Multi-virtual power plant-oriented power distribution network security space decoupling generation method and system

Abstract

The application discloses a power distribution network security space decoupling generation method and system facing multiple virtual power plants, and belongs to the technical field of power distribution network safe operation. The method comprises the following steps: acquiring topological connection relationship data and whole network operation state data of a power distribution network; constructing a double-path graph attention teacher model, which respectively outputs whole network node voltage risk features and branch power risk features; constructing a separated student model composed of multiple sub-networks, each of which adopts an improved input convex neural network to process corresponding local power injection data and output local mapping values of global prediction risks, and then aggregates the local mapping values into global prediction risks; using a projection gradient descent algorithm to generate an adversarial sample inputting the separated student model, and taking the whole network node voltage and branch power risk features as a supervision signal to perform adversarial distillation training on the separated student model; and converting the trained separated student model into an independent linear constraint set to generate a power distribution network operation security space.

Description

Technical Field

[0001] This invention belongs to the field of distribution network safety operation technology, specifically relating to a method and system for decoupling and generating distribution network safety space for multiple virtual power plants. Background Technology

[0002] As the scale of virtual power plants expands, their profit-driven and random regulatory behaviors can easily trigger physical safety risks such as distribution network overreach. To address this, distribution network operators typically constrain virtual power plants by publishing a distribution network operation safety space, ensuring overall system security without exposing underlying network privacy parameters. However, in scenarios where multiple virtual power plants coexist within the same distribution network, a core contradiction arises between strong physical coupling and independent decision-making.

[0003] On the one hand, the global coupling characteristics of power flow in the distribution network mean that the power regulation of any virtual power plant will dynamically affect the safety margin of other entities; on the other hand, each virtual power plant, as an independent stakeholder, requires its regulation space to have complete decoupling and independence, and to be free from interference from the real-time behavior of other entities.

[0004] Existing safety space decoupling methods have significant limitations. Among model-driven methods, structural decoupling, due to its reliance on linear assumptions, struggles to characterize realistic non-convex safety boundaries; while parametric decoupling (such as the running envelope technique) suffers from a severe "curse of dimensionality," making it unsuitable for the multi-node aggregation scenarios of modern virtual power plants. Although data-driven methods reduce parameter dependence, traditional black-box models lack structural constraints and cannot mathematically guarantee multi-space independence, making them highly susceptible to inducing hidden adversarial power grabbing and power flow overrun risks during parallel operation. Furthermore, the input convex neural networks introduced to meet the requirements of convex optimization are prone to fitting distortion and misjudgment when dealing with complex, non-monotonic physical boundaries due to strict non-negative weight constraints.

[0005] Therefore, there is an urgent need in this field for a novel technical solution that can generate an absolutely decoupled, convex geometrically oriented, and physically secure independent operating space for multiple virtual power plants in a strongly coupled system, while concealing the underlying parameters of the distribution network, and supporting proactive defense against hidden power interference between the main bodies and fair allocation of differentiated capacity. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the present invention aims to provide a method and system for decoupling and generating security spaces in distribution networks with multiple virtual power plants, thereby solving the problems in existing technologies.

[0007] The objective of this invention can be achieved through the following technical solutions: A method for decoupling and generating the security space of a distribution network with multiple virtual power plants includes the following steps: Acquire data on the topological connections of the distribution network, as well as the overall network operation status data, which includes local power injection data of active and reactive power. Construct a dual-path graph attention teacher model to independently map and output the voltage risk characteristics of all network nodes and the power risk characteristics of branches; A separate student model consisting of multiple sub-networks is constructed. Each sub-network uses an improved input convex neural network to independently process the local power injection data of its corresponding virtual power plant and output the local mapping value corresponding to the global prediction risk. The local mapping values ​​of each global prediction risk are then aggregated into the global prediction risk. Adversarial examples are generated using the projective gradient descent algorithm and used as input to the split student model. The network-wide node voltage risk characteristics and branch power risk characteristics are used as supervision signals, and the split student model is trained by adversarial distillation in combination with the global predicted risk. The trained discrete student model is transformed into an independent set of linear constraints through a reverse allocation strategy, generating a distribution network operation safety space for independent control of each virtual power plant.

[0008] A distribution network security space decoupling generation system for multiple virtual power plants executes the above method, including: The data acquisition module is used to acquire topological connection data of the distribution network, as well as network-wide operation status data including active and reactive power local power injection data. The teacher model construction module is used to construct a dual-path graph attention teacher model, which independently maps and outputs the voltage risk characteristics of all network nodes and the power risk characteristics of branches. The student model construction module is used to construct a separate student model composed of multiple sub-networks. Each sub-network adopts an improved input convex neural network, independently processes the local power injection data of its corresponding virtual power plant, and outputs the local mapping value of the corresponding global prediction risk. Then, the local mapping values ​​of each global prediction risk are aggregated into a global prediction risk. The adversarial distillation training module is used to generate adversarial examples using the projective gradient descent algorithm as input to the split student model, and uses the network-wide node voltage risk characteristics and branch power risk characteristics as supervision signals, combined with the global prediction risk, to perform adversarial distillation training on the split student model. The constraint explicit allocation module is used to transform the trained discrete student model into an independent set of linear constraints through a reverse allocation strategy, thereby generating a distribution network operation safety space for independent control of each virtual power plant.

[0009] A computer storage medium storing a readable program, which, when executed, instructs a computing device to perform the above-described method for generating a distribution network security space for multiple virtual power plants.

[0010] The beneficial effects of this invention are: 1. This invention innovatively employs a dual-path graph attention network as a global physical information agent, combined with a split-input convex neural network to construct a distillation architecture. This mechanism overcomes the technical bottleneck of traditional supervised learning, which struggles to balance nonlinear power flow boundaries and model independence. Under the condition of needing only the basic operating status of the distribution network without requiring precise impedance parameters for the entire network, it generates a safe operating range with local linear superposition characteristics for each virtual power plant. This enables each entity to operate safely solely based on its own decisions, completely eliminating the strong global physical coupling constraints imposed by complex power grid topologies.

[0011] 2. This invention addresses the spatial encroachment and mutual interference problems easily induced during parallel optimization by multiple independent entities by designing an active adversarial learning strategy based on projective gradient descent. By selectively introducing the worst-case cooperation pattern into the neural network optimization path for defensive reinforcement, the model can effectively identify and avoid potential power conflict blind spots. This method effectively overcomes the deficiency of conventional black-box models where safety defenses are easily compromised under unknown and adverse operating conditions, achieving a robust and rigorous internal approximation of the global physical safety boundary.

[0012] 3. This invention constructs a reverse solution mechanism that transforms implicit risk assessment into explicit independent linear constraints, enabling precise customization of boundary parameters based on the actual installed capacity and resource characteristics (such as source, load, hybrid, and other dominant attributes) of each virtual power plant. Compared to the traditional conservative control strategy of blindly tightening boundaries, this method, without sacrificing the overall safety baseline of the distribution network, greatly taps into the available power capacity hidden by redundant constraints, thereby significantly improving the economy and flexibility of multi-virtual power plant aggregation response. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart of the multi-virtual power plant safety space decoupling generation method of the present invention; Figure 2 This is a diagram showing the comparison between the actual voltage risk value and the teacher's model output; Figure 3 This is a schematic diagram illustrating the degree of fit under different attack intensities; Figure 4 This is a diagram illustrating the false negative rate under different decision thresholds; Figure 5 This is a diagram illustrating the false alarm rate under different attack intensities; Figure 6 This is a diagram illustrating the false negative rate under different attack intensities. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Example 1 like Figure 1 As shown, the method for decoupling and generating the security space of a distribution network with multiple virtual power plants includes the following steps: S1, obtain the topology connection data of the distribution network, as well as the network-wide operation status data including active and reactive power local power injection data; From a set theory perspective, this paper mathematically defines the problem of decoupling and generating the safety space of a distribution network operation under a distribution network containing multiple virtual power plants, clarifying the relationship between the coupled safe and feasible space of the entire network and the independent Cartesian product space representing the safety space of the distribution network operation. Based on this, a data-driven overall framework for generating the safety space of the distribution network operation is constructed.

[0017] Consider a power distribution system with N nodes, whose topology is interwoven with... K Each virtual power plant is independent of the others. Define a set. This is the set of virtual power plant indexes. For the... k A virtual power plant, whose managed set of nodes is denoted as . Furthermore, the jurisdictions of each entity do not overlap, thus satisfying the requirement. The controllable resource state vector of the entire network is denoted as... (Including active and reactive power injection for all nodes in the network). Correspondingly, the first... k Local control variables of a virtual power plant Defined as a global vector In the node set Subvectors on.

[0018] The physical safety boundary of a distribution network is implicitly defined by nonlinear AC power flow equations and operational constraints such as node voltage and line thermal stability. This defines the physical feasible space of the entire network. The set of all operating points that satisfy the safety constraints: (1) In the formula, This represents a highly nonlinear and non-convex power flow constraint mapping.

[0019] Due to the existence of physical coupling, the first k The safe operation of a virtual power plant depends not only on its own decisions. It is also dynamically constrained by the real-time behavior of other virtual power plants, resulting in Geometrically, it presents itself as a complex non-convex form where all dimensions are tightly intertwined and cannot be directly separated. The essence of generating the distribution network operation safety space is to directly construct a set of structurally independent distribution network operation safety spaces within the feasible space of the entire network. From a topological perspective, this problem is equivalent to finding a set of locally decoupled subspaces. , so that the combined space they form Satisfying the Cartesian product structure: (2) Under this structure, each local space Only by the first k Local power variables of a virtual power plant The definition is independent of its external variables. This segregated definition based on a subset of variables mathematically ensures that the feasible space of the entire network can be decomposed into the direct product of independent subspaces of each virtual power plant, thus satisfying the independent control requirements of multiple virtual power plants. Simultaneously, to ensure physical security, a Cartesian product space is generated. It must be a strict subset of the original globally feasible space, i.e., satisfy the inner approximation condition: (3) This condition implies a defense requirement against the risk of adversarial coupling, meaning that for any combination of sampling within the local subspace, even if all virtual power plants are operating simultaneously under the worst combination of conditions at the boundary limits, the overall network state after superposition must still be strictly within the physical safety boundary.

[0020] In summary, the core of the problem of generating the safety space for distribution network operation lies in learning a set of mapping functions that have the structural decoupling shown in equation (2) and satisfy the physical safety constraints shown in equation (3) through a data-driven approach.

[0021] Subsequently, this invention constructs a hierarchical generation framework for the safety space of distribution network operation, comprising: a global physical sensing layer, a decoupled distribution network operation safety space generation layer, and a distribution network operation safety space differential allocation layer, and executes the following steps S2~S4 respectively: S2, construct a dual-path graph attention teacher model, taking the topology connection relationship data and the network operation status data as input, and independently map and output the network node voltage risk characteristics and branch power risk characteristics respectively; Specifically, the global physical perception layer (teacher model) employs a dual-path graph attention neural network architecture, using the active and reactive power injected into nodes as input features and a binary adjacency matrix containing only connectivity relationships as topological constraints. Adaptive feature aggregation is performed using a graph attention mechanism, and dynamically generated attention coefficients implicitly represent the electrical coupling strength between nodes, thereby achieving a high-fidelity fit to the physical laws of the entire network under impedance parameter-free conditions. At the output end, the model features a dual-path parallel structure, independently mapping the safety risks of the voltage state of all network nodes and the power state of branches, providing global safety supervision signals for the student model.

[0022] S3. Construct a separate student model consisting of multiple sub-networks. Each sub-network uses an improved input convex neural network to independently process the local power injection data of its corresponding virtual power plant and output the local mapping value of the corresponding global prediction risk. Then, the local mapping values ​​of each global prediction risk are aggregated into a global prediction risk. Adversarial examples are generated using the projection gradient descent algorithm as input to the separate student model. The voltage risk characteristics of all network nodes and the power risk characteristics of branches are used as supervision signals. The separate student model is trained by adversarial distillation in combination with the global prediction risk to defend against the power coupling risk between virtual power plants. Specifically, the decoupled distribution network operation safety space generation layer (student model) is responsible for directly constructing the distribution network operation safety space that satisfies the Cartesian product structure. It employs a separate input convex neural network architecture. K The model consists of physically unconnected subnetworks, each handling only the local power injection from its corresponding virtual power plant. By performing risk superposition and combination operations at the aggregation layer, the model constructively forces the output function of the entire network to satisfy the form of separable variables, thus ensuring that the generated distribution network operation safety space initially possesses Cartesian product topological properties. Simultaneously, the non-negativity constraint of the internal connection weights of the improved input convex neural network ensures the convex geometric characteristics of the distribution network operation safety space. To defend against the implicit power coupling risk between virtual power plants, a projection gradient descent adversarial training strategy is introduced during training. This proactively defends against the worst coupling samples to achieve a robust internal approximation of the grid safety boundary.

[0023] S4, the trained separate student model is transformed into an independent set of linear constraints through a reverse allocation strategy, generating a distribution network operation safety space for independent control of each virtual power plant; Specifically, the distribution network operation safety space differential allocation layer (differentiated distribution network operation safety space generation) is responsible for transforming each input convex neural network into an explicit, proportionally allocated constraint expression. Through a reverse allocation method, the appropriate risk intercept is solved in reverse, based on the type of different virtual power plants and fairness weights. This mechanism transforms implicit risk assessment into explicit distribution network operation safety space risk thresholds, supporting differentiated customization and fair allocation of multiple distribution network operation safety spaces.

[0024] In the aforementioned hierarchical distribution network operation safety space generation framework, the teacher model plays a crucial role as the physical safety referee for the entire network. Its core task is to replace the computationally complex traditional power flow equation solution process and establish a high-precision differentiable mapping from the injected power at all network nodes to the multi-dimensional safety risk state of the system. Given the significant differences in the physical formation mechanisms of node voltage and branch power, a single feature extraction network cannot adequately consider both types of heterogeneous features. Therefore, this invention designs an independent dual-path graph attention neural network architecture, constructing separate node voltage sensing and branch power sensing channels, thus separating the physical tasks during the feature extraction stage.

[0025] The model constructs a graph structure input based solely on the physical connections of the distribution network. The input feature matrix for all nodes in the network is defined. Active power of nodes With reactive power The splicing, that is: (4) Simultaneously, construct a binary adjacency matrix. As a topological constraint, this matrix only represents physical connectivity and does not depend on specific admittance parameters. During feature aggregation, As a structural mask, it forces information to be transmitted only between physically connected nodes, thereby introducing an inductive bias consistent with the power grid topology.

[0026] To achieve dual-path modeling of node voltage and branch power, the model is designed with two encoder channels with completely independent parameters: a voltage encoder and a power encoder, both of which adopt a graph attention neural network structure.

[0027] To address the limitation of only inputting a binary adjacency matrix without explicit impedance parameters, the model utilizes a self-attention mechanism to compensate for the missing parameters. l The node feature update rule for a layer is formalized as follows: (5) In the formula, The weight matrix is ​​a learnable linear transformation. For adjacency matrix Defined neighbor set, It is a non-linear activation function. This is a residual term, ensuring the effective propagation of deep features. Attention coefficient. The nodes were dynamically measured. j For nodes i The influence weights are calculated based on feature concatenation and linear mapping: (6) (7) This mechanism enables the model to automatically learn the implicit line impedance weights from the running data, achieving reconstruction of physical laws under parameter-free conditions.

[0028] The high-dimensional features extracted by the dual encoders are fed into independent decoders and mapped to specific physical risk indicators. ① The node voltage risk vector of the unbalanced nodes in the system. The first... n After peacekeeping n The dimensions are the safety risks of exceeding the upper and lower voltage limits, respectively. The label is defined as the directed distance of the actual operating point relative to the safety boundary, retaining the directional information of exceeding the limit. Specifically: (8) In the formula, and These are the upper and lower voltage safety limit vectors, respectively. This is the normalization factor for the node voltage vector. y When y > 0, it indicates that the limit has been exceeded; when y ≤ 0, it indicates that the safety margin is within the limit.

[0029] ② Branch power risk vector. Corresponding to the entire network. n Thermal stability of the branch. The label is defined as the normalized distance of the apparent power amplitude relative to the line capacity: (9) In the formula, This is the upper limit vector for branch power safety. This is the normalization factor for the branch power vector.

[0030] During training, the model employs a multi-task joint loss function. Given that the input features already contain complete graph topology connectivity information, to balance the learning of voltage and power constraints, the loss function is defined as the direct sum of the voltage risk prediction error and the line load risk prediction error: (10) In the formula, M For batch size, For the predicted output of the teacher model, This is a true risk label calculated based on the standard AC power flow equation.

[0031] By minimizing the joint loss, the dual-path graph attention neural network can simultaneously approximate the safety risks of node voltage and branch power in a high-dimensional risk space, providing an accurate global supervision signal for the subsequent decoupling generation of the distribution network operation safety space to guide student models.

[0032] To achieve strictly convex generation of the safety space for power distribution network operation while ensuring computational efficiency, this invention proposes a student model based on a dual-path split-input convex neural network. Similar to the teacher model, the student model also establishes independent evaluation networks for node voltage and branch power to address the nonlinear characteristics of the two distinct physical constraints.

[0033] To achieve power decoupling among multiple virtual power plants, this invention constructs an additive function based on the risk output of a multi-input convex neural network. By using a discrete input convex neural network, the global safety risk of the power grid is distributed across the operational safety space of each distribution network, achieving structural decoupling of the operational safety spaces of multiple distribution networks. Simultaneously, by forcing the network weights to be non-negative using the input convex neural network, the convexity of the output with respect to the input is guaranteed.

[0034] For those classified as K In this invention, the power state vector of the entire network is decomposed into local state vectors, representing a distribution network with virtual power plants. The definition of the first... k The local risk mapping of a virtual power plant is as follows: , is used to represent a multi-layer input convex neural network.

[0035] To eliminate complex power coupling between virtual power plants and enable plug-and-play operation of the distribution network safety space when multiple virtual power plants coexist, this invention constructs a global risk function using a linear superposition method: (11) In the formula, For belonging to the first k Local power injection sub-vectors of a virtual power plant; This is the parameter set for the subnetwork; This represents a global risk deviation.

[0036] To address the multi-quadrant power operation and non-convex characteristics of distribution networks, the following key improvements were made to the original input convex neural network characterizing the safety space of distribution network operation: First, to maintain convexity, the original input convex neural network structure requires non-negative weights, which leads the network to tend to fit monotonically non-decreasing functions, making it difficult to directly handle the non-monotonic boundaries caused by bidirectional power fluctuations in the distribution network. Therefore, this invention expands the connections of the input vector to obtain a new input vector. : (12) This operation maps the original input space to a high-dimensional symmetric space, enabling the network to capture the nonlinear contributions of both positive and negative power fluctuations to risk while maintaining non-negative weight constraints.

[0037] Furthermore, unlike the ReLU activation function commonly used in traditional input convex neural networks, this invention employs the Softplus function. During the training phase of an input convex neural network, because the weights are strictly restricted to non-negativity, directly using the ReLU activation function can easily lead to the vanishing gradient problem, making it extremely difficult for deep networks to converge. Therefore, this invention uses Softplus during training to maintain a non-zero gradient across the entire network, ensuring the model's ability to learn complex boundaries. Subsequently, to strictly satisfy the combinatorial property of convex functions, the weight matrices of all hidden and output layers... W All have undergone Softplus function reparameterization: (13) The above formula forces W ≥ 0. To further overcome the gradient decay problem of traditional input convex neural networks and improve the model's ability to represent high-dimensional complex boundaries, this invention introduces a hybrid approach of directly connected input paths and convexity-preserving residual connections in the input convex neural network architecture. Unlike the direct addition in conventional residual networks, the inter-layer recursive formulas of input convex neural networks must be specially designed to satisfy convexity constraints. l Neuron activation status in layer +1 Defined as: (14) In the formula, The inter-layer weights of the main path are constrained to be non-negative. The input weights are directly connected and have no sign restrictions, allowing the network to directly capture the linear features of the input. The residual path weights are strictly constrained to be non-negative. This is the residual scaling factor.

[0038] The residual term is not a simple identity mapping, but a non-negative linear transformation of the features of the previous layer. This design retains the advantage of residual networks in promoting gradient propagation, while mathematically guaranteeing the global convexity of the entire network output with respect to the input.

[0039] Due to the adoption of a split-input convex neural network structure, each local network... The inability to directly perceive the status of other virtual power plants may lead to a blind spot regarding cross-regional coupling risks. To compensate for this deficiency, this invention introduces an adversarial training mechanism based on projective gradient descent, which enhances the model's defensive robustness through proactive power attacks between different virtual power plants. In each training iteration, instead of directly using the original samples, adversarial examples are generated using the projective gradient descent algorithm. The attack aims to identify covertly coupled operational scenarios that the teacher model deems extremely dangerous, but the current student model misjudges as safe. The attack's objective function is... The difference between the teacher model risk and the student model risk: (15) To solve this problem, this invention employs a multi-step projective gradient descent iterative algorithm to generate adversarial examples. Let... For the original input, the first t The iterative update formula for each step is as follows: (16) In the formula, For attack step size; The sign function is used to extract the direction of the gradient. It is the standard operation for projective gradient descent adversarial attacks and is convenient for efficiently finding extrema under infinite norm constraints. For the objective function with respect to the input The gradient; This is a projection operator used to truncate the updated samples to be centered on the original samples. - Within the neighborhood, and simultaneously satisfying both physical upper and lower bound constraints: (17) Through multiple rounds of iteration, the final generated This represents the worst-case scenario of a zero-sum game between virtual power plants. This sample exposes the weaknesses of current decoupling strategies, forcing student models to cover these extreme conditions in subsequent training.

[0040] To strictly ensure that the generated distribution network operation safety space is an inner approximation of the physical safety domain, an asymmetric safety loss function is used in the training, imposing a high penalty on the risk of underreporting unsafe power points as safe. Total loss function. Defined as: (18) (19) (20) (twenty one) In the formula, The L1 norm is used to fit the true value of the physical risk as an L1 fitting term. Compared to mean squared error, it is more robust to outliers, enabling the model to learn the main trends in risk distribution; This is a penalty item for safety violations, specifically targeting cases of unreported risks, and includes... As a safety buffer margin, when When this occurs, the term produces a non-zero gradient; For fairness regularization, this term is introduced to prevent a situation where some virtual power plants bear all the risks while others experience unrestricted degradation. It constrains the mean output of each local sub-network. The degree of dispersion.

[0041] To address the imbalance in rights allocation in multi-virtual power plant scenarios, this invention proposes a differentiated distribution network operation safety space generation mechanism based on a reverse allocation strategy. First, an improved input convex neural network linear constraint transformation method based on convexity-preserving residuals is derived. Then, according to preset types and fairness ratios, the reverse allocation strategy for distribution network operation safety space risk thresholds is detailed, accurately calculating the precise risk thresholds for each distribution network operation safety space.

[0042] Although the Softplus activation function guarantees smooth gradient propagation during training, its nonlinear characteristics hinder the linear constraint transformation of the input convex neural network. To address this, this invention proposes a method for linear transformation of input convex neural network constraints based on the linear lower bound of the Softplus function, and also derives an explicit output error for model analysis (derivation details are in Appendix B).

[0043] For a neuron in the l-th layer of the network (l≥1), the pre-activation input of the main path is defined as... The residual path input is : (twenty two) (twenty three) Using the two asymptotes of the Softplus function, the neuron's output state is derived. The two linear lower bounds are given. The inactive state is the residual baseline; when the main path input is suppressed, the activation function tends to zero, and the output is supported only by the residual term. (twenty four) The activation state is the superposition of residuals. When the main path input is in the response region, the activation function tends to undergo a linear identity transformation, and the output is a linear superposition of the main path and the residual terms. (25) Based on the above analysis, specific network parameters are substituted into the final risk threshold. By restricting the network output, the forward computation process of the input convex neural network can be transformed into a set of standard linear inequalities: (26) (27) In the formula, , and The residuals, interlayer weights, and input weights, respectively, all satisfy the non-negativity constraint. A vector consisting entirely of 1s represents a vector. Each component is subject to a scalar threshold. Constraints.

[0044] Ultimately, it can be transformed into a compact form: (28) In the formula, and These are the linearized compact form coefficient vectors, This is the risk prediction output for the linearized model.

[0045] It is worth noting that the above linearization process utilizes the lower bound property of convex functions, leading to the predicted values... It is always slightly smaller than the risk value of the original input convex neural network. Based on the derivation in Appendix B, the maximum linearization error of a single layer of neurons is... The upper bound of the total network error after multi-layer accumulation is To minimize this deviation and ensure that the generated distribution network operating safety space strictly meets the safety requirements of the original physical model, this invention introduces a safety compensation margin. The final risk threshold is adjusted to construct the following robust linear constraint: (29) This constraint ensures that as long as the output of the linear model satisfies: (30) Then the real physical risk must satisfy This provides a conservative boundary for subsequent risk parameter calculations.

[0046] In multi-virtual power plant scenarios, different entities exhibit significant physical heterogeneity due to differences in resource composition. Therefore, it is necessary to allocate differentiated risk thresholds for different types of virtual power plants to generate a distribution network operation safety space that satisfies the principle of fairness among these differences. The following section uses active power capacity as an example to illustrate the proportion of active power capacity in each virtual power plant. As a proportional allocation rule, and taking three typical virtual power plant types—source type, load type, and hybrid type—as examples, this paper introduces the method for generating differentiated safety spaces for distribution network operation.

[0047] First, based on the allocation coefficients and following the conservative principle of minimizing the normalized capacity, the minimum value of the normalized unit power benchmark is calculated within the set of virtual power plants of the same type. This ensures that the benchmark setting is suitable for the virtual power plants with the most limited physical regulation capabilities within the set, thereby avoiding the unsolvable risk caused by an excessively high benchmark. Then, based on the determined unit benchmark, differentiated power boundaries are set for each virtual power plant, and a risk threshold is applied using an input convex neural network. By utilizing the strict monotonicity of the aggregated active power value, the precise risk threshold can be solved using the inverse binary search method. .

[0048] The reverse binary search process is shown in the table below: Through the above process, a set of mathematical constraints that are structurally decoupled, convex, linear, and satisfy the differential allocation requirements can be obtained. For any virtual power plant in the distribution network, its corresponding distribution network operation safety space is a standardized linear model that can be directly embedded into the solver. Specifically: (33) in, and for: (34) in, This is the local injection power vector. These represent the active power vector and the reactive power vector, respectively. This is the expanded input feature vector. This is an improved input convex neural network model function. Superscripts indicate variables related to node voltages. s Superscript indicates a variable related to branch power. These represent the parameter sets (including all weights and biases) of the voltage ICNN and branch power ICNN models, respectively. Let $L$ be the node voltage risk mapping value and branch power risk mapping value output by the k-th virtual power plant. $L$ represents the total number of layers in the neural network. This is a hidden variable in the last layer. This is the weight matrix from the last layer of hidden variables to the output layer. For input The weight matrix is ​​directly passed through to the output layer. This is the bias vector for the output layer. This represents the safety boundary threshold or the upper limit of the global predicted risk for the k-th sub-network. It is a vector of all 1s, used for dimension alignment. This is the index of the network layer. For the first l The hidden state variables of the layer (in the formula) (Corresponding to the function of the ReLU activation function). These are the weight matrix and bias vector of the input layer. For the first l The hidden state of layer -1 to the... l The main connection weight matrix of the layer (to ensure the convexity of the network with respect to the input, ICNN usually requires that the elements of this matrix be non-negative, i.e. ). For the original input Jump to the first l Layer weight matrix. For the first l -1st floor to the l The residual connection weight matrix of the layer. For the first l The bias vector of the layer. These are the weighting coefficients or adjustment hyperparameters for the residual connection terms.

[0049] The hyperparameters used in the above process are explained as follows: In this embodiment, uniform data sampling is performed within a specific power range, generating a total of 34,000 sets of sample data, which are divided into training and test sets in an 8:2 ratio, with a batch size of 1024. In the teacher model, the node voltage path contains 3 layers of GAT (hidden layer dimension 64) and a 256-dimensional MLP, while the branch power path contains 5 layers of GAT (hidden layer dimension 128) and a 1024-dimensional MLP. Dropout is 0.01 for both, and the Adam optimizer (learning rate 1e-5) is used for iterative training based on mean squared error loss for 2000 epochs. In the student model, the hidden layer sizes for the node voltage path and the branch power path are [512, 256, 256, 128] and [256, 512, 512, 128], respectively, and the activation function is Softplus with a smoothing factor of 300. The residual scaling factor was 0.9. During the adversarial training phase, a PGD attack strategy was employed (5 iterations, step size 0.1, maximum perturbation 0.5), with a learning rate set to 6e-5 and a total of 2000 iterations. The loss function used was an L1 norm-based safety margin loss, with an explicit safety margin of 0.05, an over-limit penalty weight of 50.0, an effectiveness regularization weight of 2.0, and an additional cluster fairness constraint with a weight of 0.1. In the risk-based allocation phase, the 97.5% safety quantile was calculated using 20% ​​of the calibration data, with a voltage safety margin of 0.089 and a branch power safety margin of 0.079.

[0050] Example 2 To verify the effectiveness of the generation method in Example 1, this example proposes the following test and analysis scheme.

[0051] This embodiment presents a simulation test system as follows: an IEEE-33 node system containing four virtual power plants: voltage limit of 12.66kV × 0.95 / 1.05, current limit of 1kA. All virtual power plants in the test system are initially configured as hybrid virtual power plants. Virtual power plant 1 contains nodes: 1, 2, 3, 10, 12, 22, 30, 31, 32; Virtual power plant 2 contains nodes: 7, 9, 11, 16, 17, 18, 26, 27, 28; Virtual power plant 3 contains nodes: 8, 15, 19, 20, 23, 24, 25; Virtual power plant 4 contains nodes: 4, 5, 6, 13, 14, 21, 29.

[0052] C1 represents the method and mechanism proposed in this embodiment.

[0053] C2 is the control group of neural network models, where C2-1 replaces the teacher model with a multilayer perceptron, and C2-2 replaces the teacher model with a graph convolutional neural network.

[0054] C3 serves as the framework control group. C3-1 uses the proposed improved input convex neural network for adversarial training to directly derive the operational safety space for each distribution network. The parameters of the improved input convex neural network are the same as in C1. C3-2 uses a graph attention neural network to train the teacher model, while the student model still employs the improved input convex neural network and corresponding separation mechanism, without adversarial training. The parameter settings are the same as in C1.

[0055] Figure 2 The results show that, for both bidirectional voltage risk and unidirectional power risk, the predicted points converge closely to the diagonal, indicating that the model does not exhibit significant systematic bias.

[0056] Figure 3 The study demonstrates the false negative rate (FRR) variations of three models under different attack intensities. C3-2 exhibits a serious security flaw, with a FRR as high as 32.07% under undisturbed conditions, and this FRR continues to increase with attack intensity. This means that if other virtual power plants experience power fluctuations, C3-2's security capabilities immediately fail. In contrast, C1 demonstrates stronger decoupling and security characteristics. C1 consistently keeps the FRR below 0.17%, and the FRR gradually decreases with increasing attack intensity. This indicates that through adversarial decoupling distillation, C1 successfully establishes a defense mechanism against extreme coupling conditions. This defense capability against the implicit power coupling risk between virtual power plants confirms physical-level operational decoupling; that is, as long as each virtual power plant controls its power within the safe operating space of the distribution network allocated by C1, the physical security of the entire network can be reliably guaranteed regardless of how other virtual power plants adjust their power.

[0057] Figure 4The false alarm rate curve primarily describes the situation where points that are initially safe are reported as unsafe. While C3-1 has a relatively high false alarm rate (i.e., high safety), its false alarm rate is as high as 43.61% and remains persistently high. This indicates that due to the lack of global physical guidance from a teacher model, C3-1 has to adopt a blindly conservative strategy, misclassifying a large number of safe power points as dangerous areas, resulting in a significant waste of power adjustable capacity. C1, while maintaining the same false alarm rate as C3-1, controls the initial false alarm rate at 8.27%, and further reduces it to 3.65% as the attack training intensity adapts. This confirms that the distribution network operation safety space generated by the proposed method is not blindly conservative, but rather an efficient decoupling space closely adhering to physical boundaries, achieving a balance between safety and economy.

[0058] Figure 5 and Figure 6 This further reveals the aforementioned performance differences. From Figure 5 Looking at the mean squared error, C1's mean squared error is higher than C3-2's. This is not a degradation of model performance, but a direct reflection of the effect of the asymmetric security loss function. In exchange for grid security, C1 actively drives the prediction boundary to conservatively shift towards the feasible region. Figure 6 Sensitivity analysis validated the effectiveness of this shift. Under clean data, when the decision threshold was tightened to the physical critical value of 0.00, the false negative rate of C3-2 surged to nearly 40%, indicating a lack of tolerance space at its boundary; while the false negative rate of C1 remained consistently below 0.25%. This result demonstrates that the proposed method has successfully constructed an implicit safety buffer in the parameter space of the input convex neural network, achieving a robust internal approximation of the physical boundary without sacrificing excessive power adjustability.

[0059] Example 3 This embodiment proposes a distribution network security space decoupling generation system for multiple virtual power plants, specifically including: The data acquisition module is used to acquire topological connection data of the distribution network, as well as network-wide operation status data including active and reactive power local power injection data. The teacher model construction module is used to construct a dual-path graph attention teacher model, which independently maps and outputs the voltage risk characteristics of all network nodes and the power risk characteristics of branches. The student model construction module is used to construct a separate student model composed of multiple sub-networks. Each sub-network adopts an improved input convex neural network, independently processes the local power injection data of its corresponding virtual power plant, and outputs the local mapping value of the corresponding global prediction risk. Then, the local mapping values ​​of each global prediction risk are aggregated into a global prediction risk. The adversarial distillation training module is used to generate adversarial examples using the projective gradient descent algorithm as input to the split student model, and uses the voltage risk characteristics of the entire network nodes and the power risk characteristics of the branches as supervision signals, combined with the global prediction risk, to perform adversarial distillation training on the split student model in order to defend against the power coupling risk between virtual power plants. The constraint explicit allocation module is used to transform the trained discrete student model into an independent set of linear constraints through a reverse allocation strategy, thereby generating a distribution network operation safety space for independent control of each virtual power plant.

[0060] Based on a similar inventive concept, embodiments of the present invention also provide a computer storage medium storing a readable program that, when run by a processor, can execute the above-described method for generating a distribution network security space for multiple virtual power plants.

[0061] Based on a similar inventive concept, this invention provides an electronic device, including: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus; The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the above-described method for decoupling and generating the security space of a distribution network for multiple virtual power plants.

[0062] Based on a similar inventive concept, embodiments of the present invention also provide a computer program product, including computer instructions, which instruct a computing device to perform the operations corresponding to the above-described method for generating a distribution network security space for multiple virtual power plants.

[0063] The methods of the present invention can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD-ROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code originally stored on a remote recording medium or a non-transitory machine-readable medium and subsequently stored on a local recording medium, downloaded via a network. Thus, the methods described herein can be processed by software stored on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components (e.g., RAM, ROM, flash memory, etc.) capable of storing or receiving software or computer code that, when accessed and executed by the computer, processor, or hardware, implements the methods described herein. Furthermore, when a general-purpose computer accesses the code used to implement the methods shown herein, the execution of the code transforms the general-purpose computer into a dedicated computer for performing the methods shown herein.

[0064] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A method for decoupling and generating the safety space of a distribution network for multiple virtual power plants, characterized in that, Includes the following steps: Acquire data on the topological connections of the distribution network, as well as the overall network operation status data, which includes local power injection data of active and reactive power. Construct a dual-path graph attention teacher model to independently map and output the voltage risk characteristics of all network nodes and the power risk characteristics of branches; A separate student model consisting of multiple sub-networks is constructed. Each sub-network uses an improved input convex neural network to independently process the local power injection data of its corresponding virtual power plant and output the local mapping value corresponding to the global prediction risk. The local mapping values ​​of each global prediction risk are then aggregated into the global prediction risk. Adversarial examples are generated using the projective gradient descent algorithm and used as input to the split student model. The network-wide node voltage risk characteristics and branch power risk characteristics are used as supervision signals, and the split student model is trained by adversarial distillation in combination with the global predicted risk. The trained discrete student model is transformed into an independent set of linear constraints through a reverse allocation strategy, generating a distribution network operation safety space for independent control of each virtual power plant.

2. The method for decoupling and generating distribution network security space for multiple virtual power plants according to claim 1, characterized in that, The process of mapping and outputting the voltage risk characteristics of all network nodes and the power risk characteristics of branches includes: Construct a binary adjacency matrix based on the topological connection relationship data; The active power and reactive power of all nodes in the network are concatenated to form an input feature matrix; Construct parameter-independent node voltage sensing channels and branch power sensing channels. Using the input feature matrix as input and the binary adjacency matrix as a structure mask, perform adaptive feature aggregation using a graph attention mechanism to generate attention coefficients that implicitly characterize the electrical coupling strength between nodes. Output the network-wide node voltage risk feature representing the directed distance of node voltage exceeding the limit, and the branch power risk feature representing the normalized distance of apparent power of the branch.

3. The method for decoupling and generating distribution network security space for multiple virtual power plants according to claim 1, characterized in that, The process by which the subnetwork processes local power injection data using an improved input convex neural network includes: The local power injection data is concatenated with its inverse data vector to construct a symmetric high-dimensional input vector. ; Softplus is used as the activation function, and the weights of the main path interlayers and residual path layers of the improved input convex neural network are both restricted to non-negative values. A hybrid interlayer recursive formula is constructed, which includes input direct connections and convexity-preserving residual connections, to generate a network expression with local linear superposition characteristics and global output convexity.

4. The method for decoupling and generating distribution network security space for multiple virtual power plants according to claim 1, characterized in that, The steps for generating adversarial examples using the projective gradient descent algorithm include: The attack objective function is to maximize the difference between the comprehensive physical risk and the global predicted risk, wherein the comprehensive physical risk is determined based on the voltage risk characteristics of the entire network nodes and the power risk characteristics of the branch. The adversarial sample representing the worst power coordination mode among multiple virtual power plants is generated by iteratively optimizing along the gradient direction of the attack target function using a multi-step projection gradient descent algorithm and truncating the updated sample in a neighborhood that satisfies the physical upper and lower limits.

5. The method for decoupling and generating distribution network security space for multiple virtual power plants according to claim 1, characterized in that, The joint loss function used in the adversarial distillation training is: in, For the total loss, For L1 fitting loss term, For safety violations and penalties, For fairness regularization, and These are the weighting coefficients for safety violation penalties and fairness regularization, respectively. For the true value of physical risk, This is the predicted output of the student model; For safety buffer margin, Indicates the first The mean of the risk output of each local subnetwork.

6. The method for decoupling and generating security space in a distribution network for multiple virtual power plants according to claim 3, characterized in that, The steps of transforming the trained segregated student model into an independent set of linear constraints using a reverse assignment strategy include: Based on the asymptote property of the activation function, two linear lower bounds are obtained for the improved input convex neural network in the non-activated and activated states, respectively. The forward computation process of each layer of neurons in the improved input convex neural network is transformed into a standard linear inequality based on the two linear lower bounds. A safety margin parameter is introduced to correct the linearization calculation error accumulated in multiple layers of the network, and the risk prediction output constraint of the improved input convex neural network is transformed into a robust linear inequality constraint containing the safety margin parameter.

7. The method for decoupling and generating distribution network security space for multiple virtual power plants according to claim 1, characterized in that, The step of generating a distribution network operation safety space for independent control of each virtual power plant includes: Based on the principle of normalized capacity minimization, the minimum value of the normalized unit power benchmark is calculated and obtained within a set of virtual power plants of the same type. The allocation coefficient is determined based on the minimum value of the unit power benchmark, the preset virtual power plant type characteristics, and the fairness ratio weight. Based on the allocation coefficients and the robust linear inequality constraints, the precise risk thresholds for each sub-network are solved in reverse, generating an explicit linear distribution network operation safety space boundary with multiple virtual power plants decoupled.

8. The method for decoupling and generating security space in a distribution network for multiple virtual power plants according to claim 2, characterized in that, The binary adjacency matrix represents the physical connectivity between nodes in the distribution network.

9. A distribution network security space decoupling generation system for multiple virtual power plants, comprising the method described in any one of claims 1-8, characterized in that, include: The data acquisition module is used to acquire topological connection data of the distribution network, as well as network-wide operation status data including active and reactive power local power injection data. The teacher model construction module is used to construct a dual-path graph attention teacher model, which independently maps and outputs the voltage risk characteristics of all network nodes and the power risk characteristics of branches. The student model construction module is used to construct a separate student model composed of multiple sub-networks. Each sub-network adopts an improved input convex neural network, independently processes the local power injection data of its corresponding virtual power plant, and outputs the local mapping value of the corresponding global prediction risk. Then, the local mapping values ​​of each global prediction risk are aggregated into a global prediction risk. The adversarial distillation training module is used to generate adversarial examples using the projective gradient descent algorithm as input to the split student model, and uses the network-wide node voltage risk characteristics and branch power risk characteristics as supervision signals, combined with the global prediction risk, to perform adversarial distillation training on the split student model. The constraint explicit allocation module is used to transform the trained discrete student model into an independent set of linear constraints through a reverse allocation strategy, thereby generating a distribution network operation safety space for independent control of each virtual power plant.

10. A computer storage medium storing a readable program, characterized in that, When the program runs, it can instruct the computing device to execute the distribution network security space decoupling generation method for multiple virtual power plants as described in any one of claims 1-8.

Images