Prediction of Lithium-Ion Batteries Capacity Based on Limited Data And Transfer Learning

Abstract

Battery capacity estimation and prediction are critical for emerging energy systems. A proposed model employs transfer learning (TL) and long short-term memory (LSTM) neural networks for reliable lithium-ion battery capacity estimation and prediction. Trained on a fixed voltage window feature, this model can generalize across different battery types and charging strategies. It uses current battery state features to estimate and predict capacities for several future cycles. With the National Aeronautics and Space Administration (NASA) and Massachusetts Institute of Technology (MIT) datasets, results show estimation errors under 3%, prediction errors below 5.57%, and a 60.61% data volume reduction in training using transfer learning for the NASA dataset. The MIT dataset presents a 91.7% chance of prediction errors staying below 4% within 200 prediction steps, demonstrating the model's potential in data-limited scenarios.

Keywords: transfer learning, Long Short-Term Memory, lithium-ion batteries, capacity prediction