A primary battery SOC prediction method based on LSTM-BP neural network

Disposable (non-rechargeable) batteries are extensively employed in critical applications such as medical devices and smart metering due to their low self-discharge rates, high energy density, long storage duration, and improved safety characteristics. Nevertheless, their output voltage progressively decreases as capacity diminishes, which adversely influences the operational stability and efficiency of associated equipment. Accordingly, precise determination of the state of charge (SOC) in disposable batteries is crucial. Traditional prediction techniques, which primarily depend on empirical models such as open-circuit voltage and internal resistance measurements, are often insufficient under complex operating environments. Although recent developments in neural networks have enabled new strategies for residual capacity prediction, existing methods predominantly rely on multi-algorithm fusion frameworks. This study proposes a neural network-based methodology for estimating the residual capacity of disposable batteries. Discharge experiments were conducted to obtain pulse load voltage data, which facilitated the construction of a comprehensive characteristic parameter system. Historical discharge data were also employed to determine the residual capacity. By integrating these datasets into an error back-propagation neural network, a correlation model was developed to characterize non-linear degradation behaviors and establish a mapping relationship between characteristic parameters and residual capacity. Experimental validation demonstrates that the proposed method achieves a RMSE of less than 5%. This approach presents a generalized solution for SOC estimation in disposable batteries and provides a foundational model for their life-cycle management.