Abstract:

Accurate estimation of the state of charge (SOC) of lithium batteries is the basis for safe and stable operation of lithium batteries. The traditional back propagation (BP) neural network has low accuracy in estimating SOC, and the recurrent neural network (RNN) is easy to fall into local optimization. To work out these problems, a recurrent neural network (ACGA-RNN) combined algorithm based on an adaptive catastrophe genetic algorithm is proposed. The adaptive catastrophe genetic algorithm (ACGA) is utilized to optimize the initial weight and threshold of RNN and improve the global search ability of the optimal value, which effectively improves the estimation preciseness of the state of charge of lithium battery. Based on the experimental data from the charge and discharge test, the proposed ACGA-RNN combined algorithm, RNN algorithm, and GA-RNN algorithm are applied for SOC estimation of the lithium-ion battery respectively. The experimental results show that the estimation mean absolute error of ACGA-RNN combined algorithm obtains the best accuracy of SOC estimation compared with the traditional RNN and GA-RNN algorithms. The mean absolute error under the DST condition is 1.74%, which is lower than the estimation accuracy of traditional RNN (3.68%) and GA-RNN (2.49%). In addition, at 45 ℃ and 0 ℃, the mean absolute error of the ACGA-RNN combined algorithm is 1.75% and 2.05%, respectively, which meets the requirements of national standards. Therefore, the proposed ACGA-RNN combined algorithm has a good application value in SOC estimation of lithium batteries.

Key words: Lithium battery, state of charge, recurrent neural network, adaptive cataclysm genetic algorithm