Green energy grid-connected storage collaborative control method based on network-type energy storage

Abstract

The application provides a green energy grid-connected storage collaborative control method based on networked energy storage, and relates to the technical field of power systems, which obtains electrical operation data and object attribute identifiers of each controlled object in a source storage system, and constructs a time series data stream; inputs the time series data stream into a capsule network model, respectively encodes the phase attribute into an angle component of a primary capsule vector, and encodes the amplitude attribute into a modulus length component; iteratively updates routing weights between the primary capsule vector and a high-level representation capsule and converges, determines a current oscillation mode, and obtains the correlation strength of each controlled object relative to the mode; accordingly generates a parameter set containing networked energy storage impedance adjustment parameters and power distribution parameters, and converts the parameter set into an interface converter control reference quantity, so as to realize grid-connected oscillation suppression and power distribution collaboration.

Description

Technical Field

[0001] This application relates to the field of power system technology, and more specifically, to a collaborative control method for green energy grid-connected storage based on grid-connected energy storage. Background Technology

[0002] With the large-scale integration of green energy sources such as photovoltaics and wind power through power electronic interfaces, the source-storage-grid system exhibits characteristics such as multiple power sources and converters in parallel, increased network impedance uncertainty, significant coupling in control links, and rapid shifts in operating conditions. In weak grids or high-penetration scenarios, grid-based energy storage typically needs to undertake tasks such as voltage support, phase / frequency establishment, transient stability maintenance, and coordination with parallel systems. The system is prone to coupled oscillations caused by the interaction between the controller and network impedance, and the dominant oscillation mode may shift with changes in topology, equivalent impedance at the grid connection point, load disturbances, and the operating point of the parallel objects.

[0003] In existing technologies, a common approach is to perform oscillation analysis and suppression based on physical models. This includes constructing small-signal state-space models or equivalent impedance models, and combining eigenvalue analysis, impedance stability criteria, Prony-type mode identification, and spectral / time-frequency analysis to identify oscillation bands or modes. Based on this, virtual impedance, damping injection, droop coefficient, or power allocation strategies are then tuned. While this approach has clear physical meaning when the model is known and the parameters are relatively stable, in real-world scenarios with multiple parallel sources, diverse control strategies, and frequent changes in network impedance, model building and parameter identification are costly. Furthermore, it is highly sensitive to unmodeled dynamics, measurement noise, asynchronous sampling, and transient disturbances, which can easily lead to delayed mode identification or overly conservative tuning results, making it difficult to maintain stable closed-loop coordination during rapid changes in operating conditions.

[0004] Another approach employs data-driven or artificial intelligence models to learn and discriminate oscillation characteristics. For example, convolutional neural networks or recurrent neural networks can be used to extract features and classify states from time-series data such as voltage and current, thereby triggering control mode switching or parameter correction. This type of method reduces the reliance on explicit physical modeling to some extent, but existing conventional models still have insufficient adaptability in grid-connected control scenarios: the downsampling / pooling operations commonly used by convolutional networks to obtain invariance may weaken the retention of fine-grained information sensitive to phase, making it difficult to achieve high-fidelity representation of the vector characteristics of coupled oscillations; although recurrent networks can model time-series correlations, their internal representations are mostly black-box scalar states, making it difficult to provide contribution positioning criteria that can be used for collaborative control allocation while identifying oscillation modes. This leads to impedance reconstruction and power allocation still tending to be uniformly tuned or based on empirical rules, which can easily result in problems such as control antagonism between parallel objects, insufficient vibration suppression, or excessive conservatism.

[0005] Therefore, the technical problem that the existing technology urgently needs to solve is how to achieve high-fidelity vector-level extraction of coupled oscillation characteristics in a complex power grid environment with multiple sources in parallel, and accurately locate the oscillation contribution source while identifying the oscillation mode, so as to carry out targeted impedance reconstruction and power distribution. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this application provides a collaborative control method for green energy grid-connected storage based on grid-connected energy storage, including:

[0007] Obtain electrical operation data of each controlled object in the source storage system and object attribute identifiers corresponding to each controlled object, and construct a time-series data stream based on the electrical operation data;

[0008] The time-series data stream is input into the capsule network model and mapped to obtain a primary capsule vector set; wherein, the phase attribute of the electrical operation data is encoded as the angle component of the primary capsule vector, and the amplitude attribute of the electrical operation data is encoded as the magnitude component of the primary capsule vector;

[0009] The routing weights are iteratively updated between the primary capsule vector set and the higher-level representation capsule, and convergence is determined when a preset convergence condition is met. The current oscillation mode of the source-storage system is determined based on the converged higher-level representation capsule, and the correlation strength of each controlled object relative to the current oscillation mode is determined based on the converged routing weights.

[0010] Based on the correlation strength and the object attribute identifier, a parameter set is generated, which includes impedance regulation parameters for grid-type energy storage objects and power allocation parameters for grid-type energy storage objects and / or power generation objects; and the parameter set is converted into the interface converter control reference quantity of the corresponding controlled object.

[0011] Optionally, the angular components encoded as the primary capsule vector include:

[0012] An instantaneous space vector transformation is performed on the three-phase voltage data in the time-series data stream to obtain orthogonal components in the stationary coordinate system;

[0013] Based on the orthogonal components, the instantaneous phase angle is determined, and the instantaneous phase angle is used as the input of the angle components;

[0014] Specifically, for the sampling time deviation of different controlled objects, the phase compensation amount is determined based on the sampling time deviation, and the instantaneous phase angle is compensated and aligned.

[0015] Optionally, the iterative update of route weights includes:

[0016] For each primary capsule vector, the set of routing weights is sparsified so that only a preset number of maximum weights are retained in the set of routing weights, and the remaining weights are either set to zero or scaled according to a preset decay factor.

[0017] Furthermore, when the change in routing weight exceeds a preset threshold for two consecutive control cycles, the parameter set is generated using the routing weight of the previous control cycle, until the change in routing weight continuously meets the preset stability condition.

[0018] Optionally, generating the impedance regulation parameters for the grid-type energy storage object includes:

[0019] The dominant oscillation frequency band is determined based on the current oscillation mode;

[0020] Generate a set of frequency band-selective impedance adjustment parameters, wherein the damping injection parameter corresponding to the dominant oscillation frequency band is configured to be greater than or equal to a preset reference damping threshold, and the voltage support parameter corresponding to the fundamental frequency band is configured to be greater than or equal to a preset voltage support lower limit.

[0021] Optionally, the method further includes:

[0022] Based on the output of the converged high-level representation capsule, a reconstructed time-series data stream is generated through a reconstruction network.

[0023] Calculate the reconstruction residual between the reconstructed time-series data stream and the original time-series data stream;

[0024] When the reconstruction residual exceeds a preset threshold, the current oscillation mode is determined to be an unknown mode, and the parameter set is switched to a preset conservative damping parameter set.

[0025] Optionally, the set of impedance adjustment parameters for generating band selectivity includes:

[0026] Select the converged high-level characterization capsule vector corresponding to the dominant oscillation frequency band, and based on the angle and magnitude values ​​of the high-level characterization capsule vector, map the high-level characterization capsule vector to endpoints on the complex plane and construct the endpoint time-series trajectory.

[0027] Calculate the angle change and magnitude change of the endpoint time-series trajectory in adjacent control cycles, respectively.

[0028] The oscillation mechanism is identified based on the angle change and the modulus change, and the parameter configuration mode is selected in the impedance adjustment parameters according to the identification result. Among them, negative damping divergent oscillation corresponds to high damping mode, and forced source periodic oscillation corresponds to high inertia mode.

[0029] Optionally, calculating the angle change and magnitude change of the endpoint time-series trajectory in adjacent control cycles includes:

[0030] Extract the angle values ​​of the high-level character capsule vectors from two adjacent control cycles, and calculate the absolute value of their difference as the angle change.

[0031] Extract the magnitude value of the high-level characterization capsule vector between two adjacent control cycles, calculate the difference between the magnitude value of the next control cycle and the magnitude value of the previous control cycle as the magnitude increment, and calculate the cumulative value of the magnitude increment within a preset number of consecutive cycles.

[0032] Optionally, the high-damping mode includes:

[0033] When the angle change is not less than the preset angle threshold within a preset number of consecutive cycles, and the cumulative value of the modulus increment is not less than the preset growth threshold, the oscillation mechanism of the current dominant oscillation frequency band is determined to be negative damped divergent oscillation.

[0034] The impedance adjustment parameters include damping injection parameters, which include active damping gain; the active damping gain is configured to be no less than a preset strong damping threshold.

[0035] Optionally, the high inertia mode includes:

[0036] When the angle change is not less than a preset angle threshold within a preset number of consecutive cycles, and the magnitude of the high-level characterizing capsule vector remains within the preset upper and lower limits of the magnitude within the preset number of consecutive cycles, the oscillation mechanism of the current dominant oscillation frequency band is determined to be forced source periodic oscillation.

[0037] The impedance adjustment parameters include inertia-related parameters; the inertia-related parameters are configured to be greater than or equal to a preset strong inertia threshold.

[0038] Optionally, the sparsification process performed on the set of routing weights corresponding to each primary capsule vector includes:

[0039] Based on the object type identifier and parallel group identifier in the object attribute identifier, mutual exclusion sparse constraints are applied to the grid-type energy storage objects belonging to the same parallel group, so that only non-zero route weights pointing to a single high-rise characterization capsule are retained in the route weight set, and the remaining route weights are limited to zero or limited to a preset minimum weight upper limit.

[0040] Intra-group aggregation is performed on the routing weight set corresponding to the grid-type energy storage objects in the same parallel group to obtain the group-level routing preference vector. Based on the group-level routing preference vector, the single high-level characterization capsule is determined so that the mutual exclusion sparse constraints of each grid-type energy storage object in the parallel group point to the same high-level characterization capsule in the same control cycle.

[0041] Based on the primary capsule vector of the parallel group intra-network energy storage object, calculate the common mode representation vector and differential mode representation vector within the group. In the sparsification process, apply mutually exclusive sparsity constraints to the routing weight sets corresponding to the common mode representation vector and the differential mode representation vector, respectively. The routing weight set corresponding to the differential mode representation vector retains only the non-zero routing weights that point to the intra-group representation capsules used to represent the coupling modes within the parallel group.

[0042] Compared with existing technologies, this application creatively encodes the phase attribute of electrical operation data as the angular component of the primary capsule vector and the amplitude attribute as the magnitude component by introducing a capsule network model and establishing a physical mapping mechanism between electrical quantities and capsule vectors. This encoding method allows the feature transmission within the neural network to retain the vector geometric properties and rotation invariance of the AC signal, enabling the extraction of high-fidelity vector features containing phase coupling information from the time-series data stream. This significantly improves the recognition accuracy of various weak or strongly coupled oscillation modes in complex power grid environments and overcomes the defect of traditional scalar neural networks that destroy the phasor structure of electrical signals.

[0043] This application utilizes the unique dynamic routing mechanism of capsule networks to directly quantify the correlation strength between each controlled object and the current oscillation mode during the iterative clustering process from primary capsules to higher-level representation capsules. This mechanism enables the system to simultaneously and accurately locate the oscillation contribution source during the same inference process of oscillation mode identification, without introducing additional causal analysis algorithms. This achieves the integration of identification and source tracing, greatly improving the response speed and computational efficiency of the control system.

[0044] Based on the aforementioned vector-level feature extraction and precise source tracing capabilities, this application can generate differentiated parameter sets according to the specific correlation strength of each controlled object, and perform targeted impedance adjustment (such as injecting specific frequency band damping) or power redistribution on grid-type energy storage objects that act as oscillation sources. This point-to-point approach avoids the system response lag or misoperation of non-oscillation source devices that may be caused by traditional global control, thereby effectively suppressing broadband oscillations while maximizing the transmission efficiency and dynamic performance of green energy grid-connected systems. Attached Figure Description

[0045] Figure 1 A flowchart illustrating the green energy grid-connected storage collaborative control method based on grid-type energy storage provided in this application embodiment;

[0046] Figure 2 A flowchart illustrating an iterative method for updating route weights provided in this application embodiment;

[0047] Figure 3This is a flowchart illustrating a method for generating impedance adjustment parameters, as provided in an embodiment of this application. Detailed Implementation

[0048] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0049] See Figure 1 The diagram shown is a flowchart of a green energy grid-connected storage collaborative control method based on grid-type energy storage provided in this application embodiment, including steps S101 to S104, wherein:

[0050] S101: Obtain the electrical operation data of each controlled object in the source storage system and the object attribute identifier corresponding to each controlled object, and construct a time-series data stream based on the electrical operation data;

[0051] S102: Input the time-series data stream into the capsule network model and map it to obtain a primary capsule vector set; wherein, the phase attribute of the electrical operation data is encoded as the angle component of the primary capsule vector, and the amplitude attribute of the electrical operation data is encoded as the magnitude component of the primary capsule vector;

[0052] S103: Iteratively update the routing weights between the primary capsule vector set and the higher-level representation capsule, and determine convergence when the preset convergence condition is met; determine the current oscillation mode of the source-storage system based on the converged higher-level representation capsule, and determine the correlation strength of each controlled object relative to the current oscillation mode based on the converged routing weights;

[0053] S104: Based on the correlation strength and the object attribute identifier, generate a parameter set, the parameter set including impedance regulation parameters for grid-type energy storage objects and power allocation parameters for grid-type energy storage objects and / or power generation objects; and convert the parameter set into the interface converter control reference quantity of the corresponding controlled object.

[0054] Regarding the above S101:

[0055] In one embodiment, the energy source and energy storage system includes a generation-side grid-connected interface and a grid-connected energy storage interface connected to the power grid via a common coupling point. The controlled objects within the energy source and energy storage system can be power electronic interface units with independent control interfaces. The controlled objects include at least grid-connected energy storage objects and generation objects; wherein, the grid-connected energy storage object is, for example, an energy storage converter or its equivalent control unit, and the generation object is, for example, a photovoltaic inverter, a wind power converter, or its equivalent control unit. To facilitate unified modeling of the parallel operation of multiple objects and subsequent distribution of collaborative parameters, each controlled object is configured with an object attribute identifier. The object attribute identifier is used to characterize the category and organizational relationship of the controlled object, including at least the object type identifier and the parallel group identifier. The object type identifier is used to distinguish between grid-type energy storage objects and power generation objects, and the parallel group identifier is used to characterize that multiple controlled objects belong to the same electrical connection cluster in physical topology, such as being connected to the same AC bus, the low-voltage side of the same step-up transformer, or the same parallel branch. This means that the objects within the group have a small electrical distance and strong coupling characteristics, thus providing a semantic basis for the object dimension for subsequent parameter synthesis, common mode / differential mode analysis, and collaborative constraints.

[0056] In this embodiment, acquiring electrical operation data includes collecting measurements characterizing the AC vector state from the grid-connected or output side of each controlled object. The electrical operation data includes at least time-series measurement data reflecting the vector state of the output voltage and / or output current of the controlled object, such as three-phase voltage data, three-phase current data, or orthogonal component data obtained by coordinate transformation from three-phase sampling. This data can be acquired through voltage and current sampling circuits, transformers, or metering modules. The acquisition location can be a common coupling point, parallel branch node, or the output terminal of the interface converter of each controlled object. To ensure the capture of oscillation characteristics over a wide frequency range, the sampling frequency should meet the sampling theorem requirements for the target oscillation frequency band (covering subsynchronous oscillation to high-frequency resonance). To ensure a consistent extraction of phase and amplitude attributes, the electrical operation data can undergo necessary calibration and preprocessing after acquisition, including range normalization, zero-drift compensation, outlier removal, or amplitude limiting, making the data from different controlled objects under the same dimension comparable.

[0057] Constructing a time-series data stream based on electrical operation data includes: using a central controller or edge computing unit to generate a corresponding time-series record for each controlled object within each control cycle or preset sampling window, and binding the time-series record with the object attribute identifier. The time-series record includes at least timestamp information, object identifier information, and electrical operation data, enabling the time-series data stream to be represented as an object-oriented data sequence that progresses over time.

[0058] In one optional implementation, the timestamp is provided by a unified time reference from the upper-level coordinating controller, or by a synchronization reference from the time synchronization module. When there is a sampling time deviation between different controlled objects, the data of each controlled object can be aligned to a unified time step. Alignment can be achieved by resampling, interpolation, keeping the most recent valid value, or aligning by nearest neighbor, so as to form a continuous time-series data stream that satisfies subsequent processing. When there is missing or abnormal data, data validity markers can be recorded and filling or discarding strategies can be adopted to ensure the availability and stability of the time-series data stream.

[0059] Regarding S102 above:

[0060] S102 involves constructing inference samples from the time-series data stream formed in S101, using the control cycle as the beat, and inputting the inference samples into the capsule network model to output a primary capsule vector set. Specifically, the cooperative controller extracts electrical operation data sequences for each controlled object according to a sliding time window to form time segments. The time segments contain at least the sampling sequences of three-phase voltage and / or three-phase current, or the sequence of orthogonal components in the stationary coordinate system obtained by transforming three-phase quantities. To balance the real-time performance of grid-connected control with the discernibility of oscillation characteristics, the sampling rate, window length, and sliding step size can be jointly set according to the target oscillation frequency band and the control cycle; for example, the control cycle is 5ms to 20ms, the sampling rate is 1kHz to 10kHz, the time window length is 128 to 512 sampling points, and the time window sliding step size is 1 to 4 control cycles. In one example scenario, the source-storage system operates in grid-connected mode at 50Hz, the control cycle of the co-controller is 10ms, and the electrical quantity sampling rate is 2kHz. Each inference sample selects a time window of 256 points with a sliding step of 2 control cycles, allowing the model to continuously output the primary capsule vector set without changing the control cycle. The input samples undergo dimensional unification processing consistent with online control, converting voltage and current to per-unit values ​​according to their respective rated values. The phase sequence is also processed to avoid numerical jumps caused by phase boundary crossings. The phase reference can be the common coupling point reference phase or the synchronous reference phase within the co-controller, and the object attribute identifier of the same controlled object is bound to the sample so that the model output can trace back to the specific controlled object, its type, and parallel organization relationship.

[0061] In this embodiment, the capsule network model includes at least a front-end temporal mapping module and a primary capsule generation module, which are connected in series. The front-end temporal mapping module is used to extract local transient and oscillatory morphological features from the input temporal segments, and can be composed of multiple levels of one-dimensional convolutional layers to form a temporal feature map. In one example implementation, the front-end temporal mapping module includes two levels of one-dimensional convolutional layers. The first convolutional layer has 32 channels and a kernel length of 3 to 7, and the second convolutional layer has 64 channels and a kernel length of 3 to 7. Normalization and nonlinear activation are configured between layers to stabilize the feature scale under different operating conditions. The primary capsule generation module is used to rearrange the temporal feature map into multiple capsule input units according to a preset group, and perform linear mapping on each capsule input unit to obtain a primary capsule vector; multiple primary capsule vectors constitute a primary capsule vector set. For example, the number of primary capsules is 16 to 64, and the vector dimension of each primary capsule is 8 to 16 dimensions; in one example implementation, the number of primary capsules is 32, and the vector dimension of each primary capsule is 10 dimensions.

[0062] In one embodiment, each primary capsule vector is generated using a hybrid method of explicit construction and learning component splicing, and the angle and magnitude components are explicitly taken from the phase-amplitude carrying subspace. Specifically, for each controlled object, the instantaneous phase angle and instantaneous amplitude are determined based on the orthogonal components of the stationary coordinate system at each sampling moment, and a phase sequence and amplitude sequence are formed within the time window as the basis for physical encoding; wherein, the instantaneous amplitude is mapped to an amplitude control quantity according to the per-unit value, so that the amplitude control quantity and the per-unit amplitude maintain a monotonic correspondence and are limited to a preset range. In an example implementation, the upper limit of the amplitude can be 1.2pu to 1.5pu. When the per-unit amplitude does not exceed the upper limit of the amplitude, the amplitude control quantity increases with the increase of the per-unit amplitude; when the per-unit amplitude exceeds the upper limit of the amplitude, the amplitude control quantity is limited to the upper limit value. Subsequently, a physically encoded two-dimensional vector is generated within a two-dimensional orthogonal subspace. The direction of this two-dimensional vector is determined by the instantaneous phase angle, and its distance from the origin is equal to the amplitude control value. This causes the two-dimensional vector to rotate with changes in the phase angle, and its length monotonically increases or decreases with changes in the amplitude control value. Thus, the angular component of the primary capsule vector is determined by the instantaneous phase angle, and the magnitude component of the primary capsule vector is determined by the length of the physically encoded two-dimensional vector, ensuring that the phase semantics and amplitude semantics have clear engineering specifications.

[0063] Meanwhile, the feature map output by the front-end time-series mapping module is mapped by the capsule input unit to obtain the learning component. The learning component is used to characterize auxiliary features related to grid-connected transients and waveform morphology differences and is used for subsequent routing processing. To avoid the learning component changing the representation of the physical amplitude by the modulus component, an amplitude constraint is applied to the learning component in one embodiment, so that the norm of the learning component does not dominate the length semantics of the physical encoded two-dimensional vector. For example, the learning component can be normalized and multiplied by a preset scaling factor, and the scaling factor can be configured to be no greater than 0.1 to 0.3, so that the learning component remains a complement to the physical encoded two-dimensional vector rather than a replacement in numerical terms. Then, the physical encoded two-dimensional vector and the learning component are concatenated dimensionally to form the primary capsule vector, where the first two dimensions correspond to the physical encoded two-dimensional vector, and the remaining dimensions correspond to the learning component. Through the above explicit construction and constraint concatenation, the angle component and the modulus component have stable physical meanings and do not suffer from non-monotonic distortion of physical amplitude due to the nonlinearity of conventional capsule compression. Thus, the primary capsule vector set can maintain both high-fidelity vector-level quantitative semantics and the ability to express transient oscillation morphology differences in subsequent routing aggregation.

[0064] In one implementation, the parameters of the capsule network model are determined through offline training and then fixed for use during online operation. Training data is obtained by slicing historical operational data, simulation data, or hardware-in-the-loop data according to the aforementioned time window and step size rules, with the input normalization caliber consistent with the online approach. Supervision information for training samples can be formed from offline identification results, simulation condition labels, or operational event annotations, used to indicate the oscillation state category or modal category corresponding to the sample. The training objective employs a combination of classification constraints and reconstruction constraints to simultaneously ensure the separability of the capsule output for state categories and the fidelity of the input vector features. The classification constraint can use interval loss to differentiate the confidence intensity of different modal categories in the high-level representation space, while the reconstruction constraint can use mean squared error or mean absolute error to constrain the reconstruction error of the decoder network for the input time series segments. The decoder network can consist of two fully connected layers and output reconstructed time series segments of the same dimension as the input. In one example implementation, the fully connected layer sizes are 128 and 256, and the output reconstructs orthogonal component sequences or phase-amplitude sequences. The weights of classification and reconstruction constraints can be configured according to the emphasis on "modal separability" and "vector fidelity" in the control scenario. For example, the weights of classification and reconstruction constraints can be in a ratio between 1:0.2 and 1:1. In an example training configuration, the Adam optimizer is used with a learning rate of 1×10⁻⁶. -4 ~3×10 -3The batch size is 32–128, the number of training rounds is 40–120, and training stops when the validation set loss no longer decreases after several consecutive rounds. The number of dynamic routing iterations is 3–5 to balance convergence and computational cost. After training, the model weights and normalized parameters are saved. In the online phase, only forward inference is performed to output the initial capsule vector set and proceed to the subsequent routing weight update and co-parameter generation process.

[0065] Regarding the above S103:

[0066] S103 includes performing dynamic routing calculations between the primary capsule vector set and the higher-level representation capsules, enabling object-level, local vectorized representations to adaptively aggregate into system-level vectorized representations capable of characterizing the oscillation modes of the source-storage system within the control cycle scale. Specifically, the cooperative controller receives the primary capsule vector set output by S102 in each control cycle and configures a set of "voting transformation parameters" corresponding one-to-one with the higher-level representation capsules for each primary capsule vector, used to map the primary capsule vector into a prediction vector pointing to different higher-level representation capsules. The prediction vector reflects the directional contribution of the controlled object to each candidate oscillation mode within the current window. To facilitate engineering implementation, the voting transformation parameters can be grouped and stored and retrieved according to object type identifiers, allowing controlled objects of the same type to share the same set of transformation parameters or share some of the parameters, thereby reducing the parameter scale and computational burden of online inference; for example, grid-type energy storage objects and power generation objects can be respectively assigned to two sets of transformation parameter sets, and each set can be configured with a set of linear mapping weights for each higher-level representation capsule to generate the corresponding prediction vector.

[0067] In one example configuration, the number of high-level characterization capsules can be 6 to 12, covering common modal families in source-storage systems, including stationary modes, low-frequency coupled oscillation modes, subsynchronous / supersynchronous oscillation modes, and grid-connected resonant-related modes. The magnitude of the output vector of each high-level characterization capsule is used to characterize the confidence strength of that mode within the current control cycle, and the direction of the output vector is used to carry the phase attitude or equivalent phasor direction information of that mode, so that subsequent control parameter generation maintains vector consistency sensitive to phase. For example, the number of high-level characterization capsules can be set to 8, and the output vector dimension of each high-level characterization capsule can be set to 16 dimensions, so that it can simultaneously characterize the "strength" and "attitude" information of the dominant mode and retain the necessary redundant dimensions to adapt to waveform morphology differences under different operating conditions.

[0068] In this implementation, the iterative update of routing weights includes routing prior initialization, weight normalization allocation, aggregation to generate high-level inputs, output normalization compression, and consistency feedback update. First, a routing prior is established and initialized to a neutral state between each primary capsule vector and each high-level representation capsule, ensuring that the initial routing weights are unbiased across candidate high-level representation capsules. Then, in each iteration, the routing weights of the same primary capsule vector pointing to different high-level representation capsules are normalized, forming an "allocation ratio" that satisfies non-negativity and overall consistency constraints. Under this allocation ratio, the predicted vectors generated by the primary capsule vector are weighted and summed to obtain the aggregated input vector for each high-level representation capsule. Vector normalization-type nonlinear compression is performed on the aggregated input vector to limit the output vector magnitude within a preset range and retain directional information, thus preventing numerical divergence caused by the aggregation amplitude increasing with the input size. Next, a consistency metric is calculated between each predicted vector and the corresponding high-level representation capsule output vector. This consistency metric is then fed back to update the routing prior, ensuring that paths aligned with the current aggregation result receive higher routing weights in the next iteration, while paths deviating significantly in direction receive lower weights. The consistency metric can be implemented using vector direction similarity, such as the inner product or normalized inner product of two vectors, to guarantee the monotonic relationship that "higher consistency leads to a more significant weight increase." The above iterations can be executed cyclically within a preset maximum number of iterations. In one example configuration, the maximum number of iterations is set to 3 to meet the real-time constraint of a 10ms control cycle. In scenarios with sudden changes in operating conditions, significant weak network oscillations, or strong modal competition, the maximum number of iterations can be set to 5 to improve aggregation stability.

[0069] In this implementation, "determining convergence when the preset convergence condition is met" is used to ensure that the route calculation outputs a stable and repeatable modal representation under real-time control constraints. The convergence condition can be jointly limited by the stability of the route weight changes and the stability of the output of the higher-level representation capsule.

[0070] For example, a routing weight stability threshold and a vector stability threshold can be defined separately: when the maximum change in the routing weight set does not exceed the preset routing weight stability threshold in two consecutive iterations, and the change in the magnitude and direction of the corresponding high-level representation capsule output vector does not exceed the preset vector stability threshold, the routing process is considered to have converged. If the above stability conditions are not met even after reaching the maximum number of iterations, the routing weight and high-level representation capsule output of the last iteration are used as the convergence result output. Taking an example scenario with a control period of 10ms, the routing weight stability threshold can be set to 0.005~0.02, and the vector stability threshold can be set to 0.002~0.01, requiring simultaneous satisfaction in two consecutive iterations to reduce the risk of routing jitter under critical oscillation conditions being amplified and transmitted to subsequent control parameters. Optionally, when online calculation detects that convergence has not been achieved and a sudden change occurs in the routing weight, the modal output and correlation strength output of the current period can be restricted to the stable result of the previous control period or restricted to the result of the last iteration to maintain the continuity of control parameter generation.

[0071] In this implementation, determining the current oscillation mode of the source-storage system based on the converged high-level characterization capsule includes selecting the dominant mode from the converged output and providing a mode identifier.

[0072] Specifically, the confidence strength of candidate modes is determined based on the magnitude of the output vector of each high-level representation capsule, and the mode corresponding to the high-level representation capsule with the largest magnitude is selected as the current oscillation mode. To avoid weak perturbations being misjudged as oscillations, a lower confidence limit for the mode can be further set: when the maximum magnitude is not lower than the lower confidence limit, the corresponding oscillation mode is output; otherwise, a stationary mode or a weakly oscillating mode is output as the current oscillation mode. For example, the lower confidence limit for the mode can be 0.2 to 0.4, where a higher threshold can be used in weak network or high-noise scenarios to suppress false triggering; the mode identifier can correspond to the mode category label or mode prototype library established during the offline training phase, so that the output of the "current oscillation mode" has interpretable engineering semantics.

[0073] In this implementation, determining the correlation strength of each controlled object relative to the current oscillation mode based on the converged routing weights includes aggregating the routing weights by object dimension to form a comparable contribution metric. Specifically, for each controlled object, the routing weights pointing from one or more primary capsule vectors corresponding to that object to the higher-level characterization capsule corresponding to the current oscillation mode are aggregated to obtain an object-level routing contribution value. Aggregation can be performed using averaging, weighted averaging, or weighted aggregation based on the primary capsule vector magnitude, so that routing paths dominated by strong features within the same object have higher weights on the contribution. To work with object attribute identifiers, object type identifiers can also be used as weighting factors during aggregation, ensuring that grid-type energy storage objects and power generation objects meet engineering comparability constraints in terms of contribution caliber. Subsequently, the object-level routing contribution values ​​of all controlled objects are normalized to obtain the correlation strength of each controlled object relative to the current oscillation mode. The correlation strength is a scalar between 0 and 1 and meets the comparability constraint of consistent total amount.

[0074] For example, the source-storage system includes 4 parallel grid-type energy storage objects and 6 power generation objects, with a control cycle of 10ms. The primary capsules have 32 outputs per object. The object-level routing contribution value can be weighted and normalized by "the weight of all primary capsules of this object pointing to the dominant high-level characterization capsule". This allows the correlation strength of the output to stably reflect "which objects have stronger aggregation consistency and more concentrated contribution to the current mode". This provides a quantitative basis for the subsequent generation of impedance adjustment parameters and power allocation parameters based on object attribute identification, avoiding unclear contribution attribution and unbalanced control allocation caused by relying on a single measurement point or fixed empirical rules.

[0075] Regarding S104 above:

[0076] In one implementation, S104 includes jointly parsing the association strength of each controlled object output in S103 with the object attribute identifier to form an object-oriented parameter set, and then distributing the parameter set as the control reference quantity of the interface converter, so that different controlled objects can perform coordinated control with consistent field caliber within the same parameter update cycle. To facilitate engineering implementation, the coordinated controller establishes an object-level data structure for each controlled object, including at least object identifier, object type identifier, parallel group identifier, rated capacity and rated voltage level, current operating boundary, and association strength. The current operating boundary is used to limit the upper limit of the amplitude that can participate in coordination in the current cycle. The current operating boundary of a grid-type energy storage object can be determined by the state of charge, DC-side power limit, temperature rise margin, current limiting margin, etc., while the current operating boundary of a power generation object can be determined by available output, power limit state, reactive current margin, etc. After receiving the association strength, the coordinated controller first performs an object-dimensional validity check and normalization consistency processing on the association strength, such as limiting outliers, using the valid value from the previous cycle for missing values, and maintaining comparability across the entire object range.

[0077] In this implementation, the collaborative controller selects parameter and constraint templates based on object attribute identifiers to ensure that objects of the same type use consistent parameter meanings and delivery interfaces. For example, the parameter set entries for grid-type energy storage objects include at least impedance regulation parameter entries and power allocation parameter entries, while the parameter set entries for power generation objects include at least power allocation parameter entries. Impedance regulation parameter entries may include at least one of damping injection-related parameters, virtual impedance-related parameters, and inertia-related parameters. Power allocation parameter entries may include at least one of active power target adjustment, reactive power target adjustment, droop coefficient, or power factor target. The specific fields of the above parameter entries can correspond one-to-one with the parameter area of ​​the interface converter controller or the offline delivery interface, enabling the parameter set output by the collaborative controller to be directly parsed and loaded.

[0078] In this implementation, the impedance regulation parameters generated for grid-type energy storage objects are used to describe the equivalent output impedance shape of the interface converter at the grid connection point and its damping / inertia-related adjustable terms. The cooperative controller selects an impedance regulation template matching the current oscillation mode and maps the correlation strength to an object-level "participation coefficient" under this template to drive differentiated impedance regulation amplitudes between objects. The participation coefficient can be determined jointly by the correlation strength and the operating boundary: when the correlation strength of a grid-type energy storage object is high and the operating boundary allows, the participation coefficient is set to a higher level, allowing it to bear a more significant equivalent damping or equivalent inertia contribution; when the correlation strength is low or the operating boundary is limited, the participation coefficient is set to a lower level, keeping its adjustment amplitude small to avoid excessive traction of unrelated objects and triggering control resistance. To ensure online control stability, the co-controller applies three types of engineering constraints to the parameters before adjusting the output impedance: First, amplitude constraints, ensuring that each parameter does not exceed a preset upper limit and is not lower than a preset lower limit; second, rate of change constraints, ensuring that the parameter changes in adjacent parameter update cycles do not exceed a preset upper limit; and third, consistency constraints, ensuring that key impedance parameters within the same parallel group maintain a consistent update direction within the allowable range, preventing objects within the group from exhibiting opposite impedance trends and thus exacerbating circulating currents and coupled oscillations. For example, when the control cycle is 10ms and the parameter update cycle is 20ms, the upper limit for a single change of damping-related parameters can be set to 0.01–0.03 per unit, the upper limit for a single change of the equivalent time constant of inertia-related parameters can be set to 0.02s–0.08s, and the upper limit for a single change of virtual impedance parameters can be set to 5%–15% of their rated configuration value, in order to suppress the direct impact of parameter jitter on the grid-connected voltage vector and circulating current.

[0079] In an example scenario, a parallel group comprises two grid-connected energy storage devices with rated capacities of 500kW and 300kW, respectively. Their current operating boundaries are "available active power margin 0.20 per unit, reactive power margin 0.15 per unit" and "available active power margin 0.10 per unit, reactive power margin 0.12 per unit," respectively. The correlation strengths of the S103 outputs are 0.78 and 0.42, respectively. The co-controller generates different participation coefficients for the two devices under the impedance adjustment template. This causes the damping injection-related parameters of the first device to be increased to a higher level from the baseline configuration, while the second device remains near the baseline configuration or only slightly increased. Simultaneously, intra-group consistency constraints are set on the virtual impedance parameters of both devices, ensuring that the dominant direction of the equivalent output impedance within the group is consistent. This reflects the "differentiated contribution positioning" while avoiding intra-group control conflict.

[0080] In this implementation, power allocation parameters are used to describe the coordinated allocation of grid-type energy storage objects and / or power generation objects at the active and reactive power levels. The coordinated controller determines the available output boundaries and priority constraints of each object based on its attribute identifiers, and incorporates correlation strength into the allocation decision under these constraints, allowing objects with higher oscillation contributions and permissible boundaries to assume clearer responsibilities in power regulation. To avoid the unreasonable neglect of capacity and boundaries caused by "allocation based solely on correlation strength," the coordinated controller can construct the allocation weights as composite weights determined by rated capacity, available margin, and correlation strength: rated capacity reflects the object's basic sharing capacity, available margin reflects the executable capacity for the current cycle, and correlation strength reflects the intensity of responsibility for the current mode. Subsequently, the composite weights are normalized across all objects to obtain the weight set used for power allocation in the current cycle, and based on this, the active power target adjustment and / or reactive power target adjustment for each object are generated. For grid-based energy storage devices, power allocation parameters can also be applied synchronously to the target bias or droop coefficient related to droop control, enabling them to complete device-level power sharing while maintaining grid support; for power generation devices, power allocation parameters can be reflected as active power limit, reactive current margin allocation, or power factor target, enabling them to participate in coordination without compromising their own control boundaries.

[0081] To prevent the allocation strategy from triggering infeasible instructions under extreme operating conditions, the co-controller performs feasibility verification and reallocation processing on the power allocation parameters. Feasibility verification includes at least: verifying whether the target adjustment amount for each object exceeds its current limiting margin or DC-side power limit; verifying whether the power target within the same parallel group significantly increases circulating current risk; and verifying whether the global power balance meets grid-connected constraints. If an object reaches its limiting boundary or is determined to be infeasible, its allocation amount is compressed to its feasible range, and the compressed and released allocation amount is reallocated among objects of the same type or within the same parallel group according to composite weights, thereby ensuring that the overall allocation is closed and executable. For example, in a scenario involving two groups of parallel grid-connected energy storage objects and multiple generator inverters, the parallel group can be used as a primary allocation unit to determine the group-level allocation first, and then subdivided within the group to the object level according to composite weights, balancing consistency within the group with differentiated responsibilities at the object level.

[0082] In this implementation, the parameter set is converted into the corresponding interface converter control reference quantities for the controlled object. This includes mapping the "impedance adjustment parameters / power allocation parameters" to the outer loop target and inner loop setpoint that the interface converter can directly execute, while maintaining the fields and units of the control reference quantities in accordance with the interface conventions. For grid-type energy storage objects, the control reference quantities may include voltage amplitude reference, frequency reference, or phase angle reference, as well as voltage reference components and current reference components in a synchronous rotating coordinate system. The impedance adjustment parameters are used to adjust the equivalent parameters of the voltage outer loop, virtual impedance element, or inertia element, so that the interface converter presents the target impedance form to voltage and current disturbances on the grid-connected side. The power allocation parameters are used to adjust the target or bias terms of frequency-active droop and voltage-reactive droop, so that its output power converges according to the cooperative allocation strategy. For a power generation object, the control reference quantity may include active power reference, reactive power reference or power factor reference, and is further converted into current command or voltage command required by the grid-connected controller; when the power generation object is grid-connected control, the allocation result can be mapped to the active and reactive components of the current reference in the synchronous rotating coordinate system and executed under current limiting constraints; when the power generation object has voltage support capability, reactive power allocation can be mapped to voltage support bias to participate in voltage regulation.

[0083] In one example implementation, the interface converter employs a voltage outer loop and a current inner loop structure. The voltage amplitude and frequency references issued by the co-controller enter the voltage outer loop to generate a voltage reference component. The current inner loop then generates a modulation reference quantity based on this and outputs it to the PWM or space vector modulation module, ultimately forming the bridge arm drive signal. To ensure the continuity of the online closed loop, the co-controller can adopt a consistent timing sequence of "parameter update—reference quantity update" when issuing control reference quantities. That is, within the same parameter update cycle, the parameter set loading is completed first, and then the outer loop target is refreshed. Furthermore, amplitude and rate-of-change constraints are applied to the reference quantity changes to prevent unexplained jumps in the control reference quantity. Through the above parameter generation, feasibility verification, and reference quantity mapping, the co-controller can output a directly implementable object-level control reference in each update cycle, enabling vibration suppression responsibility and power sharing to form an interpretable, executable, and consistent co-control closed loop at the object level with the current oscillation mode.

[0084] For example, this application preferably uses time-series quantities such as grid connection point voltage, current and power as inputs, and outputs high-level characterization capsule vectors and routing weight sets through capsule networks to characterize the current dominant oscillation mechanism and the correlation strength of each grid-type energy storage object with respect to the mechanism, and generates impedance adjustment parameters and power distribution parameters accordingly, forming control reference quantities and limiting constraints that can be sent to each grid-connected interface converter controller.

[0085] In this preferred configuration, the input timing segments are constructed using a fixed sampling rate and a fixed window length. The sampling rate is preferably 500 Hz, the sliding window length is preferably 2 s, and the sliding step size is preferably 0.2 s, resulting in each window containing 1000 sampling points. The number of input channels is preferably 6 channels, representing the grid connection point voltage amplitude, grid connection point current amplitude, grid connection point phase angle deviation, frequency deviation, active power, and reactive power. The phase angle deviation is processed to ensure phase continuity to avoid cross-cycle jumps; the amplitude channel is normalized and limited to a preset amplitude upper limit; the power channel is normalized to its capacity; and the frequency deviation is normalized by a preset scaling factor after being limited. The resulting input tensor is preferably a 6×1000 two-dimensional tensor, which serves as the input to the capsule network.

[0086] In this preferred configuration, the front-end temporal mapping module of the capsule network adopts a two-level 1D convolutional cascade structure, and the layers and connections are fixed as follows: The first convolutional layer preferably has 64 output channels, a kernel length preferably of 9, and a stride preferably of 1; the second convolutional layer preferably has 128 output channels, a kernel length preferably of 9, and a stride preferably of 1; both convolutional layers use linear rectified activation functions, and batch normalization is performed after each convolutional output to improve training stability. The convolutional output is flattened to a fixed length and then enters the primary capsule generation module. The number of primary capsules is preferably 32, and the dimension of each primary capsule vector is preferably 8-dimensional; a learnable transformation matrix is ​​used to map from the primary capsule vectors to the higher-level representation capsule vectors. The number of higher-level representation capsules is preferably 6, and the dimension of each higher-level representation capsule vector is preferably 16-dimensional. The dynamic routing process is preferably iterated three times. The routing coefficients are updated using normalization in each iteration to maintain comparability of the weight sets. The high-level representation capsule vectors undergo nonlinear compression to ensure their magnitude falls within a preset upper and lower limit range, and the magnitude is used as a confidence index for the oscillation mechanism corresponding to the high-level representation capsule. To improve interpretability and avoid meaningless amplification, this preferred configuration also imposes an upper limit constraint on the amplitude of the learning components and preset upper and lower limit constraints on the magnitude of the high-level representation capsule vectors.

[0087] In this preferred configuration, to enhance training reproducibility, a decoder is set up to constrain the reconstruction of the input tensor. The decoder adopts a two-level fully connected structure, with the first fully connected layer preferably having 128 neurons and the second fully connected layer preferably having 256 neurons. The output layer dimension is consistent with the 6×1000 input tensor, and the mean squared error is used as the reconstruction loss. The total loss is composed of a weighted sum of classification loss and reconstruction loss. The classification loss is used to constrain the magnitude of the high-level representation capsule vector corresponding to the correct oscillation mechanism to be greater than a preset positive threshold and to constrain the magnitude of other mechanisms to be less than a preset negative threshold. The reconstruction loss is used to constrain the high-level representation capsule vector to faithfully represent the input time sequence. The weight coefficient of the reconstruction loss is preferably 0.0005 to avoid the reconstruction term dominating the classification term. The training optimizer preferably adopts an adaptive moment estimation optimizer, and the initial learning rate is preferably 1×10⁻⁶. -3 The batch size is preferably 128, the maximum number of training rounds is preferably 80, and an early stopping strategy is adopted to avoid overfitting. The early stopping patience value is preferably 10 rounds. To reduce training variance, the parameters are initialized with a fixed random seed, and the training samples are randomly shuffled in each training round.

[0088] In this preferred configuration, the training data preferably consists of three parts: the first part is historical operation records, the second part is electromagnetic transient or electromechanical transient simulation data, and the third part is disturbance injection data from the hardware-in-the-loop platform. All three parts of the data are sliced ​​according to the aforementioned sampling rate and window length to generate sample windows, and divided into training, validation, and test sets in an 8:1:1 ratio. The generation of supervision labels preferably employs a combination of offline identification and rule mapping: offline modal identification is performed on the voltage, frequency, or power oscillation sequence corresponding to each window. The identification method is preferably a parameter identification method based on exponentially decaying sine waves, outputting the dominant mode frequency and damping ratio, and generating mechanism labels by combining grid-connected control mechanism criteria. The mechanism labels are preferably divided into six categories: steady-state, forced-source periodic oscillation, negatively damped self-excited oscillation, grid-connected resonance, power loop coupled oscillation, and subsynchronous or near-synchronous correlated oscillation. The preferred criteria for rule mapping are as follows: when the dominant mode damping ratio is greater than 0, the frequency falls within a preset low-frequency bandwidth, and the external disturbance characteristics are significant, it is labeled as a forced source periodic oscillation; when the dominant mode damping ratio is less than 0 and the amplitude self-increases under the condition of no continuous external disturbance, it is labeled as a negatively damped self-excited oscillation; when the dominant mode frequency falls within the preset frequency band of grid resonance and the voltage-current phase relationship meets the resonance criterion, it is labeled as a grid resonance; when the phase difference and gain characteristics between power loop related quantities meet the preset coupling criterion, it is labeled as a power loop coupled oscillation; the rest that meet the steady-state criterion are labeled as steady-state, and samples that cannot be classified into the above categories are entered into the subsynchronous or semi-synchronous related oscillation category according to the preset priority or manual review. To ensure that the labels and samples are aligned, the dominant mode frequency band and mechanism labels output by offline identification are written into the sample label table using the window index as the key, and are read by key during training data loading.

[0089] In this preferred configuration, for the preferred mapping rule from the set of routing weights to impedance regulation parameters and power allocation parameters, firstly, after dynamic routing convergence, for each grid-type energy storage object, the set of routing weights pointing to each high-level characterization capsule vector is obtained from the primary capsule vector. Then, taking the high-level characterization capsule currently determined to be the dominant oscillation mechanism as the target, the routing weight pointing to that target capsule is extracted as the object's original association strength. Subsequently, the original association strengths of all objects participating in coordination within the same control cycle are normalized to obtain normalized association strengths. The preferred normalization method is sum-based normalization with a lower limit truncation of the minimum value to avoid numerical instability. Further, to reflect the impact of capacity and operational margin on coordinated allocation, the normalized association strength is preferably multiplied by the capacity weight and margin weight to obtain a composite weight, and the composite weight is again sum-based normalized to obtain the participation coefficient. The capacity weight is preferably the proportion of the object's rated power to the total rated power of the participating objects, and the margin weight is preferably the proportion of the object's current available adjustment margin to its rated adjustment margin, limited to a preset upper and lower limit range. Both impedance adjustment parameters and power allocation parameters are linearly mapped using participation coefficients as interpolation coefficients. The impedance adjustment parameters include at least a virtual resistance or damped injection gain term, and the power allocation parameters include at least the active power increment to be absorbed or released and the reactive power increment. Upper and lower limit constraints and rate of change constraints are applied to the mapping results to ensure that the control quantity is implementable and does not introduce abrupt changes.

[0090] In an example scenario, there are three grid-type energy storage objects in parallel groups, with rated powers of 10 MW, 5MW, and 3 MW, respectively, and a margin weight of 1 for each. After dynamic routing convergence, the original correlation strengths of the three objects pointing to the high-level characterization capsule of the dominant oscillation mechanism target are 0.62, 0.31, and 0.07, respectively, which remain unchanged after sum-normalization. The capacity weights are 0.556, 0.278, and 0.167, respectively. The composite weights, calculated as "correlation strength × capacity weight" and then normalized, yield participation coefficients of approximately 0.779, 0.195, and 0.026, respectively. If the upper and lower limits of the damping injection gain are set to 0.05 and 0.25 respectively, the damping injection gains of the three objects will be approximately 0.206, 0.089, and 0.055 respectively. If the upper and lower limits of the virtual resistance are set to 0.02 pu and 0.08 pu respectively, the virtual resistances of the three objects will be approximately 0.0667 pu, 0.0317 pu, and 0.0216 pu respectively. If the active power increment to be shared in this control cycle is 3 MW, the active power increments allocated to the three objects will be approximately 2.337 MW, 0.584 MW, and 0.079 MW respectively. Before the power increment is issued, a rate of change limit is applied to the power increment, for example, limiting the single-cycle change to no more than 10% of the rated power of the object, to ensure smooth controller execution.

[0091] Optional, see Figure 2 The flowchart of an iterative route weight update method provided in this application embodiment includes steps S201 to S202, wherein:

[0092] S201: Perform sparsification processing on the routing weight set corresponding to each primary capsule vector, so that only a preset number of maximum weights are retained in the routing weight set, and the remaining weights are either set to zero or scaled according to a preset decay factor.

[0093] S202: Furthermore, when the change in routing weight exceeds a preset threshold for two consecutive control cycles, the parameter set is generated using the routing weight of the previous control cycle until the change in routing weight continuously meets the preset stability condition.

[0094] Optionally, in the scenario of multi-object parallel collaboration in a source-storage system, electrical operation data is often independently collected by different interface converter controllers or measurement units. Due to differences in sampling clocks, communication links, and buffer scheduling, there may be time deviations in the sampling times of the three-phase voltages of different controlled objects. This time deviation is equivalent to an additional phase shift in phasor representation, causing inconsistencies in the phase attributes at the same physical moment between objects. If this phase attribute is directly used as the input of the angular component of the primary capsule vector, the model may misjudge the phase difference introduced by asynchronous sampling as a coupling phase difference between objects or an oscillating phase drift, thereby affecting the reliability of subsequent routing consistency measurement, mode identification, and contribution measurement.

[0095] Based on this, this application introduces phase compensation alignment driven by sampling time deviation, ensuring that the angle components entering the capsule network have a consistent time caliber across objects. In this optional embodiment, the cooperative controller performs a space vector transformation on the three-phase voltages in the time-series data stream to obtain orthogonal components in the stationary coordinate system, and determines the instantaneous phase angle sequence accordingly. This instantaneous phase angle sequence is used as one of the inputs to the angle components of the primary capsule vector. The continuity, de-scrambling, and necessary smoothing of the phase sequence can be achieved using existing engineering methods to ensure the usability of the phase sequence; the above processing is not considered a limiting focus of this application.

[0096] The distinguishing feature of this application is that the cooperative controller first determines the sampling time deviation of different controlled objects relative to a unified time reference, and generates a phase compensation amount accordingly to compensate and align the instantaneous phase angle. The unified time reference can be provided by the cooperative controller's master clock, the timestamp of the common coupling point measurement unit, or a reference phase source agreed upon within the system. The sampling time deviation can be obtained based on the timestamp alignment result and the time synchronization message result; in the absence of a reliable timestamp, the deviation estimate can also be obtained through alignment search between the object's phase sequence and the reference phase sequence, and a consistency check is performed on the deviation estimate to avoid misalignment caused by occasional noise. The phase compensation amount can be generated according to the "phase shift corresponding to the time deviation," and combined with the real-time estimate of the grid frequency to adapt to operating conditions with slight frequency drift. To avoid abrupt changes in the angle component caused by compensation injection, the cooperative controller can use periodic gradual updates to the phase compensation amount, ensuring that the compensated phase angle is continuous in time.

[0097] For example, in a scenario with 50Hz grid connection, 2kHz sampling rate, and 10ms control cycle, the sampling time deviation can be concentrated in the range of tens to hundreds of microseconds. The cooperative controller can configure the upper limit of the allowable deviation to be 200μs to 500μs and the transition time of the compensation gradient to be 2 to 5 control cycles to achieve a balance between alignment effectiveness and phase continuity. When the sampling time deviation is detected to exceed the allowable upper limit, or the deviation exhibits unstable jumps within multiple consecutive control cycles, the cooperative controller can mark the controlled object as a time-synchronization anomaly and set an "invalid flag" or "low-confidence flag" for the object's angle component. In subsequent routing calculations and contribution measurements, a preset low-weight coefficient or the effective angle component of the previous cycle is used to replace it, thereby reducing the interference of the time-synchronization anomaly object on modal recognition and contribution localization.

[0098] Optionally, to suppress jitter caused by the dispersed allocation of routing weights under disturbances or non-stationary input conditions, which propagates to the downstream parameter set, the controller performs sparsification processing on the routing weight set corresponding to each primary capsule vector, concentrating routing allocation on a small number of high-weight destinations. Specifically, the controller can select K maximum weights as reserved items within the routing weight set corresponding to the same primary capsule vector, while the remaining weight items are limited to zero or below a preset minimum weight upper limit. In implementations requiring numerical continuity, non-reserved items can be reduced to a small fraction of their original values ​​by a preset attenuation factor, and the reserved items can be normalized, ensuring a consistent interpretation of the sparsified weight set across different control cycles. For example, in a configuration with 6 to 12 candidate destinations, K is a preset number, which can be 1 to 3, the attenuation factor can be 0.1 to 0.4, and the minimum weight upper limit can be 0.02 to 0.05, to achieve concentrated routing, weaken tail terms, and reduce disturbances to the parameter set.

[0099] Furthermore, to prevent the frequent switching of the dominant destination under critical oscillation conditions after the sparsified routing weights have led to rapid changes in the parameter set, the controller introduces a stability maintenance strategy: when the change in routing weights over two consecutive control cycles exceeds a preset threshold, the parameter set is generated using the routing weights from the previous control cycle, until the change in routing weights continuously meets the preset stability condition. Here, the change in routing weights can be obtained by measuring the difference in the weight set, for example, by aggregating the changes in the magnitude of each weight item to form an overall change index; the preset threshold is used to limit the allowable range of weight fluctuations, and the preset stability condition is used to limit the number of consecutive cycles required for the weights to return to stability and the retention of the dominant destination. For example, with a control cycle of 10ms and 3 routing iterations, the change threshold can be in the range of 0.01 to 0.03, and the number of stable cycles M can be 3 to 8 control cycles; and a criterion of "the dominant destination remains unchanged within the stable window or the number of switching does not exceed the preset number" can be added to avoid the weights oscillating around the threshold and triggering frequent switching.

[0100] In one alternative implementation, after the stability-maintaining strategy is triggered, the controller can continue to iteratively update the routing weights within the current cycle to promote convergence. However, before the stability condition is met, these weights are only used for internal iterations and not for parameter set generation. Alternatively, the controller freezes the routing weights used for parameter generation during the stability-maintaining strategy and maintains them as the weights of the previous cycle. Through the above sparsity processing and stability-maintaining strategy, under the condition that multi-source sensing data fluctuates over time and parallel coupling leads to non-stationary inputs, the propagation of routing weight dispersion and jitter to the downstream control parameter determination stage can be suppressed. This reduces the risk of frequent switching and instability of control output caused by parameter set jumps and improves the continuity and robustness of cooperative control.

[0101] Optional, see Figure 3 The present application provides a flowchart of a method for generating impedance adjustment parameters, including steps S301 to S302, wherein: S301: determining the dominant oscillation frequency band based on the current oscillation mode; S302: generating a set of frequency band-selective impedance adjustment parameters, wherein the damping injection parameter corresponding to the dominant oscillation frequency band is configured to be not less than a preset reference damping threshold, and the voltage support parameter corresponding to the fundamental frequency band is configured to be not less than a preset voltage support lower limit.

[0102] In one implementation, when the controlled object is a grid-type energy storage object, after receiving the current oscillation mode output by S103, the cooperative controller further determines the dominant oscillation frequency band corresponding to that mode. To ensure that this process has an engineering-feasible and definite standard, the cooperative controller can pre-establish a "mode-frequency band" mapping table or a mode prototype library, so that the mode identifiers obtained during the offline training phase are stored corresponding to their typical frequency band ranges; when a certain mode identifier is output online, its corresponding frequency band range is directly read as the dominant oscillation frequency band. The cooperative controller can also fine-tune the upper and lower boundaries of the frequency band near the center of the frequency band, combined with the currently observed frequency band concentration degree. The fine-tuning can be based on the spectral peak concentration range or energy concentration range obtained by conventional frequency domain statistics of grid connection point voltage, current, or power disturbances, so as to ensure that the dominant frequency band covers the main energy components of the mode while avoiding the introduction of unrelated frequency bands with excessively wide bandwidth.

[0103] For example, in a scenario with a 50Hz grid connection, a 2kHz sampling rate, a 10ms control period, and a 256-point model inference window, the dominant oscillation frequency band can be configured according to the modal prototype library as one of the following types: low-frequency coupled oscillation mode corresponding to 10Hz~30Hz; subsynchronous related mode corresponding to 30Hz~55Hz; and grid-connected side resonant related mode corresponding to 55Hz~120Hz. For a specific mode, the upper and lower bounds of the frequency band can also be further set as a bandwidth rule of "5Hz~15Hz above and below the center frequency" to adapt to the frequency band drift caused by different short-circuit ratios or changes in external equivalent impedance. The fundamental frequency band can be configured as 45Hz~55Hz according to engineering practice, as the main operating frequency band for voltage support and steady-state regulation.

[0104] In one implementation, the cooperative controller generates a band-selective set of impedance regulation parameters, emphasizing vibration suppression within the dominant oscillation band and voltage support within the fundamental band. The impedance regulation parameter set includes at least two types of parameters: damping injection parameters and voltage support parameters. Damping injection parameters enhance the equivalent damping of the grid-type energy storage device within the dominant oscillation band, while voltage support parameters ensure that its voltage regulation capability within the fundamental band is not weakened. Band selectivity can be achieved by configuring a band selection unit in the impedance shaping channel. This band selection unit can be a bandpass / bandstop filter structure, a frequency-divided parallel impedance branch structure, or an equivalent band weighting structure, ensuring that disturbance components from the dominant oscillation band primarily enter the damping compensation channel, while voltage deviations within the fundamental band primarily enter the voltage support channel.

[0105] In this embodiment, to ensure the effectiveness of oscillation suppression, the cooperative controller configures the damping injection parameters corresponding to the dominant oscillation frequency band to be no less than a preset reference damping threshold. The damping injection parameters can adopt a field form that matches the device control structure, such as including at least one of the following: equivalent resistance term in virtual impedance, gain term for active power damping injection, or weighting coefficient for damping compensation branches. For example, the reference damping threshold can be configured in the range of 0.03 to 0.10 per unit, or in the range of 0.5 to 2.0 per gain, and can be configured in stages according to the rated capacity of the grid-type energy storage object and the short-circuit ratio at the grid connection point; when the short-circuit ratio is low or the external equivalent impedance is large, the reference damping threshold is set to a higher level to improve the oscillation suppression margin. To avoid a decrease in voltage support capability due to damping injection, the cooperative controller simultaneously constrains the voltage support parameters within the fundamental frequency band to be no less than the lower voltage support limit. The voltage support parameters can adopt at least one field form of voltage loop adjustment gain, upper limit constraint of reactive power droop coefficient, or weight of voltage support branches.

[0106] For example, the lower limit of voltage support can be configured to be "not less than 80% to 100% of the default value", or the critical gain of the voltage loop can be limited to above a preset minimum value to ensure that the fundamental steady-state voltage regulation capability is not weakened.

[0107] In one implementation, the co-controller writes the impedance adjustment parameter set into the impedance shaping or control gain configuration module and applies engineering boundary constraints to the parameter writing to ensure the smoothness and safety of online operation. For example, with a control cycle of 10ms and a parameter update cycle of 20ms, the single update step size of damping-related parameters can be limited to the order of 0.01 to 0.03, and the single update step size of voltage support-related parameters can be limited to 1% to 3% of the default value. Upper and lower limits are set for parameter changes to ensure they remain within the device's allowed stable range. This allows sufficient damping injection to suppress oscillations within the dominant oscillation frequency band, while maintaining a voltage support configuration no lower than the voltage support lower limit within the fundamental frequency band.

[0108] Optionally, to avoid misclassifying the current oscillation mode as a known mode when it does not belong to the preset mode set, thus leading to improper impedance adjustment parameter configuration, the cooperative controller can perform a consistency check on the original time-series data stream through the reconstruction network after obtaining the converged high-level characterization capsule output, and trigger an unknown mode fallback strategy if the check fails. Specifically, the cooperative controller inputs the high-level characterization capsule output to the reconstruction network to generate a reconstructed time-series data stream, and calculates the reconstruction residual between the reconstructed time-series data stream and the original time-series data stream; when the reconstruction residual exceeds a preset threshold, the current oscillation mode is determined to be an unknown mode, and the impedance adjustment parameter set is switched to a preset conservative damping parameter set. The preset threshold can be configured using an upper limit of residual obtained from steady-state operating data statistics, and hysteresis bands for entry and exit can be set to reduce false triggering.

[0109] For example, the reconstruction residual threshold can be configured in the range of 0.10 to 0.25 per unit, and a switchover is only triggered if the threshold is exceeded for 2 to 5 consecutive update cycles. As for the recovery criterion, the fallback can be removed only after the residual falls back to below the lower threshold for several consecutive cycles.

[0110] The conservative damping parameter set serves as a safety configuration for unknown modes. Its field definitions are consistent with the impedance adjustment parameter set, including at least a minimum level for the damping injection parameters, limits on the rate of change of parameters, and minimum retention constraints for voltage support parameters. This ensures that the control link can continue to output executable control reference quantities and maintain stability margins without changing the interface after switching. For example, the conservative damping parameter set can fix the damping injection parameters at a mid-to-high level near the reference damping threshold and further tighten the parameter update step size to prioritize vibration suppression and stability boundaries under unknown modes. Only when the subsequent reconstruction residuals continuously meet the recovery conditions can a return from the conservative damping parameter set to the parameter set based on known modes be allowed, thus reducing the impact of frequent switching on the continuity of control output.

[0111] Optionally, to enable the band-selective impedance adjustment parameter set to be configured differently for different oscillation mechanisms, the controller selects the converged high-level characterization capsule vector corresponding to the dominant oscillation frequency band when generating the band-selective impedance adjustment parameter set, and constructs endpoint time-series trajectories for mechanism identification based on the angle and magnitude values ​​of the high-level characterization capsule vector. The high-level characterization capsule vector is a vector output after aggregating the features of the dominant oscillation frequency band, where the angle value is used to characterize the phase evolution direction in that frequency band, and the magnitude value is used to characterize the oscillation intensity or energy level in that frequency band. The controller converts the angle and magnitude values ​​into complex plane endpoints using a polar-to-planar coordinate mapping method, forming an endpoint sequence with the control period as the sampling granularity, and connects them in chronological order to obtain the endpoint time-series trajectory, so as to interpretably characterize the oscillation evolution pattern without relying on an accurate system model.

[0112] In one implementation, the controller calculates the angle change and magnitude change of the endpoints in adjacent control cycles to characterize the dynamic evolution of the dominant oscillation frequency band. The angle change can be obtained from the phase difference between the endpoints of adjacent cycles, and conventional phase continuity processing can be used to avoid artifacts caused by boundary jumps. The magnitude change can be obtained from the amplitude difference between the endpoints of adjacent cycles, and smoothing can be performed within a short window to improve noise immunity. To balance real-time performance and robustness, the controller can statistically analyze the angle change and magnitude change within a preset sliding window to obtain trend quantities for mechanism discrimination, such as the mean and cumulative increment of the magnitude change, and the consistency of the angle change (consistent rotation direction or stable mean).

[0113] For example, in a scenario with a control cycle of 10ms, the mechanism discrimination window length can be set to 20–60 control cycles to cover at least several effective oscillation cycles of the dominant oscillation frequency band; the significance threshold for angle change can be configured in the range of 2°–8°, the growth threshold for modulus change can be configured in the range of 0.003–0.02 per unit, and the cumulative increment threshold can be configured in the range of 0.02–0.10. These thresholds can be set in tiers based on the short-circuit ratio at the grid connection point, the range of external equivalent impedance changes, and the device noise level.

[0114] In one implementation, the controller identifies the oscillation mechanism based on the changes in angle and modulus, and selects a parameter configuration mode from the impedance adjustment parameters according to the identification results. When the modulus change shows a continuous positive increase within a continuously preset window and the cumulative increment exceeds a first threshold, while the angle change exhibits a consistent rotation direction or a stable mean, the controller identifies the current oscillation mechanism as negatively damped divergent oscillation and selects a high-damping mode. In the high-damping mode, the controller performs an increase configuration on the damping injection parameter corresponding to the dominant oscillation frequency band, ensuring that the damping injection parameter is not lower than the reference damping threshold, and simultaneously tightens the parameter change rate limit related to this frequency band to suppress further amplitude divergence. For example, the damping injection parameter can be increased by 10% to 50% based on the reference threshold, and the single update step size of this parameter is tightened to 30% to 70% of the reference step size to avoid abrupt changes in control commands caused by the intervention of high damping.

[0115] When the modulus variation exhibits periodic alternations around zero within a continuously preset window, and its absolute mean value is less than a second threshold, while the angle variation shows approximately constant steps or coincides with the period of the dominant oscillation band, the controller identifies the current oscillation mechanism as forced source periodic oscillation and selects a high inertia mode. In high inertia mode, the controller performs an enhancement configuration on the inertia-related parameters corresponding to the dominant oscillation band, ensuring that the inertia-related parameters are not lower than the preset high inertia lower limit, and can maintain the damping injection parameter at the reference level or adjust it in small increments to improve the phase-following stability to external periodic disturbances. For example, the inertia-related parameters can be increased by 20% to 80% based on the default value, or the equivalent inertia time constant can be increased by 0.05s to 0.30s, while keeping the damping injection parameter not lower than the reference threshold but not significantly increased, to avoid unnecessary conservative constraints on steady-state power distribution.

[0116] To avoid frequent mode switching caused by the mechanism criterion being near a threshold, the controller can set hysteresis and stability confirmation rules for high-damped and high-inertia modes. This includes using different entry and exit thresholds, or requiring the mechanism criterion to be met consecutively a preset number of times before mode switching. For example, the entry criterion can require satisfaction for 2-3 consecutive windows, and the exit criterion can use a lower threshold and require it to be maintained for several consecutive windows to reduce jitter caused by transient noise or short-term disturbances. Simultaneously, parameter increases in high-damped or high-inertia modes only affect the parameters corresponding to the dominant oscillation frequency band; parameters in other frequency bands remain at the baseline configuration or are adjusted by a small amount to maintain band selectivity and avoid unnecessary impact on the fundamental frequency band voltage support capability. Through this method, the controller can identify the oscillation mechanism in real-time at the terminal based on the phase and amplitude evolution of the capsule vector, and select a parameter configuration mode that matches the mechanism in the impedance adjustment parameters, thereby improving the targeting and robustness of the oscillation suppression strategy.

[0117] In one optional implementation, to improve the robustness of angle changes and modulus increments under noise, measurement glitches, or transient model output conditions, the controller can perform validity checks and outlier suppression on the angle and modulus values ​​before calculating the changes. The validity check may include at least: whether the high-level characterization capsule vector satisfies the convergence criterion, whether the data integrity flag for the control cycle is valid, and whether the modulus value falls within a preset physically feasible range. When a convergence criterion is not met, the modulus value exceeds the physically feasible range, or a sudden jump occurs, the controller can mark the angle or modulus value for that cycle as invalid and use the valid value from the previous cycle, or apply amplitude limiting and gradual change constraints to the changes for that cycle to avoid trend statistical distortion caused by a single anomaly.

[0118] For example, in a scenario where the control period is 10ms and the dominant frequency band characteristics are sensitive, the change limit can be configured to "not exceed 2 to 3 times the historical steady-state fluctuation in a single period", and a protection strategy of triggering weight reduction or suspending mechanism discrimination is triggered for continuous abnormal periods.

[0119] Furthermore, when the dominant oscillation frequency band switches, the modality recognition result changes, or the controller enters a new operating phase, the controller can reset or cold-start update the cumulative statistics of the modality length increment, so that the cumulative window only reflects the amplitude evolution trend within the current frequency band or the current phase, thereby reducing the risk of misjudgment caused by cross-scenario superposition.

[0120] For example, the cumulative window length can be configured to cover the number of control cycles covering several effective oscillation cycles of the dominant frequency band. In the implementation of 50Hz grid connection and 10ms control cycle, the window length can be 20 to 60 control cycles. In the first 3 to 5 control cycles after frequency band switching, a gradual introduction (from small window to target window) is adopted to complete the cold start, so as to suppress the statistical mutation at the moment of switching.

[0121] In one embodiment, the controller uses the angle change, magnitude increment, and their trend statistics as input features for oscillation mechanism identification, and combines them with other statistics of the endpoint time-series trajectory for subsequent parameter configuration mode selection. For example, when external disturbances cause a sustained small-amplitude drift in the endpoint phase and a sustained increase in the magnitude trend, this feature combination can be used to indicate a sustained divergence tendency in the amplitude, thus providing a basis for subsequently selecting a parameter configuration mode with stronger damping. When a periodic forced source disturbance causes the angle change to exhibit a stable periodic oscillation while the magnitude trend remains within the confined band, this feature combination can be used to indicate that the amplitude does not exhibit sustained divergence but has a stable periodic drive, thus providing a basis for subsequently selecting a higher inertia or a more conservative band gain configuration.

[0122] The aforementioned characteristic aperture can be consistently interpreted within the control cycle and is adaptable to phase boundaries and noise disturbances, thereby supporting the reliable configuration of the oscillation mechanism identification and impedance adjustment parameter set.

[0123] Optionally, to address the potential divergence of the dominant oscillation frequency band when grid-connected energy storage devices are connected to weak grids or when there are significant changes in external equivalent impedance, the controller, after determining the dominant oscillation frequency band and obtaining the statistical quantities of angle change and modulus length trend, can further perform patterned configuration of negative damping divergence discrimination and impedance adjustment parameters within this frequency band to achieve rapid suppression. Specifically, the controller uses a preset number of consecutive cycles as the judgment window length and sets a significance threshold for angle change and a trend threshold for modulus length growth as divergence criterion thresholds, respectively. These thresholds can be stored as configurable parameters and can be set in different levels according to the short-circuit ratio at the grid connection point, the line impedance range, the operating stage, or historical calibration results.

[0124] For example, in a 10ms control cycle implementation, the decision window length can be 30–50 control cycles; the angle change significance threshold can be configured in the range of 2°–8°; the modulus growth trend threshold can be configured as "cumulative increase exceeding 3–10 times the noise amplitude of the steady-state baseline," and combined with different hysteresis rules for entry / exit thresholds to distinguish between transient oscillations caused by short-term impacts and continuous divergence caused by insufficient equivalent damping. When the angle change continuously meets the requirement of not less than the angle threshold within the decision window and the modulus trend statistic simultaneously meets the requirement of not less than the growth threshold, the controller determines that the oscillation mechanism of the dominant frequency band is negative damped divergent oscillation, and configures the impedance adjustment parameter to a high-damping mode to enhance the equivalent damping in this frequency band.

[0125] In high-damping mode, the impedance adjustment parameters may include at least damping injection parameters and their rate of change constraint parameters. The damping injection parameters may correspond to active damping gain or equivalent damping weight, used to adjust the intensity of the active damping injection, enabling the controller to introduce an additional damping component into the active channel to change the damping characteristics of the equivalent output impedance. The controller can configure the active damping gain to be no less than a preset strong damping threshold and employ a smooth intervention strategy for gain updates to avoid output jitter or secondary excitation caused by sudden gain changes. For example, the strong damping threshold can be set to 1.2 to 1.8 times the default damping configuration, and the single update step size can be limited to 5% to 15% of the target increase, while maintaining effective amplitude limiting and rate of change constraints on the drive command signal to ensure that the intervention process of high-damping mode is consistent with the actuator interface and safety boundaries.

[0126] On the other hand, in one embodiment, to identify forced source periodic oscillations driven by external periodic disturbances, the controller can, after determining the dominant oscillation frequency band and obtaining the corresponding converged high-level characterization capsule vector, jointly discriminate the sustained significance of angle changes and the limitation of modulus length. Specifically, the controller uses a preset number of consecutive cycles as the discrimination window, forms a sequence of angle changes cycle by cycle within the window, and determines whether it continuously reaches the significance threshold. At the same time, it monitors cycle by cycle whether the modulus length always remains within the preset upper and lower limits of modulus length to confirm that the oscillation amplitude is within the limited range and does not show cumulative growth or significant decay over time. For example, the discrimination window length can be 20 to 60 control cycles; the modulus length limitation band can be set to "10% to 25% of the allowable bandwidth above and below" the current steady-state operating point, and can be adaptively adjusted according to the device's rated current margin, grid connection point voltage constraints, and historical stable operation statistical range. When both of the above conditions are met, the controller determines that the oscillation mechanism of the dominant frequency band is forced source periodic oscillation, thereby avoiding misjudging the forced steady-state oscillation of "continuous phase swing but limited amplitude" as negative damped divergent oscillation and triggering over-damped injection.

[0127] After determining that the oscillation mechanism is a forced source periodic oscillation, the controller can switch the impedance adjustment parameters to a high inertia mode to reduce the sensitivity of the network control to external periodic excitation and suppress the amplification and propagation of phase oscillation in the control link. Parameters in high inertia mode may include at least one of the following: virtual inertia coefficient, equivalent inertia time constant, frequency change rate suppression gain, or inertia filtering time constant of the frequency / phase estimation stage. The controller configures the above inertia-related parameters to be greater than or equal to a preset high inertia threshold and can perform a smooth transition during the update process.

[0128] For example, the high inertia threshold can be set to 1.2 to 2.0 times the default inertia configuration, or the equivalent inertia time constant can be increased by 0.05 to 0.30 seconds, and the transition period can be set to 5 to 20 control cycles. This ensures a smooth response while avoiding unnecessary continuous lag in power distribution and voltage support. When subsequent monitoring shows that the angle change decreases below the exit threshold, or the module length exceeds the restricted band and shows a continuous growth / decrease trend, the controller can exit the high inertia mode and revert to the default parameter configuration or other configuration mode that matches the current operating conditions. This maintains the robustness and safety of the control strategy when the forced source disturbance subsides or the restricted conditions are no longer met.

[0129] In an alternative implementation, see Figure 2 The sparsification step S201 shown is designed to adapt to the engineering characteristics of multiple grid-type energy storage objects operating in parallel within the same parallel group in the source-storage system. When performing sparsification on the routing weight set corresponding to each primary capsule vector, the cooperative controller introduces intra-group consistency constraints based on object attribute identifiers. The object attribute identifier includes at least an object type identifier and a parallel group identifier. Based on this, the cooperative controller can identify the set of grid-type energy storage objects belonging to the same parallel group and apply mutually exclusive sparsity constraints within that set. This prevents the objects within the group from generating dispersed votes for different candidate modes, which could lead to inconsistent vibration suppression directions or power responsibility allocation in the downstream parameter set within the same update cycle.

[0130] Specifically, the cooperative controller first divides the grid-type energy storage objects into several parallel groups based on the parallel group identifier, and obtains the routing weight set of each object within the group in each control cycle. The routing weight set can be a normalized allocation result obtained from dynamic routing iteration, used to characterize the preference strength of the object (or the primary capsule vector corresponding to the object) towards each higher-level characterization capsule. To form a "consistent single direction within the group", the cooperative controller performs intra-group aggregation on the routing weight sets of each grid-type energy storage object within the same parallel group to obtain a group-level routing preference vector.

[0131] In this context, intra-group aggregation can be achieved by averaging the weight set item by item, by weighting the average based on the object's rated capacity / available margin, or by weighting based on the primary capsule vector magnitude, so that the group-level routing preference vector can reflect the overall preference and consistency of the parallel group for candidate modes in the current control cycle.

[0132] After obtaining the group-level routing preference vector, the cooperative controller determines a single high-level characterization capsule based on this vector, ensuring that all grid-type energy storage objects within the parallel group consistently point to this single high-level characterization capsule within the same control cycle. For example, the cooperative controller can select the high-level characterization capsule with the highest preference degree in the group-level routing preference vector as the current direction for that parallel group. To avoid frequent switching when preferences are not significant, the cooperative controller can set a preference confidence lower limit or a threshold of "minimum difference between the first preference and the second preference." If the threshold is not met, the direction result from the previous control cycle is used, or a minimum hold time and hysteresis rule are adopted to ensure that the direction within the group remains stable under disturbance boundary conditions.

[0133] After identifying a single high-level characterization capsule, the collaborative controller applies mutually exclusive sparsity constraints to networked energy storage objects belonging to the same parallel group, ensuring that only non-zero route weights pointing to the single high-level characterization capsule are retained in the routing weight set corresponding to each object. The remaining route weights can be set to zero to form strict mutual exclusion sparsity; or they can be limited to a minimum preset weight limit, suppressing the remaining weights within a very small range to reduce numerical fluctuations while retaining a very weak backup path for smooth transitions under abnormal operating conditions. The minimum preset weight limit can be stored as a configurable parameter in the controller parameter area and can be configured according to the number of candidate high-level characterization capsules, the number of routing iterations, and real-time requirements.

[0134] Furthermore, to distinguish between the common behavior of "parallel groups participating in external modes as a whole" and the "inconsistent behavior caused by internal coupling within parallel groups," the cooperative controller can also construct common-mode and differential-mode representation vectors within the parallel group based on the primary capsule vectors of the grid-type energy storage objects within the parallel group. During sparsification, mutually exclusive sparsity constraints are applied to the routing weight sets corresponding to the common and differential modes, respectively. The common-mode representation vector is used to characterize the common trends in phase and amplitude evolution of objects within the parallel group. It can be obtained by aggregating the primary capsule vectors of each object according to preset rules, such as averaging or weighting by capacity / margin. Normalization and amplitude limiting can be performed after aggregation to maintain numerical domain stability. The differential-mode representation vector is used to characterize the relative deviations and inconsistencies between objects within the group. It can be composed of the residuals of each object's primary capsule vector relative to the common-mode representation vector, or the difference vectors between object pairs. The amplitude of the residuals can be suppressed and outliers removed to prevent single measurement spikes from being amplified in the differential-mode channel.

[0135] In this optional implementation, the cooperative controller applies a mutually exclusive sparse constraint to the set of routing weights corresponding to the common mode representation vector, making them consistent with the aforementioned group-level routing preferences and pointing to a single high-level representation capsule. This ensures that the parallel group as a whole maintains consistency in its identification and contribution allocation to the external dominant mode within the same control cycle.

[0136] Meanwhile, the cooperative controller applies a specific mutually exclusive sparsity constraint to the set of routing weights corresponding to the differential mode representation vectors, ensuring that the differential mode channels retain only non-zero routing weights pointing to intra-group representation capsules used to characterize the coupling modes within parallel groups. Intra-group representation capsules can be "intra-group coupling capsules" reserved in the higher-level representation capsule set, used to characterize coupling mode features caused by factors such as circulating current, impedance inconsistencies, control parameter inconsistencies, or communication delay differences within parallel groups. In an optional configuration, the cooperative controller can configure one or a small number of intra-group representation capsules for each parallel group, or configure shared intra-group representation capsules for different coupling types, to provide an interpretable representation of the intra-group coupling mechanism without significantly increasing computational overhead.

[0137] By employing the aforementioned sparsity processing—including group-level preference determination of a single direction, mutual exclusion sparse constraints within parallel groups, and common-mode / differential-mode splitting and separate constraints—the system can suppress the dispersed routing of objects within a group to different candidate modes during parallel operation of multiple network objects. This avoids downstream impedance adjustment parameters and power allocation parameters from clashing or frequently switching within the same update cycle. Simultaneously, the differential-mode channel centrally routes inconsistencies within the group to the group's characterization capsule, enabling the collaborative controller to separate the vibration suppression responsibility of the external dominant mode from the suppression responsibility of inconsistencies in the internal coupling during the parameter generation phase. This improves the consistency and robustness of control reference quantities within the parallel group and reduces the risk of stability margin degradation caused by control jitter.

[0138] For example, the online sampling configuration for training and online deployment of the capsule network model can be as follows: the control period is 10 ms, the electrical quantity sampling rate is 2000 Hz, the sliding time window length is 256 sampling points, and the sliding step size is 2 control periods; the input channels use orthogonal component sequences in a stationary coordinate system and convert them to per-unit values ​​according to the rated values, and the phase sequence is processed to avoid cross-boundary jumps. The model structure configuration is as follows: the front-end temporal mapping module uses two-level one-dimensional convolutional layers, the first convolutional layer has 32 channels and a kernel length of 5, the second convolutional layer has 64 channels and a kernel length of 5, and normalization and nonlinear activation are used between layers; there are 32 primary capsules, and the primary capsule vector dimension is 10 dimensions, of which the first 2 dimensions are phase-amplitude physical encoding two-dimensional vectors, and the remaining 8 dimensions are amplitude-constrained learning components with a scaling factor of 0.2; there are 8 high-level representation capsules, and the output vector dimension of each high-level representation capsule is 16 dimensions; the dynamic routing iteration count is 3. The training configuration is as follows: training samples are obtained by slicing historical running data and simulation data according to the above window length and step size, and the training set and validation set are divided in an 8:2 ratio; classification constraints and reconstruction constraints are jointly optimized, with the classification constraint weight set to 1.0 and the reconstruction constraint weight set to 0.5; the Adam optimizer is used, with a learning rate of 0.001, a batch size of 64, and 80 training epochs, and training is stopped when the validation set loss does not decrease for 10 consecutive epochs; after training, the model weights and normalized parameters are fixed, and only forward inference is performed in the online phase to output the primary capsule vector set and the high-level representation capsule vector.

[0139] For example, for oscillation mechanism identification, parameter mode selection, and abnormal backoff strategy, the angle change threshold is set to 0.08 rad, and the preset number of continuous cycles is 5 control cycles; the cumulative growth threshold for the magnitude increment used to determine negative damping divergence is set to 0.04; and the upper and lower limits of the magnitude used to determine forced source periodic oscillation are set to [0.45, 0.55]. In high-damping mode, the reference damping threshold for active damping gain is set to 0.08, and the strong damping threshold is set to 0.15. In high-inertia mode, the preferred inertia-related parameter is the virtual inertia time constant T_in of the grid-type energy storage, and the strong inertia threshold is set to 1.5 s. When entering high-inertia mode, T_in is configured to be no less than 1.5 s and limited to the range of [1.5 s, 3.0 s]. To avoid a decrease in fundamental voltage regulation capability due to vibration suppression, the preferred voltage support parameters are voltage loop gain K_v or equivalent voltage support weight w_v, with the lower limit of voltage support set at K_v ≥ 0.8 or w_v ≥ 0.6. Upper and lower limit constraints and rate of change constraints are applied to these parameters before online transmission. The reconstructed residual is calculated using the normalized mean square error (MSE) caliber. Where x_i is the original time-series data stream sampling point, To reconstruct the sampling points of the timing data stream, a reconstruction residual threshold of 0.12 is set. When e > 0.12, it is determined to be an unknown mode and the system switches to a conservative damping parameter set. Example fields of the conservative damping parameter set are: damping injection gain of 0.10, virtual resistance of 0.03 pu, and power distribution increments are limited to 0 or within ±2% of rated power to prioritize system stability and control continuity when the mode is uncertain.

[0140] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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 or all of the technical features therein. 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 this application.

Claims

1. A collaborative control method for green energy grid-connected storage based on grid-type energy storage, characterized in that, include: Obtain electrical operation data of each controlled object in the source storage system and object attribute identifiers corresponding to each controlled object, and construct a time-series data stream based on the electrical operation data; The time-series data stream is input into the capsule network model and mapped to obtain a primary capsule vector set; wherein, the phase attribute of the electrical operation data is encoded as the angle component of the primary capsule vector, and the amplitude attribute of the electrical operation data is encoded as the magnitude component of the primary capsule vector; The routing weights are iteratively updated between the primary capsule vector set and the higher-level representation capsule, and convergence is determined when a preset convergence condition is met. The current oscillation mode of the source-storage system is determined based on the converged higher-level representation capsule, and the correlation strength of each controlled object relative to the current oscillation mode is determined based on the converged routing weights. Based on the correlation strength and the object attribute identifier, a parameter set is generated, which includes impedance regulation parameters for grid-type energy storage objects and power allocation parameters for grid-type energy storage objects and / or power generation objects; and the parameter set is converted into the interface converter control reference quantity of the corresponding controlled object.

2. The green energy grid-connected storage collaborative control method based on grid-type energy storage according to claim 1, characterized in that, The angular components encoded as the primary capsule vector include: An instantaneous space vector transformation is performed on the three-phase voltage data in the time-series data stream to obtain orthogonal components in the stationary coordinate system; Based on the orthogonal components, the instantaneous phase angle is determined, and the instantaneous phase angle is used as the input of the angle components; Specifically, for the sampling time deviation of different controlled objects, the phase compensation amount is determined based on the sampling time deviation, and the instantaneous phase angle is compensated and aligned.

3. The green energy grid-connected storage collaborative control method based on grid-type energy storage according to claim 1, characterized in that, The iterative update of routing weights includes: For each primary capsule vector, the set of routing weights is sparsified so that only a preset number of maximum weights are retained in the set of routing weights, and the remaining weights are either set to zero or scaled according to a preset decay factor. Furthermore, when the change in routing weight exceeds a preset threshold for two consecutive control cycles, the parameter set is generated using the routing weight of the previous control cycle, until the change in routing weight continuously meets the preset stability condition.

4. The green energy grid-connected storage collaborative control method based on grid-type energy storage according to claim 1, characterized in that, Generating the impedance regulation parameters for the grid-type energy storage object includes: The dominant oscillation frequency band is determined based on the current oscillation mode; Generate a set of frequency band-selective impedance adjustment parameters, wherein the damping injection parameter corresponding to the dominant oscillation frequency band is configured to be greater than or equal to a preset reference damping threshold, and the voltage support parameter corresponding to the fundamental frequency band is configured to be greater than or equal to a preset voltage support lower limit.

5. The green energy grid-connected storage collaborative control method based on grid-type energy storage according to claim 1, characterized in that, The method further includes: Based on the output of the converged high-level representation capsule, a reconstructed time-series data stream is generated through a reconstruction network. Calculate the reconstruction residual between the reconstructed time-series data stream and the original time-series data stream; When the reconstruction residual exceeds a preset threshold, the current oscillation mode is determined to be an unknown mode, and the parameter set is switched to a preset conservative damping parameter set.

6. The green energy grid-connected storage collaborative control method based on grid-type energy storage according to claim 4, characterized in that, The set of impedance adjustment parameters for generating band selectivity includes: Select the converged high-level characterization capsule vector corresponding to the dominant oscillation frequency band, and based on the angle and magnitude values ​​of the high-level characterization capsule vector, map the high-level characterization capsule vector to endpoints on the complex plane and construct the endpoint time-series trajectory. Calculate the angle change and magnitude change of the endpoint time-series trajectory in adjacent control cycles, respectively. The oscillation mechanism is identified based on the angle change and the modulus change, and the parameter configuration mode is selected in the impedance adjustment parameters according to the identification result. Among them, negative damping divergent oscillation corresponds to high damping mode, and forced source periodic oscillation corresponds to high inertia mode.

7. The green energy grid-connected storage collaborative control method based on grid-type energy storage according to claim 6, characterized in that, The calculation of the angle change and magnitude change of the endpoint time-series trajectory in adjacent control cycles includes: Extract the angle values ​​of the high-level character capsule vectors from two adjacent control cycles, and calculate the absolute value of their difference as the angle change. Extract the magnitude value of the high-level characterization capsule vector between two adjacent control cycles, calculate the difference between the magnitude value of the next control cycle and the magnitude value of the previous control cycle as the magnitude increment, and calculate the cumulative value of the magnitude increment within a preset number of consecutive cycles.

8. The green energy grid-connected storage collaborative control method based on grid-type energy storage according to claim 7, characterized in that, The high-damping mode includes: When the angle change is not less than the preset angle threshold within a preset number of consecutive cycles, and the cumulative value of the modulus increment is not less than the preset growth threshold, the oscillation mechanism of the current dominant oscillation frequency band is determined to be negative damped divergent oscillation. The impedance adjustment parameters include damping injection parameters, which include active damping gain; the active damping gain is configured to be no less than a preset strong damping threshold.

9. The green energy grid-connected storage collaborative control method based on grid-type energy storage according to claim 7, characterized in that, The high inertia modes include: When the angle change is not less than a preset angle threshold within a preset number of consecutive cycles, and the magnitude of the high-level characterizing capsule vector remains within the preset upper and lower limits of the magnitude within the preset number of consecutive cycles, the oscillation mechanism of the current dominant oscillation frequency band is determined to be forced source periodic oscillation. The impedance adjustment parameters include inertia-related parameters; the inertia-related parameters are configured to be greater than or equal to a preset strong inertia threshold.

10. The green energy grid-connected storage collaborative control method based on grid-type energy storage according to claim 3, characterized in that, The process of sparsifying the routing weight set corresponding to each primary capsule vector includes: Based on the object type identifier and parallel group identifier in the object attribute identifier, mutual exclusion sparse constraints are applied to the grid-type energy storage objects belonging to the same parallel group, so that only non-zero route weights pointing to a single high-rise characterization capsule are retained in the route weight set, and the remaining route weights are limited to zero or limited to a preset minimum weight upper limit. Intra-group aggregation is performed on the routing weight set corresponding to the grid-type energy storage objects in the same parallel group to obtain the group-level routing preference vector. Based on the group-level routing preference vector, the single high-level characterization capsule is determined so that the mutual exclusion sparse constraints of each grid-type energy storage object in the parallel group point to the same high-level characterization capsule in the same control cycle. Based on the primary capsule vector of the parallel group intra-network energy storage object, calculate the common mode representation vector and differential mode representation vector within the group. In the sparsification process, apply mutually exclusive sparsity constraints to the routing weight sets corresponding to the common mode representation vector and the differential mode representation vector, respectively. The routing weight set corresponding to the differential mode representation vector retains only the non-zero routing weights that point to the intra-group representation capsules used to represent the coupling modes within the parallel group.

Images