Methods and systems for assessing the health status of energy storage devices in multi-mode grid connection scenarios

Abstract

This invention relates to the field of energy storage system monitoring technology, and discloses a method and system for assessing the health status of energy storage devices under multi-mode grid connection scenarios. The method includes acquiring historical multi-source heterogeneous data and real-time operating parameters of the energy storage unit; based on the historical multi-source heterogeneous data, performing operating condition identification, autoencoder training, and cluster analysis to obtain an ideal health status vector center library for each scenario; based on the real-time operating parameters and the ideal health status vector center library, performing feature mapping and distance calculation to obtain the state deviation; and based on the state deviation, performing time-series trend fitting and multi-dimensional feature matching to obtain the final health level. This method can achieve accurate assessment of the health status of energy storage devices and early warning of degradation under complex operating conditions.

Description

Technical Field

[0001] This invention relates to the field of energy storage system monitoring technology, and in particular to a method and system for assessing the health status of energy storage devices in multi-mode grid connection scenarios. Background Technology

[0002] Currently, with the increasing penetration of new energy sources such as wind and solar power into the power grid, energy storage systems, as core equipment for smoothing fluctuations and ensuring the stability of power supply, have attracted much attention regarding their safety and reliability. To ensure that energy storage units maintain optimal operating conditions in complex power grid environments over the long term, prognostics and health management (PHM) technology for energy storage devices has become a hot research topic in the industry. Constructing an efficient and accurate fault detection and health management system is of significant practical importance for timely detection of potential risks, avoidance of catastrophic failures, and reduction of operational and maintenance efficiency.

[0003] In existing technologies, health status assessment of energy storage devices is typically limited to setting fixed static alarm thresholds for single key performance indicators such as voltage, temperature, and internal resistance. This method often monitors each parameter in isolation, triggering an alarm only when the parameter value absolutely exceeds a preset safety threshold. However, energy storage devices face switching between various grid connection modes in actual operation, such as peak shaving and frequency regulation, resulting in significant differences in the normal fluctuation range of parameters under different operating conditions. Traditional methods ignore the complex nonlinear coupling relationships between the multi-physics parameters within the device under different operating modes, making it difficult to distinguish between parameter fluctuations under normal operating conditions and parameter drift caused by early faults.

[0004] Therefore, existing technologies have technical problems that cannot effectively adapt to dynamic changes in operating conditions under multi-mode grid connection scenarios. Summary of the Invention

[0005] This invention provides a method and system for assessing the health status of energy storage devices in multi-mode grid connection scenarios, in order to solve the technical problem that existing technologies cannot effectively adapt to dynamic operating condition changes in multi-mode grid connection scenarios.

[0006] Firstly, to address the aforementioned technical problems, this invention provides a method for assessing the health status of energy storage devices in multi-mode grid-connected scenarios, comprising:

[0007] Historical operation data is acquired, and the historical operation data is processed by time alignment and health filtering to obtain a historical health synchronization dataset;

[0008] The historical health synchronization dataset is processed by working condition identification and label addition to obtain a scenario-based training dataset;

[0009] Based on the scenario-based training dataset, an autoencoder model is constructed and iteratively trained to obtain a feature encoding model.

[0010] Based on the feature encoding model and the scenario-based training dataset, feature space construction and clustering are performed to obtain a health status benchmark library;

[0011] Real-time operating parameters are obtained, and the real-time operating parameters are feature-mapped using the feature encoding model. Distance is calculated by combining the health status benchmark library to obtain the state deviation.

[0012] Based on a preset time window, a trend fitting analysis is performed on the state deviation to obtain the trend slope;

[0013] Based on the state deviation and the trend slope, multi-dimensional feature fusion and level matching are performed to determine the health status level.

[0014] Secondly, the present invention provides a health status assessment system for energy storage devices in multi-mode grid connection scenarios, comprising:

[0015] The data processing module is used to acquire historical operation data and perform time alignment and health filtering on the historical operation data to obtain a historical health synchronization dataset.

[0016] The working condition processing module is used to perform working condition identification and label addition processing on the historical health synchronization dataset to obtain a scenario-based training dataset.

[0017] The model training module is used to construct an autoencoder model and perform iterative training based on the scenario-based training dataset to obtain a feature encoding model.

[0018] The library construction module is used to construct and cluster the feature space based on the feature encoding model and the scenario-based training dataset to obtain a health status benchmark library.

[0019] The deviation calculation module is used to obtain real-time operating parameters, perform feature mapping on the real-time operating parameters using the feature encoding model, and perform distance calculation in combination with the health status benchmark library to obtain the state deviation degree.

[0020] The trend analysis module is used to perform trend fitting analysis on the state deviation according to a preset time window to obtain the trend slope;

[0021] The level determination module is used to perform multi-dimensional feature fusion and level matching based on the state deviation and the trend slope to determine the health status level.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] (1) This invention constructs a vector center library of ideal health status for different scenarios by acquiring historical multi-source heterogeneous data of energy storage units and performing operating condition identification and tag addition. This scenario-aware benchmark modeling method effectively overcomes the limitations of traditional methods that rely solely on fixed thresholds and cannot adapt to multi-mode switching scenarios such as peak shaving and frequency regulation. It can dynamically eliminate the interference of normal parameter fluctuations under different operating conditions on the evaluation results, and significantly improve the evaluation accuracy and environmental adaptability of energy storage devices in complex grid-connected environments.

[0024] (2) This invention extracts low-dimensional feature vectors by constructing an autoencoder model and calculates its real-time state deviation. By utilizing the nonlinear feature extraction capability of deep learning, it can keenly capture hidden features where single indicators such as voltage and temperature have not yet reached the hard alarm threshold, but the internal multi-physics coupling relationship has undergone slight degradation. This achieves a highly sensitive early warning of early performance degradation signs of energy storage equipment and solves the technical problem of delayed early warning in traditional monitoring methods.

[0025] (3) This invention determines the health status level by performing time-series trend fitting and frequency domain transformation on the state deviation, combined with multi-dimensional features such as ambient temperature. This comprehensive diagnostic mechanism that integrates time and frequency domains can not only quantify the gradual aging rate of the equipment through the trend slope, but also identify specific fault modes such as abnormal increase in internal resistance through the fluctuation main frequency. It realizes the leap from simple numerical monitoring to refined level diagnosis, and provides a scientific basis for formulating targeted operation and maintenance strategies. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of a method for assessing the health status of energy storage devices in a multi-mode grid connection scenario, provided in the first embodiment of the present invention.

[0027] Figure 2 This is a schematic diagram of the structure of a health status assessment system for energy storage devices in a multi-mode grid connection scenario provided by the second embodiment of the present invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Reference Figure 1 The first embodiment of the present invention provides a method for assessing the health status of energy storage devices in a multi-mode grid connection scenario, comprising the following steps:

[0030] S11, acquire historical operation data, and perform time alignment and health filtering processing on the historical operation data to obtain a historical health synchronization dataset;

[0031] S12, Perform work condition identification and label addition processing on the historical health synchronization dataset to obtain a scenario-based training dataset;

[0032] S13, Based on the scenario-based training dataset, construct an autoencoder model and perform iterative training to obtain a feature encoding model;

[0033] S14. Based on the feature encoding model and the scenario-based training dataset, feature space is constructed and clustering is performed to obtain a health status benchmark library.

[0034] S15, obtain real-time operating parameters, use the feature encoding model to perform feature mapping on the real-time operating parameters, and combine the health status benchmark library to perform distance calculation to obtain the status deviation.

[0035] S16, perform trend fitting analysis on the state deviation according to the preset time window to obtain the trend slope;

[0036] S17. Based on the state deviation and the trend slope, perform multi-dimensional feature fusion and level matching to determine the health status level.

[0037] In step S11, historical running data is acquired, and time alignment and health filtering are performed on the historical running data to obtain a historical health synchronization dataset, including:

[0038] The battery management system data sequence, the energy storage converter data sequence, and the energy management system instruction sequence are acquired, and the battery management system data sequence, the energy storage converter data sequence, and the energy management system instruction sequence are integrated to obtain the historical operating data;

[0039] The historical operation data is resampled and filled to obtain a multidimensional time series data set;

[0040] Feature parameters are extracted and calculated from the multidimensional time-series data set to obtain thermodynamic property parameter values;

[0041] If the thermodynamic characteristic parameter value exceeds the preset safety threshold, abnormal data removal processing is performed to obtain the historical health synchronization dataset.

[0042] It should be noted that the acquisition of data sequences from the Battery Management System (BMS), the Energy Storage Converter (PCS), and the Energy Management System (EMS) is achieved through a communication gateway connected to the energy storage power station. This gateway subscribes to high-frequency messages (including cell voltage and temperature, typically at 100 milliseconds) from the BMS, medium-frequency messages (including AC output power and current, typically at 1 second) from the PCS, and low-frequency messages (including dispatch commands and grid frequency, typically at 5 seconds) from the EMS. These data streams from different sources, at different frequencies, and with independent local timestamps are aggregated into the same database, thus constituting the aforementioned historical operational data.

[0043] It is worth noting that the database in question is a time-series database pre-built during the system deployment phase. Its construction process involves first creating a data storage table structure and defining key column attributes, specifically including "Global Timestamp" (a 64-bit integer accurate to milliseconds), "Source ID" (a string used to distinguish BMS / PCS / EMS), "Metric Type" (an enumerated value, such as 'Cell_Temp', 'AC_Power'), and "Measurement Value" (a double-precision floating-point number). Subsequently, an inverted index is built with the "Global Timestamp" as the primary key to support efficient time-range queries on massive amounts of historical data.

[0044] It should be noted that the resampling and padding of the historical operational data aims to address the inconsistencies in time granularity between multi-source data. This process first establishes a unified global time reference. Specifically, the system synchronizes its local clock via Network Time Protocol (NTP) to the start time of the data acquisition task. Using the anchor point, a continuous standard discrete time series is generated according to a preset resampling interval (e.g., 1 second). This serves as a unified timeline for the entire system. For high-frequency BMS data (such as temperature), a maximum value aggregation method is used to extract the maximum value of all sampling points within each 1-second time window as the representative value for that second, capturing the most extreme heat dissipation conditions during that time period. For low-frequency EMS data (such as scheduling instructions), forward padding is used, meaning that before the next new instruction arrives, the current value is filled with the value of the previous known instruction to ensure the continuity of instruction status on the timeline. After this processing, all data is aligned to the same second-level time grid, forming a synchronized matrix, which is the aforementioned multi-dimensional time-series data set.

[0045] It should be noted that the feature parameter extraction and calculation of the multidimensional time-series data set involves traversing each time record in the set and parsing out the key thermodynamic indicators. In this embodiment, the highest temperature of a single battery cell is extracted as the thermodynamic characteristic parameter value. The specific steps of this extraction operation are as follows: the system parses the data frame uploaded by the BMS containing the temperatures of all individual cells in the entire battery cluster and obtains the temperature value array. ,in The total number of individual cells; an iterative comparison method is used to initialize the maximum value variable. The first element of the array It is compared sequentially with the remaining elements in the array. If the current element is greater than... Then update The result obtained after traversal This represents the highest temperature of a single battery cell at that moment. If the thermodynamic characteristic parameter value exceeds a preset safety threshold, it indicates that the battery is in an abnormal high-temperature operating state (such as a risk of thermal runaway), and this data point cannot be used as a sample for constructing a health benchmark. Therefore, the system performs abnormal data removal processing, physically deleting records for all dimensions at that moment from the dataset. The remaining data records that do not exceed the limits constitute the historical health synchronization dataset.

[0046] It is worth noting that the preset safety threshold (e.g., 60 degrees Celsius) is determined based on statistical analysis of the safe operating boundaries in the battery specifications and historical thermal runaway accident data. Specifically, temperature distribution data of this battery model during historical operation is collected, and combined with the electrolyte decomposition temperature critical value provided by the battery manufacturer, 90% of this critical value (leaving a 10% safety margin) is selected as the safety threshold to ensure that the selected data is strictly within the safe and healthy operating range of the battery.

[0047] For example, in a record with a timestamp of 14:30:15, the system extracts a maximum battery cell temperature of 61.5 degrees Celsius. Since this value exceeds the preset safety threshold of 60 degrees Celsius, the system determines this data point to be invalid and removes it from the multidimensional time-series data set. However, at 14:30:16, the temperature drops to 59.8 degrees Celsius, and this record is retained and stored in the historical health synchronization dataset.

[0048] In step S12, the historical health synchronization dataset is processed for working condition identification and label addition to obtain a scenario-based training dataset, including:

[0049] The historical health synchronization dataset is subjected to electrical parameter analysis and filtering to obtain a smooth electrical state data stream;

[0050] Differential calculation processing is performed on the smoothed electrical state data stream to obtain the current change rate;

[0051] The operating mode code is extracted from the historical health synchronization dataset, and the current change rate and the operating mode code are matched with the preset operating condition identification rules to obtain the initial scene label;

[0052] The initial scene labels are subjected to time-series correction processing to obtain the scene-based training dataset.

[0053] It should be noted that the electrical parameter parsing and filtering of the historical health synchronization dataset first involves parsing the current and voltage time series uploaded by the power storage converter (PCS) from the dataset. Considering the measurement noise that may be introduced by the high-frequency switching operations of power electronic equipment, the system uses a moving average filtering method to preprocess these sequences. Specifically, a width of... A sliding window for time current value its smoothed value Through the formula

[0054]

[0055] The calculations show that this process effectively filters out random spikes or jitters at individual time points, making the trends in current and voltage more explicit and generating the smooth electrical state data stream.

[0056] It should be noted that differential calculation processing is performed on the smoothed electrical state data stream to quantify the dynamic characteristics of current changes. The system uses a first-order forward differential method to calculate the rate of change of current for each time slice. The specific calculation formula is as follows:

[0057]

[0058] in and These are the smoothed current values ​​for the current time and the next time, respectively. The sampling time interval (e.g., 1 second). Calculated... It can intuitively reflect the transient changes in the power throughput of the energy storage unit, namely the current change rate.

[0059] It should be noted that the matching process between the current change rate and the operating mode code and the preset operating condition identification rules is implemented using decision tree logic. The system extracts the operating mode code (e.g., "101" represents the charging mode, and "201" represents the discharging mode) from the EMS data, and searches for it in the preset rule base in conjunction with the calculated current change rate.

[0060] In one feasible implementation, the pre-built rule base is constructed during the system initialization phase. The construction process involves: first, collecting a large amount of historical operation logs and dividing the data into different original operating condition segments (such as charging, discharging, and resting segments) based on the scheduling instructions issued by the EMS; then, statistically analyzing the current change rate data within each operating condition segment and plotting a frequency distribution histogram; finally, identifying the physical dynamic boundaries of different operating conditions based on distribution characteristics (e.g., the current change rate is concentrated near 0 during steady-state operation, while it is distributed over a larger range during the ramp-up phase), and defining these statistical boundaries as discrimination thresholds in the rule base. The rule base defines the physical characteristic boundaries of different operating conditions; for example, a rule might be defined as follows: if the operating mode code is "101" and... (i.e., the stable current change rate threshold), then it is determined to be a constant current charging stage; if the operating mode code is "101" and If the rate of change of the ramp-up current (i.e., the threshold) is reached, then the system is determined to be in the charging power ramp-up phase. The tag corresponding to the successfully matched rule is the initial scenario tag.

[0061] It is worth noting that the width of the sliding window The determination of (e.g., 5) is based on the signal-to-noise ratio (SNR) analysis of historical electrical data. By comparing the denoising effect and signal delay under different window sizes, a window value that maximizes the SNR while keeping the delay within an acceptable range is selected. The threshold in the preset operating condition identification rule, namely the stable current change rate threshold... and the threshold of the rate of change of ramp current The determination of the threshold is based on statistical distribution analysis of a large amount of historical normal operation data. Specifically, current change rate data during historical constant current charging phases are collected, and the 95% confidence interval of its distribution is calculated. The upper bound of the interval is used as the stable current change rate threshold. (e.g., 2A / s) to ensure that the rules can cover the vast majority of normal operating conditions.

[0062] It should be noted that the timing correction processing of the initial scene labels aims to eliminate label flickering caused by instruction switching gaps or brief power grid disturbances. This embodiment uses a minimum duration threshold filtering method. The system sets a minimum operating condition duration threshold (e.g., 10 seconds) and iterates through the initial label sequence. If a continuous scene label sequence (e.g., standby) is found to have a duration less than the threshold, and the preceding and following segments are all from the same dominant scene (e.g., constant current discharge), then the short-duration label is determined to be noise and corrected to the label of the preceding and following dominant scenes. The corrected, continuous, and stable label sequence on the timeline, together with the original physical data, constitutes the scene-based training dataset.

[0063] It is worth noting that the minimum operating condition duration threshold is determined based on statistical analysis of the response characteristics of historical scheduling instructions. Specifically, the duration distribution of all valid operating condition states (i.e., successfully executed and sustained operating conditions) in historical operations is statistically analyzed, their cumulative distribution function (CDF) is calculated, and the 5th percentile of this distribution is selected as the minimum operating condition duration threshold. This setting ensures that most transient invalid tags caused by signal jitter can be filtered out, while retaining genuine short-term operating condition operations (such as short-term frequency modulation instructions).

[0064] For example, if the operating mode at a certain moment is coded as "101", the smoothed current changes from 250.5A to 251.2A, and the calculated rate of change... The current rate is 0.7 A / s. Since 0.7 A / s is less than the preset stable current change rate threshold of 2 A / s, the system initially marks it as constant current charging according to the rules. Subsequently, the system detects a standby tag that lasts only 1 second in the tag sequence. Since 1 second is less than the minimum operating time threshold of 10 seconds, the standby tag is corrected to constant current charging, thus ensuring the purity of the training data.

[0065] In step S13, based on the scenario-based training dataset, an autoencoder model is constructed and iteratively trained to obtain a feature encoding model, including:

[0066] The scenario-based training dataset is numerically normalized to generate a standardized high-dimensional feature matrix;

[0067] Construct an autoencoder model structure, and input the standardized high-dimensional feature matrix into the autoencoder model structure for mapping and reconstruction processing to obtain the reconstructed high-dimensional feature matrix;

[0068] Error calculation processing is performed on the reconstructed high-dimensional feature matrix and the standardized high-dimensional feature matrix to obtain the difference loss value;

[0069] If the difference loss value is lower than the preset convergence threshold, training is stopped and the encoding network is extracted to obtain the feature encoding model.

[0070] It should be noted that performing numerical normalization on the scenario-based training dataset is a crucial step in eliminating the influence of different physical dimensions (such as hundreds of volts of voltage versus single digits of the rate of change of current) on model weight updates. This embodiment employs the min-max normalization method. For each feature dimension in the dataset... Its normalized value Through the formula

[0071]

[0072] The calculation yielded, where and These are the minimum and maximum values ​​of the feature in the training set, respectively; if... - If the value of a feature is 0 (i.e., the feature is a constant value), then that feature is removed from the training feature set and will not participate in subsequent model training. After this processing, all feature values ​​are linearly mapped to the interval [0,1], forming the standardized high-dimensional feature matrix.

[0073] It's important to note that building an autoencoder model structure involves constructing a symmetric neural network that includes an encoder and a decoder. The encoder is responsible for compressing the high-dimensional input vector into a low-dimensional latent vector. Its structure includes an input layer (the number of nodes equals the feature dimension). Several hidden layers and bottleneck layers (number of nodes) The decoder is responsible for restoring the latent vectors to high-dimensional vectors, and its structure is mirror-symmetric to that of the encoder. In this embodiment, both the encoder and decoder contain two hidden layers. The ReLU activation function is used to solve the gradient vanishing problem, and the Sigmoid function is used in the output layer to match the normalized data range [0,1].

[0074] It should be noted that inputting the standardized high-dimensional feature matrix into the autoencoder model structure for mapping and reconstruction is a process of performing forward propagation computation. Specifically, the operation involves first inputting the vector... After layers of linear transformation and nonlinear activation by the encoder, the calculation formula is as follows:

[0075]

[0076] Obtain low-dimensional latent vectors (Mapping process); subsequently, the latent vector The image is fed into the decoder and undergoes multiple transformations. The calculation formula is as follows:

[0077]

[0078] Obtain an output vector with the same dimensions as the input. (Reconstruction processing), namely, reconstructing the high-dimensional feature matrix. Wherein... This is the weight matrix. This is the bias vector.

[0079] It is worth noting that the weight matrix The initial settings are based on the He initialization strategy. Specifically, for the number of connected input nodes... The layers have weight parameters ranging from a mean of 0 and a standard deviation of 0. The bias vector is randomly sampled from a normal distribution. This setting is based on the fact that when used with the ReLU activation function selected in this embodiment, it can maintain the consistency of the variance of the output values ​​of each layer of the network, thereby effectively preventing the gradient vanishing or exploding problem in the early stage of training of deep networks. It is then initialized as a vector of all zeros.

[0080] It should be noted that error calculation is performed on the reconstructed high-dimensional feature matrix and the standardized high-dimensional feature matrix to quantify the quality of the model's reconstruction capability. This embodiment uses the root mean square error (RMSE) as the loss function. For models containing... The difference loss value of a dataset of samples. Through the formula

[0081]

[0082] The calculation yielded, where These are the original standardized values. To reconstruct the values, the system uses the Adam optimizer for iterative training. The specific process of updating the weights is as follows: First, the loss function is calculated according to the chain rule. Relative to each weight parameter in the network gradient,

[0083]

[0084] Then, the first moment estimate of the gradient is calculated using the Adam algorithm. and second-order moment estimation Finally, according to the updated formula,

[0085]

[0086] Update weights, where The learning rate is used. When the calculated difference loss value remains below a preset convergence threshold (e.g., 0.01) for a consecutive preset number of epochs (e.g., 10 epochs), or when the loss on the validation set no longer decreases, an early stopping mechanism is triggered, and the model training is considered complete. At this point, the decoder part is removed, and only the trained encoder network is retained, which is the feature encoding model.

[0087] It is worth noting that the learning rate... The initial setting value (e.g., 0.001) is determined based on a learning rate scan test of historical training data. Specifically, during the model pre-training phase, the learning rate is linearly increased from a preset minimum to a maximum value, and the curve of the validation set loss changing with the learning rate is recorded. The learning rate corresponding to the point with the largest absolute value of the first derivative of this curve is calculated as the optimal initial value that enables the model to converge the fastest and without oscillation.

[0088] It is worth noting that the number of nodes in the bottleneck layer of the self-encoder The determination of (e.g., 3) is based on the prediction of the cumulative contribution rate of the features using principal component analysis (PCA). The specific calculation process involves first calculating the covariance matrix of the original data. Then on Perform eigenvalue decomposition and solve the characteristic equation. Obtain the eigenvalue sequence Finally, calculate the cumulative contribution rate.

[0089]

[0090] Select to make The smallest Value as The reference value is used. The determination of the preset convergence threshold (e.g., 0.01) is based on the statistical analysis of the model reconstruction error distribution. By testing the anomaly detection accuracy (F1-score) at different thresholds on the validation set, the error value that makes the F1-score optimal is selected as the threshold.

[0091] It should be noted that the number of nodes in the bottleneck layer of the self-encoder The value of 3 is an example; in practical applications, the optimal value is... The value depends on the specific energy storage device model, the number of selected feature dimensions, and the correlation between features; those skilled in the art can verify this through cross-validation or by observing the reconfiguration error as it progresses. The curve of value change (elbow rule) is used to determine the optimal value for a specific application scenario. The selection principle for the value is to reduce the dimensionality as much as possible and avoid overfitting while ensuring the effectiveness of feature extraction. The PCA cumulative contribution rate method mentioned above is an effective reference method, but it is not the only limiting method.

[0092] For example, an input vector containing 12 dimensions includes features such as the highest voltage of a single cell, the lowest voltage of a single cell, the average voltage of a single cell, the highest temperature of a single cell, the lowest temperature of a single cell, the average temperature of a single cell, the total current of the battery cluster, the total voltage of the battery cluster, the rate of change of current, active power, reactive power, and the grid frequency. The original value of the total voltage feature of the battery cluster is 675V (range 600-750V), which is normalized to 0.5. This 12-dimensional vector is compressed into a 3-dimensional vector by an encoder. After reconstruction by the decoder, the reconstructed value of the total voltage dimension of the battery cluster is 0.51. The overall root mean square error is calculated to be 0.008. Since 0.008 is lower than the preset convergence threshold of 0.01, and this condition has been met for 10 consecutive epochs, training ends, and the encoder is extracted for subsequent feature extraction.

[0093] In step S14, based on the feature encoding model and the scenario-based training dataset, feature space construction and clustering are performed to obtain a health status benchmark library, including:

[0094] The scenario-based training dataset is input into the feature encoding model for dimensionality reduction mapping to obtain historical low-dimensional feature vectors.

[0095] The historical low-dimensional feature vectors are grouped to obtain scene-specific vector clusters;

[0096] The scene-based vector clusters are subjected to core cluster identification processing to obtain the core clusters;

[0097] The centroid of the core cluster is calculated to obtain the initial health state center vector;

[0098] Establish the mapping relationship of the initial health state center vector to obtain the health state benchmark library.

[0099] It should be noted that inputting the scenario-based training dataset into the feature encoding model for dimensionality reduction mapping involves using the encoder network trained in S13 as a feature extractor. The system iterates through each standardized high-dimensional vector in the scenario-based training dataset. It is input into the encoder network and processed through matrix operation formulas. Perform forward propagation computation, where and These are the weight matrix and bias vector of the encoder bottleneck layer, respectively. This is the activation function. The output vector is calculated. This refers to the historical low-dimensional feature vector (e.g., 3-dimensional), which discards redundant information from the original data and retains the core features that best represent the device state. Grouping these historical low-dimensional feature vectors involves traversing the dataset, reading the scene label (e.g., constant current charging) attached to each data point, and grouping all feature vectors with the same label into the same set, thereby constructing multiple independent scene-specific vector clusters.

[0100] It should be noted that the core cluster identification process performed on the scenario-based vector clusters aims to extract the core region representing steady-state health from clusters containing noise and transitional data. This embodiment uses the DBSCAN clustering algorithm. For each scenario-based vector cluster, the algorithm calculates its core region centered on each vector point. The number of points in the neighborhood. If the number of points is greater than... If a point is found to be a core point, it is marked as such. All density-connected core points and their boundary points together form a high-density cluster. The algorithm automatically filters out outliers with insufficient density (usually corresponding to unstable states during condition switching), and the final retained main cluster is the core cluster.

[0101] It is worth noting that the neighborhood radius of the DBSCAN algorithm... The determination of (e.g., 0.05) is based on the K-distance map analysis of the distance distribution between point pairs in the low-dimensional feature space. This is achieved by calculating the distances from all points to their k-th ... Find the nearest neighbor distances, plot the sorted distance curves, and select the distance corresponding to the elbow point of the curve as the nearest neighbor distance. The best estimate. The minimum number of points. The determination of (e.g., 10) is usually set to more than twice the feature dimension to ensure the statistical significance of the cluster.

[0102] It should be noted that the centroid calculation for the core cluster is performed using the arithmetic mean method to determine the standard health state in this scenario. For clusters containing... vectors The core cluster, its centroid Through the formula

[0103]

[0104] The calculated centroid vector is geometrically located at the center of the high-density region, representing the most typical and stable health characteristic under this operating condition, i.e., the initial health state center vector. Establishing a mapping relationship refers to constructing a memory hash map. The system traverses all scenarios, using the scenario's unique identifier (e.g., "mode_101") as the key, and then uses the calculated centroid vector... The hash table is stored as a value. After construction, the hash table is serialized into a binary file and stored on the local disk, thus forming the health status benchmark library.

[0105] For example, in a constant current charging scenario, the system extracts 1000 3D feature vectors. The DBSCAN algorithm identifies 950 points belonging to the core cluster and removes 50 outliers. The average of these 950 points is calculated to obtain the centroid vector. The system stores the key-value pair {"mode_101_charge":[0.84,0.13,-0.46]} in the benchmark library. During subsequent online evaluations, the benchmark vector can be obtained simply by searching for "mode_101_charge".

[0106] In step S15, real-time operating parameters are obtained, and feature mapping and distance calculation are performed based on the feature encoding model and the health status benchmark library to obtain the state deviation, including:

[0107] Obtain real-time operating parameters and perform scene analysis according to preset operating condition recognition rules to obtain the current operating scene identifier;

[0108] The real-time operating parameters are input into the feature encoding model for feature extraction to obtain the current low-dimensional feature vector.

[0109] Based on the current operating scenario identifier, the ideal health state center vector is obtained by retrieving from the health state benchmark library;

[0110] The distance between the current low-dimensional feature vector and the center vector of the ideal healthy state is calculated to obtain the state deviation.

[0111] It should be noted that real-time operating parameters are obtained by periodically reading the current physical quantities from the local controller of the energy storage unit via the Modbus TCP communication protocol. These physical quantities specifically include individual battery voltage, individual battery temperature, AC side current of the energy storage converter, AC side voltage of the energy storage converter, active power, and grid frequency. Scene parsing based on preset operating condition identification rules is a crucial step in ensuring the consistency between online evaluation and offline modeling logic. This operation directly calls the preset operating condition identification rules determined in step S12 to logically match the real-time collected current change rate with the operating mode code. For example, if the real-time data meets the condition of "mode code 101 and current change rate less than 2A / s", it is determined that the current condition is constant current charging, and the generated identifier (such as "mode_101_charge") is the current operating scene identifier.

[0112] It should be noted that inputting the real-time operating parameters into the feature encoding model for feature extraction first requires data preprocessing identical to that in S13. The system calls the normalization parameters saved during the training phase of S13 (i.e., the normalization parameters for each feature in the training set). and The real-time data undergoes max-min normalization to ensure that the distribution of the input data is consistent with that during model training. Subsequently, the normalized vector is input into the feature encoding model obtained in S13. This model is the encoder network in an autoencoder, and its forward propagation computation process is completely consistent with the mapping process described in S13, outputting the activation values ​​of neurons in the bottleneck layer. This output value is the highly abstract current low-dimensional feature vector.

[0113] It should be noted that the retrieval from the health status benchmark database is implemented using a key-value lookup operation. The system uses the current operating scenario identifier as the key to search in the health status benchmark database (hash table structure) constructed in S14, and returns the corresponding centroid vector value. This vector represents the characteristic expression of the device's historical best health status under this specific operating condition, i.e., the ideal health status center vector.

[0114] It should be noted that the distance metric calculation between the current low-dimensional feature vector and the ideal health state center vector is implemented using the Euclidean distance formula. Let the current feature vector be... The ideal center vector is Then the state deviation degree Through the formula

[0115]

[0116] The calculated value quantifies the degree of displacement of the current device state from the ideal healthy state in the feature space, which is the state deviation.

[0117] It is worth noting that the system employs a truncation strategy to handle numerical overflow during normalization. Specifically, the operational logic is as follows: for any real-time feature value... If it is less than the minimum value of the training set If it is greater than the maximum value of the training set, then its normalized value is forcibly set to 0; If the input value is not found in the range of extreme outliers, its normalization value is forcibly set to 1; otherwise, it is calculated normally. This strategy prevents the explosion of input values ​​caused by extreme outliers, thereby avoiding uncontrollable fluctuations in the neural network inference output.

[0118] For example, the system collects data in real time and identifies the current scenario as constant current charging. The normalized real-time data is input into the encoding model, and the current feature vector is output. The system retrieves the ideal center vector corresponding to constant current charging from the benchmark library. Calculate the Euclidean distance between them.

[0119]

[0120] The value of 0.059 represents the current state deviation.

[0121] In step S16, a trend fitting analysis is performed on the state deviation according to a preset time window to obtain the trend slope, including:

[0122] The state deviation is truncated to obtain the original deviation sequence;

[0123] The original deviation sequence is smoothed to obtain a smoothed deviation numerical sequence;

[0124] Regression analysis is performed on the numerical sequence of smooth deviation to obtain the trend slope.

[0125] It should be noted that the sequence truncation processing of the state deviation is implemented based on a first-in-first-out (FIFO) circular buffer. The size of this buffer is determined by the length of the preset time window. Whenever a new state deviation value is calculated in step S15, the system pushes it to the end of the buffer while removing the oldest data point, thus maintaining a fixed-length observation window that slides over time. All data points extracted from the buffer and arranged in chronological order constitute the original deviation sequence.

[0126] It should be noted that smoothing the original deviation sequence aims to eliminate high-frequency random noise caused by instantaneous fluctuations in the power grid or sensor quantization errors, highlighting long-term trends. This embodiment uses the Simple Moving Average (SMA) method. A smoothing window size is set. ( (less than the preset time window length), for the first in the sequence Points its smoothed value Through the formula

[0127]

[0128] The calculation is obtained. This process is repeated for all data points in the sequence that meet the conditions, and the resulting new sequence is the numerical sequence of the smooth deviation.

[0129] It is worth noting that the determination of the preset time window (e.g., 60 seconds) is based on the dynamic analysis of the timescale of early fault evolution in energy storage units. By analyzing the characteristic evolution duration before voltage micro-short circuits and capacity drops occur in historical fault cases, a duration that can cover the period of early fault characteristic manifestation is selected as the window length. The smoothing window size... The determination of (e.g. 5) is based on the spectral analysis of the deviation sequence, selecting a window width that can filter out the main high-frequency noise components without causing significant phase delay.

[0130] It should be noted that the regression analysis performed on the aforementioned smooth deviation numerical sequence uses the least squares method to fit a linear trend line. A linear equation is then established. ,in The deviation after smoothing. For relative time (e.g., counting from 0). The slope Let be the intercept. The system minimizes the sum of squared errors. Solve for the optimal parameters This parameter The rate and direction of change of state deviation per unit time are directly quantified, which is the trend slope.

[0131] For example, the system sets the time window to 60 seconds. The buffer extracts the past 60 state deviation values ​​to form the original sequence. After smoothing with a moving average of width 5, the sequence fluctuations decreased. Linear regression was then performed, and the fitted linear equation was... The slope This indicates that the deviation is continuously increasing at a rate of 0.0005 per second, which is the slope of the trend.

[0132] In step S17, based on the state deviation and the trend slope, multi-dimensional feature fusion and level matching are performed to determine the health status level, including:

[0133] Obtain the current ambient temperature data;

[0134] The original deviation sequence is subjected to frequency domain transformation to obtain the main frequency of the fluctuation.

[0135] The deviation of the state is statistically processed to obtain the deviation amplitude;

[0136] The main frequency of the fluctuation, the magnitude of the deviation, the trend slope, and the current ambient temperature data are combined into a vector to obtain a quantized state feature vector.

[0137] The quantized state feature vector is matched with a preset health level classification rule library to obtain the health status level.

[0138] It should be noted that the current ambient temperature data is obtained by reading the real-time values ​​from the temperature sensor array deployed on the surface of the energy storage unit. The original deviation sequence is subjected to frequency domain transformation to identify hidden periodic oscillations in the deviation signal (which often correspond to internal resistance instability or electrochemical side reactions). This embodiment employs the Fast Fourier Transform (FFT) algorithm. First, the original deviation sequence obtained in S16 is processed to remove the DC component (subtract the mean); then, the FFT algorithm is applied to convert the time-domain signal into a frequency-domain spectrum; finally, the frequency component with the largest energy amplitude (excluding 0Hz) is searched in the spectrum, and this frequency is the dominant frequency of the fluctuation.

[0139] It should be noted that the amplitude statistical processing of the state deviation is to calculate the dispersion of the sequence. This embodiment uses the standard deviation formula.

[0140]

[0141] in This represents the deviation value in the sequence. This is the mean. Calculated... The fluctuation energy of the deviation is quantified, which is the deviation amplitude. Vector combination processing involves concatenating the trend slope calculated in S16, the dominant frequency of the fluctuation calculated in this step, the deviation amplitude, and the acquired current ambient temperature data into a multi-dimensional vector in a predefined order. That is, the quantized state feature vector.

[0142] It is worth noting that the predefined order is determined during the feature engineering configuration in the system initialization phase. The construction process involves first building an initial feature set containing dozens of candidate features (such as voltage variance and current skewness) during the offline modeling phase, and then using a random forest classifier for feature selection. This classifier consists of a preset number of... The system consists of 100 Classification and Regression Trees (CART), each built based on a subset of historical training data resampled using the Bootstrap method. When evaluating the algorithm, the system calculates the sum of the Gini Impurity decreases for each candidate feature across all decision tree nodes in the forest, using the following formula:

[0143]

[0144] in Features Importance rating For nodes The reduction in Gini impurity before and after splitting. Finally, the system is based on... The system numerically selects the four features with the highest importance scores (i.e., the aforementioned main frequency of fluctuation, deviation amplitude, trend slope, and ambient temperature), and solidifies the mapping relationship between the physical meaning of these four features and the vector dimensions into the system's feature index configuration file. This ensures that the feature vector structure constructed during online inference remains strictly consistent with the dimension definitions in the rule base, avoiding misjudgments caused by misalignment of feature order.

[0145] It should be noted that the matching process between the quantized state feature vector and the preset health level classification rule base is implemented using a multi-dimensional spatial region determination method. The rule base stores several rules, each defining a health level (such as early-stage capacity decay and abnormal increase in internal resistance) within a range of values ​​across the four dimensions mentioned above. The system traverses the rule base to determine the current vector. The system determines which rule-defined superrectangular area a character falls into, and then outputs the corresponding level label.

[0146] It is worth noting that the construction of the preset health level classification rule base is based on cluster analysis and fault tree inverse relation of battery aging test data throughout the entire life cycle. Specifically, a large amount of battery operation data from new cells to scrap is collected, and the above four features are extracted: dominant frequency of fluctuation, deviation amplitude, trend slope, and ambient temperature. The K-Means clustering algorithm is used to divide the feature space into several clusters with significant physical meaning. Combined with the battery disassembly analysis results, the centroid of each cluster is associated with specific physical degradation phenomena. For example, if a cluster corresponds to a disassembly sample with significant SEI film thickening, the cluster is labeled with abnormally increased internal resistance; if it corresponds to a sample with active lithium loss, it is labeled with early capacity decay. Finally, the numerical boundaries of each cluster in four-dimensional space are determined to form the rule base.

[0147] For example, the system obtains the current ambient temperature as 25 degrees Celsius. FFT analysis of the original deviation sequence reveals the main peak at 0.8 Hz. The standard deviation of the sequence is calculated, yielding an amplitude of 0.45. Combined with the trend slope of 0.0001 obtained from S16, a feature vector is constructed. There is a rule definition in the rule base, "If..." and and If so, it is determined as "abnormal increase in internal resistance". Because of the vector... All components of the energy storage unit meet the conditions of this rule, and the system finally determines the health status level of the energy storage unit as an abnormal increase in internal resistance.

[0148] In summary, this invention deeply integrates data-driven deep feature extraction with the logical constraints of physical operating conditions by constructing a full-process evaluation method that proceeds from standardized processing and operating condition scenario identification of multi-source heterogeneous data, to feature space construction and health benchmark library generation based on autoencoders, and finally to real-time deviation analysis and health level determination using fused time-frequency domain features. This invention innovatively establishes ideal health benchmarks for different scenarios, effectively solving the technical problems of assessment lag, high false alarm rate, and difficulty in capturing early degradation characteristics caused by reliance on fixed thresholds in existing technologies for multi-mode grid-connected scenarios. This significantly improves the accuracy, scenario adaptability, and timeliness of operation and maintenance early warning for energy storage equipment health status assessment.

[0149] Reference Figure 2 The second embodiment of the present invention provides a health status assessment system for energy storage devices in a multi-mode grid connection scenario, comprising:

[0150] The data processing module is used to acquire historical operation data and perform time alignment and health filtering on the historical operation data to obtain a historical health synchronization dataset.

[0151] The working condition processing module is used to perform working condition identification and label addition processing on the historical health synchronization dataset to obtain a scenario-based training dataset.

[0152] The model training module is used to construct an autoencoder model and perform iterative training based on the scenario-based training dataset to obtain a feature encoding model.

[0153] The library construction module is used to construct and cluster the feature space based on the feature encoding model and the scenario-based training dataset to obtain a health status benchmark library.

[0154] The deviation calculation module is used to obtain real-time operating parameters, perform feature mapping on the real-time operating parameters using the feature encoding model, and perform distance calculation in combination with the health status benchmark library to obtain the state deviation degree.

[0155] The trend analysis module is used to perform trend fitting analysis on the state deviation according to a preset time window to obtain the trend slope;

[0156] The level determination module is used to perform multi-dimensional feature fusion and level matching based on the state deviation and the trend slope to determine the health status level.

[0157] It should be noted that the energy storage device health status assessment system provided in the embodiments of the present invention is used to execute all the process steps of the energy storage device health status assessment method in the multi-mode grid connection scenario of the above embodiments. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0158] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a health status assessment program for energy storage devices in multi-mode grid connection scenarios. When the processor executes the computer program, it implements the steps in the embodiments of the above-described health status assessment methods for energy storage devices in various multi-mode grid connection scenarios, for example... Figure 1 The step S11 shown. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the above system embodiments, such as the data processing module.

[0159] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0160] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0161] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device via various interfaces and lines.

[0162] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0163] If the modules / units integrated into the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or system capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0164] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0165] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for evaluating the health state of energy storage equipment in a multi-mode grid-connected scenario, characterized in that, The method comprises the following steps: acquiring historical operation data, and performing time alignment and health screening processing on the historical operation data to obtain a historical health synchronization data set; performing working condition recognition and label adding processing on the historical health synchronization data set to obtain a scenario training data set; constructing an auto-encoder model according to the scenario training data set and performing iterative training to obtain a feature encoding model; performing feature space construction and clustering processing according to the feature encoding model and the scenario training data set to obtain a health state benchmark library; acquiring real-time operation parameters, performing feature mapping on the real-time operation parameters by using the feature encoding model, and performing distance calculation in combination with the health state benchmark library to obtain a state deviation degree; performing trend fitting analysis on the state deviation degree according to a preset time window to obtain a trend slope; performing multi-dimensional feature fusion and grade matching according to the state deviation degree and the trend slope to determine a health state grade; wherein the feature space construction and clustering processing according to the feature encoding model and the scenario training data set to obtain the health state benchmark library comprises: inputting the scenario training data set into the feature encoding model for dimension reduction mapping processing to obtain a historical low-dimensional feature vector; performing grouping processing on the historical low-dimensional feature vector to obtain a scenario vector cluster; performing core cluster group recognition processing on the scenario vector cluster to obtain a core cluster group; performing centroid calculation on the core cluster group to obtain an initial health state center vector; establishing a mapping relationship of the initial health state center vector to obtain the health state benchmark library; wherein the trend fitting analysis on the state deviation degree according to a preset time window to obtain a trend slope comprises: performing sequence interception processing on the state deviation degree to obtain an original deviation sequence; performing smoothing processing on the original deviation sequence to obtain a smoothed deviation numerical sequence; performing regression analysis on the smoothed deviation numerical sequence to obtain the trend slope; wherein the multi-dimensional feature fusion and grade matching according to the state deviation degree and the trend slope to determine a health state grade comprises: acquiring current environmental temperature data; performing frequency domain transformation processing on the original deviation sequence to obtain a fluctuation main frequency; performing amplitude statistical processing on the state deviation degree to obtain a deviation amplitude; performing vector combination processing on the fluctuation main frequency, the deviation amplitude, the trend slope and the current environmental temperature data to obtain a quantitative state feature vector; and performing matching processing on the quantitative state feature vector and a preset health grade division rule library to obtain the health state grade.

2. The method for health state estimation of energy storage device in multi-mode grid-connected scenario according to claim 1, characterized in that, The acquiring historical operation data, and performing time alignment and health screening processing on the historical operation data to obtain a historical health synchronization data set comprises: acquiring a battery management system data sequence, a storage energy converter data sequence and an energy management system instruction sequence, and integrating the battery management system data sequence, the storage energy converter data sequence and the energy management system instruction sequence to obtain the historical operation data; Resample and fill the historical operation data to obtain a multi-dimensional time series data set; Extract and calculate feature parameters from the multi-dimensional time series data set to obtain thermodynamic characteristic parameter values; If the thermodynamic characteristic parameter values exceed a preset safety threshold, perform abnormal data elimination processing to obtain the historical health synchronization data set.

3. The method of claim 1, wherein, The working condition recognition and label adding processing of the historical health synchronization data set obtains a scenario training data set, including: Perform electrical parameter analysis and filtering processing on the historical health synchronization data set to obtain a smooth electrical state data stream; Perform difference calculation processing on the smooth electrical state data stream to obtain a current change rate; Extract a running mode code from the historical health synchronization data set, and match the current change rate and the running mode code with a preset working condition recognition rule to obtain an initial scene label; Perform time sequence correction processing on the initial scene label to obtain the scenario training data set.

4. The method of claim 1, wherein, The autoencoder model is constructed and iteratively trained according to the scenario training data set to obtain a feature encoding model, including: Perform numerical normalization processing on the scenario training data set to generate a standardized high-dimensional feature matrix; Construct an autoencoder model structure, and input the standardized high-dimensional feature matrix into the autoencoder model structure for mapping and reconstruction processing to obtain a reconstructed high-dimensional feature matrix; Perform error calculation processing on the reconstructed high-dimensional feature matrix and the standardized high-dimensional feature matrix to obtain a difference loss value; If the difference loss value is lower than a preset convergence threshold, stop training and extract an encoding network to obtain the feature encoding model.

5. The method for health state estimation of energy storage device in multi-mode grid-connected scenario according to claim 1, characterized in that, The real-time operation parameters are obtained, and feature mapping and distance calculation are performed according to the feature encoding model and the health state benchmark library to obtain a state deviation, including: Obtain real-time operation parameters, and perform scene analysis according to a preset working condition recognition rule to obtain a current running scene identifier; Input the real-time operation parameters into the feature encoding model for feature extraction processing to obtain a current low-dimensional feature vector; According to the current running scene identifier, retrieve an ideal health state center vector from the health state benchmark library; Perform distance measurement calculation on the current low-dimensional feature vector and the ideal health state center vector to obtain the state deviation.

6. A health state evaluation system for energy storage devices in a multi-mode grid-connected scenario, characterized in that, The method comprises the following steps: A data processing module is configured to obtain historical operation data, and perform time alignment and health screening processing on the historical operation data to obtain a historical health synchronization data set; A working condition processing module is configured to perform working condition recognition and label adding processing on the historical health synchronization data set to obtain a scenario training data set; A model training module is configured to construct an autoencoder model and perform iterative training according to the scenario training data set to obtain a feature encoding model; A library construction module is configured to perform feature space construction and clustering processing according to the feature encoding model and the scenario training data set to obtain a health state benchmark library; The deviation calculation module is configured to obtain real-time operation parameters, perform feature mapping on the real-time operation parameters by using the feature coding model, and perform distance calculation in combination with the health state benchmark library to obtain a state deviation degree; The trend analysis module is configured to perform trend fitting analysis on the state deviation degree according to a preset time window to obtain a trend slope; The grade determination module is configured to perform multi-dimensional feature fusion and grade matching according to the state deviation degree and the trend slope to determine a health state grade.

Images