–Golay (SG) and mean filtering are used to reduce the noise and retain the fluctuation features of the data. The curves filtered by SG are used to extract differential features that obtain the relationship between temperature and voltage-current. Second, input the original data, filtered data, and differential features into two data enhancement algorithms, adding random noise and quantization, thereby increasing the diversity of the data. Finally, the iTransformer model is utilized to estimate the SOT of the target domain based on the transfer strategy. In most experiments, the RMSE is less than 1.0°C, and in many more challenging estimation cases, the RMSE does not exceed 2.0°C. Compared to other comparative models, this method demonstrates higher accuracy and better generalization ability in general. However, it performs poorly in estimating critical peak temperatures in few cases.