TL;DR
This paper introduces a novel battery grouping method that leverages the entire charge/discharge time series of individual battery cells. By employing wavelet denoising, Dynamic Time Warping (DTW) for similarity calculation, and Affinity Propagation (AP) for clustering, the method effectively groups batteries with similar dynamic performance, leading to improved battery pack consistency and longevity.
What Problem Does This Paper Address?
Electric vehicles (EVs) rely on battery packs, which are assemblies of individual battery cells. However, inherent differences in cell characteristics can lead to uneven charging/discharging, premature degradation, and reduced overall pack performance and lifespan. Traditional battery grouping methods often rely on single static characteristics (like capacity or internal resistance) or a limited set of dynamic characteristics, which fail to fully capture a battery's complex behavior. This paper addresses the need for a more comprehensive battery grouping strategy that considers the full dynamic performance profile of each cell.
Methods
The proposed method tackles battery grouping as a time series clustering problem. The core steps are:
- Data Acquisition: The study uses an embedded system to simultaneously measure terminal voltages, temperatures, and currents for multiple batteries during their full charge and discharge cycles.
- Wavelet Denoising: To handle noise inherent in the collected data, a wavelet denoising technique (specifically using the "sym8" wavelet with 5 levels of decomposition) is applied to smooth the charge/discharge sequences.
- Dynamic Time Warping (DTW) Distance Calculation: Since battery charge/discharge sequences can vary in length and timing of key events, the traditional Euclidean distance is insufficient. DTW is used to compute the similarity between sequences by allowing non-linear time warping, effectively aligning feature points even if they occur at different times. This results in a distance matrix.
- Similarity Matrix Generation: The DTW distances are normalized to create a similarity matrix, where higher values indicate greater similarity between battery sequences.
- Affinity Propagation (AP) Clustering: The Affinity Propagation algorithm is employed to cluster the batteries based on the generated similarity matrix. AP is chosen because it's an unsupervised method that doesn't require pre-specifying the number of clusters, making it suitable for unknown battery groupings. It works by passing messages between data points to identify exemplars (cluster centers).
Datasets
The paper utilizes experimental data from 94 new LiMn2O4 lithium-ion batteries with a nominal capacity of 2 Ah. The charge and discharge curves were recorded under specific conditions:
- Charging: Constant current of 1.2 A until cutoff voltage, followed by constant voltage charging until current dropped to 0.02 A.
- Discharging: Constant current of 1.2 A.
The authors also mention building two battery packs, each with four cells, for a second experiment to demonstrate the practical implications of their grouping method.
Key Insights and Ideas
- Full Time Series is Crucial: The study strongly suggests that using the entire charge/discharge sequence provides a more accurate representation of battery behavior than relying on single static or limited dynamic characteristics.
- DTW for Temporal Flexibility: Dynamic Time Warping is essential for comparing time series data where temporal alignment is not fixed, which is common in battery performance data.
- AP for Unsupervised Grouping: Affinity Propagation proves effective for unsupervised clustering of battery time series, avoiding the need to pre-determine the number of groups.
- Wavelet Denoising Enhances Accuracy: Pre-processing the raw data with wavelet denoising significantly improves the reliability of the subsequent similarity calculations and clustering.
Key Results
- Superior Performance: The proposed AP-based battery grouping method consistently outperformed K-means based spectral clustering.
- Charge Curves: For charge curves, the AP algorithm achieved a best silhouette index of 0.9583 with 6 clusters, compared to 0.8452 for spectral clustering. The AP similarity matrix was also observed to be more diagonalized, indicating better separation.
- Discharge Curves: For discharge curves, the AP algorithm achieved a best silhouette index of 0.9523 with 5 clusters, outperforming spectral clustering's best of 0.9213 with 2 clusters.
- Practical Validation: In a real-world experiment with two battery packs, cells grouped using the proposed method (based on similar charge curves) showed consistent performance over many cycles. In contrast, cells grouped using traditional methods (based on similar terminal voltages) exhibited inconsistent behavior within the pack.
Limitations and Future Work
- Charge vs. Discharge: The paper notes that different clusters are obtained when using charge series versus discharge series for the same batteries. This suggests that combining both charge and discharge information might be beneficial.
- Constant Rate Assumption: The current study used constant charge/discharge rates (0.6C). Future research could investigate the impact of different charge/discharge rates on the performance of grouped battery packs.
- Extension to Combined Series: The authors suggest that extending the clustering to combine both charge and discharge time series is a straightforward next step.
Summary updated