Session: 09-17-01: AI for Energy

Paper Number: 164484

Deep Learning-Based Prediction of Thermal Runaway Temperature of Lithium Iron Phosphate Batteries Under Multivariate Influences 

Safety prediction and early warning systems for lithium iron phosphate (LFP) batteries play a critical role in ensuring the reliable operation of energy storage systems. However, the risks associated with thermal runaway, a dangerous phenomenon where the battery temperature rapidly increases, make accurate prediction even more urgent. As artificial intelligence (AI) technologies continue to advance, deep learning methods have recently been used in predicting and preventing thermal runaway events in battery systems. In this study, we conduct six sets of thermal abuse tests on LFP cylindrical batteries to build a temperature dataset composed of the battery cells, heater, and the surrounding environment. These tests are performed under three different heating rates (60W, 80W, and 100W) and three distinct states of charge (SOC): 50%, 75%, and 100%. To analyze the collected time series data (the temperature variations), we apply convolutional neural networks and long short-term memory networks (CNN-LSTM). The goal is to understand how temperature changes occur during thermal runaway and to investigate the impact of heating rate and SOC on these changes. Additionally, we use the rate of temperature rise as an indicator to detect thermal runaway events and to predict key thermal runaway temperature parameters.

The results from the thermal abuse tests show that the initial temperature at which thermal runaway occurs for LFP batteries is primarily within the range of 230°C to 250°C. Notably, a higher SOC is associated with a lower initial temperature for thermal runaway, while a higher heating rate results in a shorter time to thermal runaway. Furthermore, the predictions made by the CNN-LSTM model reveal a strong correlation between SOC and cathode temperature (0.94) and between heating rate and temperature (0.91). The algorithm performs best when only SOC varies, with a coefficient of determination (R²) of 0.93. The predicted thermal runaway time shows a minimal average error of just 3 seconds, while the temperature prediction error is 7.9°C on average.

This study successfully combines thermal abuse experiments with deep learning techniques to explore the effects of heating rate and SOC on the thermal runaway characteristics of LFP batteries. The results provide an efficient method for predicting the actual thermal runaway temperature and offer valuable insights into how AI-driven approaches can be utilized to predict and prevent thermal runaway events. Ultimately, this research establishes a solid foundation for the integration of deep learning methods into thermal runaway prediction systems, supporting the broader application of AI in energy storage systems.

Presenting Author: Yuyao Cao Tsinghua university

Presenting Author Biography: Ph.D. candidate at the School of Safety Science, Tsinghua University. Her research focuses on early gas safety monitoring of lithium battery thermal runaway and the use of artificial intelligence for predicting thermal runaway events.

Authors:

Yuyao Cao Tsinghua university
Ya Peng China University of Mining and Technology-Beijing
Zhenxiang Tao China University of Mining and Technology-Beijing
Hui Zhang Tsinghua University