Lithium-ion batteries gradually lose capacity during operation due to complex, nonlinear degradation processes that depend on operating conditions. Accurate prediction of the remaining useful life is therefore essential for battery design and for ensuring the safety, reliability, and cost-effectiveness of battery systems.
An LSTM neural network was used to predict the capacity degradation of nine lithium-ion cells aged under different cycling conditions. Input features were calculated for each cycle from current, voltage, and temperature measurements, with the aim of investigating the performance of computationally efficient machine learning methods. Two approaches were evaluated: continuous capacity prediction based on all previously available operating data, and prediction of the subsequent degradation trajectory based on the first 200 operating cycles. The model was evaluated using a leave-one-battery-out cross-validation approach and the RMSE, MAE, and R² metrics. In continuous capacity prediction, the model achieved an MAE of 1.34 and an RMSE of 1.65. It also predicted the end of life of the battery cells with relatively high reliability.
In contrast, predicting the subsequent degradation trajectory based on the first 200 cycles proved unsuitable, with an MAE of 8.56 and an RMSE of 9.77. After an initial deviation, the predicted capacity remained approximately constant and did not follow the actual capacity decline.
|
