In the pursuit of global carbon neutrality goals, energy storage systems have emerged as pivotal components within modern power grids. Among these, lithium-ion energy storage batteries stand out due to their high energy density and efficiency. However, the gradual degradation of these energy storage batteries over time poses significant challenges for system reliability and safety. Accurately estimating the State of Health (SOH) of energy storage batteries is therefore critical for optimizing their operational lifespan and mitigating risks. Traditional SOH assessment methods often rely on complete charge-discharge cycles or complex models that are difficult to apply in real-world, grid-connected scenarios where operational data is fragmented and incomplete. This study addresses this gap by proposing a novel data-driven approach for SOH estimation of energy storage batteries using only fragmented charging data segments, enhancing both accuracy and engineering adaptability.
We begin by analyzing actual operational data from a grid-connected photovoltaic energy storage power station. The energy storage batteries in such stations typically undergo partial state-of-charge (SOC) cycles, and full capacity tests are rarely performed, making direct SOH monitoring impractical. Our analysis revealed that during charging operations, a stable current phase exists below 3.41 V, lasting over 30 minutes, which is consistently present across different days. This observation, combined with an understanding of lithium iron phosphate (LFP) energy storage battery degradation mechanisms, guided our feature selection. The degradation in LFP energy storage batteries is primarily attributed to loss of active lithium and anode active material, which manifests as shifts in characteristic peaks in the incremental capacity (dV/dQ) curves during charging. These shifts are most pronounced in the voltage region below 3.41 V. Consequently, we hypothesize that the voltage evolution within this specific charging fragment holds strong correlation with the overall health of the energy storage battery.
To validate this hypothesis and construct a robust dataset, we designed laboratory experiments simulating the identified grid conditions. We used 20 Ah LFP energy storage batteries, cycling them within an SOC window of 20% to 100% at a constant 0.5 C rate, mirroring the stable current phase observed in the field. Periodic reference capacity tests were conducted to obtain the true SOH label, defined as:
$$SOH = \frac{Q_C}{Q_{new}} \times 100\%$$
where \(Q_C\) is the current maximum available capacity and \(Q_{new}\) is the nominal capacity. Since the cycling was between 20% and 100% SOC, the measured capacity \(Q_s\) was converted to an equivalent full-cycle capacity \(Q_d\) for accurate SOH calculation using:
$$Q_d = \frac{Q_s}{0.8}$$
The core feature extracted from each simulated cycle was the sequence of charging voltage differences within the 30 minutes preceding the battery voltage reaching 3.41 V. Given that field data logging intervals are often 1 minute, we investigated different sampling intervals to balance information richness and practicality. For each cycle, we constructed three feature vectors:
- F1: Vector of 29 consecutive 1-minute voltage differences (\(\Delta V_{1min}\)).
- F2: Vector of 10 consecutive 3-minute voltage differences (\(\Delta V_{3min}\)).
- F3: Vector of 6 consecutive 5-minute voltage differences (\(\Delta V_{5min}\)).
These feature sets, derived from fragmented charging data, were intended to capture the subtle voltage changes indicative of aging in the energy storage battery. A summary of the data structure for the 1-minute interval is presented below. Similar tables were constructed for the 3-minute and 5-minute intervals.
| Cycle Index | \(\Delta V_1\) (mV) | \(\Delta V_2\) (mV) | … | \(\Delta V_{29}\) (mV) | True SOH (%) |
|---|---|---|---|---|---|
| 1 | 0.2 | 0.2 | … | 0.9 | 100.0 |
| 2 | 0.8 | 0.1 | … | 1.2 | 99.8 |
| … | … | … | … | … | … |
| 12300 | 7.2 | 5.2 | … | 1.5 | 92.1 |
Prior to model training, all feature vectors and target SOH values were normalized to a [0, 1] range using min-max normalization to ensure stable and efficient neural network training:
$$X’ = \frac{X – X_{min}}{X_{max} – X_{min}}$$
For modeling the complex, non-linear relationship between these fragmented voltage difference features and the SOH of the energy storage battery, we employed an Artificial Neural Network (ANN). Specifically, a Backpropagation (BP) neural network was chosen for its powerful non-linear mapping capabilities. However, standard BP networks are prone to converging to local optima and are sensitive to initial weight and bias values. To overcome these limitations, we integrated a Genetic Algorithm (GA) to optimize the initial parameters of the BP network, creating a hybrid GA-BP model. The GA performs a global search for a superior set of initial weights and thresholds before the BP network’s gradient-based fine-tuning begins, leading to faster convergence and better performance.
The architecture of our GA-BP neural network for energy storage battery SOH estimation is summarized below:
| Layer | Number of Nodes | Activation Function | Remark |
|---|---|---|---|
| Input Layer | 29, 10, or 6 | Linear | Depends on feature vector (F1, F2, F3) |
| Hidden Layer 1 | 10 | Hyperbolic Tangent (tanh) | Parameters optimized by GA |
| Hidden Layer 2 | 10 | Hyperbolic Tangent (tanh) | Parameters optimized by GA |
| Hidden Layer 3 | 10 | Hyperbolic Tangent (tanh) | |
| Output Layer | 1 | Linear | Outputs estimated SOH |
The GA optimization process involved encoding the network’s weights and thresholds into chromosomes, applying selection, crossover, and mutation operations over generations to evolve a population towards minimizing the initial prediction error. The fitness of a chromosome was evaluated by the mean squared error (MSE) of the network initialized with its decoded parameters on a subset of training data. The workflow is described by the following sequence:
- Encoding: Encode all initial weights and thresholds of the BP network into a chromosome.
- Population Initialization: Generate an initial population of random chromosomes.
- Fitness Evaluation: For each chromosome, decode it into a network, perform a forward pass on training data, and calculate MSE as fitness.
- Genetic Operations: Apply selection (roulette wheel), crossover (single-point), and mutation (random bit flip) to create a new generation.
- Termination: Repeat steps 3-4 until a maximum generation is reached or fitness plateaus.
- BP Training: Use the best chromosome’s parameters to initialize the BP network, then train using gradient descent with momentum.
The objective function for the BP training phase was the Mean Squared Error between the predicted SOH \(\hat{y}\) and the true SOH \(y\):
$$J(\mathbf{w}) = \frac{1}{N} \sum_{i=1}^{N} (\hat{y}_i – y_i)^2$$
where \(\mathbf{w}\) represents the network’s weight vector.
We constructed three separate GA-BP models, denoted M-1, M-2, and M-3, corresponding to feature sets F1 (1-min), F2 (3-min), and F3 (5-min) respectively. Data from three 20 Ah energy storage batteries, totaling over 12,000 cycles, was used. The dataset was split 80:20 into training and testing sets. Furthermore, an additional 1200 cycles from a separate 20 Ah energy storage battery were held out as a completely independent validation set to rigorously assess the models’ generalization ability. The performance metrics used were Mean Absolute Percentage Error (MAPE) and Root Mean Square Error (RMSE), defined as:
$$MAPE = \frac{1}{N} \sum_{n=1}^{N} \left| \frac{Y(n) – S(n)}{S(n)} \right| \times 100\%$$
$$RMSE = \sqrt{ \frac{1}{N} \sum_{n=1}^{N} (Y(n) – S(n))^2 }$$
where \(Y(n)\) is the predicted SOH and \(S(n)\) is the true SOH for sample \(n\).
The estimation results on the independent validation set are presented graphically and quantitatively. All three models demonstrated high accuracy, successfully tracking the SOH degradation of the energy storage battery using only the fragmented charging data. The model M-1, utilizing 1-minute voltage differences, achieved the lowest error, highlighting that higher temporal resolution in the feature data captures more nuanced aging signatures. The quantitative comparison is as follows:
| Model Name | Feature Parameter | MAPE (%) | RMSE |
|---|---|---|---|
| M-1 | 1-minute voltage differences | 0.37 | 0.4565 |
| M-2 | 3-minute voltage differences | 0.40 | 0.5095 |
| M-3 | 5-minute voltage differences | 0.44 | 0.5224 |
While the model performed excellently on the 20 Ah energy storage batteries it was trained on, a critical test was its applicability to other energy storage battery models with different capacities, such as the 260 Ah cells used in the actual power station. Direct application of model M-1 to 20 fresh 260 Ah energy storage batteries yielded a maximum error of 5.52%, indicating a capacity-dependent bias or feature distribution shift. To enhance the model’s universality without requiring massive new datasets, we employed a transfer learning strategy. We froze the weights and thresholds of the first two hidden layers of the pre-trained M-1 model, which presumably learned general feature representations related to LFP energy storage battery aging. We then used a small dataset of only 60 cycles from 260 Ah energy storage batteries to retrain (fine-tune) only the final hidden layer and the output layer. This process effectively adapted the model to the new cell format. The results were markedly improved, reducing the maximum estimation error for the 260 Ah energy storage batteries from 5.52% to 1.89%. The transfer learning objective can be formalized as minimizing a composite loss:
$$L_{total} = \lambda L_{source}(W_f) + (1-\lambda) L_{target}(W_t)$$
where \(L_{source}\) is the loss on the original 20Ah data for frozen weights \(W_f\), \(L_{target}\) is the loss on the new 260Ah data for tunable weights \(W_t\), and \(\lambda\) is a weighting factor close to 0 in our fine-tuning phase.
The ultimate validation of our approach lies in its deployment for bulk SOH estimation on in-service energy storage batteries. Using the transferred model, we processed one month of 1-minute interval operational charging data from a cluster of 224 in-service 260 Ah energy storage batteries at the photovoltaic station. The model processed the fragmented data from each cell and generated SOH estimates. The results showed a tight distribution of SOH values around 95%, which is plausible for batteries in their early operational life and consistent with station maintenance records. This batch estimation process, requiring only standard charging data snippets, demonstrates exceptional engineering adaptability and provides a practical tool for fleet management of energy storage batteries.
Our discussion centers on the significance of using engineered features from fragmented data. The choice of the voltage difference in the 3.41V region is not arbitrary; it is underpinned by the degradation dynamics of LFP energy storage batteries. The dominant aging modes cause measurable changes in the voltage profile during this specific charging phase. By focusing on this fragment, the method becomes feasible for real-world applications where full cycles are unavailable. The GA-BP model provides a robust framework for learning the complex mapping, and the transfer learning extension dramatically broadens its utility across different energy storage battery packs. A comparison of the information content in different sampling intervals suggests that for LFP energy storage batteries with flat voltage plateaus, finer sampling (1-minute) preserves critical differential information that coarser intervals may smooth over, explaining the superior performance of M-1. Furthermore, the method’s computational efficiency after training allows for near-real-time SOH monitoring of thousands of energy storage battery cells, a necessity for large-scale storage facilities.
In conclusion, this research presents a comprehensive framework for accurate and practical SOH estimation of lithium-ion energy storage batteries. By analyzing real grid operation patterns and battery aging mechanisms, we identified a highly informative but logistically feasible data fragment—the charging voltage curve below 3.41 V. We developed a GA-optimized BP neural network model that achieves remarkable accuracy (MAPE 0.37%) in estimating SOH from this fragmented data for 20 Ah energy storage batteries. More importantly, we demonstrated that through a targeted transfer learning approach, the model can be efficiently adapted to different capacity cells (260 Ah), reducing estimation errors by over 65%. Finally, the successful batch application to a live cluster of energy storage batteries in a photovoltaic power station validates the method’s strong engineering adaptability. This work provides a viable pathway for implementing continuous, non-invasive health monitoring in grid-scale energy storage systems, ultimately contributing to safer, more efficient, and longer-lasting energy storage battery deployments. Future work will explore the application of this fragmented data approach to other lithium-ion chemistries and investigate online learning techniques to allow the model to adapt continuously to evolving aging trajectories of energy storage batteries in the field.