Abstract
The novel method for estimating the State of Charge (SOC) of individual batteries within a LiFePO4 battery cluster, leveraging memory feature fusion. By integrating cluster-level and individual-level features, along with a trainable memory mechanism, the proposed method achieves fine-grained SOC estimation. Experimental validation on real-world data demonstrates a significant reduction in estimation error compared to methods that ignore cluster effects.
Keywords: LiFePO4 battery; SOC estimation; memory feature fusion; cluster effects; battery management system (BMS)
1. Introduction
Energy storage systems, particularly those utilizing LiFePO4 batteries, play a crucial role in modern power grids and renewable energy integration. Accurate estimation of the SOC of each battery within a cluster is essential for efficient energy management and prolonging battery life. However, accurate SOC estimation faces challenges due to variations among individual batteries and the complex interactions within the cluster.
In this paper, we propose a memory feature fusion method for SOC estimation in LiFePO4 battery clusters. Our method integrates cluster-level and individual-level features through a class attention mechanism and utilizes a trainable memory module to capture typical battery behavior patterns. The contributions of this paper are as follows:
- We consider the overall operation effect of the energy storage cluster, extending the model input with both individual and cluster features.
- We introduce a trainable memory mechanism to memorize key individual features, improving the estimation accuracy and fine-grained representation of each battery in the same cluster.
2. Methodology
2.1 Overview
The proposed method consists of five main components: an individual shared encoder, a cluster encoder, a memory module, a feature fusion layer, and an SOC estimator. The overall network architecture.
2.2 Feature Engineering
The input features are divided into individual and cluster perspectives. Individual features include voltage, temperature, and shared current data for each battery in the cluster. Cluster features are constructed using expert knowledge, such as mean, variance, skewness, kurtosis, and peak-to-peak values, based on battery group operation measurements.
Table 1: Cluster Hand-crafted Features Table
| Constructed Feature | Calculation Formula |
|---|---|
| Mean | N1∑i=1Nfi |
| Variance | N1∑i=1N(fi−mean)2 |
| Skewness | N1∑i=1N(stdfi−mean)3 |
| Kurtosis | N1∑i=1N(stdfi−mean)4 |
| Peak-to-Peak | max(fi)−min(fi) |
2.3 Memory Module
To address the inconsistencies among battery cells, such as differences in internal resistance, initial capacity, and maximum capacity, a memory module is employed. The memory module consists of M memory items, representing typical battery cell patterns, initialized randomly using a Gaussian distribution before training.
The encoded individual features are input into the memory module as queries, and the memory items are updated based on the read and update operations. To ensure the diversity and typicality of memory features, compactness and separation losses are introduced.
2.4 Feature Fusion
After obtaining cluster features (g_z), individual features (c_z), and memory features (m_z), a feature fusion operation is performed using a cluster attention mechanism, inspired by self-attention.
2.5 Model Training
The fused features (Z) are input into the SOC estimator to obtain the estimated SOC values. The model is trained end-to-end using a combination of Mean Squared Error (MSE) loss, compactness loss, and separation loss.
3. Experimental Results
3.1 Experimental Dataset and Evaluation Criteria
A real-world LiFePO4 battery dataset was used for experimental validation. The dataset consists of charging data from a battery cluster consisting of 224 series-connected battery cells, each with voltage, temperature, and shared current measurement points. Five rounds of charging data were used for training, and one round was used for testing.
3.2 Comparison with Baseline Methods
To evaluate the effectiveness of the proposed method, it was compared with Support Vector Regression (SVR), Multilayer Perceptron (MLP), and Long Short-Term Memory (LSTM) networks. The results are shown in Table 2.
Table 2: Comparison of CMCF Method with Other Methods
| Method | MSE (×10^-4) | MAPE (%) | Max Error (%) | FAST (×10^-4) |
|---|---|---|---|---|
| SVR | 5.94 | 3.13 | 2.96 | 1.38 |
| MLP | 2.99 | 7.81 | 9.41 | 3.88 |
| LSTM | 2.36 | 2.16 | 4.86 | 2.97 |
| CMCF | 1.52 | 1.72 | 3.70 | 1.87 |
The proposed method (CMCF) outperforms the baseline methods in all evaluation metrics, with at least a 71.37% reduction in MSE compared to the worst-performing method (SVR).
3.3 Ablation Study
To further validate the effectiveness of each component of the proposed method, an ablation study was conducted. The results are shown in Table 3.
Table 3: Ablation Study Results
| Method | MSE (×10^-4) | MAPE (%) | Max Error (%) | FAST (×10^-4) |
|---|---|---|---|---|
| w/o CF | 2.99 | 7.81 | 9.41 | 3.88 |
| w/o Memory | 2.36 | 2.16 | 4.86 | 2.97 |
| w/o Cluster Att | 3.36 | 2.62 | 3.88 | 3.82 |
| w/o CF & Memory | 5.94 | 3.13 | 2.96 | 1.38 |
| CMCF | 1.52 | 1.72 | 3.70 | 1.87 |
The results demonstrate that each component of the proposed method contributes to improving the SOC estimation accuracy. In particular, the cluster features play a crucial role, as removing them significantly degrades performance.
4. Conclusion
The memory feature fusion method for SOC estimation in LiFePO4 battery clusters. By integrating cluster-level and individual-level features and utilizing a trainable memory mechanism, the proposed method achieves fine-grained SOC estimation. Experimental results on a real-world dataset demonstrate the effectiveness of the proposed method, with significant improvements in estimation accuracy compared to baseline methods. The ablation study further validates the contribution of each component of the proposed method.
The proposed method provides a new perspective for SOC estimation in energy storage battery clusters, with potential applications in power grid management, renewable energy integration, and battery life prolongation.