In the context of the national push for clean energy and the construction of a new power system, energy storage has played a significant role as a crucial means to balance power supply and demand and promote the integration of renewable energy. The state of the li ion battery, being the core of battery and energy storage management systems, has garnered widespread attention, and its performance remains a key constraint for large-scale deployment. As a new type of green and environmentally friendly battery, the li ion battery boasts characteristics such as high specific energy, long cycle life, high output power, and tolerance for a wide range of charge-discharge rates. The thermal behavior of a li ion battery during operation significantly impacts its performance and safety. Therefore, monitoring the charge-discharge behavior of li ion batteries is essential.
This paper focuses on the li ion batteries within electrochemical energy storage power stations as the research subject. It records their charge-discharge behavior, conducting comparative analyses of voltage inconsistency, temperature inconsistency, State of Charge (SOC) inconsistency, SOC jumps in battery packs and clusters, voltage differentials, temperature differentials, and post-charge/discharge voltage differences. After each charge or discharge cycle, data from the process is calculated and analyzed. If anomalies are detected, that particular charge or discharge cycle is flagged as abnormal, and the results are stored in a data table. If no anomalies are found, the cycle is marked as normal.
By leveraging real-time monitoring data from an electrochemical energy storage monitoring platform, this study conducts an in-depth analysis of the consistency within battery packs at energy storage stations. First, operational data of the batteries is collected and organized through the platform, identifying key parameters such as voltage, temperature, and SOC of the li ion batteries to characterize consistency metrics. Second, evaluation feature data that reflects battery pack consistency is extracted. Finally, research is conducted on applying the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) clustering algorithm to screen for abnormal individual battery cells. Experiments prove that this algorithm can efficiently extract key electrical characteristic quantities from the operational data of energy storage stations, accurately assess the consistency status of battery packs, and rapidly pinpoint potentially faulty individual cells. This research holds significant practical value for enhancing the operational stability of energy storage stations, enabling timely judgment of battery consistency, and achieving precise screening of abnormal li ion battery cells.
Energy Storage Station Data Structure and Preprocessing Algorithm
1. Battery Pack Data Structure in Energy Storage Stations
The energy storage system of a station comprises multiple battery stacks. Each stack contains several battery clusters, and each cluster consists of hundreds of series-connected lithium iron phosphate (LiFePO4) li ion battery cells. The station is equipped with voltage monitoring, temperature monitoring, and SOC data monitoring devices to track the performance parameters of battery stacks, clusters, and individual cells in real-time.
2. Data Preprocessing
The energy storage station stores data at 15-minute intervals. For all battery cells, 12 data points for voltage, temperature, and SOC are collected. Every 3 hours is treated as one data set, capturing the voltage, temperature difference, and SOC difference among cells within a battery pack. The designed battery data preprocessing algorithm flow primarily involves three key steps: file transcoding, data extraction, and data imputation.
First, battery data files are transcoded to UTF-8 format to facilitate subsequent extraction and analysis. Then, based on the needs of the subsequent analysis, the voltage, temperature, and SOC values of individual li ion battery cells required for consistency analysis are extracted.
3. Missing Value Handling
During data acquisition, missing data points and sample rows can occur due to system hardware failures and data transmission issues. Therefore, handling missing values is necessary to ensure data segments are as complete as possible. As data can be missing in various ways (single value, single row, multiple rows, etc.), multiple methods are employed accordingly.
The specific imputation strategy is as follows: When a missing value appears in the j-th field of the i-th data row, first, determine if the timestamps for the rows with missing data are continuous, then analyze the missing time period (multiple row gaps). If the time data remains continuous and the missing data time span is less than 30 seconds (3 data rows), regardless of the number of missing values (including entire row gaps), interpolation is used for filling. The specific steps are: select 10 points near the missing value (defaulting to data from row i-5 to row i+5, adjusting if the gap is near the start or end of the data) and apply Newton’s interpolation method for filling. However, if the timestamps are not continuous, or the missing time exceeds 30 seconds, a deletion method is used, directly removing the entire row of data.
| Step | Action | Purpose/Logic |
|---|---|---|
| 1. File Transcoding | Convert all data files to UTF-8 format. | Standardize format for reliable extraction and analysis. |
| 2. Data Extraction | Extract voltage (V), temperature (T), and SOC values for individual li ion battery cells. | Prepare the specific data needed for consistency analysis. |
| 3. Missing Value Handling | Check timestamp continuity and gap duration. | Determine the appropriate method (imputation or deletion). |
| 3a. Imputation (Newton’s Method) | Applied if time is continuous & gap < 30s. | Fill missing values using nearby data points to preserve data integrity. |
| 3b. Deletion | Applied if time is NOT continuous OR gap ≥ 30s. | Remove unreliable data segments where interpolation would be inaccurate. |
Data Encoding
Electrochemical energy storage stations generate massive amounts of data with complex device configurations, necessitating a highly efficient and stable processing system to handle this vast data volume at the station level. It is crucial to transmit important data information stably and efficiently to centralized control platforms over limited power bandwidth. On average, a 1 MW energy storage unit generates about 10,000 data points. A 100 MW energy storage station would produce around 1 million data points. Energy storage data is typically transmitted at a per-second rate, especially for cell-level data, where second-level concurrency can reach several hundred thousand points. Furthermore, energy storage stations contain numerous types of equipment, and even for the same type of device, different manufacturers may have significant differences in functionality, data format, and communication protocols.
Therefore, for energy storage stations dealing with millions of data points and complex equipment, an efficient, easy-to-process, and easily extensible modeling and transmission method is required. An efficient data organization method must also be designed to transmit data to cloud or other platforms with acceptable latency over limited power network bandwidth. The adopted solution employs a Message Queuing and JSON/Protobuf communication model. This model can handle second-level big data transmission and features high-compression-ratio, high-security data encoding. It has been validated through operation at multiple hundred-megawatt energy storage sites, proving to be a secure and efficient mode.
1. Data Transmission Network Topology
When transmitting operational data from an energy storage station, data is first collected from the secondary systems (PCS and BMS) located in Security Zone I. The monitoring system in Zone I then forwards this data to a data aggregation server in Zone II. The data traverses a正向 isolation device to reach a data computation server in Zone III. This data server aggregates all real-time monitoring data from the entire station and uploads it to the cloud platform via a dedicated power line, passing through a firewall. The topology is illustrated in the text.
2. Data Classification
In the design of data classification for energy storage stations, the IDEF0 modeling method is applied. Data classification uses the systems engineering modeling tool IDEF0. It employs a top-down, business-attribute-oriented approach to identify and partition the hierarchical structure of the energy storage station business domain, ultimately identifying the smallest business units. The main elements of IDEF0 modeling are boxes and arrows. Boxes represent activities that perform certain functions, while arrows represent real information or objects required by or produced from the activity. A basic IDEF0 functional activity consists of five elements: Activity, Control, Mechanism, Input, and Output. Control refers to the mechanism and conditions for executing the activity.
| Element | Symbol | Description |
|---|---|---|
| Activity | Box | The function or process being performed. |
| Input | Arrow entering left side | Data or objects consumed by the activity. |
| Output | Arrow leaving right side | Data or objects produced by the activity. |
| Control | Arrow entering top | Conditions, rules, or standards governing the activity. |
| Mechanism | Arrow entering bottom | Resources (e.g., systems, people) that perform the activity. |
The IDEF0 method builds the system’s functional model through a top-down, layer-by-layer decomposition. Based on data classification theory and oriented towards business attributes, the energy storage station business domain is hierarchically identified and partitioned. The data is determined to have a five-level hierarchical architecture comprising Subject Domain Group, Subject Domain, Business Object, Data Entity, and Attribute, thereby establishing the subordination and association relationships of its categories.
3. Method for Data Encoding in Electrochemical Energy Storage Stations
Data encoding for electrochemical energy storage stations is the process of systematically coding data at various hierarchical levels based on the station’s data classification system. This encoding should follow basic principles such as uniqueness, matching, expandability, and conciseness. The encoding method preferably uses a hierarchical coding approach. The format for energy storage station encoding is as shown in the text.
| Code Segment | Length | Description / Example |
|---|---|---|
| Station/Company Code | Varies | Unique identifier for the station or owning company (e.g., ‘ESS_001’). |
| Device Type Code | Varies | Code for the type of equipment (e.g., ‘BMS’ for Battery Management System). |
| Hierarchical Location Code | Varies | Encodes the physical/logical hierarchy (e.g., ‘STK01_CLU05’ for Stack 1, Cluster 5). |
| Data Point/Attribute Code | Varies | Specific parameter or attribute (e.g., ‘Cell_Voltage’, ‘Temp’, ‘SOC’). |
| Instance Identifier | Varies | Identifies the specific instance (e.g., cell number ‘012’). |
Battery Pack Consistency Evaluation Algorithm
The consistency among individual li ion battery cells within a battery pack affects the overall system’s performance, efficiency, safety risks, battery lifespan, and the stable operation of the energy storage station. By regularly assessing the consistency state of the batteries, potential issues can be detected early, allowing for corresponding measures to balance battery usage, improve system safety, extend battery pack life, ultimately reduce maintenance costs, and ensure the reliability and economic viability of the energy storage system.
1. Selection of Characteristic Quantities for Battery Pack Data
First, the energy storage monitoring platform takes a sample point every 15 minutes, collecting 12 data points for voltage, temperature, and SOC for all batteries. Every 3 hours is considered one data set. The temperature difference, voltage difference, and SOC difference for batteries within a pack are calculated using the following formulas:
Let $T_n^i$, $V_n^i$, and $S_n^i$ represent the temperature, voltage, and SOC, respectively, of the $i$-th li ion battery cell at the $n$-th collection point within a pack containing $N$ cells.
The average values for the pack at point $n$ are:
$$ \overline{T_n} = \frac{1}{N} \sum_{i=1}^{N} T_n^i $$
$$ \overline{V_n} = \frac{1}{N} \sum_{i=1}^{N} V_n^i $$
$$ \overline{S_n} = \frac{1}{N} \sum_{i=1}^{N} S_n^i $$
The deviation of each cell from the pack average at point $n$ is:
$$ \Delta T_n^i = T_n^i – \overline{T_n} $$
$$ \Delta V_n^i = V_n^i – \overline{V_n} $$
$$ \Delta S_n^i = S_n^i – \overline{S_n} $$
2. Consistency Evaluation Algorithm
After calculating the relevant deviation values for each individual li ion battery cell, these deviations are compared against set thresholds: voltage threshold ($V_{th}$, e.g., 0.1 V), temperature threshold ($T_{th}$), and SOC threshold ($SOC_{th}$). Exceeding a threshold is considered a discrepancy. The cell’s corresponding data is then recorded and saved.
For the judgment rule regarding SOC jumps in battery packs or clusters, the monitoring platform calculates the per-minute SOC difference for a pack or cluster. If the SOC difference exceeds a threshold (e.g., 5%), it is considered a jump. The identification number of the pack or cluster experiencing the jump is stored in a data table, recording the anomaly type as “SOC jump.” Simultaneously, the platform calculates voltage differentials and temperature differentials. That is, for the same timestamp and within the same battery cluster, the difference between the highest cell voltage and the lowest cell voltage, and the difference between the highest cell temperature and the lowest cell temperature. If the voltage differential exceeds its threshold or the temperature differential exceeds its threshold (e.g., 10°C), it is considered a voltage differential out-of-limit or temperature differential out-of-limit, triggering an alarm.
When an alarm is received for a specific battery pack, an algorithm for screening abnormal cells within that pack becomes necessary. By analyzing each li ion battery cell within the pack individually, the underperforming cells can be identified. Repairing or replacing these weaker li ion battery cells can effectively improve the efficiency and safety of the entire battery pack.
Algorithm for Screening Abnormal Individual Cells Within a Pack
Clustering is an unsupervised learning method that can partition a dataset containing various samples into multiple categories, called “clusters,” based on some intrinsic relationship (similarity or distance), without requiring data labels. Clusters are typically unrelated to each other. The goal of clustering is to group behaviorally similar samples into the same cluster based on feature similarity, while dispersing dissimilar samples into different clusters.
1. DBSCAN Clustering Algorithm
DBSCAN (Density-Based Spatial Clustering of Applications with Noise) is a density-based clustering algorithm used to partition sample points in a dataset into clusters with similar features, effectively handling noise data. This algorithm identifies cluster shapes based on the density around sample points, does not require pre-setting the number of clusters, and is relatively effective at handling non-convex cluster shapes and noise points. The core idea of the DBSCAN algorithm is to define a radius ($\epsilon$) and a minimum number of samples ($MinPts$) to classify sample points with sufficient density into a cluster. Specifically, the steps of the algorithm are as follows:
Initialization: Set the two parameters: neighborhood radius ($\epsilon$) and minimum points ($MinPts$).
Process:
1) Select an unvisited sample point from the dataset.
2) Compute the $\epsilon$-neighborhood: Calculate the distance from the selected point to all other points, identifying points within distance $\epsilon$.
3) Identify Core Points: If the number of points in the $\epsilon$-neighborhood is $\geq MinPts$, mark the current point as a core point.
4) Expand Cluster: Expand the cluster from this core point by adding all points in its $\epsilon$-neighborhood to the cluster.
5) Mark Visited: Mark all points added to the cluster as visited.
6) Repeat: Repeat steps 1-5, selecting the next unvisited point, until all points are visited.
7) Identify Noise: Points that were visited but are not part of any core point’s cluster are marked as noise (outliers).
8) Output: Return the cluster assignments, including core points, border points (points in a cluster but not core points), and noise points.
2. Feature Selection and Clustering Analysis
To assess the consistency of li ion battery cells within a pack for a given day, the battery data from that day for the energy storage station is selected. For clustering analysis of the batteries, one or more characteristic quantities representing consistency status must be selected. The choice of features directly affects the accuracy of clustering results. Therefore, the Z-score standardization method is used to extract feature quantities from the battery data.
Assume a battery pack has $N$ cells, and there are $P$ data collection points over one day. Each li ion battery cell $i$ has $P$ temperature, voltage, and SOC values. The Z-score for cell $i$ at collection point $p$ is calculated for each parameter. This Z-score, denoted as $Z_{T,p}^i$, $Z_{V,p}^i$, and $Z_{S,p}^i$ for temperature, voltage, and SOC respectively, indicates how much the cell’s value at that point deviates from the pack’s mean, measured in standard deviations. The average Z-score for cell $i$ across all $P$ points is then computed for each parameter, providing a single daily measure of its deviation in temperature ($\overline{Z_T}^i$), voltage ($\overline{Z_V}^i$), and SOC ($\overline{Z_S}^i$).
$$ \overline{Z_T}^i = \frac{1}{P} \sum_{p=1}^{P} Z_{T,p}^i = \frac{1}{P} \sum_{p=1}^{P} \frac{T_p^i – \mu_{T,p}}{\sigma_{T,p}} $$
$$ \overline{Z_V}^i = \frac{1}{P} \sum_{p=1}^{P} Z_{V,p}^i = \frac{1}{P} \sum_{p=1}^{P} \frac{V_p^i – \mu_{V,p}}{\sigma_{V,p}} $$
$$ \overline{Z_S}^i = \frac{1}{P} \sum_{p=1}^{P} Z_{S,p}^i = \frac{1}{P} \sum_{p=1}^{P} \frac{S_p^i – \mu_{S,p}}{\sigma_{S,p}} $$
where $\mu_{T,p}, \mu_{V,p}, \mu_{S,p}$ are the pack mean and $\sigma_{T,p}, \sigma_{V,p}, \sigma_{S,p}$ are the pack standard deviation for temperature, voltage, and SOC respectively at point $p$.
These daily average Z-scores for each li ion battery cell are then used to form feature datasets. For instance, using $\overline{Z_V}^i$ and $\overline{Z_T}^i$ as two features, a two-dimensional clustering of the cells is performed using the DBSCAN algorithm. In the resulting cluster plot, the horizontal axis represents the voltage deviation (inconsistency) of a cell, and the vertical axis represents its temperature deviation. Each cell is represented as a point.
After DBSCAN clustering, if a point is classified as a noise point (outlier), it indicates that this li ion battery cell’s data is relatively distant from the data of other cells in the pack, suggesting the cell may be faulty. The final fault identification results are then displayed on the energy storage monitoring platform for further inspection and maintenance action.