A novel variable activation function-long short-term memory neural network for high-precision lithium-ion battery capacity estimation

Capacity estimation of lithium-ion batteries is significant to achieving the effective establishment of the prognostics and health management (PHM) system of lithium-ion batteries. A capacity estimation model based on the variable activation function-long short-term memory (VAF-LSTM) algorithm is proposed to achieve the high-precision lithium-ion battery capacity estimation. By re-selecting each activation function, the proposed algorithm avoids the low estimation accuracy caused by the fixed activation function of the long short-term memory (LSTM) algorithm, and meanwhile, it can effectively speed up the convergence. The algorithm inputs consider two correlation coefficients so that the health factor with the highest correlation coefficient is chosen as the network input. The experimental data used for the experimental validation is the NASA public battery data under different temperature operating conditions. The validation results show that the estimation accuracy of the VAF-LSTM algorithm under different training sets is greatly improved compared with the traditional LSTM algorithm and the back propagation (BP) algorithm, and the average estimation accuracy can reach more than 97.5%. The improvement of estimation accuracy is also clearly demonstrated under the MAE, MSE, and RMSE. Therefore, the capacity estimation model will provide an important reference role in high-precision battery management systems.