In the pursuit of accurate and engineering-adaptive state of health (SOH) estimation for energy storage cells, this study addresses the challenges posed by the performance degradation and consistency variations in lithium-ion batteries used in large-scale energy storage systems. The increasing deployment of energy storage cells in renewable energy integration, such as photovoltaic power stations, necessitates reliable SOH assessment to enhance system efficiency and mitigate safety risks. Traditional methods, including model-based approaches like electrochemical and equivalent circuit models, often suffer from complexity, poor robustness, and parameter identification errors. Data-driven techniques, while promising, typically rely on laboratory data without considering actual operational conditions, leading to reduced practicality in real-world applications. This research leverages fragmented data from energy storage cell operations to develop a robust SOH estimation framework, combining genetic algorithm (GA)-optimized back propagation (BP) neural networks and model transfer learning for improved accuracy and adaptability.
The foundation of this work lies in analyzing operational data from a 50 MW/100 MWh energy storage power station, which utilizes 260 Ah lithium iron phosphate energy storage cells. These energy storage cells are configured in series and parallel arrangements to form battery clusters integrated into the grid. Operational data, recorded at 1-minute intervals, revealed that charging often terminates when the state of charge (SOC) reaches 100% or individual cell voltage hits 3.65 V, resulting in incomplete charging for some energy storage cells. Specifically, analysis of charging cessation data showed that cell voltages typically range between 3.41 V and 3.45 V at full charge, indicating that charging data below 3.41 V are consistently available. Moreover, a stable charging phase with a current of approximately 0.5 C exists for over 30 minutes before voltage cutoff, providing a reliable data segment for SOH estimation. This insight guided the selection of charging voltage differences within 30 minutes prior to reaching 3.41 V as key feature parameters for evaluating the health of energy storage cells.
To simulate these operational conditions, laboratory experiments were conducted using 20 Ah lithium iron phosphate energy storage cells with LiFePO4 cathodes and graphite anodes. The cells underwent cyclic testing in a controlled environment at 25°C ± 2°C, with voltage limits of 2.5 V to 3.65 V and a SOC range of 20% to 100%. A constant current of 0.5 C was applied during charging, and discharge was controlled by releasing 80% of the rated capacity to mimic real-world usage. Capacity calibration was performed every 100 cycles to track degradation, with SOH defined as: $$ SOH = \frac{Q_C}{Q_{new}} \times 100\% $$ where \( Q_C \) is the current maximum available capacity and \( Q_{new} \) is the rated capacity. For the 20 Ah energy storage cells, the measured capacity in the 20% to 100% SOC range was converted to an equivalent full capacity using: $$ Q_d = 0.8 \times Q_s $$ where \( Q_d \) is the equivalent capacity for SOC 0% to 100% cycling and \( Q_s \) is the measured capacity in the 20% to 100% SOC range. This conversion ensured accurate SOH labeling for model training.
The degradation mechanism of lithium iron phosphate energy storage cells was investigated through differential voltage analysis (dV/dQ), which revealed two characteristic peaks (P1 and P2) associated with active lithium loss and anode active material loss. As SOH decreases, the P2 peak shifts leftward, and the distance between P1 and P2 reduces, indicating accelerated degradation. This behavior is most prominent in the charging voltage range below 3.41 V, particularly in the 30-minute window before reaching this voltage. Consequently, the voltage differences during this interval were chosen as feature parameters for SOH estimation. Three time intervals were evaluated: 1-minute, 3-minute, and 5-minute voltage differences, denoted as \( \Delta V_1, \Delta V_2, \ldots, \Delta V_i \), where \( i \) depends on the interval (e.g., 29 for 1-minute, 10 for 3-minute, and 6 for 5-minute intervals). The feature extraction process involved normalizing the data to a [0,1] range using: $$ X’ = \frac{X – X_{\text{min}}}{X_{\text{max}} – X_{\text{min}}} $$ where \( X \) is the original voltage difference value, \( X_{\text{min}} \) is the minimum value, and \( X_{\text{max}} \) is the maximum value. This preprocessing step ensured that all inputs were scaled appropriately for neural network training.
The dataset comprised cyclic data from three 20 Ah energy storage cells (LFP20-1#, LFP20-2#, LFP20-3#), with initial capacities around 19.3 Ah. After 4,100 cycles, capacities degraded to approximately 17.7 Ah, corresponding to SOH values between 91.45% and 92.26%. The feature parameters were organized into tables for each time interval, as shown below:
| Sample | ΔV1 (mV) | ΔV2 (mV) | … | ΔV29 (mV) |
|---|---|---|---|---|
| 1 | 0.2 | 0.2 | … | 0.9 |
| 2 | 0.8 | 0.1 | … | 1.2 |
| … | … | … | … | … |
| 12,300 | 7.2 | 5.2 | … | 1.5 |
| Sample | ΔV1 (mV) | ΔV2 (mV) | … | ΔV10 (mV) |
|---|---|---|---|---|
| 1 | 5.9 | 4.7 | … | 0.8 |
| 2 | 8.1 | 5.9 | … | 0.7 |
| … | … | … | … | … |
| 12,300 | 16.1 | 7.5 | … | 4.6 |
| Sample | ΔV1 (mV) | ΔV2 (mV) | ΔV3 (mV) | ΔV4 (mV) | ΔV5 (mV) | ΔV6 (mV) |
|---|---|---|---|---|---|---|
| 1 | 9.0 | 6.9 | 5.2 | 7.2 | 9.9 | 6.5 |
| 2 | 11.2 | 7.4 | 5.3 | 4.0 | 7.5 | 9.3 |
| … | … | … | … | … | … | … |
| 12,300 | 21.7 | 7.8 | 4.0 | 0.5 | 4.3 | 1.8 |
The GA-BP neural network was employed to model the nonlinear relationship between these voltage differences and SOH. The BP neural network architecture consisted of an input layer with nodes corresponding to the number of voltage differences (e.g., 29 for 1-minute intervals), three hidden layers with 10 nodes each, and an output layer with a single node representing SOH. The initial weights and thresholds were optimized using a genetic algorithm to prevent local minima and enhance global search capability. The GA process involved encoding the weights and thresholds as chromosomes, followed by selection, crossover, and mutation operations to evolve the population towards optimal values. The fitness function minimized the mean squared error between predicted and actual SOH. The combined GA-BP approach accelerated convergence and improved prediction accuracy, as described by the flowchart where input features \( X_n = [\Delta V_1, \Delta V_2, \ldots, \Delta V_i] \) are processed to output SOH, with errors backpropagated for weight updates.
For model training, the dataset was split into 80% for training and 20% for testing. Additionally, 1,200 cycles of data from separate energy storage cells were used for validation. The model’s performance was evaluated using mean absolute percentage error (MAPE) and root mean square error (RMSE): $$ E_{MAP} = \frac{1}{N} \sum_{n=1}^{N} \left| \frac{Y(n) – S(n)}{S(n)} \right| \times 100\% $$ $$ E_{RMS} = \sqrt{ \frac{1}{N} \sum_{n=1}^{N} (Y(n) – S(n))^2 } $$ where \( Y(n) \) is the predicted SOH, \( S(n) \) is the actual SOH, and \( N \) is the number of samples. The results demonstrated that the model with 1-minute interval voltage differences achieved the highest accuracy, with MAPE of 0.37% and RMSE of 0.4565, outperforming the 3-minute and 5-minute interval models. This superiority is attributed to the finer temporal resolution capturing more detailed degradation patterns in energy storage cells.
| Model | Feature Parameter | MAPE (%) | RMSE |
|---|---|---|---|
| M-1 | 1-minute voltage difference | 0.37 | 0.4565 |
| M-2 | 3-minute voltage difference | 0.40 | 0.5095 |
| M-3 | 5-minute voltage difference | 0.44 | 0.5224 |
To enhance the model’s generalizability to larger energy storage cells, such as the 260 Ah cells used in the power station, model transfer was implemented. Initially, the M-1 model trained on 20 Ah cells showed poor performance when applied directly to 260 Ah cells, with a maximum error of 5.52%. This was addressed by leveraging transfer learning: the first two hidden layers of the pre-trained M-1 model were frozen, and the last hidden and output layers were fine-tuned using 60 additional datasets from 260 Ah energy storage cells. This approach reduced the maximum error to 1.89%, demonstrating the effectiveness of transfer learning in adapting models across different capacities of energy storage cells without requiring extensive retraining data.
For engineering validation, the transferred model was applied to estimate SOH for a cluster of 224 energy storage cells in the photovoltaic power station. Using 1-minute charging voltage differences from operational data, the model batch-estimated SOH values clustered around 95%, confirming its practicality and robustness in real-world scenarios. This application underscores the model’s ability to handle fragmented data from energy storage cells under varying operational conditions, providing a scalable solution for large-scale energy storage systems.
In conclusion, this study presents a fragmented data-based SOH estimation method for energy storage cells that combines operational data analysis with advanced machine learning. The use of charging voltage differences below 3.41 V, optimized through GA-BP neural networks, achieves high accuracy, while model transfer extends applicability to diverse energy storage cell types. The framework offers significant improvements in engineering adaptability, enabling reliable health monitoring of energy storage cells in renewable energy installations. Future work could explore real-time implementation and integration with battery management systems for dynamic SOH tracking.