A power system intelligent emergency control method and system for data loss
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-14
- Publication Date
- 2026-08-11
Smart Images
Figure CN122553149A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power system control technology, specifically to an intelligent emergency control method and system for power systems with missing data. Background Technology
[0002] The safe and stable operation of power systems relies on real-time data provided by wide-area measurement systems. In actual operation, especially when system faults cause transient fluctuations, the measurement data received by the control center often suffers from packet loss or interruptions due to sensor failures or communication network congestion, resulting in data loss conditions.
[0003] Faced with incomplete measurements, existing technologies typically perform numerical interpolation directly on the missing data in the physical space. This interpolation method struggles to adapt to the nonlinear characteristics of high-dimensional data, easily disrupting the inherent algebraic topological constraints of the power grid, and consequently leading to errors in subsequent system state assessments. Furthermore, system dynamic trajectory prediction based on this lack of underlying physical constraints, if using conventional discrete-time series or purely data-driven models, is prone to trajectory divergence when faced with boundary conditions or severe data gaps, failing to provide accurate forward-looking conditions.
[0004] Furthermore, after the system assessment identifies an instability risk and generates an emergency control strategy, the instructions may encounter downlink communication hardware blockage during the physical delivery phase, making some controlled nodes controlless zones. Existing emergency control strategies often struggle to implement adaptive equivalent adjustments to address the issue of control instructions failing to execute due to communication interruptions. If control demands cannot be reasonably transferred after a critical node control failure, the originally planned stable control actions may fail, potentially causing disordered diffusion of control actions in other areas, making it difficult to guarantee the effectiveness of emergency control in the power system under adverse communication conditions. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention provides an intelligent emergency control method and system for power systems with missing data, solving the technical problems of low accuracy in state prediction when measurement data is missing and the inability to execute emergency control commands due to communication interruptions.
[0006] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of the present invention provides an intelligent emergency control method for power systems with missing data, comprising the following steps: Acquire local observation data of the power system and divide the power grid nodes into healthy observation sets, data missing sets, and control blind zone sets; Local observation data belonging to the health observation set are mapped to the latent space, and a dynamic prediction model of the power grid latent space with embedded physical equations is constructed. The model is then used to predict the latent state of the mapped system by time integration. In the latent space, virtual mirror nodes corresponding to the nodes in the missing data set are generated, a topological energy potential field is constructed, and the virtual mirror nodes are driven to undergo adaptive evolution through the global smoothness of the topological energy potential field. When the total energy of the topological energy potential field exceeds the safety threshold, an ideal emergency control vector is generated based on the energy descent gradient. Identify the available and non-executable actions in the ideal emergency control vector, perform equivalent reconstruction of the action space for the non-executable actions, and find alternative combinations of actions; The available control actions are combined with the alternative combined actions to generate a comprehensive emergency control command vector, which is then issued for execution.
[0007] Furthermore, the step of dividing the power grid nodes into a health observation set, a data missing set, and a control blind zone set includes: obtaining the original full state vector of the power grid nodes; assigning nodes with normal communication and complete data to the health observation set, assigning nodes that cannot receive measurement data to the data missing set, and assigning nodes that cannot receive and execute control commands to the control blind zone set; and extracting the original full state vector belonging to the health observation set as the local observation data.
[0008] Furthermore, the construction of the power grid implicit space dynamic prediction model with embedded physical equations includes: adopting a normal differential equation framework, embedding Kirchhoff's laws and generator swing equations as physical inertial priors into the model's implicit layer; solving the initial value problem of the ordinary differential equations containing the node admittance matrix of the system's real-time topology changes and the control vector at the current moment, and outputting the implicit space prediction state at a future set time.
[0009] Furthermore, the step of driving the virtual mirror node to adaptively evolve through the global smoothness of the topological energy potential field includes: defining a generalized Lyapunov energy function representing the transient energy of the system in the current hidden state, and constructing the topological energy potential field; calculating the Hessian matrix of the topological energy potential field with respect to the full hidden state, and updating the state of the virtual mirror node through spatial gradient descent by combining the local potential field curvature smoothness penalty term; and stopping the iteration to complete the autonomous coupling when the difference between the states of the virtual mirror node in adjacent iterations is less than a preset error threshold.
[0010] Furthermore, the step of generating the ideal emergency control vector based on the energy descent gradient includes: constructing an optimization objective function, which includes the predicted energy at the end of the look-ahead control time window and the control cost and its weight coefficients for penalizing over-control behavior; solving the optimization objective function to minimize the weighted sum of the predicted energy and the control cost, and outputting the ideal emergency control vector.
[0011] Furthermore, the search for alternative combination actions includes: orthogonally decomposing the ideal emergency control vector into the available control actions falling within the non-blind zone interval and the non-executable actions falling within the control blind zone set interval; when the non-executable actions exist, triggering an equivalent reconstruction mechanism to search for the alternative combination actions in the normal communication area.
[0012] Furthermore, the triggering equivalent reconfiguration mechanism searches for the alternative combination of actions in the normal communication area, specifically by: calculating the ideal total energy reduction corresponding to the ideal emergency control vector; minimizing the difference between the actual total energy reduction generated by the available control actions and the alternative combination of actions and the ideal total energy reduction, and constructing an equivalent reconfiguration optimization problem in combination with a sparsity penalty coefficient; solving the equivalent reconfiguration optimization problem under the condition of satisfying physical capacity constraints to obtain the alternative combination of actions.
[0013] Furthermore, the generation and execution of the integrated emergency control command vector includes: concatenating the available control actions with the alternative combination actions to generate the integrated emergency control command vector; sending the integrated emergency control command vector to the physical actuators through the normal communication topology to drive the execution of generator disconnection, load unloading, or reactive power compensation operations; and updating the physical state of the power grid and entering the next control cycle after the physical actuators have acted.
[0014] A second aspect of this invention provides an intelligent emergency control system for power systems experiencing data loss, comprising multiple modules sequentially connected in communication to fully execute the aforementioned method. The system specifically includes: The dynamic sensing module is used to acquire local observation data of the power system and divide the power grid nodes into a healthy observation set, a data missing set, and a control blind zone set. The latent space prediction module is used to map local observation data belonging to the health observation set to the latent space, construct a dynamic prediction model of the power grid latent space with embedded physical equations, and use the model to perform time integral prediction of the mapped system latent state. The potential field evolution module is used to generate virtual mirror nodes corresponding to nodes in the missing data set in the latent space, construct a topological energy potential field, and drive the virtual mirror nodes to perform adaptive evolution through the global smoothness of the topological energy potential field. An energy optimization module is used to generate an ideal emergency control vector based on the energy descent gradient when the total energy of the topological energy potential field exceeds a safety threshold. The equivalent reconstruction module is used to identify the available control actions and non-executable actions in the ideal emergency control vector, perform equivalent reconstruction of the action space for the non-executable actions, and find alternative combination actions. The instruction synthesis and execution module is used to combine the available control actions with the alternative combination actions to generate a comprehensive emergency control instruction vector and issue it for execution.
[0015] Furthermore, in the aforementioned intelligent emergency control system for power systems facing data shortages, the equivalent reconfiguration module is specifically used for: The ideal emergency control vector is orthogonally decomposed into available control actions and non-executable actions. When a non-executable action exists, the ideal total energy reduction corresponding to the ideal emergency control vector is calculated. With the objective of minimizing the difference between the actual total energy reduction generated by the available control actions and the alternative combined actions and the ideal total energy reduction, an equivalent reconstruction optimization problem is constructed by combining a sparsity penalty coefficient. The equivalent reconstruction optimization problem is solved under the condition of satisfying the physical capacity constraint to obtain the alternative combined actions.
[0016] This invention provides an intelligent emergency control method and system for power systems experiencing data shortages. It offers the following advantages: 1. To address the technical problem that traditional numerical interpolation easily violates the algebraic topology constraints of the power grid when processing local observation data, this invention maps multi-source incomplete measurements to a latent space in a dimension-reduced manner, constructing a dynamic prediction model of the power grid with embedded physical priors. By using Kirchhoff's laws and the underlying dynamic equations as the boundary logic for the integral derivation of ordinary differential equations, the trajectory divergence phenomenon that easily occurs in pure data-driven models under boundary conditions is alleviated, and the fidelity of the system's continuous prediction trajectory is improved when facing data gaps.
[0017] 2. To address the problem of unobservable states of specific nodes due to communication interruptions, this invention introduces a smoothness penalty mechanism based on the generalized Lyapunov energy potential field and the curvature of the local Hessian matrix in the latent space. This mechanism utilizes the coupling characteristics of the power grid's energy spatial distribution to drive the virtual mirror node to adaptively evolve along the potential field gradient, achieving the restoration and autonomous coupling of the latent state of the breakpoint under the condition of missing local measurements.
[0018] 3. To address the risk of emergency control commands failing to execute due to downlink hardware blockage, this invention performs spatial orthogonal decomposition of the control vector to isolate unexecutable actions. Based on the power transfer and diffusion paths implied by the grid node admittance matrix, and aiming to minimize the approximation error between the actual and ideal energy reduction, it reconstructs alternative combined actions within the healthy communication region, incorporating action sparsity constraints. This mechanism prevents the disordered diffusion of control actions and smoothly transfers control requirements through spatial electrical coupling, ensuring the implementation of interception commands. Attached Figure Description
[0019] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a system structure block diagram of the present invention; Figure 3 This is a schematic diagram of the power grid node state isolation and deep nonlinear coding network structure of the present invention; Figure 4 This is a calculation diagram of the neural ordinary differential equations embedded in the physical priors of the present invention; Figure 5 This is a schematic diagram illustrating the adaptive evolution principle of the virtual mirror node in the topological energy potential field according to the present invention. Figure 6 This is a physical schematic diagram of the orthogonal decomposition and spatial equivalent reconstruction of the control blind zone action of the present invention. Detailed Implementation
[0020] The technical solutions in 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.
[0021] It should be noted that the local observation data sequence obtained by the control center in this embodiment includes electrical characteristic components such as voltage amplitude, phase angle, and active and reactive power. These components are all derived from publicly available or authorized industrial measurement systems within the power grid. They are purely physical status data of equipment operation and do not contain any sensitive privacy information involving the personal identity of users. Furthermore, the data acquisition and usage process fully complies with relevant data security and legal requirements.
[0022] Please see the appendix Figure 1 - Appendix Figure 6 This invention provides a method and system for intelligent emergency control of power systems with missing data. The method for intelligent emergency control of power systems with missing data may include the following steps: Acquire local observation data of the power system and divide the power grid nodes into healthy observation sets, data missing sets, and control blind zone sets; Local observation data belonging to the health observation set are mapped to the latent space, and a dynamic prediction model of the power grid latent space with embedded physical equations is constructed. The model is then used to predict the latent state of the mapped system by time integration. In the latent space, virtual mirror nodes corresponding to the nodes in the missing data set are generated, a topological energy potential field is constructed, and the virtual mirror nodes are driven to undergo adaptive evolution through the global smoothness of the topological energy potential field. When the total energy of the topological energy potential field exceeds the safety threshold, an ideal emergency control vector is generated based on the energy descent gradient. Identify the available and non-executable actions in the ideal emergency control vector, perform equivalent reconstruction of the action space for the non-executable actions, and find alternative combinations of actions; The available control actions are combined with the alternative combined actions to generate a comprehensive emergency control command vector, which is then issued for execution.
[0023] In general, the actual physical dynamics of a power system follow nonlinear differential-algebraic equations. Let the actual full state vector of the power grid be... ,in For the overall system state dimension, the time variable is... The control input vector accepted by the system is ,in Let be the dimension of the controllable device. The continuous evolution of the system follows a differential equation: ; In this formula, This represents the derivative of the system's total state with respect to time. Nonlinear functions characterizing the underlying physical laws of power systems, It is the node admittance matrix that contains the real-time connection relationships of the system.
[0024] It should be noted that existing power system models are typically represented as a system of differential-algebraic equations containing differential and algebraic variables (such as node voltages and injected currents). In the above formula expression of this invention, to accommodate subsequent continuous mapping to the implicit space and modeling of neural constant differential equations, the traditional differential-algebraic equations are compactly reconstructed and simplified. Specifically, the full state vector... It is a generalized set of states, which internally includes electrical characteristic components such as voltage magnitude, phase angle, and power of each node (i.e., it incorporates the representation of traditional algebraic variables); while nonlinear functions The physical operation logic implicitly includes the network node admittance matrix. The process of eliminating variables in the power flow network equations and algebraic variables is used to abstract the overall dynamics of the system into the unified ordinary differential equation form mentioned above.
[0025] In actual engineering operation environments, the control center acquires observation sequences through remote sensors. Due to communication failures or data loss, the measurement equations acquired by the control center are expressed as follows: ; In this formula, The sequence of observation data received by the control center. For the observation function, A time-varying Boolean mask vector characterizing the connectivity of a communication link, with symbol... This represents the Hadamard product operation, which involves multiplying corresponding elements one by one. This refers to measurement noise during the sensor and transmission process.
[0026] When the mask vector When a specific element in the data is zero, the measurement data of the corresponding node is completely lost, forming local observation data. This invention acquires this local observation data and analyzes the data message status of each node. Based on the integrity of the messages and the bidirectional communication handshake status, all nodes in the power grid topology are physically divided into three subsets: a healthy observation set with complete data interaction, a missing set with nodes that can only send data or are completely disconnected, resulting in an unknown status, and a control blind zone set with measurement capabilities but blocked downlink control links.
[0027] Faced with incomplete measurements, directly performing numerical interpolation in a high-dimensional physical space containing noise and missing discontinuities would violate the system's algebraic topological constraints. Therefore, this invention establishes a latent space mapping mechanism to compress and map local observation data belonging to the healthy observation set to a low-dimensional continuous manifold space. The state variables in the latent space are defined as follows: The latent space eliminates physical observation noise and extracts the essential dynamic characteristics of the system. Within this space, a dynamic prediction model (used to characterize the system's environmental state transitions and dynamic evolution laws) is constructed using a neural ordinary differential equation framework, forcibly embedding the system's underlying physical equations as prior knowledge into the model's deductive logic. Using this dynamic prediction model, continuous-time integral predictions are performed along the time axis to predict the system's future latent state trajectory, outputting the state evolution trend within the look-ahead control time window.
[0028] To address the issue of missing data, this invention constructs corresponding virtual mirror nodes for missing nodes within the latent space topology. These virtual mirror nodes lack real sensor data support, and their state evolution is entirely governed by the physical coupling and energy diffusion patterns of adjacent healthy nodes. Based on this, a topological energy potential field characterizing the system's stability margin is constructed, and a global smoothness penalty factor for the entire potential field distribution is calculated. This penalty factor acts as a driving force, compelling the virtual mirror nodes to adaptively iteratively evolve along the direction of potential field gradient descent until the distribution of the entire network's topological energy eliminates drastic singularities and discontinuous abrupt changes, achieving autonomous coupling based on physical laws.
[0029] After completing the latent state derivation of the system, the total energy integral value of the topological energy potential field is calculated in real time. The total energy represents the degree to which the current power grid operating point deviates from the static safe and stable equilibrium point. When the total energy exceeds the set safety threshold, the system is deemed to face the risk of instability or even cascading collapse. At this time, the descent gradient of energy with respect to the control action variable is calculated in the latent space, and guided by this gradient, an ideal emergency control vector that can theoretically suppress the energy rise most quickly is generated.
[0030] However, the generated ideal emergency control vector encounters a control blind zone when physically deployed. This invention accurately identifies available control actions and actions that fall into the blind zone and cannot be executed by performing spatial orthogonal decomposition of the ideal action vector. For unexecutable actions, an equivalent reconfiguration mechanism is triggered. This reconfiguration mechanism searches for alternative combination actions within the topology region with normal communication. Its core objective is to ensure that the actual energy reduction of the entire system after the execution of the alternative combination action is highly consistent with the energy reduction that the original unexecutable action should have produced. Finally, the reconfigured alternative combination action is concatenated with the original available actions to generate a comprehensive emergency control command vector, which is then deployed to the physical actuators through a healthy downlink communication link, thereby achieving closed-loop and equivalent emergency stability control of the power system under harsh data environments.
[0031] In this embodiment, accurately dividing the data dimensions of the power grid nodes is a prerequisite for constructing a high-dimensional latent space manifold. At the control center, the original full state vector of the power grid nodes is obtained, and this state vector is defined as... ,in This represents the total number of physical nodes in the power system topology. Representing the Each node It contains electrical characteristic components including voltage amplitude, phase angle, and active and reactive power.
[0032] Based on the packet flag bits reported by the underlying communication gateway device, the uplink measurement link state parameters of each node are extracted. and downlink control link state parameters The link status parameter is a Boolean value, and its determination criteria are as follows: within a set time window (e.g., 20 milliseconds), when the packet loss rate is less than 5% and the communication delay is less than the control period, a value of 1 indicates that the link is connected and the packet verification passes; conversely, a value of 0 indicates that the link is blocked or there is severe data loss. Based on the above parameters, the node set is divided into absolutely physically isolated groups: those that meet the following conditions are considered. and The nodes are assigned to the health observation set. ; will satisfy The nodes are strictly divided into sets of missing data. ; will satisfy but The nodes are assigned to the control blind zone set. .
[0033] In this invention, after the physical attribute boundaries of the nodes are calibrated, a diagonal mask matrix is constructed. A node is considered healthy if and only if it belongs to the health observation set. When the corresponding diagonal element is 1, the rest are 0. Multiplying this mask matrix by the original full state vector, the original full state vector belonging to the health observation set is masked and zeroed out, resulting in local observation data with fixed-dimensional features. Its mathematical expression is: ; In this formula, To effectively filter out the same-dimensional masked observation matrix after the broken link nodes (i.e., matrix dimension preserved) (Unchanged, but the data of the broken chain node is set to 0), providing uncontaminated and dimensionally aligned initial measurement stimuli for subsequent models.
[0034] As a specific embodiment, the deep nonlinear coding network adopts a multilayer perceptron architecture, specifically including an input layer, three fully connected hidden layers, and an output layer. The number of neurons in the hidden layers is set to 128, 64, and 32 respectively. The Tanh hyperbolic tangent activation function is used between hidden layers to ensure the continuous differentiability of the gradient. The final output is a hidden state vector with a dimension of 16. For the framework of God's frequent differential equations, the function The forward propagation process specifically involves: transferring the hidden state... The process is divided into a pure data-driven branch and a physics-prior branch. Specifically, for the pure data-driven branch, the input data format is a one-dimensional feature sequence tensor, which is the 16-dimensional hidden state vector Z(t) reconstructed into a one-dimensional tensor format of [batch size, number of channels, sequence feature length] (e.g., 1 channel and 16 sequence length) before being input into the network. The pure data-driven branch extracts temporal evolution features through a two-layer 3×3 one-dimensional convolutional network. For the physics-prior branch, the input data format is an algebraic matrix and vector form. The physics-prior branch directly calls the hard-coded Kirchhoff current equations and rotor motion equations of the power system, and uses the encoded transformation matrix to convert the real-time updated admittance matrix. After mapping (in the form of a two-dimensional square matrix) to the latent space to obtain the admittance mapping matrix, it is then compared with the latent state. The physical residual gradient is obtained by multiplying the gradient vectors (in one-dimensional column vector format). Finally, the gradient vectors output from the two branches are weighted and summed with a fixed weight ratio of 0.7 and 0.3, and used as the final slope input of the ordinary differential equation integrator, thus completing the forced embedding of the physical prior.
[0035] To enable the deep nonlinear coding network to map from a high-dimensional physical space to a low-dimensional latent space, offline training is required before the system goes live. The training dataset uses historical real SCADA / WAMS measurement data from the power grid and corresponding fault simulation data. The training loss function is weighted by the mean square error between the predicted and actual states, and a penalty term for violating Kirchhoff's laws on physical constraints. The specific weighting formula is expressed as follows: In this formula, For the total loss function, This represents the mean squared error between the predicted state and the actual state. This represents the network node imbalance residual penalty term calculated based on Kirchhoff's laws. These are the weight coefficients for the physical constraints. The Adam optimizer is used for parameter updates, with the initial learning rate set to [value missing]. This continues until the loss function converges to a stable value on the validation set.
[0036] Traditional discrete-time series forecasting models cannot capture the continuous dynamic evolution trajectory of power systems, which possesses strong physical rigidity. Therefore, a deep nonlinear coding network is employed to process the local observation data... The initial hidden state representation is obtained by mapping to a continuous, low-dimensional hidden space that eliminates measurement noise. ,in This is the current observation time.
[0037] Furthermore, a dynamic prediction model for the hidden space of the power grid with embedded physical equations is constructed. This model employs a neural network ordinary differential equation (NDE) framework, parameterizing the time derivative of the system's hidden states as a continuous deep neural network. To prevent trajectory divergence in the purely data-driven model under boundary conditions, Kirchhoff's laws and the generator swing equations are forcibly embedded as physical inertial priors into the hidden layer structure of this NDE framework. The NDE is defined as follows: ; In this formula, For the hidden state, the continuous-time derivative is... For weight parameters The physical embedded neural network function is constructed, and the weight parameters are... These represent the learnable weights and biases in the neural network, which are specifically obtained by offline training and iterative updates of the network using historical operation and fault simulation data before the system goes live. This is the system control vector at the current moment. The node admittance matrix is used to characterize the real-time topological connectivity of the system.
[0038] Combining the network forward propagation process described above, the physically embedded neural network function The specific structural expression is as follows: ; In this specific structure, This represents the output of the residual gradient extracted by the one-dimensional convolutional network in the pure data-driven branch. This represents the physical residual gradient output calculated using physical equations such as the real-time admittance matrix in the physical prior branch.
[0039] The function Within the hidden computational graph, a set of physical residual evaluation paths (i.e., the physical prior branches mentioned above) are connected in parallel through structured design. Specifically, when solving the above ordinary differential equations, the Kirchhoff's laws prior are obtained by calling the admittance mapping matrix mapped to the hidden space and the hidden state. This is achieved through matrix multiplication. It's important to note that this multiplication is not the traditional admittance-voltage multiplication in physical space, but rather utilizes the network topology connections carried by the admittance mapping matrix to represent the hidden states that incorporate complex electrical characteristics. Feature aggregation is performed on the topological graph space. This algebraic operation equivalently calculates the potential mapping relationship between the injected current and the power balance residual of the grid nodes in the latent space. The generator swing equation prior is calculated in real time by using the power flow transmission power derived from the above network balance residuals to calculate the imbalance between it and the equivalent mechanical power of the nodes, and then obtaining the rate of change of the equivalent rotor inertial kinetic energy. This path uses the residual values calculated based on the two major physical laws as the physical guiding gradient (i.e., the physical prior component in the aforementioned formula), forcing it to participate in the time-dependent calculation. continuous derivative In the calculation, this architecture ensures that the direction of the state gradient output of the neural network is always constrained within the feasible solution space manifold that conforms to the physical laws of the actual power system.
[0040] In this embodiment, for a future set time The state prediction is rigorously transformed into a process of solving an initial value problem of an ordinary differential equation. Given the initial hidden state... The system inputs control commands and the admittance matrix, which includes real-time updates of the circuit breaker's opening and closing status, are used to perform continuous-time integration along the time manifold using a fourth-order Runge-Kutta numerical integrator, outputting the latent space predicted state at a future set time. Its integral solution formula is expressed as: ; In this formula, The integral process is a continuous-time variable. This integral process fully preserves the physical continuity of the system's dynamics, overcomes the truncation error problem that is easily generated by the fixed-step discrete model when dealing with the high-frequency transient and violent fluctuations of the power grid, and outputs a high-fidelity latent state look-ahead trajectory for subsequent energy potential field analysis.
[0041] In this embodiment, to address the incomplete system state space caused by missing data sets, physical law-driven data completion and state evolution are implemented in the latent space. A node topology graph is extracted from the latent space, which maintains strict isomorphism with the physical node connection topology of the real power system. Specifically, the nodes in this topology graph correspond one-to-one with the physical power grid nodes, and the connection edges and weights between nodes are determined by the node admittance matrix of the aforementioned real-time system topology connection relationship. The unique characteristic is that the adjacent node relationships in the latent space directly correspond to the node relationships with direct electrical connections in the physical power grid. Based on this isomorphic topology, nodes in the missing data set are directly mapped to virtual mirror nodes in the latent space. Let the full state vector of the latent space be... ,in The deterministic hidden state vector determined by the mapping of the health observation set (i.e., the locally effective hidden state mentioned above) ), For the hidden state vector to be evolved corresponding to the virtual mirror node, its data dimension is the same as... The nodes have the same feature dimensions (e.g., 16 dimensions). The initial value and its dynamic evolution will be obtained by iteratively solving the subsequent topological energy objective functional.
[0042] In this invention, a generalized Lyapunov energy function characterizing the transient energy of the system in the current hidden state is constructed for the hidden space, thereby forming a network-wide topological energy potential field. It should be noted that this network-wide topological energy potential field is mapped using the aforementioned positive semi-definite admittance mapping matrix. The network topology connection weights are structurally constructed. The total energy of the global situation is not a simple one-dimensional algebraic sum of the energies of isolated nodes, but rather is determined through the system's hidden state vector. Mapping matrix with admittance Matrix multiplication (i.e., in the following formula) ) and nonlinear integral terms (i.e., in the following formula) This calculates the sum of the interactive coupling potential energies between all electrically connected adjacent node pairs in the latent space. The total energy scalar value is... A high-dimensional potential energy surface is constructed in the multidimensional hidden state variable space. The spatial distribution and gradient of this surface in the directions of each node state constitute the global topological energy potential field driving the evolution of the virtual mirror node. This energy function equivalently maps the generator rotor kinetic energy and the magnetic / electric potential energy transmitted between network nodes in the physical space of the power grid to the hidden space. The generalized Lyapunov energy function is defined as follows: ; In this formula, Latent space state The corresponding scalar energy value; it needs to be clarified that the 16-dimensional hidden state vector output by the encoding network mentioned above specifically refers to the feature subset extracted from the locally effective observation data (i.e., the deterministic hidden state of the healthy node in this section). ); and the formula used in this formula and subsequent energy potential field calculations This refers to the state of health in the latent stage. Hidden state of the virtual mirror node to be evolved The system-level full hidden state vector (i.e., the one defined above) is formed by completely piecing together the physical topology of the power grid. ); positive semidefinite admittance mapping matrix From the nodal admittance matrix of physical space The mapping is obtained through a linear transformation, and the mapping relationship is as follows: ,in This is the encoding transformation matrix from physical space to latent space. Specifically, this matrix is obtained by calculating the Jacobian matrix of the input physical feature vector at the current steady-state operating point of the aforementioned deep nonlinear coding network. It is used to accurately characterize the linearized topological mapping relationship of the nonlinear space at local operating points; static safe and stable equilibrium state. The data is obtained by mapping the measured data during normal steady-state operation of the system through an encoder. The specific mapping process is as follows: Complete node voltage amplitude, phase angle, and active and reactive power measurement data under historical rated operating conditions are extracted as a reference input vector. This vector is then input into the previously trained deep nonlinear coding network (i.e., the encoder mentioned here, which specifically adopts a multilayer perceptron architecture containing an input layer, three fully connected hidden layers with 128-64-32 nodes, and an output layer, with Tanh activation functions used between layers). Through the forward propagation operation of this multilayer perceptron network, the dimensionality-reduced 16-dimensional hidden space reference state vector is directly output. This vector represents the stable state where the power angle and voltage of the corresponding power grid are within the rated operating range. The nonlinear mapping function... The hyperbolic tangent function is used to characterize the nonlinear saturation characteristics of transient power exchange between nodes, and its expression is: ; Let be the state integration variable during the integration process. After this nonlinear integral term is calculated, it is added to the aforementioned matrix multiplication term to jointly constitute the generalized Lyapunov energy scalar value of the current system. This is used as the basis for subsequent differentiation calculations of the Hessian matrix. And the basic input for constructing the evolution objective functional.
[0043] Virtual mirror nodes lack the support of real measurement data, and their state evolution is constrained by the surrounding topological potential field. The calculation of the topological energy potential field with respect to the total hidden states... The Hessian matrix is a mathematical expression for the energy potential field. This matrix reflects the curvature distribution and second-order rate of change of the energy potential field in multidimensional space. ; In this formula, The calculated Hessian matrix has dimensions corresponding to the dimensions of the full set of hidden states. and These represent the hidden state vectors respectively. The first in The and the first Each component.
[0044] Combining the Hessian matrix, a global smoothness penalty term based on the curvature of the local potential field is defined. By globally summing the traces of the local Hessian matrices of all nodes, this penalty term is used to constrain the continuity of adjacent physical nodes in the latent space energy field, avoiding non-physical state abrupt changes caused by data breakpoints, and realizing the adaptive evolution of the virtual mirror nodes driven by the global smoothness of the topological energy potential field. An evolutionary objective functional containing the total energy and the smoothness penalty term is constructed: ; In this formula, For the evolution objective functional composed of the state variables of the virtual mirror nodes, The penalty weight constant for adjusting the curvature smoothing intensity typically ranges from [0.1, 1.0], and is preferably set to 0.5 in this embodiment; Tr The trace operation is used to measure the overall distortion of the local potential field.
[0045] The state of the virtual mirror node is updated using spatial gradient descent. Before starting the iteration, the hidden state to be solved is set. Initial iteration starting point To ensure rapid convergence of gradient descent and prevent getting trapped in local minima, the initial iteration starting point... The specific acquisition method is as follows: First, extract the historical cached hidden state of the corresponding physical node in the missing data set during the last monitoring period before the communication interruption, as the initial value; when the historical cached data is unavailable or expired, extract all adjacent healthy observation nodes with direct physical topology connections to the virtual mirror node, and calculate the arithmetic mean of their hidden state vectors as the initial value. In each iteration, maintain the healthy hidden state. Absolutely fixed, calculate the evolution objective functional pair The first-order partial derivative is used to update the state variables of the virtual mirror node along the negative gradient direction. The update rule for the next iteration is expressed as: ; In this formula, and The first Round and number The hidden state of the virtual mirror node after each iteration. For evolutionary step size, Let be the gradient operator with respect to the variable to be evolved.
[0046] The evolution iteration stops when the state difference of the virtual mirror node between adjacent iterations tends to converge or reaches the preset maximum allowable computation time. A preset error threshold is set to... and the maximum number of iterations (For example, set to 50 iterations to ensure the iteration process is completed strictly within the emergency control communication cycle of hundreds of milliseconds), calculate the second-order norm distance between the state vectors of two adjacent iterations. The stopping criterion is: or .
[0047] When any of the above conditions are met, it indicates that the forces acting on the virtual mirror node in the topological energy potential field have reached dynamic equilibrium or have touched the hard real-time computing boundary. At this point, iteration stops, confirming the completion of the autonomous coupling between the virtual mirror node and the real physical network. This process utilizes the physical rigidity of the power grid's energy distribution to accurately reconstruct and complete the hidden state of the missing node even when local sensor data is completely lost.
[0048] It is worth noting that the virtual mirror nodes in the latent space maintain a strict one-to-one correspondence with the missing data nodes of the real physical network in terms of the index dimension. The complete set of latent states generated after evolution... Instead of using a decoder to inversely map back to a high-dimensional physical space, it is directly used as the sole state input for subsequent topological energy potential field calculations, thereby enabling subsequent system-level transient stability assessments and control command generation to be completed directly on a low-dimensional manifold, significantly reducing online computation latency.
[0049] In this embodiment, after completing the state autonomy coupling of the virtual mirror node in the hidden space, the transient stability margin of the overall power system is evaluated in real time. The current moment is extracted. Full hidden state Substituting this into the aforementioned generalized Lyapunov energy function, the total energy of the topological energy potential field at the current moment can be calculated. Pre-configure a safety threshold corresponding to the system stability boundary. When the total energy calculated in real time meets the requirements... If the system determines that its current operating condition has crossed the safety manifold boundary, it directly triggers the generation mechanism of the system-level emergency control sequence.
[0050] In this invention, to ensure the power grid state trajectory returns to the static, safe, and stable equilibrium point in the shortest possible time, an optimization objective function is constructed to solve for emergency action commands. The look-ahead control time window is defined as... ,in This is the set prediction window duration. Let... Let be the control vector to be solved by the system during this period. Construct an optimization objective function that includes the predicted energy, control cost, and their weighting coefficients at the end of the look-ahead control time window. Its mathematical expression is: ; In this formula, To characterize the scalar optimization objective function of the overall control benefits, For the end of the forward control time window Predicting energy in a time-based system To apply control vector The corresponding control cost, To predict the energy weighting coefficient, To control the cost weighting coefficient.
[0051] In this optimization objective function, the energy is predicted. The value of depends strictly on the aforementioned latent space dynamic prediction model under specific control input. The results of subsequent continuous-time integral derivation. Control cost. Used to penalize excessive control behavior, strictly limiting the damage to system topology integrity caused by large-scale load shedding or generator disconnection operations. The mathematical formula for calculating the control cost function is: ; In this formula, The control cost matrix is a positive definite diagonal matrix, where each element on the diagonal maps to the hardware losses and economic penalty weights of different physical actuators in the power system.
[0052] The above model is optimized to solve for the objective function, minimizing the weighted sum of the predicted energy and the control cost. Utilizing the continuous differentiability of the latent space dynamic prediction model's state evolution, the mapping relationship between the total energy and the control variable is calculated. Based on the adjoint sensitivity analysis algorithm, an adjoint equation that evolves backwards from the original ordinary differential equation is constructed. The adjoint state is solved using continuous backpropagation within a deep learning framework, thereby calculating the objective function with respect to the control vector. Energy descent gradient operator Within the feasible region constraints formed by the upper and lower limits of the nameplate capacity of all physical actuators in the network, the gradient descent method is used to perform spatial optimization iteration on the control vector along the opposite direction of the energy descent gradient. The iteration step size is preferably set to 0.02, and the maximum number of iterations is set to 50.
[0053] When the second norm of the energy descent gradient is less than the set convergence threshold (e.g.) When the maximum number of iterations is reached, it indicates that the minimum point of the objective function has been found. At this point, the iterative update operator is terminated, and the optimal control solution matrix that minimizes the weighted sum is extracted. This optimal control solution matrix is defined and output as the ideal emergency control vector. This ideal emergency control vector represents the baseline control strategy that can achieve energy suppression most quickly under purely ideal operating conditions, ignoring anomalies in the underlying communication network and hardware blind spots.
[0054] In this embodiment, although the generated ideal emergency control vector theoretically possesses the best stability suppression effect, due to the existence of the control blind zone set, some control commands cannot physically reach the controlled device. The topology information of the control blind zone set obtained in the previous steps is extracted to construct a diagonal mapping matrix for the control link state. When the downlink control link of a physical node is connected, its diagonal elements are 1; when it belongs to the control blind zone set, its diagonal elements are 0. Using this mapping matrix, the ideal emergency control vector is orthogonally decomposed into the available control actions falling within the non-blind zone interval. And the unexecutable actions falling within the control blind zone set interval. The mathematical expression for this orthogonal decomposition process is as follows: ; ; In this formula, This is the identity matrix that matches the dimension of the control vector. Since the projection matrix satisfies... This ensures, from a mathematical perspective, that the decomposed usable action subspace and the non-executable action subspace are strictly orthogonal, thus avoiding the redundant calculation of action energy.
[0055] Calculate the norm of the unexecutable action vector. When the following conditions are met... When this condition occurs, it indicates that there is an emergency interception action within the system that cannot be implemented due to a communication channel interruption. In response to this condition, the system triggers an equivalent reconfiguration mechanism to find alternative combinations of actions within the normal communication region. First, based on the latent space generalized Lyapunov energy function, the ideal total energy reduction corresponding to the complete application of the ideal emergency control vector is calculated. : ; In this formula, Let be the initial total transient energy of the system at the moment it is disturbed. The total system energy predicted at the end of the look-ahead control time window is given under the drive of an ideal emergency control vector sequence.
[0056] In this invention, due to the strong physical coupling and diffusion characteristics of the power grid's spatial energy field, the removal of a single power imbalance device within a blind zone can be effectively replaced by coordinating the operation of multiple adjacent generators or load nodes within the normal communication area. The physical basis for this is that, based on the node admittance matrix of the power grid, the electrical distance and power transfer distribution factor (PTDF) between nodes are calculated. Healthy nodes with closer distances and higher PTDF exhibit higher sensitivity to power imbalances in blind zone nodes, thus effectively absorbing the unbalanced energy that should be controlled in the blind zone through the electrical coupling path. For heterogeneous control actions such as generator shedding, load shedding (active power regulation), and reactive power compensation (reactive power regulation) within the action space, the system constructs a unified energy per-unit value mapping benchmark, uniformly converting the actual physical output of different types of actions into an equivalent contribution to the decrease in the system's generalized Lyapunov energy function. Based on this, the replacement combination action to be solved within the normal communication area is set as follows: By using the forward extrapolation of the hidden space dynamic prediction model, the actual total energy reduction resulting from the combined action of available control actions and alternative actions is calculated. The optimization objective is to minimize the difference between the actual total energy reduction and the ideal total energy reduction. Combined with a sparsity penalty coefficient, an equivalent reconstruction optimization problem is constructed: ; In this formula, To reconstruct the squared error term, which characterizes the approximation accuracy of the equivalent energy-suppressed trajectory; The constant coefficient is the sparsity penalty coefficient; The L1 norm is introduced to replace the action vector. The purpose of introducing the L1 norm is to force the action commands of non-critical control nodes to converge to zero, thereby preventing excessive divergence of reconstruction commands from causing large-scale and frequent switching across the entire system.
[0057] When numerically optimizing the equivalent reconfiguration problem, the solution space must be limited to the ultimate carrying capacity of the power system equipment. The static physical capacity of each control station within the normal communication area is extracted, and a minimum control capacity lower bound vector for the actuators is constructed. With the maximum adjustable capacity upper limit vector And apply physical capacity constraints: .
[0058] Under the condition of satisfying the above physical capacity constraints, the equivalent reconstruction optimization problem containing L1 norm non-smooth terms is solved iteratively using a sequential quadratic programming algorithm or an interior point method, and the converged alternative combination action is output. This reconfiguration process successfully transferred the control requirements accumulated in the blind zone to a cluster of devices with complete control links, based on the inherent topological admittance relationship of the power grid.
[0059] Finally, in the final physical scheduling execution phase, the available control actions are concatenated. The alternative combination action after optimization verification Generate a comprehensive emergency control command vector The algebraic composition logic representation of this synthesized instruction vector is as follows: .
[0060] After synthesis, the integrated emergency control command vector is sent to the corresponding physical actuators through an undamaged normal communication topology and security gateway. Through the drive circuit at the relay protection layer, generator disconnection, load unloading, or rapid activation / deactivation of reactive power compensation devices are precisely executed. After the physical actuators complete the relay switching operation, the control system actively refreshes the global power grid sensor messages, updates the power grid physical state cache array, and seamlessly enters the next control cycle, achieving all-weather steady-state defense under harsh communication conditions.
[0061] To further verify the feasibility of this invention, a set of preferred hyperparameter configurations for the above algorithm model are given in a practical application scenario of an IEEE 39-node system in a certain region: the evolution step size of the latent space virtual nodes is set as follows: The preset error threshold for adjacent iterations is set to This ensures autonomous convergence is achieved within a 100-millisecond control period. Safety threshold. The system's historical static stability margin is calibrated to 15.0 pu. The specific calibration process is as follows: During the system's offline phase, multiple sets of node state data for critical stability and instability conditions are extracted from the power grid's historical operation or simulation data. These data are then substituted into the aforementioned generalized Lyapunov energy function to calculate the total energy value. The critical energy value that distinguishes the stability and instability boundaries is taken as the system's fixed safety threshold, which is confirmed to be 15.0 pu in the IEEE 39-bus system scale of this embodiment. When constructing the equivalent reconfiguration optimization problem, to balance the energy reduction and control costs, the predicted energy weight coefficient is set to... The control cost weighting coefficient is set to The optimal sparsity penalty constant coefficient for the L1 norm is... Under these parameters, the system can precisely limit the load shedding action to no more than three key nodes, effectively avoiding excessive dispersion of the action.
[0062] Furthermore, those skilled in the art should understand that the technical solution of the present invention is not only applicable to the aforementioned IEEE 39-node transmission network, but also to power systems of different scales (such as provincial backbone power grids with hundreds of nodes) or different voltage levels (such as distribution networks containing a large number of distributed power sources). It only requires adaptive adjustment of the full state vector according to the physical topology of the actual power grid. Dimensions and node admittance matrix The order of the deep nonlinear coding network is used to proportionally expand the number of hidden layer nodes and the output hidden state vector of the multilayer perceptron based on the node size. The dimensions (e.g., expanding to 128 dimensions for a scale of hundreds of nodes) are adjusted proportionally to the system's baseline capacity, while the security threshold is adjusted accordingly. With the core algorithm parameters, the latent space mapping and reconstruction algorithm architecture disclosed in this invention can be directly reused, which has broad system applicability and scalability.
[0063] In this embodiment, based on the aforementioned disclosed intelligent emergency control method for power systems with missing data, this invention provides an intelligent emergency control system for power systems with missing data, used to fully implement the above-mentioned algorithm flow and physical operation steps. This system, as a central decision-making platform, is deployed in the power grid dispatch and control center and includes multiple core logic computing modules connected in sequence. The system includes a dynamic sensing module, whose physical input is connected to the power grid wide-area measurement system and data acquisition and monitoring gateway. This module acquires local observation data sequences of the power system under disturbed conditions, and based on the real-time parsed bidirectional handshake state of the message link, strictly divides the physical mapping space of all network nodes into a healthy observation set, a data missing set, and a control blind zone set, and sends the division results of the above sets and the corresponding data to the hidden space prediction module through the data bus.
[0064] In this invention, the system further includes a latent space prediction module and a potential field evolution module. The latent space prediction module receives local observation data belonging to the health observation set, projects it onto a low-dimensional continuous latent space through a nonlinear mapping function, and constructs a dynamic prediction model of the power grid latent space. Utilizing this neural network constant differential equation framework embedding Kirchhoff's laws and generator swing equations, it solves the differential initial value problem containing the real-time physical admittance matrix, thereby performing accurate continuous-time integral prediction of the mapped system latent state. The potential field evolution module receives the model parameters and latent state trajectories output by the latent space prediction module. In the latent space topology, it generates corresponding virtual mirror nodes for nodes lacking sensor data and defines a generalized Lyapunov transient energy function to construct the entire network topology energy potential field. This module extracts the Hessian matrix to calculate the global smoothness penalty term for the local potential field curvature, driving the virtual mirror nodes to undergo adaptive autonomous evolution along the spatial negative gradient direction until the state differences converge and approximate the true power grid state manifold.
[0065] To generate and verify the defense interception command, the system is configured with an energy optimization module and an equivalent reconstruction module. The energy optimization module receives the full system predicted hidden state trajectory, including the autonomous evolution results of virtual mirror nodes, output by the potential field evolution module, and performs real-time numerical integration of the energy of the full system predicted trajectory. When the total energy of the topological energy potential field exceeds the pre-set safety threshold manifold boundary, it calculates and outputs the theoretically optimal ideal emergency control vector based on the energy-with-control-variable descent gradient operator. The equivalent reconstruction module receives this vector and, combined with the control blind zone set mapping relationship transmitted by the dynamic sensing module, performs a rigorous orthogonal decomposition of the physical control dimension using the state mask matrix of the control blind zone. ; In this vector algebraic expression, For a complete ideal emergency control vector, For available control actions falling within the non-blind zone, An unexecutable action that results in physical layer instruction blocking due to falling into a control blind zone set interval.
[0066] The equivalent reconstruction module further incorporates an optimization engine. When an unexecutable action is determined to exist, this module calculates the ideal total energy reduction corresponding to the complete application of the ideal control command sequence. Subsequently, the available control actions are combined with the alternative actions to be solved. The objective is to minimize the approximation error between the actual total energy reduction resulting from the synergistic effect and the ideal total energy reduction. This is achieved by incorporating an L1-norm sparsity penalty term to suppress excessive dispersion of control actions, and performing spatial optimization within the constraints of the upper and lower limits of the actual adjustable physical capacity of generators or loads at each control station, outputting alternative combination actions that satisfy the equivalent control effect. .
[0067] The system includes an instruction synthesis and execution module. The output of this module establishes an instruction delivery channel directly with the power grid physical actuators on the substation side via a secure isolation gateway and a normally operating downlink communication network. At the data layer, this module performs vector concatenation operations to merge the available control actions. Alternative combination actions obtained from reconstruction Generate the final integrated emergency control command vector. Upon receiving the command message, the physical actuator drives the underlying circuit breaker tripping circuit to quickly perform generator disconnection, precise load unloading, or dynamic reactive power compensation operations. After the physical switch contacts have completed their flipping action, the system refreshes the entire network sensor status array again, completing the physical closed loop and entering the next emergency control and monitoring cycle.
[0068] In this embodiment, the dynamic sensing module, latent space prediction module, potential field evolution module, energy optimization module, equivalent reconstruction module, and instruction synthesis and execution module included in the above system are all software program modules. These software program modules are solidified and integrated into the non-volatile memory of an industrial-grade electronic device with high-performance matrix operation capabilities in the form of instruction code. The electronic device includes a high-speed communication bus, a network communication interface, non-volatile memory, and a multi-core central processing unit (such as an AI server with a tensor computing unit). The multi-core central processing unit strictly completes the aforementioned full state partitioning, partial differential equation solving, Hessian matrix updating, and optimal iterative reconstruction operations by calling and executing the program code of each module loaded in the memory. In addition, the present invention also provides a computer-readable storage medium, which persistently stores the control program implementing the above methods. Those skilled in the art should understand that the mathematical matrix transformation, topological evolution logic, and underlying hardware execution network detailed in the embodiment constitute an inseparable engineering closed-loop whole. Any equivalent parameter replacement and architecture adjustment based on the same underlying physical evolution mechanism and energy reconstruction idea are strictly within the protection scope of the technical solution of the present invention.
[0069] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended technical solutions and their equivalents.
Claims
1. A method for intelligent emergency control of power systems with missing data, characterized in that, include: Acquire local observation data of the power system and divide the power grid nodes into healthy observation sets, data missing sets, and control blind zone sets; Local observation data belonging to the health observation set are mapped to the latent space, and a dynamic prediction model of the power grid latent space with embedded physical equations is constructed. The model is then used to predict the latent state of the mapped system by time integration. In the latent space, virtual mirror nodes corresponding to the nodes in the missing data set are generated, a topological energy potential field is constructed, and the virtual mirror nodes are driven to undergo adaptive evolution through the global smoothness of the topological energy potential field. When the total energy of the topological energy potential field exceeds the safety threshold, an ideal emergency control vector is generated based on the energy descent gradient. Identify the available and non-executable actions in the ideal emergency control vector, perform equivalent reconstruction of the action space for the non-executable actions, and find alternative combinations of actions; The available control actions are combined with the alternative combined actions to generate a comprehensive emergency control command vector, which is then issued for execution.
2. The intelligent emergency control method for power systems with missing data according to claim 1, characterized in that, The division of power grid nodes into a health observation set, a data missing set, and a control blind zone set includes: Obtain the original full state vector of the power grid node; Nodes with normal communication and complete data are assigned to the health observation set, nodes that cannot receive measurement data are assigned to the data missing set, and nodes that cannot receive and execute control commands are assigned to the control blind zone set. The original full state vector belonging to the health observation set is truncated as the local observation data.
3. The intelligent emergency control method for power systems with missing data according to claim 1, characterized in that, The construction of the power grid implicit space dynamic prediction model with embedded physical equations includes: We adopt the framework of ordinary differential equations and embed Kirchhoff's laws and the generator swing equation as physical inertial priors into the hidden layer of the model. Solve the initial value problem of the ordinary differential equation that includes the node admittance matrix and the control vector at the current time, which involves real-time topology changes of the system, and output the latent space prediction state at a future set time.
4. The intelligent emergency control method for power systems with missing data according to claim 1, characterized in that, The process of driving the virtual mirror node to adaptively evolve through the global smoothness of the topological energy potential field includes: Define a generalized Lyapunov energy function that characterizes the transient energy of the system in the current hidden state, and construct the topological energy potential field; The Hessian matrix of the topological energy potential field with respect to the full hidden state is calculated, and the state of the virtual mirror node is updated by spatial gradient descent, combined with the local potential field curvature smoothness penalty term. When the difference in the state of the virtual mirror node between adjacent iterations is less than a preset error threshold, the iteration stops and autonomous coupling is completed.
5. The intelligent emergency control method for power systems with missing data according to claim 1, characterized in that, The generation of the ideal emergency control vector based on energy descent gradient includes: Construct an optimization objective function, which includes the predicted energy at the end of the look-ahead control time window and the control cost and its weighting coefficients for penalizing over-control behavior; Solve the optimization objective function to minimize the weighted sum of the predicted energy and the control cost, and output the ideal emergency control vector.
6. The intelligent emergency control method for power systems with missing data according to claim 1, characterized in that, The action of finding alternative combinations includes: The ideal emergency control vector is orthogonally decomposed into the available control actions that fall within the non-blind zone interval and the non-executable actions that fall within the control blind zone set interval; When the aforementioned unexecutable action exists, an equivalent reconfiguration mechanism is triggered to find the alternative combination action in the normal communication area.
7. The intelligent emergency control method for power systems with missing data according to claim 6, characterized in that, The triggering equivalent reconstruction mechanism searches for the alternative combination of actions in the normal communication area, specifically as follows: Calculate the ideal total energy reduction corresponding to the ideal emergency control vector; The objective is to minimize the difference between the actual total energy reduction and the ideal total energy reduction resulting from the available control actions and the alternative combined actions. An equivalent reconstruction optimization problem is constructed by combining the sparsity penalty coefficient. The equivalent reconstruction optimization problem is solved under the condition of satisfying the physical capacity constraint to obtain the alternative combination action.
8. The intelligent emergency control method for power systems with missing data according to claim 1, characterized in that, The generation and execution of the integrated emergency control command vector includes: The available control actions are combined with the alternative combined actions to generate the integrated emergency control command vector; The integrated emergency control command vector is sent to the physical actuators through the normal communication topology to drive the execution of generator disconnection, load unloading or reactive power compensation operations; After the physical actuator is activated, the physical state of the power grid is updated and the next control cycle begins.
9. An intelligent emergency control system for power systems with missing data, characterized in that, For implementing a smart emergency control method for power systems with missing data as described in any one of claims 1 to 8, the system comprises the following components connected in sequence via communication: The dynamic sensing module, whose input end is connected to the power grid measurement system, is used to acquire local observation data of the power system and divide the power grid nodes into a healthy observation set, a data missing set, and a control blind zone set. The latent space prediction module is used to map local observation data belonging to the health observation set to the latent space, construct a dynamic prediction model of the power grid latent space with embedded physical equations, and use the model to perform time integral prediction of the mapped system latent state. The potential field evolution module is used to generate virtual mirror nodes corresponding to nodes in the missing data set in the latent space, construct a topological energy potential field, and drive the virtual mirror nodes to perform adaptive evolution through the global smoothness of the topological energy potential field. An energy optimization module is used to generate an ideal emergency control vector based on the energy descent gradient when the total energy of the topological energy potential field exceeds a safety threshold. The equivalent reconstruction module is used to identify the available control actions and non-executable actions in the ideal emergency control vector, perform equivalent reconstruction of the action space for the non-executable actions, and find alternative combination actions. The instruction synthesis and execution module, whose output is connected to the power grid physical actuator, is used to combine the available control actions with the alternative combined actions to generate a comprehensive emergency control instruction vector and issue it for execution.
10. The intelligent emergency control system for power systems with missing data according to claim 9, characterized in that, The equivalent reconstruction module is specifically used for: The ideal emergency control vector is orthogonally decomposed into the available control actions and the non-executable actions; when the non-executable actions exist, the ideal total energy reduction corresponding to the ideal emergency control vector is calculated; The objective is to minimize the difference between the actual total energy reduction and the ideal total energy reduction resulting from the available control actions and the alternative combined actions. An equivalent reconstruction optimization problem is constructed by combining the sparsity penalty coefficient. The equivalent reconstruction optimization problem is solved under the condition of satisfying the physical capacity constraint to obtain the alternative combined actions.