材料科学与工程学报2025，Vol.43Issue(6)：946-954,9.DOI:10.14136/j.cnki.issn1673-2812.2025.06.010
基于多模态神经网络的锂电池安全状态预测
Safety State Prediction of Lithium-ion Batteries Based on Multimodal Neural Network
摘要
Abstract
Commercial lithium-ion battery energy storage systems commonly suffer from issues such as inadequate monitoring methods,inaccurate state assessment,and delayed risk warning,posing severe safety challenges.To address these issues,the research proposes a State of Safety(SOS)prediction method for lithium-ion batteries by integrating multimodal feature data and Multi-Layer Perceptron(MLP).Through collaborative monitoring by sensors and the Battery Management System(BMS),ten types of state physical quantities during the operation of lithium-ion batteries were acquired as input parameters for the neural network.Through the research on the learning mechanism of the neural network algorithm,a correlation coefficient between the input physical quantities and the learning results was introduced to increase the learning rate of highly relevant physical quantities,optimize the model's learning ability,and significantly improve the prediction accuracy.Conventional charge-discharge condition data at different rates,as well as extreme condition test data involving various overcharge rates,ambient temperatures,and heat source powers,were employed as the training set.Data from other working conditions outside the training set served as the test set.The SOS values predicted by the MLP were compared with those calculated based on the measured state parameters.The test results show that the SOS estimation errors under different working conditions are all less than 6%,verifying the model's high accuracy and universality.The thermal runaway experiment validates that the time when SOS reaches the warning threshold is 11 minutes earlier than the thermal runaway trigger time,demonstrating that this safety state prediction method enables ultra-early and precise risk warning.This study provides a multimodal fusion-based dynamic warning model for precise safety states assessment and ultra-early risk warning in lithium-ion battery energy storage systems,offering critical technical support for constructing a proactive defense-oriented energy storage safety system.关键词
锂离子电池/多模态数据/多层感知神经网络/学习率优化/安全状态预测Key words
Lithium-ion battery/Multimodal monitoring data/Multi-layer perceptron/Learning rate optimizing/Safety state prediction分类
通用工业技术引用本文复制引用
刘炜,孟凡敏,朱尤祥,王羽平,车俊辉,付亮,王锐,郭鲁斌,杨东辉,屠芳芳,姜媛媛,蔡佳文,相佳媛,严莉..基于多模态神经网络的锂电池安全状态预测[J].材料科学与工程学报,2025,43(6):946-954,9.基金项目
国家电网有限公司总部管理科技项目(520627230016) （520627230016）
