Battery charging curve prediction via Fourier graph neural network fused prompt features

Complete charging curves offer valuable insights into battery states, including current capacity and capacity degradation, which are critical for ensuring the safety and longevity of battery-powered applications. However, harsh temperatures, dynamic operations and different battery materials pose challenges on entire charging curves estimation. In this work, we use a Fourier Graph Neural Network (FourierGNN) which could spatially-temporally accomplish time series forecasting from graph perspective, to estimate full constant-current charging curves using only random small segments of the curves. To enhance estimation accuracy and generalization, we incorporate battery degradation knowledge extracted from a large language model (LLM). The LLM processes structured prompts—including battery data descriptions, instructions, and dataset statistics—and encodes them into informative vectors that retain both the provided information and the model's learned knowledge. By integrating these vectors with FourierGNN's outputs, our method achieves highly accurate, generalizable, and robust charging curve estimations, yielding an average state of health (SOH) RMSE of 0.86 % and a MAX error of 4.24 % on the open-source Oxford dataset. Further validation across diverse datasets spanning varying current rates and temperatures demonstrates the method's robustness.