A network-constructed energy storage black start control method and device

Abstract

The application discloses a network-constructing type energy storage black start control method and device, the method comprising: obtaining current state parameters of each battery cluster in an energy storage system and historical state data before power grid failure; inputting the current state parameters and the historical state data into a dynamic feature extraction network to generate dynamic feature vectors corresponding to each battery cluster; based on the dynamic feature vectors, calculating the similarity of each battery cluster to a preset ideal starting feature through a contrast learning model to generate a dynamic priority score of each battery cluster; adjusting a black start competition time threshold according to the dynamic priority score and the severity of the power grid failure; selecting a battery cluster with the highest dynamic priority score as a main starting unit through a distributed consensus algorithm to control the main starting unit to supply power to an energy storage converter to complete black start. The method effectively balances the speed and accuracy of black start and significantly improves the response capability of the energy storage system under different power grid conditions.

Description

Technical Field

[0001] This application relates to the field of energy storage system technology, and in particular to a black start control method and device for grid-type energy storage. Background Technology

[0002] In the field of energy storage system technology, black start refers to the process of starting up and restoring part or all of the power supply by using an energy storage system without relying on the support of the external power grid after the power grid fails. It is crucial for ensuring the reliability and stability of the power system.

[0003] Current black-start control technologies have many shortcomings. In terms of selecting the battery cluster starting unit, most rely on simple rules or experience-based judgments, such as determining the main starting unit based on a single indicator like the remaining capacity and initial voltage of the battery cluster. This makes it impossible to flexibly adjust according to actual grid fault conditions, and it is difficult to ensure the safe and stable operation of the energy storage system under various fault scenarios. Summary of the Invention

[0004] The main technical problem addressed in this application is to provide a black start control method and device for grid-connected energy storage, which effectively balances the speed and accuracy of black start and significantly improves the coping ability of energy storage systems under different grid operating conditions.

[0005] To address the aforementioned technical problems, this application provides a technical solution: a black-start control method for grid-connected energy storage. This method includes: acquiring the current state parameters of each battery cluster in the energy storage system and historical state data prior to a grid fault; inputting the current state parameters and historical state data into a dynamic feature extraction network to generate dynamic feature vectors corresponding to each battery cluster; calculating the similarity between each battery cluster and a preset ideal start-up feature using a comparative learning model based on the dynamic feature vectors to generate a dynamic priority score for each battery cluster; adjusting the black-start competition time threshold according to the dynamic priority score and the severity of the grid fault; wherein the black-start competition time threshold is the maximum time limit allowing each battery cluster to become the main start-up unit; selecting the battery cluster with the highest dynamic priority score as the main start-up unit using a distributed consensus algorithm, and controlling the main start-up unit to supply power to the energy storage converter to complete the black start.

[0006] The method further includes: pre-training a dynamic feature extraction network; pre-training the dynamic feature extraction network includes: acquiring training samples under historical black start scenarios; wherein, the training samples include state parameter samples of each battery cluster and corresponding historical state data samples before the power grid fault; normalizing the training samples and labeling them with ideal start feature labels; extracting multi-dimensional spatiotemporal features through a convolutional neural network, and optimizing the network parameters by combining a contrastive loss function, so that the feature vectors are clustered in the embedding space according to the type of power grid fault.

[0007] The contrast loss function employs an event-triggered dynamic temperature parameter adjustment mechanism, specifically including: reducing the temperature parameter in response to the detection of a grid fault of voltage collapse; and increasing the temperature parameter in response to the detection of a grid fault of frequency oscillation. The adjustment range of the temperature parameter is a preset threshold range.

[0008] The generation of dynamic priority scores for each battery cluster includes: constructing a composite weighting function based on the battery cluster's health score, response speed score, and load matching score; the composite weighting function and the health score are positively correlated with the aging degree of the battery cluster; the health score is calculated based on the capacity decay rate, cycle count, and internal resistance change rate; the response speed score is calculated based on the deviation of the voltage ramp-up rate from the historical benchmark value; and the load matching score is generated based on the matching degree between the load demand curve recorded before the grid fault and the output characteristics of the battery cluster.

[0009] Among them, adjusting the black start competition time threshold according to dynamic priority scoring and grid fault severity includes: establishing a grid fault severity classification model, which classifies fault levels based on the product of voltage drop amplitude and rate, frequency deviation duration and load change gradient; and adjusting the strategy according to the preset time threshold mapped by the fault level.

[0010] The method further includes dividing the energy storage system into several logical super-units according to electrical topology and communication links, and exchanging feature vectors among the battery clusters in each logical super-unit through a shared bus.

[0011] The process involves selecting the battery cluster with the highest dynamic priority score as the main startup unit using a distributed consensus algorithm, and controlling the main startup unit to supply power to the energy storage converter to complete the black start. This includes performing the following operations on each logical superunit: extracting the dynamic priority score of each battery cluster within the logical superunit; generating a superunit-level startup priority using a weighted voting mechanism, where the weight of the weighted voting is calculated based on the electromagnetic cluster capacity ratio plus a communication quality correction factor; and prioritizing the startup of the logical superunit with the highest startup priority.

[0012] To address the aforementioned technical problems, another technical solution adopted in this application is: providing a grid-based energy storage black-start control device, which includes: a data acquisition module, a feature extraction module, a scoring module, an adjustment module, and a control module. Specifically, the data acquisition module acquires the current state parameters of each battery cluster in the energy storage system and historical state data prior to grid faults; the feature extraction module inputs the current state parameters and historical state data into a dynamic feature extraction network to generate dynamic feature vectors corresponding to each battery cluster; the scoring module calculates the similarity between each battery cluster and a preset ideal start-up feature using a comparative learning model based on the dynamic feature vectors, generating a dynamic priority score for each battery cluster; the adjustment module adjusts the black-start competition time threshold according to the dynamic priority score and the severity of the grid fault; wherein, the black-start competition time threshold is the maximum time limit allowing each battery cluster to become the main start-up unit; and the control module selects the battery cluster with the highest dynamic priority score as the main start-up unit using a distributed consensus algorithm, and controls the main start-up unit to supply power to the energy storage converter to complete the black start.

[0013] Unlike existing technologies, the advantages of this application are as follows: By acquiring the current state parameters of each battery cluster in the energy storage system and historical state data before grid faults, and inputting these into a dynamic feature extraction network to generate dynamic feature vectors, the real-time state and potential startup capability of the battery clusters are comprehensively and accurately reflected. Based on the dynamic feature vectors, a contrastive learning model is used to calculate the similarity between each battery cluster and the preset ideal startup features, generating a dynamic priority score. The black-start competition time threshold is adjusted according to the dynamic priority score and the severity of the grid fault, fully considering the actual situation of the grid fault and the startup suitability of each battery cluster. The distributed consensus algorithm ensures an effective balance between the speed and accuracy of black-start in complex situations where multiple battery clusters participate in the competition, significantly improving the energy storage system's ability to cope with different grid operating conditions. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating one embodiment of the black-start control method for grid-type energy storage in this application.

[0015] Figure 2 This is a schematic diagram of the structural framework of one embodiment of the grid-type energy storage black start control device of this application. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions 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, 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.

[0017] The terms "comprising" and "having," and any variations thereof, used in the embodiments of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the steps or units listed, but may optionally include other steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0018] To facilitate understanding of this embodiment, a detailed description of a grid-type energy storage black-start control method disclosed in this embodiment of the invention will be provided first. For example... Figure 1 As shown, this method includes the following steps.

[0019] Step S11: Obtain the current state parameters of each battery cluster in the energy storage system and the historical state data before the grid fault.

[0020] Among them, the current state parameters of the battery cluster are used to reflect the current working status of the battery cluster, such as the battery voltage, current, temperature, etc.; the historical state data before the grid fault can include the grid operation data and relevant grid parameters during the period before the grid fault.

[0021] In some application scenarios, the state parameters of the battery cluster can be collected in real time by the battery management system, including voltage, current, temperature, state of charge, and internal resistance, with a sampling frequency of 10Hz. Historical grid data can be captured by a snapshot taken by a Supervisory Control and Data Acquisition (SCADA) system at the last moment before a fault, recording grid frequency, voltage amplitude, load curves, etc., and stored in non-volatile memory.

[0022] In other application scenarios, the state parameters of the battery cluster can be indirectly calculated through electrochemical impedance spectroscopy. Historical state data before a grid fault can be captured by a synchronous phasor measurement unit, which captures high-frequency sampling data at the moment before the fault. The moment before the fault can be set as needed, for example, 10ms.

[0023] In some embodiments, after obtaining the current state parameters of the battery cluster and the historical state data before the grid fault, the obtained data can be preprocessed. For example, the state parameters of the battery cluster can be subjected to sliding window parallel filtering to eliminate instantaneous noise. The historical state data before the grid fault is stored using a lossless compression algorithm.

[0024] Furthermore, historical state data prior to grid failures can be distributed and backed up to adjacent management units that are electrically isolated from this energy storage system.

[0025] In some specific embodiments, the current state parameters of the acquired battery cluster may include: voltage: 720V±5V; temperature: 35℃; state of charge: 65%; and historical data before the grid fault may include the instantaneous frequency before the fault: 49.5Hz; and load demand: 500kW. It is understood that the data acquisition frequency can be adjusted as needed according to the type of electromagnetic cluster in the energy storage system.

[0026] Step S12: Input the current state parameters and historical state data into the dynamic feature extraction network to generate dynamic feature vectors corresponding to each battery cluster.

[0027] Specifically, the dynamic feature extraction network aims to extract key features from a large amount of input data that can effectively characterize the state of the battery cluster. These features are combined to form a dynamic feature vector, which is used to characterize the startup capability of the battery cluster.

[0028] In some embodiments, the dynamic feature extraction network can be a convolutional neural network. In addition to using convolutional neural networks for feature extraction, recurrent neural networks and their variants, such as long short-term memory networks and gated recurrent units, can also be used. These networks have unique advantages in processing time series data and can better capture the dynamic changes in the data. They can be selected as needed.

[0029] In some applications, the dynamic feature extraction network employs a convolutional neural network. Its input layer contains 20-dimensional data, consisting of voltage, temperature, and internal resistance data collected over 10 seconds at a frequency of 2Hz, along with grid frequency and load data. This data encompasses key information about the battery cluster and the power grid. The hidden layers employ a three-layer fully connected structure with 256, 128, and 64 nodes respectively, using the Rectified Linear Unit (ReLU) activation function. The Rectified Linear Unit effectively addresses the vanishing gradient problem, improving the network's training efficiency. The output layer outputs a 32-dimensional feature vector, which is used for subsequent comparative learning to evaluate the similarity between the battery cluster and the ideal startup features.

[0030] In some specific embodiments, the battery cluster timing data has a window set to 10 seconds, with the voltage gradually changing from 720V to 715V within 10 seconds, the temperature fluctuating between 34℃ and 36℃, and the internal resistance gradually increasing from 0.12Ω to 0.15Ω. In the power grid data, the frequency is 49.5Hz, and the load abrupt change gradient is 200kW / s. After processing by the dynamic feature extraction network, the output feature vector is, for example, [0.87, -0.23, 0.45, ..., 0.12], 32-dimensional, ranging from [-1, 1]. In practical applications, the feature dimensions can be dynamically trimmed according to the embedded system's resources, such as memory size and processor performance, for example, adjusted to 16 or 64 dimensions to balance computational load and feature representation capability.

[0031] In some embodiments, the method further includes: pre-training a dynamic feature extraction network; pre-training the dynamic feature extraction network includes: acquiring training samples under historical black start scenarios; wherein, the training samples include state parameter samples of each battery cluster and corresponding historical state data samples before the power grid fault; normalizing the training samples and labeling them with ideal start feature labels; extracting multidimensional spatiotemporal features through a convolutional neural network, and optimizing the network parameters by combining a contrastive loss function, so that the feature vectors are clustered in the embedding space according to the type of power grid fault.

[0032] Specifically, the training of the dynamic feature extraction network aims to accurately extract multi-dimensional spatiotemporal features from battery cluster state parameters and historical power grid data. By optimizing network parameters through a contrastive learning framework, the network can effectively capture key information in the data, enabling feature vectors of similar fault scenarios to aggregate in the embedding space, thus providing strong support for subsequent calculation of dynamic priority scoring of battery clusters.

[0033] In historical black-start scenarios, the dimensions of parameters such as voltage, temperature, and internal resistance of battery clusters vary. Normalizing these parameters can unify them to the same scale, eliminating the impact of dimensional differences and optimizing network training performance. Labeling can be based on actual successful start-up cases, labeling them with "ideal start-up features." For example, scenarios with voltage stability greater than 95% and response latency less than 200ms at start-up can be marked as positive samples. This gives the network a clear learning objective during training, knowing which feature combinations are more conducive to successful battery cluster start-up. For example, a hybrid structure of 3 convolutional layers and 2 Long Short-Term Memory (LSTM) networks can be used. Convolutional layers can extract spatial features of the data, such as the distribution of voltage differences between battery clusters, revealing spatial patterns of different battery clusters. LSTM excels at capturing time-series characteristics, such as the trend of voltage ramp-up rate over time, remembering long-term dependencies, thus better handling data with time-series characteristics. In some embodiments, the comparison loss function employs an event-triggered dynamic temperature parameter adjustment mechanism; specifically, it includes: reducing the temperature parameter in response to detecting that the type of grid fault is voltage collapse; and increasing the temperature parameter in response to detecting that the type of grid fault is frequency oscillation; wherein the adjustment range of the temperature parameter is a preset threshold range.

[0034] Specifically, voltage collapse faults in the power grid can be detected through real-time sampling of the grid voltage, calculation of the effective value, determination of the rate of voltage drop, and threshold comparison. Real-time voltage sampling calculates the real-time effective value of each phase voltage and the rate of voltage change. If this rate of change is less than a preset value for a preset number of consecutive sampling periods, it indicates that the voltage drop is too rapid, triggering a voltage collapse warning and indicating a potential risk of voltage collapse. Threshold comparison: The calculated effective voltage value is compared with the nominal voltage. If the effective voltage value is less than a preset multiple of the nominal voltage, and this low voltage state lasts for more than a preset time period, a voltage collapse fault is determined to have occurred. The detection of grid fault types can also be set using existing detection methods and are not limited here.

[0035] Based on the detected grid fault type, the temperature parameter in the contrast loss function is dynamically adjusted to change the degree of clustering of feature vectors in the embedding space, thereby improving the model's ability to distinguish features of different fault types.

[0036] Step S13: Based on the dynamic feature vector, calculate the similarity between each battery cluster and the preset ideal start-up features through a contrastive learning model, and generate a dynamic priority score for each battery cluster.

[0037] Specifically, the contrastive learning model is used to measure the similarity between the dynamic feature vector and the preset ideal start-up features. The higher the similarity, the closer the battery cluster is to the ideal start-up conditions, thus obtaining a higher dynamic priority score.

[0038] In some embodiments, a contrastive learning model can be used to calculate the cosine similarity between the feature vector and the ideal startup feature, with the similarity ranging from [0,1]. A similarity greater than a first threshold is considered high priority, and a similarity less than a second threshold is considered low priority, thereby quickly determining the startup priority of the battery cluster. The second threshold is less than the first threshold; the first threshold can be 0.7, and the second threshold can be 0.3.

[0039] In some embodiments, generating a dynamic priority score for each battery cluster includes: constructing a composite weighting function based on the battery cluster's health score, response speed score, and load matching score; the composite weighting function and the health score are positively correlated with the aging degree of the battery cluster; wherein, the health score is calculated based on the capacity decay rate, cycle count, and internal resistance change rate; the response speed score is calculated based on the deviation of the voltage ramp-up rate from the historical benchmark value; and the load matching score is generated based on the matching degree between the load demand curve recorded before the grid fault and the output characteristics of the battery cluster.

[0040] Furthermore, the dynamic priority scoring comprehensively considers the battery cluster's health, response speed, and load matching, and is calculated through a composite method. The weights are dynamically adjusted as the batteries age to more reasonably assess the priority of battery clusters during black start-up.

[0041] The health of the battery cluster is calculated as the sum of the products of its capacity decay, current internal resistance, and cycle count, each with its respective weight. Capacity decay reflects changes in the battery's energy storage capacity, internal resistance changes affect charge / discharge efficiency, and cycle count reflects the battery's usage level. These three factors comprehensively measure the battery's health. Response speed is calculated based on the normalized standard deviation of the voltage ramp-up rate compared to a historical benchmark. Standard deviation normalization maps the voltage ramp-up rate data to a standardized range, more accurately reflecting its deviation from the historical benchmark, thus assessing the battery cluster's response speed to startup commands. Load matching is the cosine similarity between the load curve and the battery cluster's output power curve at the moment before the grid fault (one minute prior). Cosine similarity measures the degree of similarity between the two curves; higher similarity indicates a better match between the battery cluster's output power characteristics and grid load requirements.

[0042] In some specific embodiments, when the grid frequency drops to 48Hz at a rate of 2Hz / s, battery cluster A has a similarity of 0.85, belonging to high priority; battery cluster B has a similarity of 0.65. According to the established fault level calculation method, with a grid frequency deviation (Δf) of 2Hz and a voltage deviation (ΔV) of 5%, the fault level is determined to be 3. At this point, according to the threshold adjustment rule, the black-start competition time threshold is shortened from the default 500ms to 150ms. Within 150ms, after comparison, battery cluster A is selected as the main starting unit. In actual systems, the similarity threshold and fault level classification need to be calibrated based on factors such as the characteristics of the energy storage system and the operating requirements of the grid to achieve the best control effect.

[0043] Step S14: Adjust the black start contention time threshold according to the dynamic priority score and the severity of the grid fault; wherein, the black start contention time threshold is the maximum time limit for each battery cluster to become the main start-up unit.

[0044] Specifically, the severity of the grid fault determines the urgency of black start, while dynamic priority scoring reflects the suitability of each battery cluster for startup. The black start contention time threshold is adjusted based on these two factors to balance the speed and accuracy of black start. The black start contention time threshold can be adjusted through piecewise linear adjustment or fuzzy control. Piecewise linear adjustment adjusts the threshold based on the percentage of frequency deviation; for example, the threshold is reduced by 50ms for every 10% frequency deviation. Fuzzy control switches the threshold based on different grid states such as "urgent," "normal," and "relaxed," providing greater flexibility to adapt to complex grid conditions.

[0045] For example, when the grid frequency drop rate is greater than 1 Hz / s, it indicates that the grid fault is relatively urgent. In this case, the black start contention time threshold is compressed to 200 ms to speed up the selection of the main start unit. When the load fluctuation is less than 5%, the grid state is relatively stable, and the threshold can be extended to 1.5 s to allow multiple rounds of negotiation in order to select the appropriate main start unit more accurately.

[0046] In some embodiments, adjusting the black-start competition time threshold based on dynamic priority scoring and grid fault severity includes: establishing a grid fault severity classification model, which classifies fault levels based on the product of voltage drop amplitude and rate, frequency deviation duration, and load mutation gradient; and adjusting the strategy according to the fault level mapping preset time threshold.

[0047] Specifically, the fault level is calculated as follows: the ratio of voltage drop to rated voltage is multiplied by a first coefficient; the ratio of frequency deviation to rated frequency is multiplied by a second coefficient; and the ratio of load mutation gradient to maximum power change rate is multiplied by a third coefficient. Here, voltage drop refers to the amount of voltage decrease when the fault occurs, frequency deviation is the difference between the actual frequency and the rated frequency, and load mutation gradient represents the rate of change of load power over time.

[0048] The fault levels can be divided into Level 1, Level 2, and Level 3 faults, from mild to severe. Level 1 faults correspond to the first time threshold adjustment strategy, Level 2 faults correspond to the second time threshold adjustment strategy, and Level 3 faults correspond to the third threshold adjustment strategy. The time range of the first time threshold adjustment strategy is shorter than that of the second time threshold adjustment strategy, and the time range of the second time threshold adjustment strategy is shorter than that of the third threshold adjustment strategy.

[0049] In one specific embodiment, the first coefficient is 0.4, the second coefficient is 0.3, and the third coefficient is 0.3. A level 1 fault is defined as a fault level less than 1, a level 2 fault as a fault level greater than or equal to 1 and less than or equal to 3, and a level 3 fault as a fault level greater than 3. The time range for the first time threshold adjustment strategy is greater than 2 seconds, the time range for the second time threshold adjustment strategy is 500 milliseconds to 1 second, and the time range for the third threshold adjustment strategy is 100 milliseconds to 300 milliseconds. If the timeout occurs, the competition is forcibly terminated, and the battery cluster with the highest current score is selected as the main startup unit, ensuring rapid system power restoration. This ensures rapid and reliable selection of the main startup unit during black start, avoiding circulating current risks and delays caused by multi-cluster competition, significantly improving system recovery efficiency and safety. In emergency situations with grid faults, the competition time can be quickly shortened, accelerating the selection of the main startup unit and restoring power in a timely manner; while when the grid state is relatively stable, the competition time can be appropriately extended to more accurately select the most suitable main startup unit.

[0050] Step S15: Select the battery cluster with the highest dynamic priority score as the main start-up unit through the distributed consensus algorithm, and control the main start-up unit to supply power to the energy storage converter to complete the black start.

[0051] Specifically, the distributed consensus algorithm ensures that, in the event of competition among multiple battery clusters, the battery cluster with the highest dynamic priority score is selected fairly and accurately as the main startup unit, and then controls it to supply power to the energy storage converter, thereby achieving a black start for the energy storage system.

[0052] Furthermore, the consensus process of the distributed consensus algorithm is as follows: Each battery cluster management unit broadcasts its calculated priority score to other units. Each unit receives the scores from its neighboring nodes and maintains a local sorted list, updating the priority status of each battery cluster in real time. When the competition times out or a majority consensus is reached, a master unit election is triggered to determine the master startup unit. The master unit closes the circuit breaker, supplies power to the energy storage converter, and starts the energy storage system.

[0053] In some specific embodiments, in an energy storage system containing five battery clusters, the priority scores of each battery cluster are [0.92, 0.85, 0.78, 0.65, 0.60], and the black-start contention threshold is set to 200ms. During the consensus process, node 1 has the highest priority score of 0.92 and receives confirmation from all nodes within 50ms, thus being directly elected as the master startup unit. Subsequently, a control signal is sent to the circuit breaker, and grid connection is completed within 200ms, successfully starting the energy storage system. In practical applications, consensus timeout, heartbeat interval, and other parameters can be dynamically configured according to the network topology of the energy storage system, such as the distribution of battery clusters and the quality of communication links, to optimize the performance of the distributed consensus algorithm.

[0054] In some embodiments, the method further includes: dividing the energy storage system into several logical super-units according to electrical topology and communication links, wherein the battery clusters in each logical super-unit exchange feature vectors through a shared bus.

[0055] Specifically, this partitioning method facilitates data interaction and collaborative control within the energy storage system. Battery clusters within each logical supercell can share feature vectors, improving information transmission efficiency and enabling more accurate assessment of the status and startup priority of each battery cluster.

[0056] In some embodiments, the battery cluster with the highest dynamic priority score is selected as the main startup unit through a distributed consensus algorithm, and the main startup unit is controlled to supply power to the energy storage converter to complete the black start. This includes performing the following operations on each logical superunit: extracting the dynamic priority score of each battery cluster within the logical superunit; generating a superunit-level startup priority using a weighted voting mechanism, wherein the weight of the weighted voting is calculated based on the electromagnetic cluster capacity ratio plus a communication quality correction factor; and prioritizing the startup of the logical superunit with the highest startup priority.

[0057] Specifically, within the logical supercell, a weighted voting mechanism is used to determine the startup priority. This weight comprehensively considers the capacity proportion of the battery clusters and the communication quality correction factor, thereby more rationally determining the startup order of each battery cluster within the supercell.

[0058] Furthermore, the base weight is obtained by dividing the capacity of a single battery cluster by the total capacity of all battery clusters within the logical supercell, and then multiplying by 100%. A larger capacity battery cluster has a higher base weight, reflecting its energy advantage during startup. The final weight is equal to the base weight multiplied by e raised to the power of 'e'. Greater communication latency results in a smaller correction factor and a lower final weight, preventing untimely or inaccurate decisions due to communication delays. The sum of the final weights of all battery clusters within the logical supercell is the startup priority of that supercell. A higher startup priority means the supercell starts up more quickly during black boot.

[0059] In the above-described scheme, this application obtains the current state parameters of each battery cluster in the energy storage system and historical state data before the grid fault, and inputs them into a dynamic feature extraction network to generate dynamic feature vectors, comprehensively and accurately reflecting the real-time state and potential startup capability of the battery clusters. Based on the dynamic feature vectors, a contrastive learning model is used to calculate the similarity between each battery cluster and the preset ideal startup features, generating a dynamic priority score. The black-start competition time threshold is adjusted according to the dynamic priority score and the severity of the grid fault, fully considering the actual situation of the grid fault and the startup suitability of each battery cluster. The distributed consensus algorithm ensures an effective balance between the speed and accuracy of black-start in complex situations where multiple battery clusters participate in the competition, significantly improving the energy storage system's ability to cope with different grid operating conditions.

[0060] Please see Figure 2 , Figure 2 This is a schematic diagram of the structural framework of one embodiment of the grid-type energy storage black start control device of this application. Figure 2 As shown, the grid-type energy storage black-start control device 20 includes: a data acquisition module 21, a feature extraction module 22, a scoring module 23, an adjustment module 24, and a control module 25. Specifically, the data acquisition module 21 acquires the current state parameters of each battery cluster in the energy storage system and historical state data before a grid fault; the feature extraction module 22 inputs the current state parameters and historical state data into a dynamic feature extraction network to generate dynamic feature vectors corresponding to each battery cluster; the scoring module 23 calculates the similarity between each battery cluster and a preset ideal start-up feature using a comparative learning model based on the dynamic feature vectors, generating a dynamic priority score for each battery cluster; the adjustment module 24 adjusts the black-start competition time threshold according to the dynamic priority score and the severity of the grid fault; the black-start competition time threshold is the maximum time limit allowed for each battery cluster to become the main start-up unit; and the control module 25 selects the battery cluster with the highest dynamic priority score as the main start-up unit using a distributed consensus algorithm, and controls the main start-up unit to supply power to the energy storage converter to complete the black start.

[0061] In some embodiments, the dynamic feature extraction network also needs to be pre-trained. Pre-training the dynamic feature extraction network includes: acquiring training samples under historical black start scenarios; wherein, the training samples include state parameter samples of each battery cluster and corresponding historical state data samples before the power grid fault; normalizing the training samples and labeling them with ideal start feature labels; extracting multidimensional spatiotemporal features through a convolutional neural network, and optimizing the network parameters by combining a contrastive loss function, so that the feature vectors are clustered in the embedding space according to the type of power grid fault.

[0062] In some embodiments, the comparison loss function employs an event-triggered dynamic temperature parameter adjustment mechanism; specifically, it includes: reducing the temperature parameter in response to detecting that the type of grid fault is voltage collapse; and increasing the temperature parameter in response to detecting that the type of grid fault is frequency oscillation; wherein the adjustment range of the temperature parameter is a preset threshold range.

[0063] In some embodiments, the scoring module 23 generates a dynamic priority score for each battery cluster, including: constructing a composite weighting function based on the battery cluster's health score, response speed score, and load matching score; the composite weighting function and the health score are positively correlated with the aging degree of the battery cluster; wherein, the health score is calculated based on the capacity decay rate, cycle number, and internal resistance change rate; the response speed score is calculated based on the deviation of the voltage ramp-up rate from the historical benchmark value; and the load matching score is generated based on the matching degree between the load demand curve recorded before the grid fault and the output characteristics of the battery cluster.

[0064] In some embodiments, the energy storage system can be divided into several logical super-units according to electrical topology and communication links, and the battery clusters in each logical super-unit exchange feature vectors through a shared bus.

[0065] In some embodiments, the control module 25 is further configured to select the battery cluster with the highest dynamic priority score as the main startup unit through a distributed consensus algorithm, and control the main startup unit to supply power to the energy storage converter to complete the black start, including: performing the following operations on each logical super unit: extracting the dynamic priority score of each battery cluster within the logical super unit; generating a super unit-level startup priority using a weighted voting mechanism, wherein the weight of the weighted voting is calculated based on the electromagnetic cluster capacity ratio plus a communication quality correction factor; and prioritizing the startup of the logical super unit with the highest startup priority.

[0066] In the above-described scheme, this application obtains the current state parameters of each battery cluster in the energy storage system and historical state data before the grid fault, and inputs them into a dynamic feature extraction network to generate dynamic feature vectors, comprehensively and accurately reflecting the real-time state and potential startup capability of the battery clusters. Based on the dynamic feature vectors, a contrastive learning model is used to calculate the similarity between each battery cluster and the preset ideal startup features, generating a dynamic priority score. The black-start competition time threshold is adjusted according to the dynamic priority score and the severity of the grid fault, fully considering the actual situation of the grid fault and the startup suitability of each battery cluster. The distributed consensus algorithm ensures an effective balance between the speed and accuracy of black-start in complex situations where multiple battery clusters participate in the competition, significantly improving the energy storage system's ability to cope with different grid operating conditions.

[0067] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the grid-type energy storage black-start control method.

[0068] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system, device, or unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0069] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0070] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A black start control method for grid-type energy storage, characterized in that, The method includes: Acquire the current state parameters of each battery cluster in the energy storage system and the historical state data before the grid fault; The current state parameters and the historical state data are input into the dynamic feature extraction network to generate dynamic feature vectors corresponding to each battery cluster; Based on the dynamic feature vector, the similarity between each battery cluster and the preset ideal start-up features is calculated through a contrastive learning model to generate a dynamic priority score for each battery cluster. Based on the dynamic priority score and the severity of the grid fault, the black start contention time threshold is adjusted; wherein, the black start contention time threshold is the maximum time limit for each battery cluster to become the main start-up unit; The battery cluster with the highest dynamic priority score is selected as the main startup unit through a distributed consensus algorithm, and the main startup unit is controlled to supply power to the energy storage converter to complete the black start. The method further includes: pre-training the dynamic feature extraction network; The pre-training of the dynamic feature extraction network includes: Acquire training samples under historical black start scenarios; wherein, the training samples include state parameter samples of each battery cluster and corresponding historical state data samples before the power grid fault; The training samples are normalized and labeled with ideal activating feature labels; Multidimensional spatiotemporal features are extracted by convolutional neural networks, and network parameters are optimized by combining a contrastive loss function, so that feature vectors are clustered in the embedding space according to the type of power grid fault. The generation of dynamic priority scores for each battery cluster includes: A composite weighting function is constructed based on the battery cluster's health score, response speed score, and load matching score; the composite weighting function and the health score are positively correlated with the battery cluster's aging degree. The health score is calculated based on the capacity decay rate, cycle count, and internal resistance change rate; the response speed score is calculated based on the deviation of the voltage ramp-up rate from the historical benchmark value; and the load matching score is generated based on the matching degree between the load demand curve recorded before the grid fault and the battery cluster output characteristics. The step of adjusting the black-start contention time threshold based on the dynamic priority score and the severity of the power grid fault includes: A fault severity classification model is established, which classifies fault levels based on the product of voltage drop amplitude and rate, frequency deviation duration, and load change gradient. The fault level is mapped to a preset time threshold adjustment strategy.

2. The black-start control method for grid-type energy storage according to claim 1, characterized in that, The contrast loss function employs an event-triggered dynamic temperature parameter adjustment mechanism; specifically, it includes: In response to the detection that the type of power grid fault is voltage collapse, the temperature parameter is reduced; In response to the detection of a power grid fault of the type described as frequency oscillation, the temperature parameter is increased; The temperature parameter is adjusted within a preset threshold range.

3. The black-start control method for grid-type energy storage according to claim 1 or 2, characterized in that, The method further includes: The energy storage system is divided into several logical super-units according to electrical topology and communication links. The battery clusters in each logical super-unit exchange the feature vectors through a shared bus.

4. The black-start control method for grid-type energy storage according to claim 3, characterized in that, The process of selecting the battery cluster with the highest dynamic priority score as the main startup unit through a distributed consensus algorithm, and controlling the main startup unit to supply power to the energy storage converter to complete the black start includes: Perform the following operations on each of the aforementioned logical supercells: Extract the dynamic priority score of each battery cluster within the logical supercell; A weighted voting mechanism is used to generate supercell-level startup priorities, wherein the weights of the weighted voting are calculated based on the electromagnetic cluster capacity ratio plus a communication quality correction factor; The logical superunit with the highest startup priority will be started first.

5. A grid-type energy storage black start control device, characterized in that, The grid-type energy storage black start control device includes: Data acquisition module: used to acquire the current status parameters of each battery cluster in the energy storage system and historical status data before grid faults; Feature extraction module: Inputs the current state parameters and the historical state data into the dynamic feature extraction network to generate dynamic feature vectors corresponding to each battery cluster; Scoring module: Based on the dynamic feature vector, the similarity between each battery cluster and the preset ideal start-up features is calculated through a contrastive learning model to generate a dynamic priority score for each battery cluster; Adjustment module: Based on the dynamic priority score and the severity of the grid fault, adjust the black start contention time threshold; wherein, the black start contention time threshold is the maximum time limit that allows each battery cluster to become the main start-up unit; Control module: Selects the battery cluster with the highest dynamic priority score as the main start-up unit through a distributed consensus algorithm, and controls the main start-up unit to supply power to the energy storage converter to complete the black start; The device is also used to: pre-train the dynamic feature extraction network; The pre-training of the dynamic feature extraction network includes: Acquire training samples under historical black start scenarios; wherein, the training samples include state parameter samples of each battery cluster and corresponding historical state data samples before the power grid fault; The training samples are normalized and labeled with ideal activating feature labels; Multidimensional spatiotemporal features are extracted by convolutional neural networks, and network parameters are optimized by combining a contrastive loss function, so that feature vectors are clustered in the embedding space according to the type of power grid fault. The generation of dynamic priority scores for each battery cluster includes: A composite weighting function is constructed based on the battery cluster's health score, response speed score, and load matching score; the composite weighting function and the health score are positively correlated with the battery cluster's aging degree. The health score is calculated based on the capacity decay rate, cycle count, and internal resistance change rate; the response speed score is calculated based on the deviation of the voltage ramp-up rate from the historical benchmark value; and the load matching score is generated based on the matching degree between the load demand curve recorded before the grid fault and the battery cluster output characteristics. The step of adjusting the black-start contention time threshold based on the dynamic priority score and the severity of the power grid fault includes: A fault severity classification model is established, which classifies fault levels based on the product of voltage drop amplitude and rate, frequency deviation duration, and load change gradient. The fault level is mapped to a preset time threshold adjustment strategy.

Images