A Method for Interval Prediction of Satellite Battery State of Health Based on Sample Entropy

Satellites need batteries to provide energy when operating in shadow regions, and lithium-ion batteries have become the batteries of choice for most satellites due to their high energy density, low self-discharge rate, and long cycle life. When a satellite battery is working in outer space, its capacity will gradually decrease as the number of cycles increases, and a certain degree of capacity recovery will occur. Due to the excellent mapping relationship between the discharge cutoff voltage and the capacity degradation of lithium-ion batteries and the fact that the sample entropy (SampEn) can sensitively capture local fluctuations, such as the recovery effect during lithium-ion battery capacity degradation, a method for interval prediction of the satellite battery state of health (SOH) based on SampEn was proposed. This method adopts a neural network model based on lower upper bound estimation (LUBE). The method uses the discharge cutoff voltage and the discharge voltage SampEn as the inputs and the battery SOH as the output for the neural network model. To improve the prediction interval coverage and reduce the prediction interval width, especially considering that the lower bound of the interval prediction often determines whether the satellite battery output power reaches the warning threshold, a modified comprehensive indicator function, the coverage width-based criterion (CWC), was constructed. Additionally, based on the nondifferentiability of this indicator function, a simulated annealing algorithm was used to optimize the neural network; at the same time, the optimal values of the interval coverage and interval width were taken into account, resulting in the lower bound of the prediction interval being closer to the actual value. Finally, test data from a NASA #18 battery were used to validate, analyze and verify the interval prediction algorithm proposed in this paper. The results were compared with those obtained from a support vector machine (SVM)-based interval prediction method.