A kind of evaluation optimization method for energy storage battery safety early warning algorithm

Abstract

The present application relates to the technical field of energy storage battery safety management, and discloses a kind of evaluation optimization methods for energy storage battery safety early warning algorithm, comprising the following steps: S1, the operation monitoring data of energy storage power station is obtained;S2, generating a synthetic fault sample with risk evolution trend;S3, generating a multi-stage early warning label for each synthetic fault sample;S4, the real playback data and the synthetic fault sample are run on the safety early warning algorithm to be measured, and the alarm level and the predicted risk object output by the safety early warning algorithm to be measured are obtained;S5, the multi-dimensional evaluation index of the safety early warning algorithm to be measured is calculated;S6, a weak sample set is constructed;S7, the safety early warning algorithm to be measured is iteratively optimized, and a degradation prevention constraint is applied during optimization.The present application can accurately simulate the physical evolution process of battery failure, finely evaluate the performance of the early warning algorithm from multiple dimensions, and continuously optimize and iterate the algorithm.

Description

Technical Field

[0001] This invention relates to the field of energy storage battery safety management technology, specifically to an evaluation and optimization method for energy storage battery safety early warning algorithms. Background Technology

[0002] Energy storage power stations typically consist of a large number of lithium iron phosphate cells, modules, battery clusters, battery cabinets, or containers. They are characterized by a large number of cells, complex operating conditions, long lifespans, inconsistent sensor quality in the field, and a scarcity of real-world safety incident samples. Existing battery safety early warning methods mostly rely on fixed thresholds, single statistical anomaly detection, or single machine learning models. While fixed threshold-based methods are simple to implement, they struggle to adapt to differences in cell characteristics, aging states, and thermal management conditions, leading to numerous false alarms or missed alarms. Statistical anomaly detection methods can adapt to operating condition fluctuations to some extent, but they fail to adequately utilize the temporal trends and physical laws governing fault evolution. Machine learning-based methods perform well on specific datasets, but their generalization ability and interpretability remain challenges.

[0003] However, the above studies all focus on improving the early warning algorithms themselves. There is still a lack of systematic solutions in the existing technology for how to systematically evaluate the performance of these algorithms, how to conduct sufficient testing and verification under the condition of scarce real fault samples, and how to find the weaknesses of the algorithms based on the evaluation results and guide their iterative optimization.

[0004] Specifically, the evaluation and optimization of energy storage battery safety early warning algorithms face several technical bottlenecks: (1) There are very few actual thermal runaway, micro-internal short circuit, and overheating of connectors accident samples, making it difficult to fully test and verify the algorithm, and its performance boundaries and applicable scope are difficult to define accurately.

[0005] (2) Traditional binary classification labels can only mark “fault / normal” and cannot evaluate key capability dimensions such as lead time, level evolution and positioning accuracy required for safety early warning.

[0006] (3) Common field problems such as sensor drift, jamming, and communication abnormalities are easily misjudged as battery risks, resulting in unnecessary downtime or maintenance costs. However, the existing evaluation system lacks a mechanism to distinguish between sensor abnormalities and battery abnormalities.

[0007] (4) Model iteration usually relies on manual experience to select samples, making it difficult to perform targeted sampling and anti-degradation verification based on the weak scenarios of the algorithm.

[0008] Therefore, there is an urgent need for an evaluation and optimization method for energy storage battery safety early warning algorithms. Summary of the Invention

[0009] The present invention aims to provide an evaluation and optimization method for safety early warning algorithms for energy storage batteries, which can accurately simulate the physical evolution process of battery failure, evaluate the performance of early warning algorithms from multiple dimensions, and drive continuous optimization and iteration of the algorithms.

[0010] The basic solution provided by this invention is: an evaluation and optimization method for safety early warning algorithms of energy storage batteries, comprising the following steps: S1, acquire the operation monitoring data of the energy storage power station, the operation monitoring data including at least cell voltage, cell temperature, current, state of charge, and health status; S2, Based on the operation monitoring data, a fault evolution function that satisfies the physical constraints of the battery is superimposed on the real healthy operation segment to generate a synthetic fault sample with a risk evolution trend. S3, generate a multi-stage early warning label for each synthetic fault sample. The multi-stage early warning label includes at least the time when the abnormality begins, the time when a first-level early warning should be triggered, the time when a second-level early warning should be triggered, and the time when a protection or shutdown recommendation should be triggered. S4, run the safety early warning algorithm under test on real playback data and the synthetic fault sample to obtain the alarm level and predicted risk object output by the safety early warning algorithm under test; S5. Based on the multi-stage early warning label and the output of the safety early warning algorithm under test, calculate the multi-dimensional evaluation index of the safety early warning algorithm under test. The multi-dimensional evaluation index includes at least the lead time index and the positioning accuracy index. S6. Based on the evaluation results of the multi-dimensional evaluation indicators, identify the weak samples that cause the evaluation results to fail to meet the standards from the synthetic fault samples, and construct a set of weak samples. S7. Based on the weak sample set, the security early warning algorithm to be tested is iteratively optimized, and anti-degradation constraints are applied during the optimization process, so that the comprehensive score of the optimized algorithm on the fixed benchmark sample set is not lower than the comprehensive score of the algorithm before optimization on the fixed benchmark sample set minus the allowable degradation tolerance.

[0011] The working principle and advantages of this invention are as follows: This invention provides an evaluation and optimization method for safety early warning algorithms in energy storage batteries. This method accurately simulates the physical evolution of battery faults, provides a multi-dimensional and refined evaluation of the early warning algorithm's performance, and drives continuous optimization and iteration of the algorithm. The key points are: First, this invention directly targets the early warning algorithm, treating it as a testable and optimizable system. Unlike existing technologies that directly assess battery status, this invention accepts any early warning algorithm as input, outputs its multi-dimensional evaluation score, weak sample set, and specific optimization directions through a standardized process. It does not limit the specific implementation of the early warning algorithm, can be deployed independently of any specific early warning algorithm, and serves algorithm developers. It is compatible with cloud models, edge models, BMS embedded algorithms, and third-party algorithm evaluation platforms.

[0012] Second, this invention utilizes a mechanism-constrained fault injection technique to overlay a fault evolution function that satisfies battery physical constraints onto real healthy operating segments, actively generating synthetic fault samples with risk evolution trends. The evolution of fault signals is constrained by electrochemical mechanisms, while background noise and operating condition changes retain the characteristics of real data, avoiding the distribution offset problem of purely simulated data. The injection start time, evolution rate, and final severity can all be freely controlled, enabling a systematic coverage of the complete evolution cycle from early weak signs to severe thermal runaway, forming a verifiable, reproducible, and scalable risk test set in situations where real accident samples are scarce.

[0013] Third, this invention defines a multi-stage labeling system that includes the moment when an anomaly begins, the moment when a first-level warning should be triggered, the moment when a second-level warning should be triggered, and the moment when a protection or shutdown recommendation should be triggered. This provides a precise time reference for evaluation and can improve the evaluation of warning capabilities from "whether an alarm can be triggered" to a more refined level of "when to trigger an alarm, whether the level matches, and whether the location is accurate".

[0014] Fourth, the lead time metric reflects sensitivity in the time dimension, while the positioning accuracy metric reflects precision in the spatial dimension. Identifying weak samples based on the evaluation results of these two metrics and performing targeted iterative optimization enables the safety early warning algorithm to improve its sensitivity to early, subtle signs while retaining its original capabilities. This ultimately results in enhanced identification capabilities for early fault modes such as micro-internal short circuits, connector overheating, and cooling degradation. Combined with degradation prevention constraints, a continuous model iteration mechanism can be established to prevent the model from improving only on new samples while degrading in existing scenarios. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating an embodiment of the evaluation and optimization method for a safety early warning algorithm for energy storage batteries according to the present invention. Figure 2 This is a schematic diagram of the voltage and temperature evolution analysis results under a simulated micro-internal short circuit scenario, representing an embodiment of the evaluation and optimization method for a safety early warning algorithm for energy storage batteries according to the present invention. Figure 3This is a schematic diagram of the multi-stage label and risk score curve analysis results of an embodiment of the evaluation and optimization method for a safety early warning algorithm for energy storage batteries according to the present invention. Figure 4 This diagram illustrates the comparative analysis results of evaluation indicators for different safety early warning algorithms under test, based on an embodiment of the evaluation and optimization method for safety early warning algorithms of energy storage batteries according to the present invention. Detailed Implementation

[0016] The following detailed explanation illustrates the specific implementation methods: The basic implementation examples are as follows: Figure 1 As shown: A method for evaluating and optimizing a safety early warning algorithm for energy storage batteries, comprising the following steps: S1 acquires operational monitoring data of the energy storage power station and establishes a multi-scale topology relationship between individual units, modules, packs, battery clusters, battery cabinets, containers, and the power station.

[0017] Specifically, the energy management system (EMS) and battery management system (BMS) of the energy storage power station collect operational monitoring data (including historical operational data) of all battery cells. This operational monitoring data includes at least cell voltage, cell temperature, current, state of charge (SOC), and state of health (SOH). Preferably, the collected raw operational monitoring data undergoes preprocessing operations such as missing value handling, outlier removal, and time alignment.

[0018] S11 is based on cells with the same topology, operating conditions, and aging stage to build a similar cell group and form a dynamic benchmark.

[0019] Specifically, in this embodiment, the energy storage power station is assumed to include a number of battery cells, and the set of battery cells is represented as follows: ; in, It represents the collection of all the battery cells in an energy storage power station; Indicates the first One battery cell; Indicates the cell serial number; This indicates the total number of battery cells.

[0020] For battery cells At any moment The running feature vector is defined as: ; in, Indicates battery cell At any moment The running feature vector; Indicates the cell voltage; Indicates the cell temperature; Indicates the current of the battery cluster or battery cabinet; Indicates the state of charge of the battery cell; Indicates the health status of the battery cell; Indicates the equivalent internal resistance of the battery cell; Indicates ambient temperature; This represents a quantity indicating the cooling capacity of the area where the battery cell is located; superscript This represents the transpose of a vector.

[0021] Furthermore, a hierarchical topology for energy storage power stations is established. The topology hierarchy from bottom to top is: single cell - module - pack - battery cluster - battery cabinet - container - power station.

[0022] Based on topological location, operating conditions, and aging status, the battery cell is constructed. Similar cell groups: ; in, Indicates battery cell At any moment Similar cell groups; Indicates a candidate reference cell; Indicates the candidate cell serial number; Indicates battery cell and Distance in the unit-module-pack-cluster-cabinet topology.

[0023] Indicates the topological distance threshold; Indicates the state of charge of the candidate battery cell; Indicates the SOC difference threshold; Indicates the health status of the candidate battery cell; This represents the SOH difference threshold.

[0024] For any feature Its dynamic benchmark is defined as a robust statistic of similar cell groups: ; in, Indicates battery cell The corresponding dynamic baseline value; This indicates the feature to be analyzed, which can be voltage, temperature, internal resistance, or rate of temperature rise. Indicates cells within a similar cell group Similar characteristics; Represents the median operator; This indicates a group of similar battery cells.

[0025] Dynamic deviation is defined as: ; in, Indicates battery cell Dynamic deviation; This indicates the current characteristic value of the battery cell; This represents the dynamic reference value for a group of similar battery cells.

[0026] This method uses dynamic deviation instead of a fixed threshold, which can eliminate baseline differences between different power plants, different operating conditions and different aging stages.

[0027] S2, based on the operation monitoring data, a fault evolution function that satisfies the physical constraints of the battery is superimposed on the real healthy operation segment to generate a synthetic fault sample with a risk evolution trend.

[0028] Select real healthy operating segments from the historical operating data of the energy storage power station. The selection criteria are as follows: (1) There are no BMS alarms, protection actions, manual maintenance records or known abnormal events in the segment; (2) The cell voltage, temperature, current, SOC, SOH and communication status data are continuous, and the missing rate and abnormal jump rate are lower than the preset data quality threshold; (3) The segment covers reproducible working conditions such as charging, discharging or resting, and the current, SOC and ambient temperature are within the normal operating range of the corresponding working conditions; (4) The dynamic deviations of voltage, temperature, internal resistance and temperature rise rate of similar cell groups are all within the healthy threshold range (such as voltage deviation not exceeding 10mV, temperature deviation not exceeding 0.5℃), and there is no continuous individual deviation.

[0029] After selecting the healthy segment, record the dynamic reference value of each cell within the window preceding the injection start time (e.g., the first 300 seconds). , , , This serves as the normal baseline before fault injection. This baseline is the starting reference point for the subsequent fault evolution curve, ensuring that the fault signal originates from the actual operating condition level, rather than from an arbitrary set value.

[0030] This step includes at least one of the following execution methods: (1) Micro-internal short circuit injection: Micro-internal short circuit refers to a localized, tiny internal short circuit inside the cell caused by diaphragm defects, metal dendrite punctures, or impurity particles. Its typical characteristics are: the short circuit current is initially small, and it grows in an S-shape over time (slow in the early stage, accelerated in the middle stage, and saturated in the later stage). This invention uses the Sigmoid function to model the evolution process of the micro-internal short circuit current.

[0031] According to the formula Construct the equivalent micro-internal short-circuit current; in, Indicates battery cell The equivalent micro-internal short-circuit current; Indicates the initial short-circuit current; Indicates the magnitude of the short-circuit current increase; This represents the Sigmoid growth function; Indicates the current time; Indicates the start time of fault injection; This represents the short-circuit evolution time constant.

[0032] The equivalent micro-internal short-circuit current is then superimposed onto the cell's state-of-charge evolution process and heat generation power. Specifically: The internal short-circuit heating power is defined as: ; in, Indicates the internal short-circuit heating power; Represents the equivalent micro-internal short-circuit current; This represents the equivalent short-circuit path resistance. During fault injection, it will... This is superimposed on the thermal balance equation of the battery cell as an additional abnormal heating term.

[0033] The aforementioned micro-internal short-circuit heating power is used as an equivalent Joule heating term. The equivalent short-circuit channel resistance can be set as a constant or an equivalent parameter that changes slowly over time, depending on the severity of the fault. This limitation is used to maintain the physical coupling relationship between the micro-internal short-circuit current, SOC drop, voltage deviation, and temperature rise in the synthesized fault sample, and does not limit the specific geometry of the actual micro-short-circuit channel.

[0034] The SOC evolution correction caused by micro-internal short circuits is as follows: ; in, Indicates the state of charge at the next moment; Indicates the current state of charge; Indicates the main circuit current; Indicates the sampling time interval; Indicates the usable capacity of the battery cell; This represents the equivalent micro short-circuit current.

[0035] The corrected SOC is used to update the state of the synthetic fault sample at the next sampling time, and input the OCV-SOC mapping relationship (the corrected SOC is substituted into the OCV-SOC mapping table to obtain the open circuit voltage corresponding to the SOC, and then combined with the main circuit current and internal resistance to calculate the terminal voltage, thereby generating the synthetic evolution curve of voltage and temperature), voltage mechanism residual calculation and subsequent risk scoring process, so that the additional power loss caused by micro-internal short circuit can be reflected in voltage deviation, mechanism residual and multi-stage early warning label simultaneously.

[0036] like Figure 2As shown in the figure, a micro-internal short-circuit fault is injected at time 18h, with a certain healthy operating segment as the background. It can be seen from the figure that the voltage curve before injection basically coincides with the reference group benchmark. After injection, the voltage gradually deviates from the benchmark and shows a downward trend. The deviation of the temperature curve increases rapidly with the increase of short-circuit current, which fully presents the typical evolution characteristics of micro-internal short-circuit fault with electrical signal leader and thermal signal lag.

[0037] (2) Fault injection due to increased connection resistance: The contact resistance increment is superimposed on the equivalent internal resistance of the battery cell. And according to the formula The additional heating power is calculated; this additional heating power is added to the cell's thermal balance equation, causing the cell's temperature to be significantly higher than that of similar cells under high current conditions. This indicates the additional heat generated due to increased connection resistance; This indicates the main circuit current.

[0038] The equivalent resistance after the connection resistance is increased is defined as: ; in, This represents the equivalent resistance after an injection connection anomaly. This represents the equivalent internal resistance of a healthy state; This indicates the increment of the contact resistance of the connector.

[0039] (3) Cooling degradation fault injection: According to the formula This reduces the equivalent heat dissipation coefficient of the battery cell; among which, , Indicates the equivalent heat dissipation coefficient after cooling degradation; Indicates the cooling capacity attenuation coefficient; It represents the equivalent heat dissipation coefficient under healthy cooling conditions.

[0040] (4) Temperature sensor drift fault injection: According to the formula An offset drift is superimposed on the temperature measurement; where... This indicates the temperature sensor reading; This indicates the actual temperature of the battery cell; Indicates the amount of sensor bias drift; This indicates the sensor's measurement noise.

[0041] For each real healthy operating segment, one or more of the above-mentioned fault modes are independently injected to generate a synthetic fault sample with fault evolution. Each synthetic fault sample contains complete time-series data, including synthetic evolution curves of features such as voltage, temperature, current, state of charge, and internal resistance.

[0042] S21, Construct a sensor trustworthiness gating system to distinguish between real battery anomalies and sensor anomalies.

[0043] Based on voltage mechanism residual and temperature mechanism residual Calculation cell At any moment Sensor reliability ;in, Represents the natural exponential function; Indicates the temperature residual scaling factor; This represents the voltage residual scaling factor; If the temperature signal rises abnormally, but the voltage, SOC, current, and the direction of temperature propagation in adjacent cells do not conform to the thermal diffusion law (i.e., only the temperature of this cell is abnormal while the temperature of adjacent cells is normal), and the thermal balance residual does not support actual heating (i.e., the actual heating power has not increased, but the measured value is too high), then... If the signal decreases, the segment will be preferentially identified as a sensor malfunction rather than a battery malfunction.

[0044] The voltage mechanism residual is based on the formula Calculated; in, Indicates the measured voltage; This represents the estimated open-circuit voltage function; Indicates the estimated state of charge; Indicates current; This indicates an estimate of the internal resistance; The temperature mechanism residual is based on the formula Calculated; in, Indicates the residual of the temperature mechanism; This indicates the measured temperature at the next moment. This indicates the temperature predicted at the next moment based on the thermal balance model.

[0045] S3, generate a multi-stage warning label for each synthetic fault sample. The multi-stage warning label includes at least the time when the abnormality begins, the time when a first-level warning should be triggered, the time when a second-level warning should be triggered, and the time when a protection or shutdown recommendation should be triggered.

[0046] Specifically, it includes the following sub-steps: S3-1, for battery cells Constructing normalized anomaly feature vectors: ; in, Represents the normalized outlier feature vector; Indicates the degree of low voltage deviation; Indicates the degree of temperature deviation; Indicates the degree of high deviation in internal resistance; Indicates the degree of abnormality in the rate of temperature rise; Indicates the degree of abnormality in mechanism residuals; superscript This represents the transpose of a vector.

[0047] The degree of low voltage deviation can be defined as follows: ; in, Indicates the degree of low voltage deviation; This represents the function that takes the maximum value. Indicates a dynamic voltage reference for a group of similar battery cells; This indicates the actual measured voltage of the battery cell; This represents a robust discrete scale for the voltage of similar battery cell groups. It represents a small positive number that is prevented from being divided by zero.

[0048] The degree of temperature deviation can be defined as: ; The degree of high deviation of internal resistance can be defined as: ; The degree of abnormal temperature rise rate can be defined as: ; ; The degree of abnormality of mechanism residuals can be defined as: ; in, , and They represent battery cells respectively. At any moment Temperature, equivalent internal resistance, and rate of temperature rise; , and These represent dynamic benchmarks for the corresponding characteristics of similar cell groups; , and These represent robust discrete scales representing the corresponding features of similar battery cell groups; and These represent the voltage mechanism residual and the temperature mechanism residual, respectively. and These represent the voltage residual scaling factor and the temperature residual scaling factor, respectively. It represents a small positive number that is prevented from being divided by zero.

[0049] S3-2, Calculate the battery body risk score and sensor anomaly score: Battery risk score is defined as: ; in, Indicates battery cell Battery body risk score; Indicates the reliability of the sensor; Represents the Sigmoid function; Indicates the bias term; Indicates the number of abnormal features; Indicates the first The weights of each anomalous feature; Indicates the first A normalized anomaly feature.

[0050] Sensor anomaly score is defined as: ; in, Indicates battery cell Sensor anomaly score; Indicates the reliability of the sensor; Represents the Sigmoid function; This represents the bias term in the sensor anomaly model; Indicates the number of sensor residual features; Indicates the first Weights of sensor residual features; Indicates the first Individual sensor residual characteristics.

[0051] Battery risk rating The higher the score, the greater the risk of actual battery cell failure; sensor anomaly score The higher the value, the more likely the current anomaly is due to a sensor malfunction.

[0052] S3-3, Determine the alarm level.

[0053] according to The size determines the alarm level. .

[0054] The alarm level function is defined as follows: ; in, Indicates battery cell The alarm level; Indicates the risk score of the battery itself; This indicates the first-level warning threshold; Indicates the level 2 warning threshold; This indicates the recommended threshold for protection or shutdown. Alarm level 0 indicates no anomaly, level 1 indicates a Level 1 warning (attention), level 2 indicates a Level 2 warning (warning), and level 3 indicates a protection or shutdown recommendation (critical). Each threshold can be adjusted according to the actual application scenario and operation and maintenance strategy.

[0055] In this embodiment, when the battery body risk score is normalized to the range of 0 to 1, the recommended initial value is [value to be filled in]. =0.30、 =0.60、 =0.85, and satisfies 0 < < < <1; In actual deployment, the above thresholds can be calibrated based on the cost of missed alarms, the tolerance for false alarms, and the protection strategy.

[0056] S3-4, generate multi-stage warning labels.

[0057] For each risk event The multi-stage warning label is defined as follows: ; in, Indicates the first One risk event; Indicates the event sequence number; The object of risk can be a battery cell, module, battery cluster, battery cabinet, or container; Indicates the type of risk; Indicates the moment when an abnormality begins to emerge; This indicates the time when a Level 1 warning should be triggered; Indicates the time when a Level II warning should be triggered; Indicates the time when protection or shutdown recommendations should be triggered; This indicates the recommended action.

[0058] Among the multi-stage early warning labels: The anomaly budding moment is the moment when the abnormal characteristics of a risk event first appear. The trigger time for a Level 1 warning is when the battery's risk score in the aforementioned risk event first reaches the first threshold. The moment; The trigger point for a Level 2 warning is when the battery's risk score first reaches the second threshold. At that moment, ; The recommended time to trigger protection or shutdown is when the battery's risk score first reaches the third threshold. At that moment, .

[0059] like Figure 3 As shown in the figure, for the same micro-internal short circuit event, the risk score exhibits a monotonically increasing trend over time, as indicated by the graph. Abnormal budding moment; The time when a Level 1 warning should be triggered; The time when a Level II warning should be triggered; When protection or shutdown recommendations should be triggered.

[0060] S4, run the safety warning algorithm under test on real playback data and the synthetic fault sample to obtain the alarm level and predicted risk object output by the safety warning algorithm under test.

[0061] In this embodiment, real playback data refers to independent time period data selected from the historical operation data of the energy storage power station collected by S1 that was not included in the selection of real healthy operation segments, including normal operating conditions and actual fault conditions (if any). Real playback data is used to evaluate the performance of the algorithm under test on actual historical data.

[0062] Synthetic fault samples are synthetic fault samples with multi-stage warning labels generated by S3, used to evaluate the sensitivity and early warning capability of the algorithm under test in controllable fault scenarios.

[0063] The two sets of data are organized in chronological order to form a unified test dataset, and each test data item is ensured to have a timestamp and cell identifier.

[0064] Furthermore, the security early warning algorithm under test is deployed on a test computing platform, using the test dataset as input, and run frame by frame at the original sampling frequency (1Hz in this embodiment). The security early warning algorithm under test can be: A method based on a fixed threshold; Methods based on statistical anomaly detection (such as box plots, etc.) Guidelines); Based on traditional machine learning methods (such as Isolation Forest, Support Vector Machine, Logistic Regression); Deep learning-based methods (such as LSTM autoencoders and Transformers). Physical model-based methods (such as electrochemical models, thermal models, and residual monitoring); Big data early warning algorithms deployed in the cloud, lightweight algorithms deployed at the edge, or real-time algorithms embedded in the BMS.

[0065] All algorithms under test were run with the same data input in the same hardware and software environment to ensure comparability of the evaluation. During the execution of each algorithm, the following output information was recorded: Battery risk score (if the safety warning algorithm under test outputs a risk probability or anomaly score, it is uniformly mapped to the range of 0 to 1). Alarm level (if the security warning algorithm under test outputs multiple alarm levels, record them directly; if it only outputs binary classification, map it to two levels: 0 and 1). Predicting risk objects (the cell number or equipment unit that the safety warning algorithm considers to be most likely to be at risk). Explanatory factors (risk attribution information output by the safety early warning algorithm under test, such as "abnormal voltage" or "excessive temperature rise").

[0066] S5. Based on the multi-stage early warning label and the output of the safety early warning algorithm under test, calculate the multi-dimensional evaluation index of the safety early warning algorithm under test. The multi-dimensional evaluation index includes at least the lead time index and the positioning accuracy index.

[0067] For each risk event in the test dataset (Its actual attributes come from the multi-stage warning tags generated by S4), and the behavior of the algorithm under test is statistically analyzed before and after the event occurs. The event detection window is defined as... ; This is the observation window before the event (300 seconds in this example). This is the post-event observation window (600s in this example).

[0068] If the algorithm under test is At least one time point within the interval must output alarm level If the event is detected, it is considered to have been detected; otherwise, it is recorded as a missed event.

[0069] If the algorithm under test detects non-event objects (i.e., not...) within the detection window... (Battery cell) output alarm level If the alarm cannot be attributed to any known risk event, it is recorded as a false alarm.

[0070] The calculation method for the lead time index is as follows: For the m-th risk event The q-th level warning, the moment when the security warning algorithm under test first reaches the q-th level warning is defined as... ; in, This indicates taking the minimum value; Indicates the candidate time; This indicates that the security early warning algorithm under test is effective in predicting risk objects. At any moment The alarm level output; Regarding the first The lead time for a Level 1 warning is defined as: ; in, Indicates the first The risk event was in the [number]th [year]. Lead time for level-one early warning; The time when the q-th level warning should be reached in the multi-stage warning label; if This indicates an early warning; if This indicates a delayed report. If... , indicating exactly at the same time.

[0071] The location accuracy index is calculated as follows: According to the formula Calculate the positioning score; in, Indicates a risk event Positioning score; Represents the natural exponential function; Indicates the true risk object The predicted risk object of the security early warning algorithm under test Topological distance; This represents the positioning distance attenuation coefficient.

[0072] when hour, , This indicates that the positioning is completely correct. When and The greater the topological distance, The closer it is to 0.

[0073] The multidimensional evaluation indicators also include at least one of the following: Grade consistency score, according to the formula Calculated; in, Indicates the score for consistency in grade levels; This indicates that the security early warning algorithm under test is at the tag time. The alarm level output for the predicted risk object; Indicates the label level; This indicates the highest alarm level.

[0074] The level consistency score measures the consistency between the alarm level output by the security alert algorithm under test at the tag time and the tag level. When the level output by the security alert algorithm under test is consistent with the tag level at the tag time, The greater the difference, the lower the score.

[0075] Stability score, based on the formula Calculated; in, Indicates the stability score; Indicates the number of sampling points within the evaluation window; Indicates the sampling time index; Indicates the sampling interval; This indicates that the security early warning algorithm under test is at time [time missing]. The alarm level output; This indicates the alarm level output by the security warning algorithm under test at the previous sampling time; This indicates the highest alarm level.

[0076] The stability score measures the degree of fluctuation in the alarm level of the security warning algorithm under test over time. The smaller the cumulative amplitude of the alarm level jump between adjacent sampling points, the higher the stability.

[0077] The interpretability score is determined by experts based on the consistency between the explanatory factors output by the safety warning algorithm under test and the actual causes of the risk. In this embodiment, at least three experts in the field of energy storage safety independently score the interpretability of each event, and the average score is taken as the interpretability score for that event. .For example, The explanatory factors accurately identify the true type and cause of risk; This indicates that the explanatory factors have partial correlation but are not complete; This indicates that the algorithm does not provide any explanatory information.

[0078] Overall score, based on the formula Calculated; in, This indicates the overall score of the algorithm; Indicates the score for advance; This indicates the accuracy score for positioning; Indicates the score for consistency in grade levels; Indicates the stability score; Indicates the interpretability score; Indicates the number of underreported cases or the underreporting rate; Indicates the number of false alarms or the false alarm rate; Indicates the number of late reports or the late reporting rate; Indicates the weight of lead time; Indicates the positioning weight; Indicates the weight of level consistency; Indicates stability weights; Indicates interpretability weight; Indicates the penalty weight for underreporting; Indicates the weight of the false alarm penalty; This indicates the penalty weight for late reporting.

[0079] S6. Based on the evaluation results of the multi-dimensional evaluation indicators, identify the weak samples that cause the evaluation results to fail to meet the standards from the synthetic fault samples, and construct a set of weak samples.

[0080] The set of weak samples is defined as follows: ; in, This represents the set of weak samples.

[0081] This represents the set of missed detection samples; including the security early warning algorithm under test within the complete event window. None of the samples generated any alarms. The missed alarms indicate that the security warning algorithm under test is completely insensitive to this type of fault, which is the most serious type of weakness.

[0082] This represents the set of false alarm samples; it includes samples where the security alert algorithm under test outputs alarms for non-event objects or during non-event time periods. False alarm samples indicate that the security alert algorithm under test is overly sensitive to normal fluctuations or has feature obfuscation issues.

[0083] This represents the set of delayed reporting samples; including the moment when the security early warning algorithm under test detected the event but issued its first alarm. The samples that failed to issue alarms before the set time for the Level 1 warning label indicate that the tested security warning algorithm is insufficient in recognizing early, subtle signs.

[0084] This represents the set of samples with location errors; it includes objects whose events were detected but which were predicted as risky by the security early warning algorithm under test. The samples with positioning errors indicate that the security early warning algorithm under test has insufficient spatial resolution capability for risk sources.

[0085] This represents a set of samples with missing or incorrect explanations; including samples where the safety warning algorithm under test detects an event but does not provide an explanation factor, or where the explanation factor does not match the actual fault type (e.g., interpretability score less than 0.4).

[0086] It represents the union of sets.

[0087] Preferably, automatic attribution analysis is performed on each of the above-mentioned weak sample categories: (1) Fault type attribution: Statistically analyze the distribution ratio of various fault types (micro-internal short circuit, increased connection resistance, cooling degradation, sensor drift) in weak samples to identify which fault type the algorithm under test is least sensitive to.

[0088] (2) Operating conditions attribution: statistical analysis of the operating conditions (charging, discharging, static, different ambient temperature ranges) of weak samples to identify under which operating conditions the performance of the algorithm under test degrades most significantly.

[0089] (3) Feature attribution: Analyze the value distribution of each normalized abnormal feature in the weak sample to identify which abnormal feature the algorithm under test is not responding well to.

[0090] (4) Time attribution: Statistically analyze the distribution of weak samples in the failure evolution stage (emergence stage, development stage, maturity stage) to identify the stage in which the algorithm under test is most likely to fail.

[0091] S7. Based on the weak sample set, iteratively optimize the security early warning algorithm to be tested, and apply anti-degradation constraints during the optimization process, so that the comprehensive score of the optimized algorithm on the fixed benchmark sample set is not lower than the comprehensive score of the algorithm before optimization on the fixed benchmark sample set minus the allowable degradation tolerance. Specifically, this includes the following sub-steps: S7-1, Review and Enhancement of Weak Samples.

[0092] The set of weak samples mined by S7 Submitted to expert review. Experts review each weak sample to confirm whether it truly falls under the category of algorithm weaknesses, ruling out false alarms caused by labeling errors or data quality issues. After passing the review, the weak samples undergo enhancement processing: (1) Mechanism enhancement: For weak samples confirmed by experts, the mechanism-constrained fault injection model based on S3 generates more similar synthetic samples with the same fault type and parameters, thereby expanding the sample size in this scenario.

[0093] (2) Weak labeling: For samples whose fault parameters cannot be accurately labeled, a semi-automatic labeling method is used to label their fault type and approximate severity, and they are added to the training set in a weakly supervised manner.

[0094] The weak samples, after review and enhancement, are merged with the original sample pool to form an updated training sample set.

[0095] S7-2, Model Retraining and Calibration.

[0096] Using the original model parameters of the security early warning algorithm under test Use the initial values ​​to retrain or calibrate the model on the updated training sample set.

[0097] Model parameter updates can be represented as minimizing the following objective function: ; in, Indicates the first Model parameters after one iteration; This represents the parameter values ​​that minimize the objective function; Indicates the parameters of the model to be optimized; The model represents a fixed benchmark sample set. The loss function on.

[0098] This refers to a fixed benchmark sample set (a standardized set of test samples selected from historical operating data that does not change with each iteration). Indicates the loss weights for weak samples; The model represents the first time. Wheel weak sample set loss function on; This represents the penalty coefficient for parameter changes; Represents the L2 norm; Indicates the first Wheel model parameters.

[0099] S7-3, Verification of anti-degradation constraints.

[0100] After the model is updated, the new model must be validated. In a fixed benchmark sample set The overall score was not significantly lower than that of the old model. .

[0101] The degradation prevention constraint is defined as follows: ; in, Indicates the first The overall score of the +1 round model on a fixed benchmark sample set; This represents the overall score of the model in the r-th round on a fixed benchmark sample set; Indicates the parameters of the new model; Indicates the parameters of the old model; This indicates the maximum allowable tolerance for rating degradation.

[0102] If the above constraints are met, the new model passes the anti-degradation verification and can proceed to the deployment phase. If the constraints are not met, the loss weights of weak samples are reduced or the regularization coefficient is increased, and the model is retrained until the anti-degradation constraints are met or the maximum number of iterations is reached.

[0103] S7-4, multi-test set comprehensive verification.

[0104] After applying anti-degradation constraints, the new model needs to be comprehensively validated on the following three test sets: (1) Fixed benchmark set The new model's performance was verified to have not significantly degraded in the original scenario.

[0105] (2) Fault injection test set: Another batch of synthetic fault samples generated independently of S3 to verify the improvement effect of the new model in various fault scenarios.

[0106] (3) Real replay test set: Real historical running data independent of the training data to verify the generalization ability of the new model under actual working conditions.

[0107] Only when the new model's overall score on all three test sets is not lower than that of the old model (allowed) The new model was only approved for field shadow run verification in S8 when it met the degradation tolerance requirement.

[0108] S8, in the field of energy storage power stations, the new model is verified online using a shadow operation mode (shadow operation refers to the model receiving real-time data streams and outputting early warning results in the actual operating environment, but the output results do not directly trigger protection actions or alarm notifications, and are only used for recording and subsequent analysis), including: The new model and the existing early warning strategy (current model) are connected in parallel to the same real-time data stream.

[0109] Record the risk score, alarm level, predicted risk object, and explanatory factor output by the new model without directly triggering protection actions.

[0110] The output of the new model is compared with the output of the existing strategy, the conclusions of manual review, and the results of subsequent actual operation.

[0111] Counterfactual replay of false alarms, missed alarms, delayed alarms, and location deviation segments is performed, and input disturbances, thresholds, and fault injection parameters are adjusted to test boundary robustness. Once the preset false alarm rate, missed alarm rate, lead time, and stability requirements are met during the shadow operation period, the system will proceed to gray-scale deployment or formal deployment.

[0112] like Figure 4 As shown, the comprehensive score distribution of the early warning algorithm (i.e., the optimized algorithm instance / new model), the fixed threshold early warning algorithm, and the statistical anomaly early warning algorithm after running on the same test dataset is illustrated. The performance differences between the different algorithms are clearly visible. Operations personnel or algorithm developers can use this to identify which algorithms perform well in which dimensions and which dimensions have weaknesses, thus enabling them to selectively choose the most suitable early warning algorithm for specific application scenarios or to perform targeted optimization of existing algorithms. Furthermore, the algorithm optimized by this method exhibits balanced and excellent performance in multiple aspects such as positioning accuracy, stability, and interpretability, achieving the highest comprehensive score.

[0113] This embodiment provides an evaluation and optimization method for safety early warning algorithms for energy storage batteries. This method can accurately simulate the physical evolution of battery faults, finely evaluate the performance of early warning algorithms from multiple dimensions, and drive continuous optimization and iteration of the algorithm. In particular: This solution overlays a fault evolution function with mechanistic constraints onto real-world healthy operating segments, proactively generating a verifiable, reproducible, and scalable risk test set covering diverse scenarios such as micro-internal short circuits, connector overheating, and cooling degradation, thus eliminating reliance on scarce real-world accident samples. Based on multi-stage early warning labels and dynamic benchmark bias, it can finely evaluate the early warning algorithm's ability to identify and locate faults throughout the entire fault evolution process. Sensor reliability gating effectively suppresses false alarms caused by sensor drift, sudden changes in operating conditions, and individual cell differences. The evaluation results automatically drive weak sample mining and targeted iterative optimization, coupled with anti-degradation constraints to ensure that the model improves on new samples without performance degradation in existing scenarios, thereby establishing a sustainable and verifiable algorithm evolution mechanism. This framework is decoupled from the specific algorithm implementation, allowing flexible adaptation to cloud models, edge models, BMS embedded algorithms, and third-party evaluation platforms. It outputs the located object, risk level, explanatory factors, and recommended actions, forming a complete closed loop from evaluation and diagnosis to operation and maintenance guidance.

[0114] The above descriptions are merely embodiments of the present invention. Commonly known structures and characteristics of the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent.

Claims

1. A method for evaluating and optimizing safety early warning algorithms for energy storage batteries, characterized in that, Includes the following steps: S1, acquire the operation monitoring data of the energy storage power station, the operation monitoring data including at least cell voltage, cell temperature, current, state of charge, and health status; S2, Based on the operation monitoring data, a fault evolution function that satisfies the physical constraints of the battery is superimposed on the real healthy operation segment to generate a synthetic fault sample with a risk evolution trend. S3, generate a multi-stage early warning label for each synthetic fault sample. The multi-stage early warning label includes at least the time when the abnormality begins, the time when a first-level early warning should be triggered, the time when a second-level early warning should be triggered, and the time when a protection or shutdown recommendation should be triggered. S4, run the safety early warning algorithm under test on real playback data and the synthetic fault sample to obtain the alarm level and predicted risk object output by the safety early warning algorithm under test; S5. Based on the multi-stage early warning label and the output of the safety early warning algorithm under test, calculate the multi-dimensional evaluation index of the safety early warning algorithm under test. The multi-dimensional evaluation index includes at least the lead time index and the positioning accuracy index. S6. Based on the evaluation results of the multi-dimensional evaluation indicators, identify the weak samples that cause the evaluation results to fail to meet the standards from the synthetic fault samples, and construct a set of weak samples. S7. Based on the weak sample set, the security early warning algorithm to be tested is iteratively optimized, and anti-degradation constraints are applied during the optimization process, so that the comprehensive score of the optimized algorithm on the fixed benchmark sample set is not lower than the comprehensive score of the algorithm before optimization on the fixed benchmark sample set minus the allowable degradation tolerance.

2. The evaluation and optimization method for a safety early warning algorithm for energy storage batteries according to claim 1, characterized in that, S2 includes at least one of the following execution methods: Micro-internal short-circuit fault injection: According to the formula Construct the equivalent micro-internal short-circuit current; in, Indicates battery cell The equivalent micro-internal short-circuit current; Indicates the initial short-circuit current; Indicates the magnitude of the short-circuit current increase; This represents the Sigmoid growth function; Indicates the current moment; Indicates the start time of fault injection; The short-circuit evolution time constant is represented; and the equivalent micro-internal short-circuit current is superimposed on the cell's state of charge evolution process and heating power; Increased connection resistance during fault injection: The increase in contact resistance is superimposed on the equivalent internal resistance of the battery cell. And according to the formula Calculate the additional heating power; where, This indicates the additional heat generated due to increased connection resistance; Indicates the main circuit current; Cooling degradation fault injection: according to the formula This reduces the equivalent heat dissipation coefficient of the battery cell; among which, , Indicates the equivalent heat dissipation coefficient after cooling degradation; Indicates the cooling capacity attenuation coefficient; It represents the equivalent heat dissipation coefficient under healthy cooling conditions; Temperature sensor drift fault injection: According to the formula An offset drift is superimposed on the temperature measurement; where... This indicates the temperature sensor measurement value; This indicates the actual temperature of the battery cell; Indicates the amount of sensor bias drift; This indicates the sensor's measurement noise.

3. The evaluation and optimization method for a safety early warning algorithm for energy storage batteries according to claim 1, characterized in that, Among the multi-stage early warning labels: The anomaly budding moment is the moment when the abnormal characteristics of a risk event first appear. The trigger time for a Level 1 warning is when the battery's risk score in the aforementioned risk event first reaches the first threshold. The moment; The trigger point for a Level 2 warning is when the battery's risk score first reaches the second threshold. At that moment, ; The recommended time to trigger protection or shutdown is when the battery's risk score first reaches the third threshold. At that moment, .

4. The evaluation and optimization method for a safety early warning algorithm for energy storage batteries according to claim 1, characterized in that, It also includes the step of building sensor trust gating: Based on voltage mechanism residual and temperature mechanism residual Calculation cell At any moment Sensor reliability ;in, Represents the natural exponential function; Indicates the temperature residual scaling factor; This represents the voltage residual scaling factor; If the temperature signal rises abnormally, but the voltage, SOC, current, the direction of temperature propagation of adjacent cells, and the thermal balance residual do not support actual heating, If the signal decreases, the segment will be preferentially identified as a sensor malfunction rather than a battery malfunction.

5. The evaluation and optimization method for a safety early warning algorithm for energy storage batteries according to claim 4, characterized in that, The voltage mechanism residual is based on the formula Calculated; in, Indicates the measured voltage; This represents the estimated open-circuit voltage function; Indicates the estimated state of charge; Indicates current; This indicates an estimate of the internal resistance; The temperature mechanism residual is based on the formula Calculated; in, Indicates the residual of the temperature mechanism; This indicates the measured temperature at the next moment. This indicates the temperature predicted at the next moment based on the thermal balance model.

6. The evaluation and optimization method for a safety early warning algorithm for energy storage batteries according to claim 1, characterized in that, The calculation method for the lead time index is as follows: For the q-th level warning of the m-th risk event, the moment when the security warning algorithm to be tested first reaches the q-th level warning is defined as follows: ; in, This indicates taking the minimum value; Indicates the candidate time; This indicates that the security early warning algorithm under test is effective in predicting risk objects. At any moment The alarm level output; Regarding the first The lead time for a Level 1 warning is defined as: ; in, Indicates the first The risk event was in the [number]th [year]. Lead time for level-one early warning; The time when the q-th level warning should be reached in the multi-stage warning label; if This indicates an early warning; if This indicates a delayed report.

7. The evaluation and optimization method for a safety early warning algorithm for energy storage batteries according to claim 1, characterized in that, The location accuracy index is calculated as follows: According to the formula Calculate the positioning score; in, Indicates a risk event Positioning score; Represents the natural exponential function; Indicates the true risk object The predicted risk object of the security early warning algorithm under test Topological distance; This represents the positioning distance attenuation coefficient.

8. The evaluation and optimization method for a safety early warning algorithm for energy storage batteries according to claim 1, characterized in that, The multidimensional evaluation indicators also include at least one of the following: Grade consistency score, according to the formula Calculated; in, Indicates the score for consistency in grade levels; This indicates that the security early warning algorithm under test is at the tag time. The alarm level output for the predicted risk object; Indicates the label level; Indicates the highest alarm level; Stability score, based on the formula Calculated; in, Indicates the stability score; This indicates the number of sampling points within the evaluation window; Indicates the sampling time index; Indicates the sampling interval; Indicates time The alarm level; Indicates the alarm level at the previous sampling time; Indicates the highest alarm level; Overall score, based on the formula Calculated; in, This indicates the overall score of the algorithm; This indicates the score for leading time; This indicates the accuracy score for positioning; Indicates the score for consistency in grade levels; Indicates the stability score; Indicates the interpretability score; Indicates the number of underreported cases or the underreporting rate; Indicates the number of false alarms or the false alarm rate; Indicates the number of late reports or the late reporting rate; Indicates the weight of lead time; Indicates the positioning weight; Indicates the weight of level consistency; Indicates stability weights; Indicates interpretability weight; Indicates the penalty weight for underreporting; Indicates the weight of the false alarm penalty; This indicates the penalty weight for late reporting.

9. The evaluation and optimization method for a safety early warning algorithm for energy storage batteries according to claim 1, characterized in that, The set of weak samples is defined as follows: ； in, Represents a set of weak samples; This represents the set of underreported samples; Represents the set of false alarm samples; This represents the set of late-reported samples; This represents the set of samples with location errors. This represents the set of samples with missing or incorrect explanations. It represents the union of sets.

10. The evaluation and optimization method for a safety early warning algorithm for energy storage batteries according to claim 1, characterized in that, In the iterative optimization, the model parameter update is represented as follows: ； in, Indicates the first Model parameters after one iteration; This represents the parameter values ​​that minimize the objective function; Indicates the parameters of the model to be optimized; The model represents a fixed benchmark sample set. loss function on; Represents a fixed reference sample set; Indicates the loss weights for weak samples; The model represents the first time. Wheel weak sample set loss function on; This represents the penalty coefficient for parameter changes; Represents the L2 norm; Indicates the first Wheel model parameters; The degradation prevention constraint is defined as follows: ; in, Indicates the first The overall score of the +1 round model on a fixed benchmark sample set; This represents the overall score of the model in the r-th round on a fixed benchmark sample set; Indicates the parameters of the new model; Indicates the parameters of the old model; This indicates the maximum allowable tolerance for rating degradation.

Images