Uniqueness of multilayer perceptron-based capacity prediction for contributing state-of-charge estimation in a lithium primary battery

Abstract

Lithium primary batteries having high power and energy densities have irreversible energy characteristics, unlike lithium secondary batteries. Therefore, their application in experiments is limited, and capacity information is difficult to derive. The capacity information serves as the initial value in the ampere-hour counting method to estimate the battery state-of-charge (SOC). Therefore, in order to estimate the SOC of a lithium primary battery, it is necessary to predict the capacity information. In this paper, we propose a method to predict the capacity of lithium primary batteries using machine learning techniques. To predict capacity, pre-measurement factors (PMFs) for lithium primary batteries are selected and correlation with capacity is analyzed. PMFs that can be measured before the lithium primary battery operates include AC impedance and open-circuit voltage, weight, and resistance, Resistances were extracted using the closed-circuit voltage experiment. These PMFs derive a correlation coefficient through correlation analysis of the capacity of lithium primary battery. However, the correlation coefficient between individual PMFs and the capacity of lithium primary battery is lower than 0.5. That is, the relationship between PMFs and capacity is not linear. Therefore, this study implements a multilayer perceptron (MLP) model that uses multiple nonlinear parameters. A total of 300 lithium primary battery data is used for model design, and the model is verified by classifying the data into training data and test data. The capacity information predicted through the MLP model was used for the SOC estimation, and the SOC estimation accuracy was analyzed by comparing it with the SOC measurement value.