Enhancing Lithium-Ion Battery Temperature Prediction with a CNN-Bi-LSTM-AM Model

The Lithium-Ion Battery Temperature Field Prediction Model Based on CNN-Bi-LSTM-AM

Abstract: This study focuses on the internal temperature field of lithium-ion batteries, aimed at addressing the temperature variations caused by complex operating conditions in new energy batteries. To manage unpredictable temperature fluctuations and long delay times, we propose an enhanced Convolutional Bidirectional Long Short-Term Memory Neural Network (CNN-Bi-LSTM-AM) model for temperature field prediction. The model integrates CNN for spatial feature extraction, Bi-LSTM for capturing temporal characteristics, and an attention mechanism to enhance the identification of key time-series features. By simulating temperature variations through a lumped model and thermal runaway model, we generate temperature field data that are utilized by the deep learning model to effectively capture the complex nonlinear relationships between temperature, voltage, state of charge (SOC), insulation resistance, current, and the internal temperature field. Performance evaluation using accuracy metrics and validation under various environmental conditions demonstrates that the model improves prediction accuracy by 1.2–2.3% compared to traditional methods (e.g., ARIMA, LSTM) with only a slight increase in testing time. Comprehensive evaluations, including ablation studies, thermal runaway tests, and computational efficiency analysis, further validate the robustness and applicability of the model. Furthermore, this study contributes to the optimization of battery life and safety by enhancing the prediction accuracy of the internal temperature field, thereby reducing resource waste caused by battery performance degradation. The findings provide an innovative approach to advancing new energy battery technology, promoting its development toward greater safety, stability, and environmental sustainability, which aligns with global sustainable development goals.

Keywords: lithium-ion battery; Convolutional Neural Network (CNN); Bidirectional LSTM (Bi-LSTM); temperature field prediction model; environmental impact

1. Introduction

In recent years, the rapid growth of the new energy industry has made energy-efficient and low-power electric vehicles (EVs) a focal point for many countries, with lithium-ion batteries emerging as a core technology for these vehicles. However, lithium-ion batteries face several challenges during charge and discharge cycles, including an increasing number of cycles, significant fluctuations in charge and discharge currents, and reduced charge and discharge durations. Additionally, overcharging can lead to charge overload, resulting in thermal runaway and increased temperature fluctuations, which pose a significant threat to battery safety. Therefore, an efficient thermal management strategy has become a critical issue.

Under complex operating conditions, parameters such as electrochemical reactions, capacity, and internal resistance undergo significant changes, leading to pronounced nonlinear polarization phenomena. This increases internal resistance and heat generation. As heat accumulates, internal temperature fluctuations within the battery intensify. If these temperature variations are not accurately predicted and controlled, they can lead to the failure of the thermal management system, endangering the battery’s operational safety. Thus, accurately predicting and controlling the temperature fluctuations of lithium-ion batteries is crucial for enhancing battery performance, ensuring the safety of electric vehicles, and extending battery lifespan.

Currently, models for studying battery thermal behavior and predicting battery temperature are primarily categorized into two types: one based on internal chemical reactions, exploring the relationship between heat generation and heat transfer, and the other using neural networks for supervised learning to achieve effective temperature predictions. Chemical reaction-based methods include electrochemical–thermal coupling models, electrical–thermal coupling models, and thermal abuse models. These methods often respond slowly under complex operating conditions, making real-time monitoring of battery abnormalities challenging.

In contrast, data-driven models can achieve higher computational efficiency by optimizing the environment and duration for training and prediction. Using historical or real-time data, self-learning models can capture latent correlations to predict the temperature on the surface of lithium-ion batteries, thus avoiding the need for complex electrochemical process modeling. Research has shown that recurrent neural networks (RNNs) perform well in temperature field prediction for various types of batteries. The cumulative research indicates that combining convolutional networks with LSTM networks significantly enhances the accuracy of temperature prediction.

2. Datasets

In practical experiments, multiple temperature sensors are installed on lithium-ion battery cells, and thermal runaway conditions are induced through cyclic charging to obtain accurate thermal runaway temperature field data. To reduce costs and improve efficiency, this study introduces the use of an electrochemical reaction model to simulate real-world data, generating temperature field data for training purposes.

2.1. Lumped Model

A lumped model is employed to simulate the normal temperature rise in temperature field data of lithium-ion batteries. This model simulates three primary effects during the battery’s operation: Ohmic effect, polarization effect, and concentration gradient effect, enabling real-time reflection of the battery’s state.

2.2. Thermal Runaway Model

Battery thermal runaway refers to uncontrollable heat generation due to a series of chain chemical reactions. The mathematical model is formulated using the Arrhenius equations for different reactions, allowing for effective simulation of temperature behaviors during thermal runaway.

3. Methods

The proposed CNN-Bi-LSTM-AM model effectively captures both local spatial features and temporal dependencies from the input time series. The network begins with an input layer that processes a sequence of battery features, including current, voltage, ambient temperature, SOC, and battery load. Subsequently, CNN extracts spatial characteristics from the raw data, while Bi-LSTM models temporal relationships by capturing both past and future dependencies at each time step.

4. Results and Analysis

A comprehensive evaluation of the proposed CNN-Bi-LSTM-AM model was carried out, comparing its performance with several other deep learning models. The results demonstrate that the CNN-Bi-LSTM-AM network outperforms the traditional models, achieving significant improvements in prediction accuracy.

5. Discussion

This study proposes a deep learning model based on the self-attention mechanism—CNN-Bi-LSTM-AM—for accurately predicting the temperature field of lithium-ion batteries. The method leverages the CNN framework to extract contextual sequence features, utilizes the Bi-LSTM framework to capture temporal dependencies, and employs a self-attention layer to enhance the similarity between input vectors.

The main contributions of this study are evident through the demonstrated effectiveness of the CNN-Bi-LSTM-AM model, which achieves satisfactory accuracy under various conditions, improves feature extraction capabilities, and shows excellent generalization capabilities. Future research will focus on developing deep learning frameworks based on attention mechanisms and optimizing hyperparameter selection methods to further enhance the performance of lithium-ion battery temperature field predictions.