Power distribution network fault control method based on multi-source data fusion

Abstract

The application discloses a power distribution network fault control method based on multi-source data fusion, relates to the technical field of power distribution network fault monitoring, and comprises the following steps: constructing a physical-logic double-layer coupled network by taking an electrical admittance topology as a physical layer and a generalized potential manifold as a logic layer; mapping an operation pressure index to a physical layer matrix corresponding to the physical layer, fusing the operation pressure index with a logic layer gradient feature, driving the logic layer to evolve in a state, solving a primary control instruction, and generating a secondary allocation vector in an optimization direction along the logic layer gradient feature; performing an orthogonal projection operation on the secondary allocation vector to a zero space of a fault section power flow sensitivity matrix, superimposing the secondary allocation vector with the primary control instruction, generating a final control instruction, and delivering the final control instruction to the physical layer. The application is used to solve the problems that in the prior art, multi-agent economic game and network frame physical safety are decoupled, heterogeneous multi-source data is difficult to cooperate, and traditional optimization iteration is lengthy and difficult to meet real-time response requirements in a power distribution network fault control scene.

Description

Technical Field

[0001] This invention relates to the field of power distribution network fault monitoring technology, and more specifically, to a power distribution network fault control method based on multi-source data fusion. Background Technology

[0002] Existing distribution network control methods typically employ optimal power flow (OPF) or distributed collaborative optimization models, which transform the electrical parameters and power flow constraints on the physical side into multidimensional mathematical constraints, and use economic indicators such as nodal electricity prices as objective functions, and solve them numerically through large-scale nonlinear programming algorithms.

[0003] However, the aforementioned control methods face significant bottlenecks in practical applications. Existing technologies often employ a sequential approach to handle "physical-economic" coordinated control, resulting in a substantial structural decoupling between multi-stakeholder economic game dynamics and network physical security. Even when using "cyber-physical system (CPS) coupling" control schemes, the coupling depth often remains at the superficial data interaction level, such as adding communication terminals, constructing network adjacency matrices, and measuring transmission delays. When dealing with faults like system active power imbalance, these methods often rely on simple proportional allocation and consistency iteration using node capacity or bandwidth. Due to the lack of deep algebraic structural isomorphism, heterogeneous operating parameters at the physical layer are difficult to convert into decision signals at the control layer in real time and without loss. Relying solely on external interface iterations not only easily leads to communication delays but also results in generated instructions that are highly susceptible to deviating from physical security constraints. To compensate for this lack of cross-domain integration, the system is forced to rely on complex global numerical optimization to explore feasible solutions, causing the computational dimension to increase exponentially with scale, making it difficult to meet millisecond-level real-time response requirements.

[0004] Furthermore, existing distributed algorithms typically linearize the strictly nonlinear AC power flow equations, transforming them into polyhedral safety boundaries, or introducing them as penalty terms into the "soft constraints" of the optimization objective. This approximation method is well-suited for steady-state control scenarios; however, in emergency situations requiring millisecond-level rapid response, such as distribution network faults, the safety of control commands depends entirely on the complete convergence of the optimization algorithm. Due to computational delays caused by multi-agent competition, multi-agent game iterations often fail to achieve instantaneous complete convergence in emergency situations. In such cases, if resource allocation commands are directly issued based on unconverged intermediate results that do not yet fully conform to the physical manifold, the control commands, lacking a rigid interception mechanism independent of the game solution state, are highly prone to deviating from the true physical safety boundary. This can lead to serious risks such as transient power mismatch, local voltage exceeding limits, and static or thermal stability limit exceeding limits in unconverged situations. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a distribution network fault control method based on multi-source data fusion, in order to solve the problems in the prior art where multi-subject economic game and network physical security are decoupled in the distribution network fault control scenario, heterogeneous multi-source data are difficult to coordinate, and the control command depends on the complete convergence of optimization, which easily leads to steady-state safety exceeding the limit under non-convergence conditions, as well as the lengthy traditional optimization iteration that is difficult to meet the real-time response requirements.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The distribution network fault control method based on multi-source data fusion includes the following steps: Using electrical admittance topology as the physical layer and a generalized potential manifold calculated based on operating pressure indicators and nodal marginal electricity prices as the logic layer, a physical-logic dual-layer coupled network is constructed. Cross-layer bidirectional interaction is performed based on this dual-layer coupled network, specifically including: mapping the operating pressure indicators to the physical layer matrix corresponding to the physical layer; fusing the physical indicators with the logic layer gradient features through graph algebra multiplication to drive the logic layer to evolve its state; in the logic layer after state evolution, solving for the primary control command to eliminate fault exceedances under the constraints of the physical equipment capacity feasible region, and optimizing the generation of secondary allocation vectors along the logic layer gradient feature direction; performing orthogonal projection operations on the secondary allocation vectors onto the null space of the fault section power flow sensitivity matrix, and superimposing them with the primary control command to generate the final control command, which is then sent to the physical layer.

[0007] The technical effects and advantages of this invention based on multi-source data fusion for coordinated control of distribution network faults are as follows: The physical-logic dual-layer coupled network constructed in this invention achieves a structured fusion of physical layer electrical admittance topological objective constraints and logic layer generalized potential manifold multi-objective optimization within the same algebraic feature space through a cross-layer bidirectional interaction mechanism of uplink state penetration and downlink orthogonal projection. It provides a rigid interception mechanism independent of the fully converged state of the optimization through the null space orthogonal projection operation, filtering out dangerous power components that may cause static or thermal instability exceedances before the final control command is issued. This reduces the hard dependence of the control system on the global fully converged state of the multi-agent game in the logic layer. It also utilizes the physical properties of spatial orthogonal complement in mathematical... The physical characteristics of the secondary dispatching feature are reduced, and the traditional sequential external data communication iteration is replaced by the bottom-level cross-layer graph algebraic analytical mapping. This not only alleviates the physical safety decoupling problem caused by the optimization manifold being detached from the physical base and avoids the risk of safety exceeding the limit under the condition of incomplete convergence, but also alleviates the control lag and model mismatch bottleneck caused by the traditional sequential data interaction. Under the premise of ensuring the convergence of primary fault elimination and meeting the real-time response and control requirements of the distribution network, the dynamic adjustment margin of the entire network resources is effectively released, and the physical decoupling, superposition and safe execution of multi-dimensional control commands are realized. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the distribution network control method based on multi-source data fusion.

[0009] Figure 2 This is a schematic diagram of the topological construction of a generalized potential manifold.

[0010] Figure 3 A heatmap showing the convergence characteristics of the physical-logic two-layer coupling. Detailed Implementation

[0011] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0012] Example 1 Please see Figure 1 To address the problems in existing technologies regarding distribution network fault control scenarios, such as the decoupling of multi-stakeholder economic game and network physical security, the difficulty in coordinating heterogeneous multi-source data, the high risk of steady-state safety exceeding limits due to non-convergence conditions caused by the dependence of control commands on fully converged optimization, and the significant increase in state dimensionality due to multi-stakeholder competition making it difficult to meet real-time response requirements, this invention presents a distribution network fault control method based on multi-source data fusion, comprising the following steps: S1 uses electrical admittance topology as the physical layer and generalized potential manifold calculated based on operating pressure index and nodal marginal electricity price as the logical layer to construct a physical-logical dual-layer coupled network. S2, based on the dual-layer coupled network, performs cross-layer bidirectional interaction, specifically including: The operational pressure indicators are mapped to the physical layer matrix corresponding to the physical layer, and then fused with the gradient features of the logic layer through graph algebra multiplication to drive the state evolution of the logic layer. In the logic layer after state evolution, the primary control command to eliminate fault overruns is solved under the constraint of the feasible region of physical device capacity, and the secondary allocation vector is generated by optimization along the gradient feature direction of the logic layer. The secondary allocation vector is orthogonally projected onto the null space of the power flow sensitivity matrix at the fault section, and then superimposed on the primary control command to generate the final control command, which is then sent to the physical layer.

[0013] This invention proposes a fault control method for distribution networks based on multi-source data fusion. First, a two-layer coupled network is constructed using electrical admittance topology as the physical layer and a generalized potential manifold calculated based on operating pressure indicators and nodal marginal electricity prices as the logic layer, establishing a manifold metric benchmark for isomorphic heterogeneous data. Second, the operating pressure indicators are mapped to the corresponding physical layer matrix, and fused with the logic layer gradient features through graph algebra multiplication to drive the logic layer state evolution. The dynamic impedance boundary of the physical network is directly embedded into the optimization space, achieving adaptive limit avoidance of the evolution trajectory. Subsequently, under the constraint of the feasible region of physical equipment capacity, the primary control command for eliminating fault limits is separated, and a secondary allocation vector is generated by optimization along the direction of the logic layer gradient features. Finally, the secondary allocation vector is orthogonally projected onto the null space of the power flow sensitivity matrix of the fault section, and superimposed with the primary control command to generate the final control command. This invention eliminates the cross-coupling of secondary dispatching characteristics to the power flow at the fault section through null space orthogonal constraints, achieving rigorous decoupling and safe execution of multi-dimensional control commands while ensuring convergence of primary fault elimination. The specific implementation is as follows: S1 uses electrical admittance topology as the physical layer and generalized potential manifold calculated based on operating pressure index and nodal marginal electricity price as the logical layer to construct a physical-logical two-layer coupled network.

[0014] To address the problem that existing Cyber-Physical Systems (CPS) modeling in distribution networks only focuses on superficial interactions such as communication terminals and data transmission, leading to a substantial decoupling of multi-agent optimization and network physical security in mathematical structure (i.e., "cross-domain decoupling"), and consequently causing a fragmentation of the system's state space, this invention constructs a physical-logical bidirectional coupled network. The physical layer (lower layer), representing objective electrical connectivity, and the logical layer (upper layer), representing the evolution of multi-agent decision-making, together constitute a two-layer coupled network. This network architecture skips the communication transmission step, establishing the topological carrier for optimization in an abstract mathematical space, providing a prerequisite operating environment for subsequently eliminating the "computation-operation separation" and realizing the physical feasibility of control commands. The specific steps for constructing the physical-logical bidirectional coupled network are as follows: S11, based on real-time measurement data, extracts line impedance to generate electrical admittance topology, which serves as the physical layer to carry the underlying power flow hard constraints; The dispatch master station acquires node measurement data and the actual opening and closing status of each branch switch in real time through the underlying feeder automation terminal (FTU / TTU) (defining switch status parameters). Based on this, the master station extracts the pre-stored line impedance parameters (branch resistance and reactance). Based on the conventional node admittance calculation method, the system dynamically generates an electrical admittance matrix in memory that represents the hard constraints of the current physical power flow of the distribution network, and establishes it as the physical layer matrix.

[0015] Specifically, the electrical admittance matrix can be expressed as equation (1): (1) In equation (1), for 3D electrical admittance matrix ( (Total number of nodes) Real-time switching status parameters between nodes (closed is...) Disconnected ), For nodes Earth-to-ground admittance For the Kronecker function (when hour ;when hour ); , These are the pre-stored line resistance parameters between nodes. , These are the line reactance parameters pre-stored between nodes.

[0016] It is important to note that the core purpose of generating the electrical admittance matrix in this embodiment is to reuse its algebraic properties as a constraint for the network structure in subsequent cross-space mapping. The real distribution network structure inherently gives the electrical admittance matrix a high degree of sparsity (i.e., the mutual admittance of nodes without direct electrical connections is strictly zero). At the physical layer, the electrical admittance matrix not only defines the objective path of current flow but also, through the rigid distribution of zero and non-zero elements, provides a topological filtering base for subsequent cross-layer graph algebraic fusion, effectively shielding the logical optimization evolution of disconnected nodes in the mathematical space and enabling precise penetration of the physical network impedance characteristics into the logical optimization space.

[0017] S12, acquire electrical operating parameters and quantify them into operating pressure indicators (physical safety risks). Traditional power distribution networks often employ a traditional external trial-and-error iterative model of "first seeking economic optimization, then physical verification." This model not only consumes enormous computing power, but also, when high-proportion distributed power sources (such as photovoltaic / wind power) experience frequent output fluctuations or topology reconfigurations, it is highly susceptible to iterative oscillations or even divergent collapse due to rank reduction or condition number deterioration of the Jacobian matrix in power flow calculations. This makes it impossible to meet the real-time scheduling requirements at the second / millisecond level. This step integrates the underlying multi-dimensional physical security risks (such as voltage deviation, harmonic distortion, etc.) into an algebraic-dimensional "operational pressure index," and equates it to "safety resistance," directly injecting it into the upper logical layer space. This constructs a pathway for physical security risks to penetrate into the logical layer (logic optimization space). Specifically: S121, Obtain electrical operating parameters; S1211, the control unit calculates the electrical operating parameters of each node in real time through automated terminals (such as phasor measurement units, PMUs, or smart meters) deployed at the physical layer nodes. These electrical operating parameters include: voltage deviation (characterizing voltage stability), harmonic distortion rate (characterizing power quality), and resource regulation margin (characterizing the node's regulation potential). Specifically, the formula for calculating the harmonic distortion rate is shown in equation (2): (2) In equation (2), Harmonic distortion rate, The amplitude of the fundamental component. For the first amplitude of second harmonic components To control the number of cycles.

[0018] The formula for the voltage deviation is shown in equation (3): (3) In equation (3), For voltage deviation, For the first Each node in the control cycle The measured voltage, This refers to the rated voltage or reference voltage corresponding to the node.

[0019] The formula for the resource adjustment margin is shown in equation (4): (4) In equation (4), To adjust resource margins, This indicates the available adjustable capacity of the node in its current operating state (the distance between the current operating point and the feasible region of the physical device capacity). This indicates the maximum adjustment capacity corresponding to the node.

[0020] S1212, after dimensionless normalization of the electrical operating parameters (voltage deviation, harmonic distortion rate, and resource adjustment margin), an operating pressure index is obtained through nonlinear fusion to characterize the tension between the current state of the node and the physical safety risk boundary. The calculation formula for the operating pressure index is shown in equation (5): (5) In equation (5), For nodes The operational pressure index characterizes the "repulsive force" of the node's physical risk on the logical manifold; , Representing voltage deviation and node respectively Voltage deviation limits (set according to national standard GB / T12325-2008 "Power Quality - Supply Voltage Deviation"); To represent nodes respectively The real-time value of total harmonic distortion and the standard limit of harmonics (defined according to the national standard GB / T14549-1993 "Power Quality and Harmonics in Public Power Grids", which defines the hard constraint line of power quality) are used to reflect the negative impact of power quality on the stability of the physical layer. To adjust the margin of resources; The preset weighting coefficient (in this embodiment, the preset value is: (used to balance the weights of voltage safety, power quality, and capacity margin in pressure synthesis). The preset minimum positive real number (taken in this embodiment) This is used to prevent numerical overflow caused by a denominator of zero, and to ensure the computational stability of engineering applications.

[0021] S13: Obtain the marginal electricity price of the node and calculate the potential gradient by vector synthesis with the partial derivative of the operating pressure index; and dynamically construct the logical layer manifold accordingly. This step uses algebraic composition to isomorphize complex physical constraints and decision objectives within the same operator space. Specifically: S131, the dispatching master station synchronously obtains the real-time node marginal electricity price and algebraically couples it with the operation pressure index to construct a comprehensive objective function. The calculation formula of the comprehensive objective function is as follows (6), and the economic cost function is as follows (7): (6) (7) In equation (6), For the comprehensive objective function, The economic cost function is represented by a quadratic cost model. Inject power decision variables (the decision variables to be optimized in the logic layer, hereinafter referred to as decision variables) into the node. The penalty weighting coefficient is used to convert the dimensionless operational pressure index into an economic cost dimension, thereby controlling the weight of physical security intervention in economic decision-making. In equation (7), For the marginal electricity price at the node, The cost coefficient operator (a preset quadratic cost term coefficient used to characterize the nonlinear marginal cost of resource allocation (such as equipment aging depreciation), the value of which is related to the system equipment capacity and operating condition level).

[0022] S132, Calculate the Euclidean gradient of the integrated objective function with respect to the decision variables, and reconstruct the Riemann metric tensor based on the Hessian matrix approximation of the running pressure index to generate the Riemann potential gradient for guiding the optimization: First, calculate the Euclidean gradient of the integrated objective function with respect to the node injection power decision variables: (8) In the formula, For nodes Euclidean gradient components; The partial derivative of economic cost; The cross derivative term represents the node. The power regulation action has a cross-derivative effect on the operating pressure indicators of other related nodes in the entire network.

[0023] Secondly, the second-order partial derivatives of the operating pressure index with respect to the decision variables are extracted to construct the metric reconstruction matrix of the Riemannian manifold. To eliminate the enormous computational overhead of inverting high-dimensional non-convex matrices and meet the second-level real-time control requirements of the power grid, this embodiment employs a local diagonalization truncation method to extract the second-order curvature of the main diagonal of the operating pressure index at this node. This curvature is then converted to positive definiteness by absolute value spectrum shifting and used as a local positive definite approximation of the Hessian matrix. Based on this, a Riemannian metric tensor is generated. The formula for calculating the Riemannian metric tensor is as follows: (9) In the formula, It is the identity matrix. The preset curvature modulation coefficient is used to calibrate the degree of curvature in the measurement space; This is an approximation of the Hessian matrix for the operating pressure index.

[0024] Finally, the Riemann potential gradient is generated by mapping the Euclidean gradient to the manifold tangent space through the inverse matrix of the Riemann metric tensor: (10) In the formula, Let be the Riemann potential gradient. Let be the inverse matrix of the Riemannian metric tensor.

[0025] It is particularly important to note that the technical essence of Riemann metric reconstruction in this embodiment lies in reshaping the physical perception of "distance" in the state space by dynamically changing the manifold tensor. When the distribution network parameters are in a stable range with sufficient physical safety margin, the operating pressure index is extremely small and the curvature is gentle. The approximate value of the Hessian matrix approaches the zero matrix, the Riemann metric tensor degenerates into an identity matrix, and the Riemann potential gradient is losslessly equivalent to the Euclidean gradient. At this time, the economic optimization of the system is not interfered with. However, when the operating parameters of the underlying physical equipment approach the risk boundary of exceeding the limit, the characteristics of the penalty function cause the diagonal elements of the Hessian matrix in the corresponding dimension to increase nonlinearly and remain positive definite. This causes the mapping coefficient of the inverse matrix of the metric tensor to decay sharply and approach zero in this dimension of the risk boundary, thereby exerting a very strong physical "soft barrier" effect on the evolution step size of this dimension. This nonlinear metric distortion mechanism enables the system's state evolution trajectory to automatically deflect along the manifold surface and avoid physical limits without external hard logic intervention, thus eliminating the generation of dangerous control commands from the bottom layer of the mathematical mechanism.

[0026] S133, the system uses the potential gradient of all network nodes as the spatial basis to dynamically construct the generalized potential manifold of the logic layer; This step involves up-mapping the local physical operational pressures of each node (such as voltage deviation, harmonic distortion rate, and resource regulation margin) to construct a generalized potential manifold with highly nonlinear geometric characteristics at the logic layer. This manifold manifests as a "high potential energy extremum field" (i.e., geometric bulges or potential barriers) in regions where physical parameters approach the limit, thus visualizing the safety state of the underlying power grid as a high-dimensional topological structure. The specific implementation steps are as follows: In one specific embodiment, a linear scalarization method is used to dynamically construct the generalized potential manifold of the logic layer. Specifically, based on the injection power decision variables of each node in the entire network, a joint policy space is constructed. The comprehensive objective functions of each node are linearly superimposed using scalars. The resulting value of the generalized potential of the entire network is used as the height potential dimension (i.e., the local geometric height of the manifold), and it is embedded into the network. The Euclidean space is used to construct the generalized potential manifold of the logic layer, thereby accurately mapping the economic optimization process of multiple agents across the network and the game process of the underlying physical boundary into a geometric surface in the metric space. The generalized potential manifold of the logic layer is shown in Equation (11): (11) (12) (13) In equation (11), For the generalized potential manifold of the logic layer, The generalized potential of the entire network is obtained by scalar summation of the comprehensive objective functions of all participating nodes. Its magnitude determines the generalized potential manifold at the logic layer. The calculation formula for the "local geometric height of the manifold" in the joint strategy space is given in Equation (12); The Euclidean space dimension constructed for the manifold, where Let be the joint policy space composed of the power injection decision variables of each node. The dimension of height potential energy is composed of the generalized potential of the entire network (derived from the local geometric height of the manifold). composition); The generalized potential value / coordinates of the entire network; A power decision variable vector is injected into the entire network to represent the global logical state of the system. Its calculation formula is shown in equation (13). The total number of nodes is kept consistent with the dimension of the physical layer electrical admittance matrix (and therefore the naming is consistent) to ensure point-to-point mapping between physical topology and logical variables.

[0027] like Figure 2 As shown, the generalized potential manifold of the entire network is mapped from the joint policy space consisting of the power injection decision variables of each node. The vertical axis represents the generalized potential value of the entire network calculated by the linear scalarization method. The surface in the figure is located at... The sharp protrusion at the coordinates originates from the gradient extrema generated by the nonlinear penalty term in the operational pressure mapping mechanism. Mathematically, this region is algebraically represented by the global Riemann gradient field. It represents the high-potential gradient field generated by Riemann metric reconstruction in the optimization space when the system approaches the physical equipment capacity safety limit (the local gradient field after Riemann metric reconstruction and local curvature modulation, i.e., by dynamically adjusting the algebraic parameters of the optimization space, changing the geometric direction of the optimization trajectory, and forcing the optimization direction to deviate from the high-risk physical region). This high-potential gradient field forms a strong nonlinear guiding constraint on the iterative trajectory of the optimization operator during the optimization process, driving the optimization operator to produce a significant geometric trajectory deflection, ensuring that the final output secondary allocation vector achieves economic optimality under linear scalarization objectives while satisfying the physical safety operation boundary. The smooth concave region of the manifold surface represents the global optimal solution space for economic cost after multi-objective optimization.

[0028] S2, based on the dual-layer coupled network, performs cross-layer bidirectional interaction, specifically including: To overcome the technical shortcomings of existing distribution network cyber-physical systems (CPS) that rely solely on communication terminals for surface data stacking, resulting in a lack of hard physical network constraints for multi-objective optimization, this embodiment executes a cross-layer bidirectional interactive closed loop based on the two-layer coupled network constructed in step S1 (step S2). The specific implementation steps are as follows: S21, map the operational stress indicators to the corresponding physical layer matrix, and fuse them with the gradient features of the logic layer through graph algebra multiplication to drive the state evolution of the logic layer; specifically including the following sub-steps: S211, Extract the physical layer matrix after mapping the operating pressure index; To address the issue that sequential collaborative control cannot obtain real-time physical topology changes and spatial distribution of operating pressure states, the system constructs an operating pressure mapping operator. Specifically, the operating pressure index is constructed as a weighted matrix element of the operating pressure mapping operator and applied to the physical layer matrix corresponding to the physical layer (in this embodiment, an electrical admittance matrix is ​​used) to generate an electrical admittance matrix after mapping the operating pressure index. The electrical admittance matrix after mapping the operating pressure index is shown in Equation (14) below: (14) (15) In equation (14), The pressure mapping operator is represented as a diagonal matrix with the operating pressure indicators of all nodes in the network as diagonal elements. In equation (15), This is the electrical admittance matrix mapped from the operating pressure index; This step uses standard matrix multiplication to transform the operational risk of each node into a dynamic weighting coefficient of the admittance of the connected electrical branches. While maintaining the original physical connectivity (matrix sparsity) of the distribution network, it effectively amplifies the "topological impedance" of high-risk nodes.

[0029] S212 uses graph algebra multiplication to fuse it with the gradient features of the logic layer across layers, driving the logic layer to evolve its state. This step introduces graph algebra multiplication to perform cross-layer spatial fusion of the physical layer electrical admittance matrix (after mapping the operating pressure) and the logic layer gradient features (Riemann potential gradient). In the underlying linear algebra space, this operation forcibly injects the sparse features of the network structure representing physical topological connectivity (the electrical admittance matrix after mapping the operating pressure index) into the evolution of the logic manifold (generalized potential manifold). This maps the real-time operating pressure index of each node into a high-potential gradient field in the multidimensional manifold space of the logic layer, ensuring that the spatial topological analytical state of the generalized potential manifold (specifically manifested as Riemannian metric reconstruction and local curvature modulation) is controlled by the real-time physical network state. Specifically: S2121, the system extracts the current logical layer potential gradient features of each node in the entire network (in this embodiment, the Riemann potential gradient is used) to form a potential gradient vector; the potential gradient vector represents the generalized potential of the entire network in... The evolution trend along each axis of the decision space. As shown in Equation (16): (16) In equation (16), This is the potential gradient vector.

[0030] In order to inject logical gradients The system employs a three-dimensional physical topology space, allowing each node in the logic layer to access the electrical connectivity and topological constraints of the entire network when injecting power decision variables. Graph algebra multiplication is used to perform matrix-vector multiplication on a matrix with physical topology properties and a gradient vector with economic-security game-theoretic properties, generating a global Riemann gradient field. This field drives the evolution and optimization of states on the generalized potential manifold towards lower potential regions. The expression for the global Riemann gradient field is as follows: (17) In equation (17), This represents the global Riemann gradient field.

[0031] This step achieves topological diffusion of the gradient through matrix multiplication, reconstructing the analytic structure of the manifold feature space (Riemannian metric reconstruction and local curvature modulation), specifically the nodes corresponding to the surge in operational pressure indicators and their associated topology. This generates real-time local numerical deflections and bulges in the global Riemann gradient field, thereby altering the gradient magnitude and spatial orientation of the global Riemann gradient field along the axis of the injected power decision variables at each node. This achieves precise penetration of physical network constraints into the logical optimization space (the evolution and optimization of states on the generalized potential manifold towards lower potential regions), providing a gradient-guided basis with hard physical boundary constraints for subsequent optimization operators. Figure 3 As shown in the figure, the convergence characteristics of the physical-logic two-layer coupled network under different operating pressure indicators are represented. As the operating pressure indicator (physical safety risk) increases, the generalized potential manifold of the logic layer is reconstructed through Riemannian metric, which adaptively guides the optimization trajectory to evolve towards a region with sufficient margin of physical safety risk boundary, effectively avoiding iterative oscillations in traditional optimization under complex operating conditions.

[0032] S22, in the logic layer after state evolution, solve the primary control command to eliminate fault over-limit under the constraint of physical device capacity feasible region, and optimize and generate secondary allocation vector along the gradient feature direction of logic layer. After the logical layer manifold completes the fusion and evolution of physical characteristics and topology driven by graph algebra multiplication, the optimal control solution needs to be found in the manifold space. In this embodiment, the primary control command and secondary allocation vector are solved in parallel or sequentially on the logical layer manifold after state evolution, overcoming the network-wide power flow oscillation and cascading limit exceedance problems caused by the short-sighted optimization of distributed nodes.

[0033] To ensure the absolute timeliness and safety of fault elimination, the system does not rely on the cross-sectional power flow sensitivity matrix described later. Instead, it directly uses the physical layer electrical admittance matrix defined in step S11 and the physical equipment capacity feasible region (active power injection upper and lower limit box constraints) of the current step to perform rigid inverse solution of the physical boundary.

[0034] Specifically, the system first identifies nodes whose values ​​exceed a preset safety threshold based on the real-time operating pressure indicators of each node calculated in step S12. These nodes are then identified as physical over-limit fault source nodes, and their node numbers are defined as follows. Subsequently, the fault source nodes in the physical layer electrical admittance matrix are reused. Based on the distribution of non-zero elements in the corresponding row, extract the set of one-step adjacent nodes (including the fault source node itself) that are directly electrically connected to the fault source node. The system directly uses the modulus of each off-diagonal mutual admittance element in the row to quantitatively characterize the local voltage-power topology response intensity of each adjacent node to the state changes of the fault source node.

[0035] Based on this, a two-dimensional sparse topology selection and control weight operator matrix is ​​constructed by extracting the sparse distribution structure corresponding to the fault source nodes in the electrical admittance matrix. The operator matrix any matrix element (line subscript is) , column subscript The assignment logic for ) is as follows: if and only if the column index Strictly equal to the fault source node number , and subscript When it belongs to the set of local adjacent nodes, the element The value is equal to the node With the fault source node The electrical admittance modulus between the nodes is divided by the sum of the electrical admittance moduli of all nodes in the local neighboring node set and the fault source node. The algebraic summation of the electrical admittance moduli between them; in all other cases, the elements The value is always cleared to zero. This operator matrix, through column locking and row amortization properties, rigidly confines the safety control domain within the fault local network.

[0036] Finally, under the hard boundary constraints of the feasible region of physical equipment capacity, the reference power adjustment vector required to forcibly pull the fault source node and its directly electrically connected adjacent nodes back to the safety margin range is calculated through direct algebraic truncation. This vector serves as the primary control command, and the calculation formula for the primary control command is as follows: (18) In the formula, The primary control command vector obtained by solving; It is a sparse topology selection and control weight operator matrix composed of the dynamically weighted electrical admittance moduli of each fault source node and its adjacent nodes; It is a physical boundary strategy vector composed of the pre-stored upper or lower limit of active power injection; This is to regulate the node power state vector at the initial moment. Solving this primary instruction involves a rigid algebraic mapping and does not participate in subsequent game iterations, thus locking in the bottom-line control quantity to ensure the safety of the grid.

[0037] Among them, the feasible range of physical device capacity This refers to a closed convex set constraint space jointly determined by the hardware physical limits of the distributed resources of each node in the distribution network, algebraically represented as a set of Cartesian products of the node power injection upper and lower bound parameters; in this embodiment, the node power injection upper and lower bound parameters (i.e., the aforementioned...) and The method involves pre-calculating and storing the inverter PQ power circle model in the traditional analytical method in conjunction with the distribution network operation safety domain (OSR) boundary theory in the memory of the dispatch master station; or, in another preferred embodiment, the data-driven convex hull algorithm can be used to extract the spatial boundary of the historical safe active power measurement data of each node.

[0038] Above the safety baseline determined by the primary control command, the system further invokes a distributed optimization operator to iteratively update the generalized potential of the entire manifold network along the descent direction (i.e., driving the policy variables of all nodes in the network to synchronously move in the opposite direction of the global Riemann gradient field) to generate a secondary allocation vector for optimizing the network's economic and operational pressure. The system employs the Projected Gradient Descent (PGD) algorithm as the optimization operator to perform iterative solutions. The next state update equation is represented by equation (19): (19) In equation (19), The aforementioned power decision variables are injected into the entire network in the first... The state vector at the next iteration; The aforementioned power decision variables are injected into the entire network in the first... The state vector at the next iteration; The preset iteration step size (in this embodiment, it is taken as...) ); The local gradient descent direction (iteration direction) is directly taken as the opposite direction of the global Riemann gradient field generated in the previous steps, thereby avoiding the distortion of the secondary mapping of feature information. To the physical device capacity feasible domain Projection operator for the feasible region of projection operations; To ensure the feasibility of the algorithm, this embodiment specifically defines the projection operation as performing a projection operation onto the feasible region of the physical device capacity under the box constraint formed by the pre-stored upper and lower limits of active power injection, and performing iterative updates. The feasible region projection operator expression corresponding to each node is as follows: (20) (twenty one) In equation (20), For feasible region projection operators, it represents an algebraic operation that forces out-of-bounds variables back into the physical safety boundary; Before performing the projection operation The intermediate power injection variables calculated at each node satisfy the algebraic relationship shown in equation (21); The local gradient descent direction vector In the The scalar component values ​​along the decision axis of each node; and They are the pre-stored numbers. The lower and upper limits of active power injection at each node.

[0039] It is worth noting that by directly introducing the global Riemann gradient field as the iteration direction into the state update equation, the optimization operator can actively exit the high-risk state region (i.e., high potential gradient field) triggered by the operating pressure index. That is, when the local physical network is overloaded by power flow, the upward mapping of the operating pressure will cause the corresponding component of the global Riemann gradient field at the overloaded node to produce a nonlinear leap in generalized potential (i.e., safety resistance surpasses positive driving force, i.e., the penalty effect derived from the operating pressure index fully dominates the potential gradient evolution based on the economic cost function). This Riemann metric reconstruction effect forms an endogenous Riemann gradient repulsion field in the manifold tangent space, forcibly guiding the iteration trajectory of the optimization operator to converge to the physical feasible region with a high safety margin, thereby quickly locking in the local minimum point that takes into account both the marginal electricity price of the node and the physical safety margin of the system. When the optimization operator state satisfies the convergence condition: the state satisfies the convergence threshold for two consecutive iterations (i.e.) , The preset convergence threshold is used. Set as When the per-unit value (pu) is reached, it is determined that the Nash equilibrium point has been reached, and the convergence state difference is output to generate the secondary allocation vector, the implementation formula of which is as follows (22): (twenty two) In equation (22), This is the secondary allocation vector, which is the difference between the final state after convergence and the initial state. This refers to the node power state at the initial moment when a transient fault occurs in the system or a control command is received.

[0040] It should be noted that when multiple nodes compete for the limited transmission capacity of the same local power grid, a heavy power flow occurs along with the local grid structure, causing local nodes to approach the line thermal stability limit or voltage safety boundary, triggering a surge in operational pressure indicators. The upward mapping of this physical state causes the safety resistance component (i.e., the penalty effect derived from the operational pressure indicators) in the global Riemann gradient field to be amplified and surpass the economic driving component (potential gradient evolution based on the economic cost function), resulting in an algebraic deflection of the gradient vector originally pointing to this heavy-load region in the feature space. This deflection mechanism forces the optimization trajectory of each node to actively exit the safety constraint critical region and transfer along the adjusted gradient direction to nodes with sufficient physical safety margins across the entire network, until the system converges in the solution space to the multi-objective optimization steady-state point of distributed resources where no node can unilaterally reduce the convergence of its individual objective function by changing the injected power.

[0041] Traditional local gradient descent only reflects the greedy optimization of isolated nodes in Euclidean space, lacking awareness of the coupling characteristics of the physical network structure. In contrast, the generalized potential descent direction extracted in this invention is essentially the natural gradient of the Riemannian manifold, taking into account electrical distance and physical boundaries. This invention uses this global manifold gradient as the iterative direction for local control at each node, overcoming the network-wide power flow oscillations and cascading limit exceedances caused by the short-sighted optimization of distributed nodes.

[0042] S23, perform orthogonal projection operation on the secondary allocation vector to the null space of the power flow sensitivity matrix of the fault section, and superimpose it with the primary control command to generate the final control command and send it to the physical layer; In practical implementation, in order to prevent the system from generating reverse power flow interference on the triggered emergency fault sections during secondary optimization (such as economic allocation or reducing operating pressure), the system executes a zero-space physical anti-limit decoupling mechanism before issuing instructions. The specific implementation steps are as follows: The current Jacobian matrix of the physical layer power grid is obtained in real time. Using the standard power flow sensitivity analysis method, a fault section power flow sensitivity matrix is ​​constructed to characterize the degree of influence of changes in injected power at all nodes on the linearity of power flow at the fault section. To filter out power components that may cause the static stability limit or thermal stability limit to be exceeded, the system performs an algebraic decomposition operation (singular value decomposition (SVD) in this embodiment) on the power flow sensitivity matrix of the faulty section. Its standard algebraic expression is: (twenty three) In equation (23), This represents the current fault section power flow sensitivity matrix at the physical layer. and These are the left orthogonal matrix and the right orthogonal matrix corresponding to the singular value decomposition, respectively. It is a diagonal matrix composed of singular values.

[0043] However, directly solving the full singular value decomposition of the high-dimensional fault section power flow sensitivity matrix would incur a huge overhead on the main station's computing power. Therefore, in actual execution, this embodiment limits the use of the analytical reconstruction algorithm of autocorrelation matrix eigenvalue decomposition (EVD) to extract the aforementioned right orthogonal matrix. The implementation steps are as follows: As shown in equation (24), construct the real symmetric autocorrelation matrix corresponding to the power flow sensitivity matrix of the fault section; As shown in equation (25), eigenvalue decomposition is performed on the real symmetric autocorrelation matrix; According to the matrix algebra isomorphism mapping mechanism, the orthogonal eigenvector matrix is ​​directly equivalent to the right orthogonal matrix in singular value decomposition, and the eigenvalues ​​and singular values ​​satisfy the analytic relationship shown in equation (26). (twenty four) (25) (26) In equation (25), For the eigenvalues The diagonal matrix formed The corresponding orthogonal eigenvector matrix; In equation (26), The aforementioned diagonal matrix composed of singular values The first in A singular value diagonal element.

[0044] Extract the singular values ​​corresponding to zero in the right orthogonal matrix (i.e., the corresponding eigenvalues ​​satisfy the following conditions). , The preset zero-space tolerance threshold is set to [value]. In this embodiment, the preset zero-space tolerance threshold is set to [value]. The family of column vectors of ) forms the null space orthogonal mapping basis. The space formed by this matrix is ​​the "safe and feasible direction space" of the system.

[0045] The generated secondary allocation vector is orthogonally projected onto the null space and superimposed with the primary control command to generate the final control command, the calculation formula of which is shown in equation (27): (27) In equation (27), For the final control commands sent to the physical layer, Let be the transpose of the null space orthogonal mapping basis matrix; This constitutes an orthogonal projection matrix operator that projects onto the null space of the sensitivity matrix.

[0046] During closed-loop execution at the physical layer, the control commands possess inherent state-variable immunity. That is, without worsening the node voltage deviation and critical section steady-state constraints at the current operating point of the distribution network, intrinsically safe, disturbance-free offsetting and substitution allocation of active power injection decision variables for all nodes in the network is achieved. This completes the cross-layer, bidirectional interactive closed loop of the physical-logical dual-layer coupled network, ensuring that the issued control commands do not compromise the intrinsic safety baseline of the physical layer.

[0047] In summary, the physical-logical dual-layer coupled network constructed in this invention achieves system control through the following cross-layer bidirectional interaction mechanism: Upward state penetration mechanism: Step S1212 quantifies the underlying dynamic electrical operating parameters into a unified operating pressure index, and explicitly embeds it as a prerequisite state parameter into the potential gradient equation of the upper logic space. Before the system approaches the hard physical red line, the Riemann high potential energy gradient field generated by reconstruction on the generalized potential manifold enables the spontaneous geometric boundary retreat of the optimization trajectory, avoiding the occurrence of limit-breaking oscillations in traditional control schemes; Downlink orthogonal projection mechanism: Step S23 performs orthogonal projection operation on the secondary allocation vector generated by game optimization to the null space of the power flow sensitivity matrix of the fault section. By utilizing the rigid physical properties of spatial orthogonal complement, all possible back crosstalk fault sections, dangerous control components that disrupt power flow balance and physical safety red line are forcibly filtered out.

[0048] The aforementioned bidirectional interaction mechanism of upward penetration of operational pressure and downward orthogonal projection of resource allocation parameters enables a rigorous and structured integration of the objective constraints of the physical layer electrical admittance topology and the economic game of the marginal electricity price and generalized potential manifold of the logic layer within the same algebraic feature space. This mechanism replaces the traditional sequential external data communication iteration with algebraic analytical mapping of the underlying cross-layer graph, thereby eliminating the inherent defect of the mathematical decoupling between the multi-agent economic game and the physical security of the distribution network in distribution network control. It breaks through the control lag and model mismatch bottlenecks caused by traditional sequential data interaction, significantly improving the overall stability and intrinsic safety fault resistance of the distribution network under complex and highly dynamic operating conditions.

[0049] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0050] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0051] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0052] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0053] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0054] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for distribution network fault control based on multi-source data fusion, characterized in that, include: A physical-logical dual-layer coupled network is constructed, with the electrical admittance topology as the physical layer and the generalized potential manifold calculated based on operating pressure indicators and nodal marginal electricity prices as the logical layer. Performing cross-layer bidirectional interaction based on the aforementioned two-layer coupled network specifically includes: The operational pressure indicators are mapped to the physical layer matrix corresponding to the physical layer, and then fused with the gradient features of the logic layer through graph algebra multiplication to drive the state evolution of the logic layer. In the logic layer after state evolution, the primary control command to eliminate fault overruns is solved under the constraints of the feasible region of physical device capacity, and the secondary allocation vector is generated by optimization along the gradient feature direction of the logic layer. The secondary allocation vector is orthogonally projected onto the null space of the power flow sensitivity matrix at the fault section, and then superimposed on the primary control command to generate the final control command, which is then sent to the physical layer.

2. The method according to claim 1, characterized in that, The calculation of the gradient features of the logical layer includes: Obtain the marginal electricity price at each node, algebraically couple it with operational pressure indicators, and construct a comprehensive objective function; Calculate the Euclidean gradient of the integrated objective function with respect to the power injection decision variables at each node; The Hessian matrix is ​​calculated based on the comprehensive objective function, and after positive definite processing, it is reconstructed into the Riemannian metric tensor of the generalized potential manifold. The Riemann potential gradient is generated by spatially mapping the Euclidean gradient to the inverse matrix of the Riemann metric tensor.

3. The method according to claim 2, characterized in that, The calculation of the gradient features of the logical layer also includes: Using the node injection power decision variable as an independent variable, the comprehensive objective function is linearly superimposed to calculate and generate the generalized potential of the entire network; A joint policy space is constructed based on the injection power decision variables of all nodes in the network. The scalar value of the generalized potential of the entire network is used as the local geometric height of the manifold to construct the generalized potential manifold.

4. The method according to claim 2, characterized in that, The formula for the comprehensive objective function is as follows: ， ， In the formula, For the comprehensive objective function, Let be the economic cost function. For nodes Operating pressure indicators Inject power decision variables into nodes. This is the penalty weighting coefficient; For the marginal electricity price at the node, This is the cost coefficient operator.

5. The method according to claim 1, characterized in that, The process of mapping operational pressure indicators to the physical layer matrix corresponding to the physical layer includes: Construct an operating pressure mapping operator with the operating pressure index as the diagonal element, and perform a standard matrix multiplication operation on it with the electrical admittance matrix to generate the electrical admittance matrix after the operating pressure index is mapped.

6. The method according to claim 5, characterized in that, The driving logic layer performs state evolution, including: The potential gradient vector is formed by using the Riemann potential gradient, and graph algebra multiplication is performed on the electrical admittance matrix mapped to the operating pressure index to generate a global Riemann gradient field, which drives the state on the generalized potential manifold to evolve and optimize towards the lower potential region.

7. The method according to claim 6, characterized in that, The process of optimizing and generating secondary allocation vectors along the gradient feature direction of the logical layer includes: Based on the preset iteration step size and the global Riemann gradient field, calculate the intermediate power injection variables at the nodes; Using the box constraints formed by the pre-stored upper and lower limits of active power, the intermediate power injection variable is mapped to the feasible region of physical device capacity through the feasible region projection operator, and iteratively updated. When the deviation between two consecutive iterations meets the convergence threshold, the difference between the convergence state and the initial state is calculated, and the secondary allocation vector is generated.

8. The method according to claim 1, characterized in that, The operating pressure index is obtained by nonlinear weighted mapping of voltage deviation, harmonic distortion rate and resource adjustment margin, wherein the resource adjustment margin is negatively correlated with the operating pressure index.

9. The method according to claim 8, characterized in that, The formula for calculating the operating pressure index is as follows: ， In the formula, For nodes Operating pressure indicators , Representing nodes respectively Voltage deviation, voltage deviation limit, To represent nodes respectively Real-time total harmonic distortion rate and harmonic standard limits. To adjust the margin of resources; The preset weighting coefficients, It is a pre-defined minimal positive real number.

10. The method according to claim 1, characterized in that, The generation of final control commands and their distribution to the physical layer specifically includes: Singular value decomposition is performed on the cross-sectional tidal current sensitivity matrix to extract a family of column vectors whose absolute singular values ​​are less than a preset tolerance threshold, forming a null space orthogonal mapping basis; Construct orthogonal projection matrix operators based on the null space orthogonal mapping basis matrix; The orthogonal projection matrix operator is multiplied by the secondary allocation vector and then superimposed onto the primary control command to generate the final control command.

Images