A BP neural network-based approach for SOC testing and interval estimation of Li-SOCL₂ batteries in smart meter applications

Disposable batteries (non-rechargeable batteries) offer notable advantages, including low self-discharge rate, high energy density, extended storage life, and enhanced safety, rendering them suitable for deployment in critical domains such as medical devices and intelligent instrumentation. As battery capacity declines, the output voltage correspondingly decreases, adversely impacting the operational stability and efficiency of dependent systems. Conventional prediction techniques primarily utilize empirical models based on open-circuit voltage and internal resistance, which lack adaptability under complex and variable operating conditions. Recent advancements in neural network methodologies have introduced novel frameworks for residual capacity prediction. However, prevailing neural network-based approaches often depend on hybrid algorithmic structures and insufficiently consider inter-batch variability among batteries. This study presents a neural network-based method for estimating the SOC of disposable batteries. A feature parameter system encompassing multiple battery batches is developed, incorporating pulse-load voltage and historical discharge data. This system is integrated with a Backpropagation Neural Network to characterize nonlinear degradation patterns and establish a mapping between the extracted features and the residual capacity range. Experimental results indicate that the proposed method achieves a Root Mean Square Error below 1%, representing a 60% improvement in prediction accuracy over conventional techniques. This approach contributes a generalized framework for residual capacity estimation in disposable batteries and provides a technical basis for their lifecycle management.