A centralized battery energy storage control system and control method

Abstract

The application discloses a centralized battery energy storage control system and a control method. The method comprises the following steps: S1: obtaining power grid scheduling target data and real-time operation state data of each battery energy storage device, extracting a global state feature vector and a device identity from the power grid scheduling target data and the real-time operation state data; S2: inputting the global state feature vector into a central strategy network to obtain a global action strategy distribution; S3: determining a target local strategy network from a preset local strategy network set according to the global action strategy distribution and the device identity; S4: inputting the global state feature vector into the target local strategy network to obtain a local power control instruction, and generating a centralized scheduling instruction according to the global action strategy distribution and the local power control instruction; and S5: issuing the centralized scheduling instruction to corresponding battery energy storage devices and responding. Through the method, the reliability of battery energy storage control can be ensured under complex power grid working conditions.

Description

Technical Field

[0001] This invention relates to the field of power system energy storage control technology, and in particular to a centralized battery energy storage control system and control method. Background Technology

[0002] Currently, with the accelerated construction of new power systems and the continuous increase in renewable energy penetration, large-scale battery energy storage systems, as a key flexibility resource, have played a core role in ancillary services such as grid frequency regulation, peak shaving and valley filling, voltage support, and black start. Centralized energy storage control architectures have become the mainstream engineering solution due to their strong global optimization capabilities, fast response to dispatch commands, and ease of unified grid management. In recent years, artificial intelligence and reinforcement learning technologies have been gradually introduced into the field of energy storage control, aiming to replace traditional rule-based control or model predictive control (MPC) with data-driven approaches to cope with complex grid conditions characterized by high dimensionality, nonlinearity, and strong coupling. Although some research has achieved intelligent policy generation for single devices or homogeneous device groups, systematic solutions are still lacking in areas such as heterogeneous device collaboration (e.g., differences in battery types, aging levels, and communication latency), policy consistency assurance (matching global optimization objectives with local execution capabilities), robustness in communication-constrained scenarios, and long-term adaptive operation. Specifically, existing centralized battery energy storage control methods still face the following key bottlenecks: disconnect between global strategy and local execution: the instructions generated by the central controller are not dynamically adapted to the individual execution capabilities of the devices (such as power limits and response speed), and the local execution units lack a strategy calibration mechanism, which can easily lead to strategy divergence (deviation between control instructions and actual execution) and tracking failure (inability to stably follow the target); thus restricting the reliability and efficiency improvement of centralized energy storage systems under complex grid conditions. Therefore, it is very important to provide a centralized battery energy storage control system and control method that can solve the above bottlenecks. Summary of the Invention

[0003] In view of the above-mentioned prior art, the present invention provides a centralized battery energy storage control system and control method, which mainly solves the technical problems existing in the background art.

[0004] To achieve the above objectives, the technical solution of this invention is implemented as follows: In a first aspect, the present invention provides a centralized battery energy storage control method, comprising the following steps: S1: Obtain grid dispatch target data and real-time operating status data of each battery energy storage device, and extract global status feature vector and device identity identifier from the grid dispatch target data and the real-time operating status data; S2: Input the global state feature vector into the central policy network to obtain the global action policy distribution; S3: Determine the target local policy network from the preset local policy network set based on the global action policy distribution and the device identity identifier; S4: Input the global state feature vector into the target local policy network to obtain the local power control instruction, and generate a centralized scheduling instruction based on the global action policy distribution and the local power control instruction; S5: The centralized scheduling instruction is sent to the corresponding battery energy storage device, and the battery energy storage device responds according to the centralized scheduling instruction.

[0005] Optionally, the power grid dispatch target data includes: target active power value, target reactive power value, frequency regulation coefficient, and start and end times of the dispatch time window; The real-time operating status data includes: device number, current state of charge, terminal voltage, charging and discharging current, battery temperature, health status index, communication delay duration, and maximum adjustable power margin.

[0006] Optionally, the step of inputting the global state feature vector into the central policy network to obtain the global action policy distribution includes: A central policy network is constructed, which includes an input layer, a hidden layer, an activation layer, and an output layer. The hidden layer consists of three fully connected neural networks with 512, 256, and 128 neurons in each layer, respectively. The activation function of the activation layer is the ReLU function, and the output layer is mapped to a preset action space using the Tanh function. The central policy network is trained to obtain a trained central policy network. The global state feature vector is input into the trained central policy network, which outputs continuous action probability density function parameters, and these parameters are used as the global action policy distribution.

[0007] Optionally, determining the target local policy network from a preset set of local policy networks based on the global action policy distribution and the device identity includes: The device identification is indexed and encoded, an index mapping table is preset, and the device index number is obtained according to the index code and the index mapping table; Extract the action expectation component corresponding to the device index number, and generate a device-level policy guidance signal based on the action expectation component; Based on the device index number and the device-level policy guidance signal, an enhanced policy matching key value is obtained. The enhanced policy matching key value is then input into a pre-trained principal component analysis dimensionality reduction layer to reduce the dimensionality of the policy matching key value, resulting in a dimensionality-reduced enhanced policy matching key value. In the set of local policy networks, find the local policy network that has the smallest Euclidean distance to the matching key value of the dimensionality-reduced enhanced policy, and determine the found local policy network as the target local policy network.

[0008] Optionally, the step of searching the local policy network set for the local policy network that has the smallest Euclidean distance to the matching key value of the dimensionality-reduced policy, and determining the found local policy network as the target local policy network, includes: A KD-tree index structure is pre-constructed, and the KD-tree nearest neighbor search algorithm is used to find the registration key value with the smallest distance to the dimensionality-reduced policy matching key value. The local policy network corresponding to the registration key value is then determined as the target local policy network.

[0009] Optionally, the step of inputting the global state feature vector into the target local policy network to obtain local power control instructions, and generating centralized scheduling instructions based on the global action policy distribution and the local power control instructions, includes: The target local policy network is a multi-layer fully connected structure, including an input layer, a hidden layer, and an output layer; The global state feature vector is input into the target local policy network, and the local power control command is output through the output layer. Extract the mean and variance components of the policies of the devices corresponding to the target local policy network from the global action policy distribution; Calculate the deviation vector between the local power control command and the mean component of the strategy, and normalize the deviation vector according to the variance component of the strategy to generate a strategy consistency index. The moving average and moving standard deviation of the strategy consistency index generated in several control cycles are calculated. If the difference between the strategy consistency index of the current cycle and the moving average exceeds a preset multiple of the moving standard deviation, it is considered as an abnormal policy deviation that triggers the calibration mechanism in advance. If the L2 norm of the policy consistency index in the current period is greater than the preset threshold, the policy distillation calibration process is initiated. The global action policy distribution is taken as the teacher policy and the target local policy network is taken as the student policy. The calibration coefficient is generated by minimizing the KL divergence loss function. The local control command is multiplied by the calibration coefficient to generate the calibrated centralized scheduling command.

[0010] Optionally, the step of sending the centralized scheduling command to the corresponding battery energy storage device, and the battery energy storage device responding according to the centralized scheduling command, includes: Based on the device index number in the calibrated centralized scheduling instruction, the target recipient of the instruction is determined, and the calibrated centralized scheduling instruction is encapsulated into a standard instruction frame, which includes an instruction identifier, target device address, power setting value, timestamp, and checksum. The standard instruction frame is sent to the battery energy storage device corresponding to the target device address through a preset communication channel; After receiving the standard command frame, the battery energy storage device parses the standard command frame, extracts the power setting value, and adjusts the power according to the power setting value.

[0011] Optionally, it further includes: a centralized control terminal receiving execution status feedback data returned by the battery energy storage device, and adopting a response strategy based on the execution status feedback data, specifically: Within a preset time window after the standard instruction frame is issued, the centralized control terminal receives the execution status feedback data returned by the battery energy storage device. The execution feedback data includes first execution feedback data and second execution feedback data: The first execution status feedback data includes: actual output power value, execution completion flag, and exception status code; The second execution feedback data includes: the current actual output power value, the remaining estimated adjustment time, and the current execution stage identifier; The centralized control terminal judges the execution status feedback data. If it is the first execution status feedback data, a response strategy is adopted according to the abnormal status code. If it is the second execution status feedback data, a dynamic evaluation is performed according to the current execution stage identifier, and a response strategy is adopted according to the dynamic evaluation result.

[0012] Optionally, if the data is a first execution status feedback, a response strategy is adopted based on the abnormal status code, including: The abnormal status codes are classified into abnormal levels according to their severity, and the abnormal levels include warning level, degraded operation level, and shutdown protection level. The centralized control unit adopts a corresponding response strategy based on the anomaly level.

[0013] Secondly, the present invention also provides a centralized battery energy storage control system, comprising: The feature extraction module is used to acquire grid dispatch target data and real-time operating status data of each battery energy storage device, and extract global status feature vectors and device identity identifiers from the grid dispatch target data and the real-time operating status data. The policy generation module is used to input the global state feature vector into the central policy network to obtain the global action policy distribution; The policy matching module is used to determine the target local policy network from a preset set of local policy networks based on the combination of the global action policy distribution and the device identity identifier. The instruction calibration module is used to input the global state feature vector into the target local policy network to obtain local power control instructions, and generate calibrated centralized scheduling instructions based on the global action policy distribution and the local power control instructions. The instruction execution module is used to send the calibrated centralized scheduling instruction to the corresponding battery energy storage device, and the battery energy storage device performs the power response action.

[0014] The beneficial effects of this invention are as follows: The centralized battery energy storage control method provided by this invention constructs a collaborative architecture of a central strategy network and local strategy networks. It extracts global state feature vectors and device identifications from grid dispatch target data and real-time operating status data. The central strategy network outputs a global action strategy distribution, and then combines the device identifications with the target local strategy network to generate local power control commands. Finally, through calibrated centralized dispatch commands, it achieves coordinated response of heterogeneous devices under dynamic operating conditions. It enables each battery energy storage device to maintain local decision-making freedom while meeting global power tracking accuracy. It also enables the control strategy to maintain suboptimal operation even when communication is interrupted or device performance degrades, thereby improving the reliability of the centralized energy storage system under complex grid operating conditions. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating a centralized battery energy storage control method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a centralized battery energy storage control system provided in an embodiment of the present invention. Detailed Implementation

[0016] To fully understand this invention, a detailed structure will be presented in the following description to illustrate the technical solution proposed by this invention. Optional embodiments of the invention are described in detail below; however, in addition to these detailed descriptions, the invention may have other embodiments.

[0017] Example 1 Please refer to the attached document. Figure 1 This invention provides a centralized battery energy storage control method, comprising the following steps: S1: Obtain grid dispatch target data and real-time operating status data of each battery energy storage device, and extract global status feature vector and device identity identifier from the grid dispatch target data and the real-time operating status data; S2: Input the global state feature vector into the central policy network to obtain the global action policy distribution; S3: Determine the target local policy network from the preset local policy network set based on the global action policy distribution and the device identity identifier; S4: Input the global state feature vector into the target local policy network to obtain the local power control instruction, and generate a centralized scheduling instruction based on the global action policy distribution and the local power control instruction; S5: The centralized scheduling instruction is sent to the corresponding battery energy storage device, and the battery energy storage device responds according to the centralized scheduling instruction.

[0018] Specifically, in centralized battery energy storage control, the power response characteristics of different battery energy storage devices vary due to differences in battery type, aging degree, thermal management capability, and communication latency. Under grid power fluctuations or frequency disturbances, if a unified centralized dispatch command is used, it is easy to cause some devices to overload or respond late, reducing the overall regulation accuracy and system stability. At the same time, if a purely distributed autonomous decision-making is adopted, the lack of global target constraints can easily lead to power allocation deviating from the economic or safety boundary, failing to meet the tracking requirements of grid dispatch commands. Existing control methods mostly rely on fixed rules or static models, which cannot achieve dynamic coordination and strategy self-evolution under conditions of heterogeneous equipment, limited communication, and sudden changes in operating conditions. Therefore, to address the aforementioned issues of inconsistent heterogeneous device responses, conflicts between global objectives and local decisions, and the lack of online evolution capabilities in control strategies, a collaborative architecture between a central strategy network and local strategy networks is constructed. This architecture extracts global state feature vectors and device identification identifiers from grid dispatch target data and real-time operating status data. The central strategy network outputs a global action strategy distribution, which is then combined with the device identification identifiers to match the target. Local strategy networks generate local power control commands, and finally, through calibrated centralized dispatch commands, collaborative responses of heterogeneous devices under dynamic operating conditions are achieved. This allows each battery energy storage device to maintain local decision-making freedom while meeting global power tracking accuracy. Furthermore, it ensures that the control strategy can maintain suboptimal operation even during communication interruptions or device performance degradation.

[0019] As an optional implementation, the power grid dispatch target data includes: target active power value, target reactive power value, frequency regulation coefficient, and start and end times of the dispatch time window; The real-time operating status data includes: device number, current state of charge, terminal voltage, charging and discharging current, battery temperature, health status index, communication delay duration, and maximum adjustable power margin.

[0020] Specifically, grid dispatch target data is obtained from the grid dispatch control platform, and real-time operating status data of each battery energy storage device is obtained from the energy storage site monitoring platform. The grid dispatch control platform is a centralized command source for issuing power regulation targets, frequency support commands, or load balancing requirements. The energy storage site monitoring platform is a local data source for collecting operating parameters of battery energy storage devices deployed in different locations.

[0021] In practical applications, grid dispatch target data is periodically acquired from the grid dispatch control platform via a standardized data interface, while real-time operating status data of each battery energy storage device is periodically read from the energy storage site monitoring platform via the site-side data acquisition channel. The data acquisition frequency is set to once every 200 milliseconds to match the response time requirements of grid frequency regulation. From the grid dispatch target data and real-time operating status data, a global state feature vector is generated by concatenating the grid dispatch target data and real-time operating status data according to a preset feature mapping rule. Its dimension is 1×N, where N = dispatch target dimension + device status dimension × number of devices. The device identification is a fixed-length identifier vector generated by uniquely encoding the device number extracted from the real-time operating status, used to distinguish the identity and role of different battery energy storage devices in the strategy network.

[0022] As an optional implementation, the step of inputting the global state feature vector into the central policy network to obtain the global action policy distribution includes: A central policy network is constructed, which includes an input layer, a hidden layer, an activation layer, and an output layer. The hidden layer consists of three fully connected neural networks with 512, 256, and 128 neurons in each layer, respectively. The activation function of the activation layer is the ReLU function, and the output layer is mapped to a preset action space using the Tanh function. The central policy network is trained to obtain a trained central policy network. The global state feature vector is input into the trained central policy network, which outputs continuous action probability density function parameters, and these parameters are used as the global action policy distribution.

[0023] Specifically, the central policy network is a deep deterministic policy gradient network deployed in the central control terminal. Its input layer dimension is consistent with the global state feature vector dimension. After the global state feature vector is input into the central policy network, it passes through the input layer, hidden layer, activation layer and output layer in sequence. During the training phase, the central policy network receives historical state-action-reward triplets through the experience replay buffer and updates the network parameters using gradient descent with the goal of maximizing the cumulative discount reward. During the runtime phase, the global state feature vector is directly input into the trained central policy network. After forward propagation, the continuous action probability density function parameters are output as the global action policy distribution. The global action strategy distribution is a multidimensional Gaussian distribution, with the mean vector dimension equal to the total number of battery energy storage devices participating in the scheduling. The covariance matrix is ​​a diagonal matrix used to characterize the expected value and uncertainty boundary of the power adjustment command of each device. The output frequency is synchronized with the global state feature vector acquisition frequency, that is, it is output once every 200 milliseconds. In practical applications, the initial weights of the central policy network are set using the Xavier initialization method, with a learning rate of 0.001, a discount factor of 0.99, and an experience replay buffer capacity of 1×10⁻⁶. 6 1 record.

[0024] As an optional implementation, determining the target local policy network from a preset set of local policy networks based on the global action policy distribution and the device identity includes: The device identification is indexed and encoded, an index mapping table is preset, and the device index number is obtained according to the index code and the index mapping table; Extract the action expectation component corresponding to the device index number, and generate a device-level policy guidance signal based on the action expectation component; Based on the device index number and the device-level policy guidance signal, an enhanced policy matching key value is obtained. The enhanced policy matching key value is then input into a pre-trained principal component analysis dimensionality reduction layer to reduce the dimensionality of the policy matching key value, resulting in a dimensionality-reduced enhanced policy matching key value. In the set of local policy networks, find the local policy network that has the smallest Euclidean distance to the matching key value of the dimensionality-reduced enhanced policy, and determine the found local policy network as the target local policy network.

[0025] Specifically, the local policy network set stores multiple local policy networks, and each local policy network corresponds to a combination of a device identity identifier and a global action policy distribution; The device identification is a pre-registered string device number. An index mapping table is pre-built, which includes the device index number corresponding to each device identification. The device index number is an integer sequence. The system searches the pre-set index mapping table based on the device identification to obtain the integer value that uniquely corresponds to the device identification, which is then used as the device index number. If the device identifier is not registered in the preset index mapping table, the current maximum index number plus 1 will be automatically assigned as the device index number, and the correspondence between the device identifier and the device index number will be appended to the preset index mapping table.

[0026] The specific process of generating device-level policy boot signals includes: When generating the mean vector, the centralized control server mandates that the first element of the mean vector corresponds to the action expectation of the battery energy storage device with device index number 1, the second element corresponds to the action expectation of the battery energy storage device with device index number 2, and the ith element corresponds to the action expectation of the battery energy storage device with device index number ith, thereby establishing a deterministic correspondence between the device index number and the position of the mean vector. Therefore, when the centralized control server outputs the global action strategy distribution, the dimension order of the mean vector is pre-arranged in ascending order according to the device index numbers of the battery energy storage devices. Therefore, the original action expectation value at the corresponding position is extracted according to the device index number i. If the action space of a single device is K-dimensional, then K consecutive elements are extracted, that is, the (i-1)×K+1 to i×K elements are extracted to form the initial guidance signal. Then, the original action expectation value in the initial guidance signal is subjected to neighborhood weighted average: the corresponding action expectations of all devices j (including i itself) are weighted and summed according to the electrical coupling weight to generate the electrical sensing guidance signal. Then, the electrical sensing guidance signal is divided element by element by its standard deviation (σ≥0.01) in the global action policy distribution covariance matrix. The normalized guidance signal is used to suppress the dominance of high uncertainty devices in policy matching to complete the normalization and generate the device-level policy guidance signal. The formula for calculating the electrical coupling weight of device j to device i is as follows:

[0027] in, This represents the electrical coupling weight of device j to device i, used for neighborhood-weighted averaging when extracting device-level policy guidance signals. The element in the i-th row and j-th column of the grid node admittance matrix represents the electrical coupling strength between device i and device j, and N represents the total number of battery storage devices participating in the dispatch. This represents the sum of the admittances of device i and all devices (including itself), used for normalization. The policy matching key is used to represent the local decision context of a specific device under the current global policy. The policy matching key is composed of the device index number as the first dimension and the device-level policy guidance signal as the subsequent dimension. If the device-level policy guidance signal is a 1×K vector, then the policy matching key dimension is 1×(K+1), and the data type is floating point. The device index number is input into the device embedding layer, and the output is a fixed-dimensional device embedding vector; the device embedding vector is concatenated with the device-level policy guidance signal to form a policy matching key value. After the strategy matching key value is obtained by concatenation, the timestamp modulo 60 value of the current control cycle and the grid frequency deviation value are additionally concatenated to form an enhanced strategy matching key value. The timestamp modulo 60 value is used to characterize the intraday cycle characteristics, and the frequency deviation value is used to characterize the grid disturbance intensity, thereby improving the generalization ability of the strategy under time-varying operating conditions.

[0028] After the enhanced policy matching key value is generated, it is input into the pre-trained principal component analysis dimensionality reduction layer to compress the dimension to a fixed length L, and the vector is stored as the registration key value of the local policy network for policy matching in subsequent running stages. In the local policy network set, all local policies in the local policy network set are traversed, and the Euclidean distance between the registration key value of each local policy and the matching key value of the dimensionality-reduced augmented policy is calculated. The local policy network with the smallest Euclidean distance is selected as the target local policy network. The registration key value is fixed and stored after the local policy network is trained and does not change during operation.

[0029] As an optional implementation, the step of searching the local policy network set for the local policy network that has the smallest Euclidean distance to the matching key value of the dimensionality-reduced policy, and determining the found local policy network as the target local policy network, includes: A KD-tree index structure is pre-constructed, and the KD-tree nearest neighbor search algorithm is used to find the registration key value with the smallest distance to the matching key value of the dimensionality-reduced enhanced policy. The local policy network corresponding to the registration key value is then determined as the target local policy network.

[0030] Specifically, the registration key values ​​of the local policy network set are pre-constructed with a KD-tree index structure. During the matching phase, the KD-tree nearest neighbor search algorithm is used to find the registration key value with the smallest distance to the dimensionality-reduced enhanced policy matching key value, which is then used as the nearest neighbor registration key value. The local policy network corresponding to the nearest neighbor registration key value is determined as the target local policy network. The search time increases logarithmically with the size of the set, making it suitable for scenarios with more than 100 devices. If the Euclidean distance between the dimensionality-reduced enhanced policy matching key value and the nearest neighbor registration key value is greater than a preset threshold of 0.5, it indicates that the current working condition exceeds the effective coverage of the existing local policy network. At this time, the online fine-tuning process of the target local policy network is triggered. The fine-tuning samples are the current state and the reward after execution. The learning rate decays linearly with distance, and the greater the distance, the stronger the decay, ensuring the conservatism of policy updates. After the online fine-tuning process is triggered and completed, the registration key value is updated to the moving average of the original registration key value and the current dimensionality-reduced enhanced policy matching key value, so that the registration key value dynamically tracks the new working condition center adapted by the policy network.

[0031] It also includes: generating fusion instructions, specifically: using the KD tree nearest neighbor search algorithm to find the three registered key values ​​with the smallest distance to the dimensionality-reduced enhanced policy matching key value, determining the top 3 local policy networks with the smallest Euclidean distance, calculating normalized weights based on the reciprocal of the distance, and performing a weighted average of the local power control instructions output by each local policy network to generate fusion instructions; It should be noted that the three local policy networks with the smallest Euclidean distance are selected for weighted fusion, rather than using a single nearest neighbor policy. This is to improve the robustness and generalization of policy output in scenarios with heterogeneous devices, sudden changes in operating conditions, or incomplete coverage of registered key values. Implicit interpolation is achieved through multi-policy consensus to avoid policy jumps caused by matching noise or blurred boundaries. The main functions of the fusion instructions are: first, to provide high-quality supervision signals during the runtime phase for subsequent policy consistency judgment and anomaly detection; and second, to serve as soft labels during the subsequent policy distillation training phase to guide the fine-tuning or updating of the target local policy network. In the process, the calibrated centralized scheduling instruction is generated by scaling the original output of the target local policy network with the calibration coefficient, i.e., "local power control instruction × calibration coefficient"; however, when the policy consistency index triggers distillation calibration, the fusion instruction or its corresponding global action policy distribution will be used as the teacher policy, and the KL divergence will be calculated with the target local policy network to generate the calibration coefficient, which will indirectly affect the final instruction generation to obtain the calibrated centralized scheduling instruction.

[0032] As an optional implementation, the step of inputting the global state feature vector into the target local policy network to obtain local power control instructions, and generating centralized scheduling instructions based on the global action policy distribution and the local power control instructions, includes: The target local policy network is a multi-layer fully connected structure, including an input layer, a hidden layer, and an output layer; The global state feature vector is input into the target local policy network, and the local power control command is output through the output layer. Extract the mean and variance components of the policies of the devices corresponding to the target local policy network from the global action policy distribution; Calculate the deviation vector between the local power control command and the mean component of the strategy, and normalize the deviation vector according to the variance component of the strategy to generate a strategy consistency index. The moving average and moving standard deviation of the strategy consistency index generated in several control cycles are calculated. If the difference between the strategy consistency index of the current cycle and the moving average exceeds a preset multiple of the moving standard deviation, it is considered as an abnormal policy deviation that triggers the calibration mechanism in advance. If the L2 norm of the policy consistency index in the current period is greater than the preset threshold, the policy distillation calibration process is initiated. The global action policy distribution is taken as the teacher policy and the target local policy network is taken as the student policy. The calibration coefficient is generated by minimizing the KL divergence loss function. The local control command is multiplied by the calibration coefficient to generate the calibrated centralized scheduling command.

[0033] Specifically, the target local policy network adopts a multi-layer fully connected structure. The input layer dimension matches the global state feature vector, the number of hidden layers is two or three, the number of neurons in each layer decreases progressively, and the number of neurons in the output layer is the same as the action space dimension of a single device. The hidden layers use linear rectified activation functions, and the output layers use bounded activation functions. After linear transformation, the output value falls into the power adjustment range allowed by the device and serves as a local power control command. The local power control command is an independent output of the target local policy network to the current state and is only used for subsequent calculation of policy consistency index to trigger the calibration mechanism. The target local policy network superimposes random noise following a Gaussian distribution at the output layer. The noise standard deviation takes a large value in the early stage of control and gradually decreases to a small stable value as the control cycle increases. The local power control command is generated by adding the original output of the target local policy network to the noise value. It is used to enhance the policy exploration capability in the early stage of operation and improve the output stability in the later stage. The policy mean component is the mean vector in the global action policy distribution, which is the value or group of values ​​that matches the device index number position corresponding to the target local policy network; if the device action space is multidimensional, multiple consecutive values ​​are extracted; the policy variance component is the diagonal element in the covariance matrix corresponding to the device position, which is used to reflect the degree of uncertainty of the device action policy.

[0034] Establish a dynamic correspondence table between device index numbers and global action policy distribution data locations; extract the policy mean component and policy variance component based on the updated correspondence table to ensure that device identity and policy distribution are always accurately aligned and adapt to dynamic topology change scenarios. The strategy consistency index is obtained by subtracting the local power control command from the strategy mean component element by element, and then dividing each difference element by the square root of the corresponding strategy variance component plus a very small constant. The strategy consistency index generated in the most recent control cycles is recorded, and its moving average and moving standard deviation are calculated. If the difference between the strategy consistency index and the moving average in the current cycle exceeds a preset multiple of the moving standard deviation, it is considered an abnormal strategy deviation, and the calibration mechanism is triggered in advance. This preset multiple can be dynamically adjusted based on historical stability data or online performance indicators. If the L2 norm of the policy consistency index is greater than the preset threshold, the policy distillation calibration process is initiated. The teacher policy is the fusion instruction or its corresponding global action policy distribution, and the target local policy network is the student policy. The calibration coefficient is generated by minimizing the KL divergence loss function. The preset threshold is a fixed value pre-set based on historical operating data or simulation experiments, used to characterize the maximum acceptable policy deviation. If the L2 norm of the policy consistency index exceeds this threshold, the policy distillation calibration process is initiated. During the calibration process, the global action policy distribution is used as the reference policy, and the current output of the target local policy network is used as the policy to be calibrated. The probability distribution of the policy to be calibrated is made to approximate the reference policy through iterative adjustment, generating calibration coefficients for scaling. The calibration coefficients are generated through iterative optimization. The preset threshold is dynamically adjusted according to the current operating mode of the power grid, and different operating modes correspond to different stability and response speed requirements.

[0035] If the calibration process is triggered in multiple consecutive control cycles, the target local policy network parameters are slightly updated online using the current state data and the calibration instructions. At the same time, the currently generated calibration coefficients are associated with and stored with the corresponding state feature range. When the input state feature of the next cycle falls into this range, the cached calibration coefficients are directly reused to avoid repeated calculations. It should be noted that the deviation of the policy consistency index from the moving average exceeding a preset multiple of the moving standard deviation is used to detect sudden changes or abnormal fluctuations in policy behavior over time. On the other hand, if the L2 norm of the policy consistency index is greater than a preset threshold, it reflects that the current local policy output has deviated from the global policy expectation beyond the tolerance range in absolute terms, which is considered a serious deviation. At this point, a policy distillation calibration process centered on minimizing KL divergence must be initiated, and the target local policy network output is explicitly aligned with the probability distribution through a teacher-student framework. Therefore, the former focuses on dynamic stability anomalies and triggers preventive calibration, while the latter focuses on static deviation amplitude exceeding the limit and triggers reconstructive distillation calibration. The two are fundamentally different in terms of triggering conditions, judgment criteria, and calibration intensity.

[0036] If a single device's action space contains multiple independent control dimensions (such as active power and reactive power), then a calibration coefficient is calculated separately for each dimension. After calibration, the centralized scheduling command is generated by multiplying each dimension of the local power control command by the corresponding dimension's calibration coefficient, thereby achieving independent calibration of multiple degrees of freedom and improving control precision.

[0037] As an optional implementation, the step of issuing the centralized scheduling command to the corresponding battery energy storage device, and the battery energy storage device responding according to the centralized scheduling command, includes: Based on the device index number in the calibrated centralized scheduling instruction, the target recipient of the instruction is determined, and the calibrated centralized scheduling instruction is encapsulated into a standard instruction frame, which includes an instruction identifier, target device address, power setting value, timestamp, and checksum. The standard instruction frame is sent to the battery energy storage device corresponding to the target device address through a preset communication channel; After receiving the standard command frame, the battery energy storage device parses the standard command frame, extracts the power setting value, and adjusts the power according to the power setting value.

[0038] Specifically, the target device address is obtained by converting the device index number through a preset address mapping table; the standard instruction frame adopts a fixed-length structure, with each field arranged in a preset order, and the checksum is a check value calculated on the main body of the instruction frame according to a preset algorithm, which is used by the receiving end to verify the integrity of the data.

[0039] The standard instruction frame includes an additional instruction priority field, with the priority dynamically set based on the urgency level identifier in the power grid dispatch target data; high-priority instructions are sent first in the communication queue to ensure timely delivery of critical adjustment instructions; priority rules can be configured according to the power grid operation procedures.

[0040] The target device address is obtained by converting the device index number through a preset address mapping table; the standard instruction frame adopts a fixed-length structure, with each field arranged in a preset order, and the checksum is a check value calculated on the main body of the instruction frame according to a preset algorithm, which is used by the receiving end to verify the integrity of the data.

[0041] The standard instruction frame includes an additional instruction priority field, and the priority is dynamically set according to the urgency level identifier in the power grid dispatch target data; the priority rules can be configured according to the power grid operation procedures.

[0042] After verifying that the standard command frame check code is correct and the timestamp has not expired, the communication interface module of the battery energy storage device writes the power setpoint to the preset target register address of the power controller; after the power controller detects the target register update, it automatically starts the closed-loop regulation process and drives the converter to adjust the output power. After receiving the power setpoint, the power controller does not jump directly, but generates a smooth transition trajectory according to the preset power change rate limit and gradually adjusts the output power. The power change rate limit can be configured according to the equipment's thermal management capabilities or grid stability requirements to avoid power step shocks.

[0043] As an optional implementation, it further includes: a centralized control terminal receiving execution status feedback data returned by the battery energy storage device, and adopting a response strategy based on the execution status feedback data, specifically: Within a preset time window after the standard instruction frame is issued, the centralized control terminal receives the execution status feedback data returned by the battery energy storage device. The execution feedback data includes first execution feedback data and second execution feedback data: The first execution status feedback data includes: actual output power value, execution completion flag, and exception status code; The second execution feedback data includes: the current actual output power value, the remaining estimated adjustment time, and the current execution stage identifier; The centralized control terminal judges the execution status feedback data. If it is the first execution status feedback data, a response strategy is adopted according to the abnormal status code. If it is the second execution status feedback data, a dynamic evaluation is performed according to the current execution stage identifier, and a response strategy is adopted according to the dynamic evaluation result.

[0044] If the feedback data is from the first execution status, then a response strategy is adopted according to the abnormal status code, including: The abnormal status codes are classified into abnormal levels according to their severity, and the abnormal levels include warning level, degraded operation level, and shutdown protection level. The centralized control unit adopts a corresponding response strategy based on the anomaly level.

[0045] Specifically, within a preset time window after the instruction is issued, the execution status feedback data returned by the battery energy storage device is received. The execution status feedback data is actively reported by the battery energy storage device after completing power adjustment, and therefore there are two situations. The first situation is that the battery energy storage device has completed power adjustment according to the instruction within the preset time window. The second situation is that the battery energy storage device has not yet completed power adjustment within the preset time window. The execution status feedback data sent by the battery energy storage device is different in the two situations. For the first situation, the first execution status feedback data is returned. The first execution status feedback data includes the actual output power value, the execution completion flag and the abnormal status code. The actual output power value is the real-time sampling value of the device power sensor. The execution completion flag is a Boolean signal. The abnormal status code is one or more of the predefined error code set, which is used to indicate abnormal types such as overcurrent, communication interruption, and execution timeout. In the second scenario, the second execution status feedback data is returned, which includes the current actual output power value, the estimated remaining adjustment time, and the current execution stage identifier. The centralized control terminal uses this data to dynamically assess the execution progress and predict in advance whether intervention is needed. The abnormal status codes of the first execution feedback data are classified according to severity, including warning level, degraded operation level, and shutdown protection level. The centralized control terminal adopts different response strategies according to the abnormality level. For example, the warning level only logs, the degraded operation level triggers power redistribution, and the shutdown protection level triggers equipment isolation and alarm. For example, if the abnormal status code in the first execution status feedback data indicates execution failure or deviation exceeding the limit, the calibrated centralized scheduling instruction is marked as an instruction to be resent, and is preferentially reissued or the policy rollback mechanism is triggered in the next control cycle.

[0046] If the first abnormal status code indicates that the power execution deviation exceeds the allowable range, the power setting value of the calibrated centralized scheduling command will be reduced by a preset ratio and reissued in the next control cycle; the reduction ratio can be dynamically adjusted according to the degree of deviation to avoid repeated failures.

[0047] If multiple consecutive cycles fail or the anomaly level reaches the shutdown protection level, the policy rollback mechanism is triggered, switching the control mode of the corresponding device to the preset safety mode. In the safety mode, the device maintains zero power output or follows the local frequency droop curve, while reporting an isolation request to the centralized control terminal.

[0048] If a device frequently fails to execute commands, its health status index will be lowered, and its power allocation weight in the global action strategy distribution will be reduced in subsequent control cycles. At the same time, a local policy network set update process will be triggered to match the device with a more conservative local policy network, thereby achieving dynamic adaptation between control policies and device capabilities.

[0049] It should be noted that the aforementioned "response strategy based on severity of abnormal status codes" and the subsequent "response strategy for specific scenarios such as execution failure, deviation exceeding limits, and continuous failure" are not two independent mechanisms, but rather a hierarchical and complementary abnormal handling system. The former provides a general classification framework, including: warning level, degraded operation level, and shutdown protection level, while the latter is a concretization and operational refinement of typical abnormal semantics within this framework.

[0050] Example 2 Please refer to the attached document. Figure 2 The present invention also provides a centralized battery energy storage control system, comprising: The feature extraction module is used to acquire grid dispatch target data and real-time operating status data of each battery energy storage device, and extract global status feature vectors and device identity identifiers from the grid dispatch target data and the real-time operating status data. The policy generation module is used to input the global state feature vector into the central policy network to obtain the global action policy distribution; The policy matching module is used to determine the target local policy network from a preset set of local policy networks based on the combination of the global action policy distribution and the device identity identifier. The instruction calibration module is used to input the global state feature vector into the target local policy network to obtain local power control instructions, and generate calibrated centralized scheduling instructions based on the global action policy distribution and the local power control instructions. The instruction execution module is used to send the calibrated centralized scheduling instruction to the corresponding battery energy storage device, and the battery energy storage device performs the power response action.

[0051] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A centralized battery energy storage control method, characterized in that, Includes the following steps: S1: Obtain grid dispatch target data and real-time operating status data of each battery energy storage device, and extract global status feature vector and device identity identifier from the grid dispatch target data and the real-time operating status data; S2: Input the global state feature vector into the central policy network to obtain the global action policy distribution; S3: Determine the target local policy network from the preset local policy network set based on the global action policy distribution and the device identity identifier; S4: Input the global state feature vector into the target local policy network to obtain the local power control instruction, and generate a centralized scheduling instruction based on the global action policy distribution and the local power control instruction; S5: The centralized scheduling instruction is sent to the corresponding battery energy storage device, and the battery energy storage device responds according to the centralized scheduling instruction.

2. The centralized battery energy storage control method according to claim 1, characterized in that, The power grid dispatch target data includes: target active power value, target reactive power value, frequency regulation coefficient, and start and end times of the dispatch time window; The real-time operating status data includes: device number, current state of charge, terminal voltage, charging and discharging current, battery temperature, health status index, communication delay duration, and maximum adjustable power margin.

3. The centralized battery energy storage control method according to claim 1, characterized in that, The step of inputting the global state feature vector into the central policy network to obtain the global action policy distribution includes: A central policy network is constructed, which includes an input layer, a hidden layer, an activation layer, and an output layer. The hidden layer consists of three fully connected neural networks with 512, 256, and 128 neurons in each layer, respectively. The activation function of the activation layer is the ReLU function, and the output layer is mapped to a preset action space using the Tanh function. The central policy network is trained to obtain a trained central policy network. The global state feature vector is input into the trained central policy network, which outputs continuous action probability density function parameters, and these parameters are used as the global action policy distribution.

4. The centralized battery energy storage control method according to claim 1, characterized in that, The step of determining the target local policy network from a preset set of local policy networks based on the global action policy distribution and the device identity includes: The device identification is indexed and encoded, an index mapping table is preset, and the device index number is obtained according to the index code and the index mapping table; Extract the action expectation component corresponding to the device index number, and generate a device-level policy guidance signal based on the action expectation component; Based on the device index number and the device-level policy guidance signal, an enhanced policy matching key value is obtained. The enhanced policy matching key value is then input into a pre-trained principal component analysis dimensionality reduction layer to reduce the dimensionality of the policy matching key value, resulting in a dimensionality-reduced enhanced policy matching key value. In the set of local policy networks, find the local policy network that has the smallest Euclidean distance to the matching key value of the dimensionality-reduced enhanced policy, and determine the found local policy network as the target local policy network.

5. The centralized battery energy storage control method according to claim 4, characterized in that, The step of searching within the local policy network set for the local policy network that has the smallest Euclidean distance to the matching key value of the dimensionality-reduced policy, and determining the found local policy network as the target local policy network, includes: A KD-tree index structure is pre-constructed, and the KD-tree nearest neighbor search algorithm is used to find the registration key value with the smallest distance to the policy matching key value after dimensionality reduction. The local policy network corresponding to the registration key value is then determined as the target local policy network.

6. The centralized battery energy storage control method according to claim 1, characterized in that, The step of inputting the global state feature vector into the target local policy network to obtain local power control instructions, and generating centralized scheduling instructions based on the global action policy distribution and the local power control instructions, includes: The target local policy network is a multi-layer fully connected structure, including an input layer, a hidden layer, and an output layer; The global state feature vector is input into the target local policy network, and the local power control command is output through the output layer. Extract the mean and variance components of the policies of the devices corresponding to the target local policy network from the global action policy distribution; Calculate the deviation vector between the local power control command and the mean component of the strategy, and normalize the deviation vector according to the variance component of the strategy to generate a strategy consistency index. The moving average and moving standard deviation of the strategy consistency index generated in several control cycles are calculated. If the difference between the strategy consistency index of the current cycle and the moving average exceeds a preset multiple of the moving standard deviation, it is considered as an abnormal policy deviation that triggers the calibration mechanism in advance. If the L2 norm of the policy consistency index in the current period is greater than the preset threshold, the policy distillation calibration process is initiated. The global action policy distribution is taken as the teacher policy and the target local policy network is taken as the student policy. The calibration coefficient is generated by minimizing the KL divergence loss function. The local control command is multiplied by the calibration coefficient to generate the calibrated centralized scheduling command.

7. The centralized battery energy storage control method according to claim 1, characterized in that, The step of issuing the centralized scheduling command to the corresponding battery energy storage device, and the battery energy storage device responding according to the centralized scheduling command, includes: Based on the device index number in the calibrated centralized scheduling instruction, the target recipient of the instruction is determined, and the calibrated centralized scheduling instruction is encapsulated into a standard instruction frame, which includes an instruction identifier, target device address, power setting value, timestamp, and checksum. The standard instruction frame is sent to the battery energy storage device corresponding to the target device address through a preset communication channel; After receiving the standard command frame, the battery energy storage device parses the standard command frame, extracts the power setting value, and adjusts the power according to the power setting value.

8. The centralized battery energy storage control method according to claim 1, characterized in that, Also includes: The centralized control terminal receives execution status feedback data returned by the battery energy storage device and adopts a response strategy based on the execution status feedback data, specifically: Within a preset time window after the standard instruction frame is issued, the centralized control terminal receives the execution status feedback data returned by the battery energy storage device. The execution feedback data includes first execution feedback data and second execution feedback data: The first execution status feedback data includes: actual output power value, execution completion flag, and exception status code; The second execution feedback data includes: the current actual output power value, the remaining estimated adjustment time, and the current execution stage identifier; The centralized control terminal judges the execution status feedback data. If it is the first execution status feedback data, a response strategy is adopted according to the abnormal status code. If it is the second execution status feedback data, a dynamic evaluation is performed according to the current execution stage identifier, and a response strategy is adopted according to the dynamic evaluation result.

9. A centralized battery energy storage control method according to claim 8, characterized in that, If the data is the first execution status feedback data, then a response strategy is adopted according to the abnormal status code, including: The abnormal status codes are classified into abnormal levels according to their severity, and the abnormal levels include warning level, degraded operation level, and shutdown protection level. The centralized control unit adopts a corresponding response strategy based on the anomaly level.

10. A centralized battery energy storage control system, characterized in that, include: The feature extraction module is used to acquire grid dispatch target data and real-time operating status data of each battery energy storage device, and extract global status feature vectors and device identity identifiers from the grid dispatch target data and the real-time operating status data. The policy generation module is used to input the global state feature vector into the central policy network to obtain the global action policy distribution; The policy matching module is used to determine the target local policy network from a preset set of local policy networks based on the combination of the global action policy distribution and the device identity identifier. The instruction calibration module is used to input the global state feature vector into the target local policy network to obtain local power control instructions, and generate calibrated centralized scheduling instructions based on the global action policy distribution and the local power control instructions. The instruction execution module is used to send the calibrated centralized scheduling instruction to the corresponding battery energy storage device, and the battery energy storage device performs the power response action.

Images