TL;DR
This paper introduces a novel BiLSTM-Transformer deep learning framework for estimating the State-of-Health (SOH) of lithium-ion batteries in electric vehicle swapping stations. By leveraging only charging data and a new "Constant Voltage Time" (CVT) feature, the model achieves highly accurate capacity degradation predictions, even under challenging cold-temperature conditions, without needing discharge data.
What Problem Does This Paper Address?
The rapid growth of electric vehicles (EVs) and the adoption of battery swapping stations present a unique challenge for battery health monitoring. Unlike traditional EVs where batteries are charged and discharged together, swapping stations separate these processes. Furthermore, batteries are constantly circulating between different vehicles and stations, leading to fragmented operational data. This fragmentation makes it difficult to accurately assess the State-of-Health (SOH), particularly the capacity degradation, which is crucial for ensuring battery safety, reliability, and efficient station operation. Cold temperatures exacerbate battery aging, further complicating SOH estimation. Existing methods often rely on both charging and discharging data, which isn't always available in a swapping scenario, or use differential features that can amplify sensor noise.
Methods
The core of this work is a novel deep learning architecture: a BiLSTM-Transformer framework.
- BiLSTM (Bidirectional Long Short-Term Memory): This component is used as an initial encoder. It excels at processing time-series data by considering information from both past and future time steps, allowing it to capture long-term dependencies and temporal patterns in battery aging.
- Transformer: Following the BiLSTM, a Transformer encoder is employed. Its key innovation is the multi-head attention mechanism, which allows the model to weigh the importance of different parts of the input sequence. This helps in identifying and extracting the most relevant aging features from the battery data.
- Constant Voltage Time (CVT) Feature: A novel feature is introduced, extracted solely from the constant voltage (CV) phase of the battery charging process. This feature, the duration of the CV phase, is chosen because it's a stable indicator of battery aging and can be reliably measured during charging, which is a primary operation in swapping stations. This avoids the need for discharge data and the complexities associated with it.
- Incremental Capacity (IC) Analysis Features: The study also incorporates features derived from Incremental Capacity Analysis (ICA), specifically the "IC Peak Voltage," which is a standard technique for battery health monitoring.
The framework combines these elements to process the extracted features and predict the battery's capacity degradation trajectory.
Datasets
The study utilized data from 12 identical lithium-ion batteries (LIBs) subjected to cyclic aging experiments. These experiments were conducted under varying conditions, including:
- Temperatures: Room temperature (25 °C) and cold temperature (0 °C).
- Charging Rates: Ranging from 3C to 5C.
The data collected included capacity decay trajectories, charging current, and temperature curves. The model was trained using the first 40% of the battery lifecycle data, with 20% used for validation to prevent overfitting. The remaining 40% of the data was used for estimation. The paper does not mention the use of any publicly available benchmark datasets.
Key Insights and Ideas
- Charging Data is Sufficient: The most significant insight is that accurate SOH estimation, particularly capacity degradation, can be achieved using only charging data. This is a major breakthrough for battery swapping stations where discharge data is often unavailable or unreliable.
- CVT Feature's Power: The newly proposed "Constant Voltage Time" (CVT) feature, derived from the charging process, demonstrates a strong correlation with battery capacity degradation. This feature is robust and avoids the noise amplification issues associated with differential methods like ICA.
- Synergy of BiLSTM and Transformer: Combining BiLSTM for temporal feature extraction and Transformer for attention-based feature selection creates a powerful framework. The BiLSTM captures the sequential nature of aging, while the Transformer focuses on the most informative aspects of that sequence.
- Cold Temperature Resilience: The model shows good performance even under low-temperature conditions (0 °C), which are known to accelerate battery aging and complicate SOH estimation. This highlights the framework's generalization capability.
- Non-linearity Handling: The framework effectively handles non-linear capacity decay trajectories, including "knee" phenomena, which are challenging for simpler models.
Key Results
- High Accuracy: The proposed BiLSTM-Transformer framework achieved a normalized Root Mean Squared Error (RMSE) of less than 1.6% across various operating conditions, including cold temperatures.
- Exceptional Performance: In ideal conditions, the normalized RMSE reached an impressive 0.11%.
- Competitive Performance: Compared to other models like SVR and standalone LSTM or Transformer networks, the BiLSTM-Transformer framework consistently outperformed them in both linear and non-linear capacity decay scenarios. For instance, for cell #5 (linear decay), the RMSE was 0.0053 Ah, and for cell #2 (non-linear decay), it was 0.0206 Ah.
- Feature Correlation: The CVT feature showed a high correlation with battery capacity (0.98), comparable to the IC Peak Voltage (0.98), while IC Peak Value had a slightly lower correlation (0.69).
Limitations and Future Work
- Peak Swapping Period Charging: During peak battery swapping periods, the short maintenance time might not allow for full charging. The authors suggest future work could explore using the duration of current change within the CV phase as a feature to address these limited charging scenarios.
- Transformer Complexity and Overfitting: The Transformer framework's complexity can lead to overfitting, especially with limited training data. Future research could involve using multi-feature approaches, incorporating physical information about electrochemical principles, and developing lightweight architectures for embedded systems.
- Dynamic Environments and Solid-State Batteries: The paper suggests investigating the model's adaptability to dynamic, real-time operating environments and its generalization to solid-state batteries as future research directions.
Summary updated