Neural ordinary differential equations and recurrent neural networks for predicting the state of health of batteries

• ODE models achieve errors of 1% in SOH and 5% in EOL estimation when predicting remaining useful life. • Models can be pretrained for improved performance. • Among all modes, augmented neural-ODE's perform best. Battery management systems require efficient battery prognostics so that failures can be prevented, and efficient operation guaranteed. In this work, we develop new models based on neural networks and ordinary differential equations (ODE) to forecast the state of health (SOH) of batteries and predict their end of life (EOL). Governing differential equations are discovered using measured capacities and voltage curves. In this context, discoveries and predictions made with neural ODEs, augmented neural ODEs, predictor-corrector recurrent ODEs are compared against established recurrent neural network models, including long short-term memory and gated recurrent units. The ODE models show good performance, achieving errors of 1% in SOH and 5% in EOL estimation when predicting 30% of the remaining battery's cycle life. Variable cycling conditions and a range of prediction horizons are analyzed to evaluate the models’ characteristics. The results obtained are extremely promising for applications in SOH and EOL predictions.