Scholarlink

One of the most important features of any battery management system is the estimation of State of Health (SOH) of a battery. Predicting battery health based on their charging-discharging cycles and managing them accordingly is very important. Our work presents an innovative hybrid model that combines a 1D convolutional neural network (CNN), bidirectional long short-term memory (BiLSTM), and an attention mechanism to estimate the state of health (SOH) from a time series dataset. To get around the lack of labelled real-world datasets, we make a realistic synthetic dataset that mimics how batteries behave in different temperatures, currents, and cycling conditions. The suggested architecture uses CNN layers to identify local features, BiLSTM to determine long-term dependencies, and attention to make it easier to understand and give more weight to important time-sensitive events. The model is very accurate, with an RMSE of 2.17% for estimating SOH on the test set and R2 scores above 0.96. The model is small and can be used in cloud-based battery management systems (BMSs) as well as in on board systems. This work provides a scalable and comprehensible framework for intelligent battery diagnostics, connecting data-driven modelling with real-time applications.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Predicting the Battery State of Health Using Deep Neural Networks: A Hybrid LSTM-Attention Model with Data Validation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Liton Balo

摘要

One of the most important features of any battery management system is the estimation of State of Health (SOH) of a battery. Predicting battery health based on their charging-discharging cycles and managing them accordingly is very important. Our work presents an innovative hybrid model that combines a 1D convolutional neural network (CNN), bidirectional long short-term memory (BiLSTM), and an attention mechanism to estimate the state of health (SOH) from a time series dataset. To get around the lack of labelled real-world datasets, we make a realistic synthetic dataset that mimics how batteries behave in different temperatures, currents, and cycling conditions. The suggested architecture uses CNN layers to identify local features, BiLSTM to determine long-term dependencies, and attention to make it easier to understand and give more weight to important time-sensitive events. The model is very accurate, with an RMSE of 2.17% for estimating SOH on the test set and R2 scores above 0.96. The model is small and can be used in cloud-based battery management systems (BMSs) as well as in on board systems. This work provides a scalable and comprehensible framework for intelligent battery diagnostics, connecting data-driven modelling with real-time applications.