Active control method for embedded dynamic congestion based on classification neural network

Abstract

The application discloses a dynamic congestion active control method based on a classification neural network analysis embedding, comprising the following steps: establishing a congestion regulation and optimization benchmark model containing multiple types of operation constraints, and performing linearization processing on nonlinear constraints in the congestion regulation and optimization benchmark model; based on power grid historical operation data and corresponding congestion safety state judgment results, a congestion safety classification neural network proxy model is constructed; the congestion safety classification neural network proxy model is reconstructed by a mixed integer linear programming, and congestion safety classification proxy constraints are determined; the congestion safety classification proxy constraints are embedded into the linearization-processed congestion regulation and optimization benchmark model to construct a mixed integer linear programming model and solve the model, and multi-period active control strategies are output; the proxy model is analyzed by a mixed integer linear reconstruction method and embedded into the linearized benchmark model, the congestion safety constraints are realized in the form of resolvable expression and fast solution, and the initiative, real-time performance and solution efficiency of the congestion prevention and control are improved.

Description

Technical Field

[0001] This invention relates to the field of power system safety and stability control, and in particular to a dynamic congestion active control method based on the analytical embedding of a classification neural network. Background Technology

[0002] In actual power grid operation, transmission congestion is usually closely related to system operating conditions. When changes occur in renewable energy output, load levels, conventional unit output, and system voltage status, the transmission capacity and congestion risk of inter-regional tie lines also change accordingly. If potential congestion is not identified and eliminated in time, it may lead to overload of critical lines, reduced safety and stability margins, and even affect the safe and stable operation of the power grid. Therefore, it is of great significance to establish proactive control and optimization methods to address power grid transmission congestion risks, and to coordinate various regulation resources such as conventional units, energy storage, renewable energy curtailment, and load dispatch in advance to form a multi-time-period coordinated control strategy.

[0003] Current power grid control methods often rely on manual experience or scheduling optimization programs to adjust power lines after detecting heavy loads or high congestion risks. This approach suffers from a certain lag when facing rapid fluctuations in renewable energy and complex dynamic security constraints. Especially in near-real-time or real-time control scenarios, if control is only initiated after congestion has approached or actually occurred, control resources may be limited by factors such as ramp rate, response time, and energy storage status, making it difficult to generate sufficient control capacity in a timely manner, thus increasing control difficulty and operating costs. To ensure the safety of control strategies, traditional congestion active control optimization models typically require multiple types of operational constraints. Among these, congestion safety constraints, which take into account the dynamic characteristics of the system, are often not simple fixed power limits but need to be determined by combining the ultimate transmission capacity, safety verification results, and system dynamic stability characteristics under different operating scenarios. In existing technologies, determining the congestion safety state usually relies on methods such as ultimate transmission capacity calculation, static safety verification, transient stability analysis, and time-domain simulation. While these methods can accurately reflect the system safety boundary, the calculation process is complex, the simulation is time-consuming, and it is difficult to repeatedly use them in multi-scenario, multi-time-period active control optimization processes. If this type of unresolvable security decision-making process is directly embedded into the optimization model, the optimization problem will be difficult to solve quickly, making it difficult to meet the timeliness requirements of online active control. Summary of the Invention

[0004] This invention provides a dynamic congestion proactive control method based on the analysis of a classification neural network. It can initiate strategy formulation before the actual occurrence of congestion based on the congestion risk perception results, and reserve action time for adjustment resources affected by ramping constraints or response delays. While avoiding online repeated safety verification and time-domain simulation, it improves the initiative, real-time performance and solution efficiency of congestion prevention and control.

[0005] This invention provides a dynamic blocking active control method based on the parsing of embedded classification neural networks, comprising: A benchmark model for blockage control optimization, which includes multiple types of operational constraints, is established, and the nonlinear constraints in the benchmark model for blockage control optimization are linearized. Based on historical power grid operation data and corresponding congestion safety status determination results, a congestion safety classification neural network proxy model is constructed. The blocking security classification neural network agent model is reconstructed using mixed integer linear reconstruction to determine the blocking security classification agent constraints; The congestion safety classification agent constraint is embedded into the linearized congestion control optimization benchmark model to construct and solve a mixed-integer linear programming model, which is used to output a multi-period active control strategy.

[0006] According to the present invention, a dynamic congestion active control method based on a classification neural network with embedded parsing, the step of establishing a congestion regulation optimization benchmark model containing multiple types of operational constraints specifically includes: To minimize the overall cost based on conventional unit power generation cost, energy storage system charging and discharging cost, renewable energy curtailment cost, and load reduction cost within the control cycle, an objective function is established for the congestion control optimization benchmark model. The various types of operational constraints in the congestion control optimization benchmark model include at least AC power flow constraints, generator operation constraints, operation safety constraints, energy storage operation constraints, and new energy and load regulation constraints.

[0007] According to the present invention, a dynamic congestion active control method based on the analytical embedding of a classification neural network is provided, wherein the nonlinear constraints include the quadratic term of the conventional unit power generation cost in the objective function, the apparent power constraint of the branch in the AC power flow constraint and the operational safety constraint.

[0008] According to the present invention, a dynamic congestion active control method based on a classification neural network with embedded parsing is provided. The step of constructing a congestion safety classification neural network surrogate model based on historical power grid operating data and corresponding congestion safety status determination results specifically includes: Historical power grid operation data is acquired to extract the operation status characteristics at each historical moment. The operation status characteristics include at least node injected active power, node injected reactive power, and the square of generator node voltage amplitude. For each historical moment, the operation status characteristics are used to determine the blocking safety and generate the corresponding status label; Using the operating state features as input and the state labels as output, a feedforward classification neural network is constructed and trained to obtain a congestion safety classification neural network surrogate model that characterizes the mapping relationship between the power grid operating state and the congestion safety state.

[0009] According to the present invention, a dynamic blocking active control method based on the parsing embedding of a classification neural network is provided. In the process of training the feedforward classification neural network, an L1 regularization term is added to the loss function to promote the sparsity of the network weight matrix.

[0010] According to the present invention, a dynamic blocking active control method based on the parsing embedding of a classification neural network is provided, wherein the step of performing mixed-integer linear reconstruction on the blocking safety classification neural network proxy model specifically includes: For the ReLU activation function in the feedforward classification neural network, 0 / 1 auxiliary variables are introduced, and the Big-M strategy is used to transform the nonlinear logic of the ReLU activation function into linearity; A boundary contraction strategy is adopted to calculate the upper and lower boundaries of the variables before activation of each hidden layer neuron in the feedforward classification neural network and implement boundary contraction to narrow the range of Big-M values. For neurons whose activation state can be determined according to the upper and lower boundaries, their activation state is fixed and the corresponding 0 / 1 auxiliary variables are omitted.

[0011] According to the present invention, a dynamic blocking active control method based on the parsing embedding of a classification neural network is provided, wherein the expression of the blocking safety classification agent constraint is: ; In the formula, For random scene indexing, For scheduling period index, For hidden layer index, This represents the total number of hidden layers. In the scene Scheduling period The input layer below receives information. In the scene Scheduling period The input feature vector, In the scene Scheduling period The next The input vector before activation of the hidden layer. In the scene Scheduling period The next The output vector after activation of the hidden layer. Here is the weight matrix of the FNN model. For bias vectors, , These are the upper and lower boundary vectors, respectively. In the scene Scheduling period Hidden layer below The corresponding 0 / 1 auxiliary variable vector, In the scene Scheduling period The following network logarithm In the scene Scheduling period The number of network segments corresponding to the normal class under the following conditions. In the scene Scheduling period The number of network pairs corresponding to the blocking class below. , These represent the weight matrix and bias vector of the output layer, respectively. For category tags.

[0012] According to the present invention, a dynamic blocking active control method based on classification neural network parsing embedded dynamic blocking active control is provided, wherein after outputting a multi-period active control strategy, a verification operation is performed on the multi-period active control strategy; The verification operations include: performing AC power flow verification, N-1 static security verification, and transient power angle stability verification on the multi-time period active control strategy.

[0013] This invention also provides a dynamic blocking active control system based on a classification neural network with embedded analytical methods, comprising: The benchmark model establishment and linearization module is used to establish a blocking control optimization benchmark model containing multiple types of operational constraints, and to linearize the nonlinear constraints in the blocking control optimization benchmark model. The data-driven proxy model building module is used to build a congestion safety classification neural network proxy model based on historical power grid operation data and corresponding congestion safety status judgment results. The mixed-integer linear reconstruction module is used to perform mixed-integer linear reconstruction on the blocking security classification neural network agent model to determine the blocking security classification agent constraints. An embedded solution and strategy output module is used to embed the blocking safety classification agent constraint into the linearized blocking control optimization benchmark model to construct and solve a mixed integer linear programming model, which is used to output a multi-period active control strategy.

[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the dynamic blocking active control method based on the parsing of a classification neural network as described above.

[0015] This invention provides a dynamic congestion active control method based on analytical embedding of a classification neural network. It establishes a congestion regulation optimization benchmark model containing multiple types of operational constraints and linearizes the nonlinear constraints in the model, thus providing a foundation for efficiently solving the originally difficult nonlinear programming problem. Based on this, and using historical power grid operation data and corresponding congestion safety status determination results, a congestion safety classification neural network surrogate model is constructed. This transforms the congestion safety determination rules, which originally relied on time-domain simulation and were not analytically expressible, into a computable data-driven surrogate model. Through mixed-integer linear reconstruction, the surrogate model is transformed into a congestion safety classification surrogate constraint that can be embedded into the benchmark model, enabling the piecewise nonlinear logic of the neural network to be recognized and processed by a mixed-integer linear programming solver. Finally, the congestion safety classification surrogate model... By embedding the constraints into the linearized baseline model, a mixed-integer linear programming model is constructed and solved uniformly, outputting multi-period active control strategies. Thus, this application utilizes a classification neural network to learn the nonlinear mapping relationship between the power grid operating state and the congestion safety state, transforming the congestion safety determination process, which originally relied on complex safety verification and time-domain simulation, into a rapidly computeable classification proxy model. Furthermore, through mixed-integer linear reconstruction, the neural network classification boundary is analytically embedded into the baseline model, achieving unified modeling and rapid solution of congestion safety constraints and multi-period active control optimization. This application can initiate strategy formulation before the actual occurrence of congestion, reserving action time for regulation resources affected by ramp-up constraints or response delays. While avoiding repeated online safety verification and time-domain simulation, it improves the initiative, real-time performance, and solution efficiency of congestion prevention and control. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating the dynamic blocking active control method based on classification neural network parsing embedded information provided by the present invention. Figure 2 This is a schematic diagram of the feedforward classification neural network structure of the dynamic blocking active control method based on the analysis of the classification neural network provided by the present invention; Figure 3 This is the wiring diagram of the IEEE 39-node test system based on the dynamic blocking active control method embedded in the classification neural network analysis provided by the present invention. Figure 4This is a comparison of the transmission power control effects of different blocking control models (with / without early warning assistance) of the dynamic blocking active control method based on the classification neural network analysis embedded in the present invention. Among them, (a) is the blocking active control with early warning assistance, and (b) is the blocking control without early warning assistance. Figure 5 This is a comparison chart of the transmission power regulation effect of tie line 16-15 under different blocking safety constraints provided by the present invention, based on the dynamic blocking active control method embedded in the classification neural network. Among them, (a), (b), (c), and (d) are the regulation effect, blocking limit and transmission power difference obtained by four different constraints corresponding to models A, B, C, and D, respectively. Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0018] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and should not be construed as limiting the scope of the invention.

[0019] In the description of this invention, it should be noted that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0020] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "connected" and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0021] See Figure 1 This application provides a dynamic blocking active control method based on classification neural network parsing embedded data, including the following steps: S110: Establish a benchmark model for blockage control optimization that includes multiple types of operational constraints, and linearize the nonlinear constraints in the benchmark model for blockage control optimization. S120: Construct a congestion safety classification neural network proxy model based on historical power grid operation data and corresponding congestion safety status determination results; S130: Perform mixed-integer linear reconstruction on the blocking security classification neural network agent model to determine the blocking security classification agent constraints; S140: A baseline model for blocking control optimization after linearization is embedded into the agent constraint of the blocking safety classification to construct and solve a mixed-integer linear programming model for outputting multi-period active control strategies.

[0022] This application establishes a congestion control optimization benchmark model containing multiple types of operational constraints and linearizes the nonlinear constraints in the model, thus providing a foundation for efficiently solving the originally difficult nonlinear programming problem. Based on this, and using historical power grid operation data and corresponding congestion safety status determination results, a congestion safety classification neural network surrogate model is constructed. This transforms the congestion safety determination rules, which originally relied on time-domain simulation and were not analytically expressible, into a computable, data-driven surrogate model. Through mixed-integer linear reconstruction, the surrogate model is transformed into congestion safety classification surrogate constraints that can be embedded into the benchmark model, enabling the piecewise nonlinear logic of the neural network to be recognized and processed by the mixed-integer linear programming solver. Finally, the congestion safety classification surrogate constraints are embedded into the linearized benchmark model. In this application, a mixed-integer linear programming model is constructed and solved in a unified manner to output a multi-period active control strategy. Thus, this application utilizes a classification neural network to learn the nonlinear mapping relationship between the power grid operating state and the congestion safety state, transforming the congestion safety determination process, which originally relied on complex safety verification and time-domain simulation, into a rapidly computable classification proxy model. Furthermore, the classification boundary of the neural network is analytically embedded into the benchmark model through mixed-integer linear reconstruction, achieving unified modeling and rapid solution of congestion safety constraints and multi-period active control optimization. This application can initiate strategy formulation before the actual occurrence of congestion, reserving action time for regulation resources affected by ramp-up constraints or response delays. While avoiding repeated online safety verification and time-domain simulation, it improves the initiative, real-time performance, and solution efficiency of congestion prevention and control.

[0023] In S110, a stochastic optimization baseline model for proactive congestion control in multiple time periods is established to address potential transmission congestion issues in inter-regional tie lines during power grid operation. This model is used to respond to proactive congestion control requirements and outputs time-series control strategies for multiple scheduling periods within a given control time window to eliminate system transmission congestion in advance.

[0024] The model needs to incorporate congestion safety constraints that consider the ultimate transmission capacity to ensure the congestion safety of the output strategy. Since event-driven proactive grid congestion control operates on a near real-time timescale, the available renewable energy sources and ultra-short-term load forecasts are relatively accurate. Therefore, this invention uses a stochastic scenario method to account for the uncertainty in renewable energy power prediction errors and assumes the accuracy of ultra-short-term load forecasts. Using the stochastic scenario method for modeling preserves the high flexibility of the optimization model, allowing for the selection of appropriate scenario generation methods based on different system characteristics and specific requirements.

[0025] The regulation resources considered in the congestion active control optimization model of this invention include: conventional unit power generation, renewable energy curtailment, and load shedding. Furthermore, in accordance with policy guidelines for strengthening grid peak-shaving energy storage and intelligent dispatch capabilities, and considering future grid-side energy storage development trends, a small number of Energy Storage Systems (ESSs) regulation resources are considered in the model, thereby enriching the resource optimization algorithm interface. Based on the above regulation resources, the congestion control optimization baseline model aims to minimize the overall cost of unit power generation and resource regulation within the control cycle, as shown below: (1); In the formula, For random scene indexing, For a set of random scenarios of new energy, For the scene The probability of occurrence; For scheduling period index, For the set of scheduling periods; For generator index, For generator sets, Indicates generator During the period Those who have made contributions , , Indicates the fuel cost coefficient; Indexing ESSs For ESSs resource collection, , These are the charging and discharging power, respectively. Adjust the cost factor for ESSs; For wind and solar new energy index, For new energy collection, Indicates in the scene Time period New energy Wasted electricity This represents the cost coefficient for abandoned electricity; For indexes that can reduce load, Indicates the set of loads that can be reduced. This indicates the amount of active power reduction in the load. This is the load adjustment cost coefficient.

[0026] for , The blocking control optimization model includes the following constraints: 1) AC power flow constraints, including nodal active power balance equations (2), nodal reactive power balance equations (3), power flow calculation formulas (4) and (5): (2); (3); (4); (5); (6); (7); In the formula, For node indexing, Represents the set of system nodes. Indicates time period node Active power at the load, Indicates the reactive power of the load. Represents a node Injected active power, To inject reactive power; Represents new energy In the scene Time period Those who have made contributions Indicates generator Effortless effort Indicates that the load can be reduced. Reactive power reduction. Representing resources With nodes The connection relationship, When resources With nodes When connected, =1, otherwise =0. Indicates a branch Active transmission power, Indicates reactive power transmission. , Representing nodes respectively , Voltage amplitude, , Indicates the voltage phase angle. Indicates a branch Electrical conductivity Indicates a branch Saturation This represents the set of system branches. , Represent the real and imaginary elements of the nodal admittance matrix, for ,have , .

[0027] 2) Generator operating constraints, including active power output upper and lower limit constraints (8), reactive power output upper and lower limit constraints (9), and active power output ramping constraints (10): (8); (9); (10); In the formula, , They represent generators There are upper and lower limits to the amount of work done. , This indicates the upper and lower limits of reactive power output. Indicates generator The maximum amount of effort and contribution required to climb the hill.

[0028] 3) Operational safety constraints, including branch power flow constraints (11), node voltage magnitude constraints (12), and slack node phase angle constraints (13): (11); (12); (13); In the formula, Indicates a branch Maximum permissible apparent power Represents the set of system branches. , Representing nodes respectively The upper and lower limits of the voltage.

[0029] 4) Energy storage operation constraints (ESSs regulation constraints), including inter-state coupling constraints (14), upper and lower limits constraints (15), final state constraints (16), upper and lower limits constraints (17) and (18) for charge / discharge, and state of charge / discharge constraints (19): (14); (15); (16); (17); (18); (19); In the formula, Indicates the energy state of ESSs. , These represent the charge / discharge efficiency, respectively. , Indicates energy storage Upper and lower limits of energy states , This represents the upper and lower limits of the energy state at the end of the scheduling cycle, which are used to reserve upward / downward adjustment capability for the next scheduling cycle. , This indicates the charging / discharging power limit.

[0030] Equation (19) is an accurate approximation of the 0 / 1 constraint of the charging / discharging state. It can avoid introducing 0 / 1 variables while accurately simulating the charging / discharging behavior of energy storage. This effect comes from the fact that the charging and discharging states have corresponding cost terms in the objective function, thus preventing the energy storage system from charging and discharging at the optimal solution.

[0031] 5) Constraints on new energy sources and load regulation, including load reduction constraints (20) and new energy curtailment constraints (21): (20); (twenty one); In the formula, Indicates time period node Active power at the load, Representing resources With nodes The connection relationship, This indicates the amount of active power reduction in the load. Indicates in the scene Time period New energy Wasted electricity Represents new energy In the scene Time period Those who have made contributions.

[0032] 6) Blocking safety constraints in every new energy scenario In addition, for each scheduling period t, a tie-line blocking safety constraint (22) needs to be set to control the blocking risk of the output strategy: (twenty two); In the formula, This represents the set of inter-regional connecting lines. Indicates the contact line Active transmission power on Indicates the running scenario Maximum transmission capacity under certain conditions. System operating scenarios. By variables , , , as well as The invention determines the blocking state based on 95% of the maximum transmission capacity. =95%. Indicates the target operating scenario The blocking determination result, This indicates that the system is functioning normally and the blocking safety constraints are met. This indicates that there is a blockage in the system. This is a dynamic value that changes depending on the different system operating scenarios.

[0033] It should be noted that the congestion safety constraint is used to characterize whether the system operating point meets the safety requirements such as N-1 static safety and transient power angle stability. Its determination process involves high-dimensional differential algebraic equation constraints, which are non-analyzable complex constraints in the benchmark model. Equation (22) is the original expression of the congestion safety constraint, which is used to limit the tie line power within the dynamic limit. This application replaces the original determination function of Equation (22) by constructing and reconstructing a congestion safety classification neural network surrogate model.

[0034] Equations (1) to (22) constitute the baseline model for active stochastic optimization of grid congestion in multiple time periods, considering AC power flow. Among them, AC power flow constraints (4) and (5) and branch apparent power constraints (11) introduce strong nonlinearity into the model, making it a difficult-to-solve non-convex nonlinear problem. Furthermore, the introduced multi-time period congestion safety constraint (22) needs to be based on the limit transmission capacity. For input scenarios The congestion safety status is determined. However, since this determination process involves the verification of multiple safety items (such as N-1 static safety and transient power angle stability), a large number of simulation calculations and time-domain simulations are required, making the optimization problems (1)~(22) impossible to solve analytically and quickly. In order to meet the timeliness requirements of online active control, this invention first linearizes the AC power flow constraints and the apparent power constraints of the branches, then constructs a congestion safety constraint proxy model, and performs mixed integer reconstruction modeling on it, thereby transforming the original problems (1)~(22) into MILP problems that can be solved quickly, so as to support online rapid decision-making.

[0035] In this step, it is also necessary to linearize the nonlinear constraints in the congestion control optimization benchmark model. The nonlinear elements in the congestion control optimization benchmark models (1) to (22) include: ① the quadratic term of power generation fuel cost in the objective function; ② AC power flow constraints (4) and (5); ③ branch apparent power constraints (11).

[0036] Specifically, the system state variables are transformed from node voltage magnitude and node voltage phase angle to a variable representation space composed of node voltage phase angle and the square of node voltage magnitude. A first-order Taylor expansion is performed on the active and reactive power loss terms of the branches to obtain a linear approximation of the AC power flow constraints. At the same time, a circular inscribed piecewise linearization method is used for the apparent power constraints of the branches, and a piecewise linearization method is used for the quadratic term of the power generation fuel cost in the objective function, so that the part other than the congestion safety constraint is transformed into an analytically solvable linear optimization form.

[0037] ① Among them, piecewise linearization can be used to handle this type of nonlinearity. Some commercial solvers, such as Gurobi, also have the ability to automatically handle this type of nonlinearity.

[0038] The following section details the linearization method for AC power flow constraints: To better capture the nonlinearity in constraints (4) and (5), the state variables can be... Representation to variable space ,Will and The linear transformation is as follows: (twenty three); (twenty four); In the formula, , They represent the linearized branches respectively. The active and reactive power transmission power, , These represent active and reactive power losses, respectively, in the variable space. The linear approximation method within is shown below: (25); (26); In the formula, It is decomposed into two terms: voltage phase angle term. With voltage amplitude term Similarly, It is also decomposed into two terms: voltage phase angle term. With voltage amplitude term ,right as well as Perform a first-order Taylor expansion: (27); (28); In the formula, , The initial voltage phase angle, used to define the system state, can be obtained through AC power flow calculations. For and The following methods can be used to approximate the result: (29); (30); In the formula, , The initial voltage amplitude value for defining the system state can be obtained through AC power flow calculation.

[0039] At this point, the original nonlinear constraints (4) and (5) have been established in the variable space. The inner part is obtained as a completely linearized approximation.

[0040] Accordingly, the voltage upper and lower limit constraints (12) need to be transformed, as shown below: (31); For the apparent power constraint (11) of branch ③, which is a nonlinear element, it has a quadratic function form and defines a convex region. It can be processed using the circular inscribed piecewise linearization method to form a set of linear constraints to approximate its expression. Based on the linearized branch... Active and reactive power transmission , Constraint (11) can be transformed into the following set of linear approximation constraints: (32); After the above linearization process, the original active control optimization model for blocking is transformed into the following problem: Objective function: Equation (1); Constraints: Equations (2)~(5), (8)~(10), (13)~(24), (31) and (32).

[0041] At this point, except for the blocking safety constraint (22), the rest of the model has been completely parsed. The following introduces the blocking safety constraint embedding fusion method based on data-driven agent modeling, which transforms the model into a mixed integer programming problem.

[0042] In S120, due to the non-analyzability of the blocking safety constraint (22), the problem requires iterative solution, which is difficult to meet the timeliness requirements of online active control. At the same time, due to the high-dimensionality of the target problem, existing non-gradient-based solution algorithms are not applicable. Therefore, this invention first regards the blocking safety constraint (22) as a classification constraint (i.e., the system operating state is constrained to be in the "normal" category), and proposes to use a data-driven method to learn the blocking safety classification rule offline, construct the mapping relationship between the system operating point and whether the system is blocked or not, and form a blocking safety classification neural network proxy model. Subsequently, this data-driven blocking safety classification rule replaces the constraint (22), which can avoid the need for repeated simulation calculations and online time-domain simulation, form a nonlinear approximation of the original problem, and improve its solution efficiency. Furthermore, the neural network blocking safety classification constraint is reconstructed by mixed integer linearization, thereby analytically encoding the blocking safety classification boundary and realizing the mixed integer linear transformation of the original problem.

[0043] Given the accurate fitting ability of neural network models in complex nonlinear prediction tasks, this application uses a feedforward neural network (FNN) model to learn the blocking safety classification rules.

[0044] In this embodiment, historical multivariate time-series operational data of the power grid are first acquired, and a corresponding system operation scenario is constructed for each historical moment. Then, according to a preset congestion state determination method, the congestion safety state of the system operation scenario at each historical moment is determined, and a corresponding state label is assigned. The state labels include two categories: normal state and congestion state. A normal state indicates that the system meets the congestion safety requirements under this operational scenario, while a congestion state indicates that the system has a transmission congestion risk or does not meet the congestion safety requirements under this operational scenario. The congestion safety state labels of each historical moment are arranged in chronological order to form a congestion event state sequence. Then, the system operation state data and the corresponding congestion safety state label for each historical moment are collected to construct a supervised learning sample set for training a feedforward neural network congestion safety proxy model.

[0045] Specifically, in this invention, the FNN blocking safety classification model uses... For input, , This represents the active and reactive power vectors injected into the node. This represents the squared vector of generator node voltage magnitudes, facilitating the embedding of the subsequent FNN blocking security agent model. Therefore, for time... The model input can be denoted as The corresponding tag can be recorded as This allows the formation of a training dataset. , Indicates the number of paired samples. Based on An FNN blocking security agent model can be trained offline, denoted as... Assume the FNN model used has The system consists of one fully connected hidden layer, one input layer, and one output layer. The first... Each fully connected hidden layer contains One neuron, , its As input, with For the output, the FNN blocking safe classification surrogate model using the ReLU activation function can be represented by the following formula: (33); (34); (35); (36); (37); In the formula, This indicates that the input layer receives information. , These represent the weight matrix and bias vector of the FNN model, respectively. The ReLU activation function acts on every neuron in every hidden layer, i.e. Operation applied to vector Each element in. , These represent the weight matrix and bias vector of the output layer, respectively. Logarithmic division of the network is performed using the softmax function. Activation, thus ultimately forming the blocking safety classification model. , For category estimation. In this invention, This is used to determine whether the system's operating state is blocked or not, and belongs to a binary classification model. The value {0, 1} represents a system that is functioning normally and where the blocking safety constraints are met, with 0 indicating that the system is blocked. Therefore, , This represents the number of network segments corresponding to the normal class. This represents the number of network pairs corresponding to the blocking class.

[0046] like Figure 2 As shown, a typical 3-layer FNN model is presented. By minimizing the training dataset The cross-entropy loss is used for training. Based on the trained FNN blocking security classification agent model, the original blocking security constraint (22) can be replaced by the following data-driven security classification constraint (38): (38); In the formula, Indicates the system operation scenario The model input variables extracted from them. ,because Constraint (38) is equivalent to the following equation: (39); In the formula, , , Representing the scene respectively The network logarithms for the normal (safe) and blocked (insecure) classes, calculated by the congestion-safety classification proxy model at time period t. Due to the FNN model... The weight matrix and bias vector (e.g.) , After offline training, all variables have been fixed. During the optimization phase, they are adjusted by inputting variables. The corresponding network logarithm can be calculated using equations (33) to (36). and Therefore, replacing the original blocking safety constraint (22) with equation (39) can avoid the need for repeated simulation calculations and online transient time-domain simulations. At this point, the original problem has been transformed into a nonlinear programming problem.

[0047] In S130, the constraints of the blocking safety classification neural network are reconstructed using mixed-integer linear programming. For the ReLU activation function in the feedforward classification neural network, a boundary contraction strategy is adopted. The upper and lower boundaries of the variables before activation of each hidden layer neuron are calculated layer by layer, and boundary contraction is implemented. When the upper and lower boundaries can determine the neuron's activation state, the corresponding state is fixed to eliminate unnecessary binary variables, thereby narrowing the optimization range of the mixed-integer linear programming model and improving the model's solution efficiency.

[0048] By using data-driven proxy modeling based on blocking security classification rules, the problem can be rewritten as: Objective function: Equation (1); Constraints: Equations (2)~(5), (8)~(10), (13)~(21), (23), (24), (31)~(36) and (39).

[0049] Based on an offline-trained FNN blocking safety classification proxy model This model can effectively accommodate congestion safety constraints that take into account the ultimate transmission capacity, without requiring complex online differential-algebraic equation derivations to achieve dynamic safety verification. However, the process performed in equation (35) Operational limitations still hinder the complete analytical solution of the model. To enable the blocking active regulation optimization model to be solved analytically quickly, thus meeting the timeliness requirements of online regulation, this section uses the Big-M strategy to reconstruct the blocking safety classification neural network using mixed-integer linearization, thereby transforming the model into MILP form. Specifically, this is achieved by introducing 0 / 1 auxiliary variables for each hidden layer. Constraint (35) can be restated as: (40); (41); (42); (43); (44); In the formula, For hidden layer The corresponding 0 / 1 auxiliary variable vector has a number of 0 / 1 variables equal to the number of neurons in that hidden layer. This is achieved by introducing auxiliary variables. If the input of the ReLU activation function If the corresponding binary variable is 0, then constraints (42) and (43) will limit the neuron's output. It is 0. Conversely, if the input to the ReLU activation function is 0... If the corresponding binary variable is 1, then constraints (40) and (41) will limit the neuron's output. equal .

[0050] Upper and lower boundary values and The boundary value needs to be sufficiently large to avoid limiting the possible values, while also needing to be sufficiently small to improve the efficiency of the optimization solution. This invention uses an interval arithmetic method to determine the appropriate boundary. The hidden layer can be calculated using the following methods: (45); (46); In the formula, This represents the index of a neuron in the hidden layer. For Hidden layers can be created by removing the inner layers in equations (45) and (46). The function is used for calculation. Therefore, it can be derived from... Initially, calculations are performed by progressively propagating forward based on equations (45) and (46) to obtain the compacted model boundary of each hidden layer of the FNN.

[0051] Furthermore, since the aforementioned mixed-integer reconstruction process introduces 0 / 1 auxiliary variables, and their number is the same as the number of neurons in the hidden layer of the feedforward neural network, there is a trade-off between the inference accuracy of the FNN model and the computational efficiency of mixed-integer programming. Therefore, in the training process of the FNN blocking safe classification surrogate model, L1 regularization can be used to improve the efficiency of the weight matrix. The sparsity of the model reduces overfitting and improves the solution efficiency of MILP problems. (47); In the formula, For the cross-entropy loss on the training dataset, the parameters Used to control the strength of regularization This represents the L1 norm.

[0052] In S140, the congestion safety classification neural network constraints after mixed-integer linearization are embedded into the linearized congestion control optimization benchmark model to form a mixed-integer linear programming model that integrates neural network analytical constraints. The model is then solved using an optimization solver, and outputs active control strategies for multiple time periods, such as conventional unit output, energy storage charging and discharging, renewable energy curtailment, and load reduction.

[0053] By reconstructing the problem using a neural network with mixed-integer linearization, and replacing constraint (33) with equations (38) to (42), the problem can be further transformed into the following mixed-integer linear programming problem: Objective function: Equation (1) Constraints: ; The aforementioned MILP problem can be solved quickly using existing mature commercial solvers. Furthermore, initial values ​​are required for the AC power flow linearization terms (23) and (24). as well as To improve the accuracy of the model solution and reduce the linearization error of network loss, an iterative solution strategy can be used. Specifically, in the number of iterations... Initial values ​​can be formed using the solution of the optimal AC power flow. as well as The optimization problem is then solved, and the AC power flow calculation program is executed based on the optimization solution to obtain the system network loss value, as well as the corresponding system voltage amplitude and voltage phase angle vector, forming the initial values ​​for the next iteration. as well as Then, the optimization problem is solved again. The iterative calculation continues until the difference in system network loss between two adjacent iterations is less than a pre-set threshold. At this point, the iterative scheme converges, and the optimization decision result of the current iteration is output.

[0054] Optionally, after outputting the multi-period active control strategy, verification operations can be performed on the multi-period active control strategy. These verification operations include: AC power flow verification, N-1 static security verification, and transient power angle stability verification. This verification operation is performed independently after the strategy is output and can serve as a reliability confirmation step before the strategy is issued, verifying the security of the control strategy. This verification operation can be selected to be performed or not performed according to actual application requirements; this application does not impose specific restrictions.

[0055] The multi-period active control strategy in this application is a coordinated action plan formulated for various regulating resources such as conventional generating units, energy storage systems, new energy power plants, and loads that can be reduced in each scheduling period within the regulation cycle. For example, it includes the active power output plan of conventional generating units, the charging and discharging power timing arrangement of energy storage systems, the amount of curtailment reduction of new energy power plants, and the allocation of load reduction. Under the premise of satisfying congestion safety agent constraints and various physical operation constraints, this strategy is obtained by solving a mixed integer linear programming model. The entire strategy is formulated before the actual occurrence of congestion, reserving sufficient action time for regulating resources with ramp-up constraints and response delays. It has clear preventive, coordinated, and executable characteristics and can be directly issued to the scheduling operation platform for execution.

[0056] The following example demonstrates the proposed proactive grid congestion control strategy that considers data-driven classification constraints: First, Section a compares and analyzes the control effects of having and not having a congestion risk early warning procedure to test the role of quantified congestion risk perception in improving the proactiveness of system control. Then, Section b verifies the effectiveness of the proposed congestion safety constraint proxy modeling method, analyzing the effect of model data fusion-driven proactive congestion control. Finally, Section c analyzes and compares the computational efficiency of model solving to verify the online timeliness of the proposed method. This embodiment is mainly based on the IEEE 39-node test system, and the system topology and resource distribution are as follows: Figure 3 As shown, the congestion warning object and the congestion safety constraint object of the model are still the transmission sections between region 1 and region 2, including tie lines 1-39, 2-3, 18-3, and 16-15. To enhance the system's flexible adjustment capability, this embodiment adds two ESSs adjustment resources at bus lines 8 and 12.

[0057] Analysis of the effect of proactive regulation based on obstruction awareness Based on the preset congestion proactive control triggering mechanism, this section first tests the auxiliary role of the congestion early warning program in system congestion control, verifying the effectiveness of the proactive control strategy. In this embodiment, the proactive control program can be triggered by a pre-built congestion risk early warning model. When the congestion risk early warning model determines that there is a potential congestion risk in the future operation state of the power grid, the subsequent congestion proactive control optimization process is initiated. The feedforward neural network congestion safety agent model embedded in the proactive control optimization model is trained using system operating condition data and the corresponding congestion event state sequence. Specifically, firstly, system operating scenario data covering a certain operating cycle is acquired as the original operating data before control; then, according to the preset congestion state determination method, the system congestion state at each time segment is determined, forming a binary congestion event state sequence composed of normal state and congestion state. For any time segment, the data vector that can characterize the system operating state at that time segment is extracted, and its corresponding congestion state determination result is used as a supervision label, thereby forming a supervised format dataset, which is used to construct a binary classification feedforward neural network congestion safety agent model. The system's operational scenario data covers a one-year operational cycle, comprising a total of 34,080 time segments. Data from a subset of months was selected for training and validation, while data from the remaining months was used for testing, to verify the classification accuracy and generalization ability of the constructed blocking security agent model.

[0058] In this embodiment, the time window length (number of scheduling periods) of the active control program is set to... =16, which is 4 hours, scheduling interval =15 minutes. Since event-driven grid congestion control is a near real-time timescale, the available ultra-short-term forecast information for renewable energy is relatively accurate. Therefore, the number of renewable energy scenarios considered in the stochastic optimization program is set to 5. Specifically, this embodiment uses a data-driven method for scenario generation. Through learning from historical renewable energy data using a generative adversarial network (GAN), it can quickly generate wind power and photovoltaic fluctuation scenarios within the scheduling window, taking into account the correlation of renewable energy output. The GAN is implemented in a Python environment based on the Tensorflow library. For example... Figure 3As shown, the congestion-prone active control resources include 10 conventional generators, 4 wind / solar power plants, and 2 energy storage systems. Furthermore, all loads are set as loads that can be reduced. The unit adjustment cost for all ESSs is set at $8 / MWh, the unit adjustment cost for renewable energy curtailment is set at $30 / MWh, and the unit adjustment cost for load reduction is set at $300 / MWh. The rated capacity of the ESSs is set at 240MWh, the upper / lower limits of energy state are 240MWh and 40MWh respectively, the maximum charge / discharge power is set at 50MW, the charge / discharge efficiency is set at 0.95, and the upper / lower limits of final energy state are 192 MWh and 120MWh respectively, reserving upward / downward adjustment capabilities for the next scheduling cycle.

[0059] First, the blocking control effect was compared with and without blocking warning assistance to test the role of blocking risk quantification perception. Both cases embedded an FNN blocking safety classification proxy model to ensure that the optimized output control strategy met the blocking safety criteria. The blocking safety classification FNN used contained three hidden layers, each with 48 neurons. The FNN blocking safety classification proxy model was trained in a Python environment using the Keras library, with the following settings: maximum number of iterations: 200, batch size: 20, loss function: cross-entropy loss, optimizer: Adam. Furthermore, to improve the efficiency of solving subsequent MILP problems, L1 regularization was used during training to improve the weight matrix of the FNN proxy model. The sparsity is achieved by setting the regularization parameter to 0.01.

[0060] The baseline power of the experimental system was set to 100 MVA, and the initial energy state of both ESSs (Energy State of Energy) was set to 220 MWh. Except for the start-up time of regulation, all other model settings were the same in the two comparative cases. This experiment solved the blocking regulation optimization model (MILP property) with embedded FNN classification surrogate constraints using the Gurobi solver in a Python environment. The regulation results corresponding to the new energy prediction scenario with the smallest distance from the actual wind and solar power output MDTW are presented. Figure 4 The changes in transmission power on tie line 16-15 before and after regulation were plotted in two cases.

[0061] like Figure 4As shown, after congestion control, the active power transmission on tie line 16-15 was effectively reduced in both comparative scenarios (with / without congestion risk awareness), and the tie line power flow corresponding to the output strategy was below the congestion limit (95% of the maximum transmission capacity), passing the preset congestion safety check. This demonstrates the effectiveness of the proposed control optimization method with embedded congestion safety neural network proxy constraints. Specifically, based on the quantified congestion risk awareness results output by the calibrated MA-LSTMT model, congestion control can initiate strategy formulation at 16:45 to proactively eliminate congestion risk and smoothly control the inter-region tie line power below the congestion limit. Figure 4 As shown in (a). It should be noted that although the power flow of the controlled tie-line is close to the congestion limit for most of the time, since the congestion limit is defined as 95% of the maximum transmission capacity in this paper, a 5% safety margin is still retained, and the system is in a safe state. In contrast, without the auxiliary early warning of the congestion risk quantification perception results, the relevant scheduling and operation personnel can only start policy formulation after observing the congestion event. Even without considering the time for online analysis of the system congestion status (accurate congestion analysis requires a large number of simulation calculations and time-domain simulations), the corresponding control policy can only be formulated at 17:15, which is 30 minutes later than the preventive proactive control. In this case, if Figure 4 As shown in (b), the system will be in a blocked state for at least one period of time (time step 16), with a significantly insufficient safety margin.

[0062] Meanwhile, without the assistance of congestion risk awareness and early warning information, the system is already exposed to security risks when a congestion event is observed online, leaving very limited time for strategy formulation and implementation. At this point, due to response delays or ramp-up constraints on relevant adjustment resources, the system is likely to suffer from temporary insufficient adjustment capacity. The total operating cost of the proactive control program with congestion risk awareness assistance in this scenario is $186,480.86, while the operating cost of passive waiting control, which performed load reduction actions in time period 17, is $188,650.77, a difference of $2,169.91. In summary, without the assistance of congestion risk awareness programs, the urgency and difficulty of online congestion control will increase significantly, and the system may face short-term insufficient adjustment (ramp-up) capacity, requiring more costly control actions to protect system security. In contrast, with the assistance of congestion risk awareness, scheduling operators can perform preventative proactive control to trigger adjustment resources with ramp-up constraints or response delays in advance, thereby gaining more adjustment time, better and smoother protection of system security, and reducing control costs. The above results verify the effectiveness of the proposed proactive grid control strategy based on congestion risk perception.

[0063] b. Analysis of the regulatory effect of fusing FNN classification constraints This section verifies the effectiveness of the proposed FNN congestion safety classification proxy modeling method in ensuring that the output strategy meets the congestion safety criterion. It primarily tests the safety verification of the output control strategy under different control criteria. The experiments mainly compare the following cases: ① Model A: Embedded FNN congestion safety classification constraints, hidden layer structure {64, 64, 64}; ② Model B: Embedded FNN congestion safety classification constraints, hidden layer structure {64}; ③ Model C: FNN constraints removed, replaced by 80% of the tie-line thermal rating; ④ Model D: FNN constraints removed, replaced by 77.5% of the tie-line thermal rating. The latter two cases are set based on commonly used empirical methods in current engineering practice. The FNN proxy model training settings are the same as in the previous section. This embodiment first tests the impact of the above four congestion safety constraints on the control effect of specific event cases, using the same congestion event cases as in section a, as well as a random scenario for new energy prediction. Based on the quantitative perception results of congestion risk, preventative proactive control was implemented in all four comparative scenarios. Using the system state at time step 14 as the initial state, system control strategies were formulated for time steps 15 to 30 (a total of 16 time segments). The initial state of system adjustment resources, optimized model parameters, and solution environment were the same as in the previous section. Based on the control and optimization results, in... Figure 5 The changes in transmission power on tie line 16-15 before and after regulation under four different models are illustrated.

[0064] like Figure 5 As shown, models A through D exhibit different transmission power regulation effects based on different congestion safety constraints. Among them, comparing model A and model B, model A's output regulation strategy tracks the congestion safety boundary better and effectively controls the tie-line transmission power to always remain within the congestion limit, such as... Figure 5 As shown in (a); however, the tie-line transmission power under Model B control exceeded the blocking limit during certain periods, such as... Figure 5 (b) As shown on the right. This is mainly because Model A embeds a more accurate blocking security proxy model with three hidden layers. Therefore, the deviation between the security proxy constraints it forms and the actual blocking security constraints is smaller, which helps to output a more accurate control strategy. It can be seen that the accuracy of the proxy model is of great significance to the accuracy of the output strategy.

[0065] Comparing Model A and Model C, the latter uses a relatively static power limit (80% thermal rating) to constrain the power flow of the tie lines, thus replacing the FNN surrogate constraint. Using a static power limit is also a common way to simplify congestion constraints in actual scheduling operations, helping to accelerate decision analysis efficiency. In this case, since the FNN surrogate constraint has been removed, the optimization model no longer contains integer decision variables, and the MILP problem is transformed into a linear programming problem. However, as... Figure 5As shown in (c), the static power limit does not consider the changing system operating conditions. Under Model C, the tie-line transmission power exceeds the congestion limit (under-regulation) in many periods. In contrast, the FNN congestion safety classification proxy model embedded in Model A can generate dynamic safety constraints based on different scenario inputs, thereby dynamically tracking the system operating conditions and helping to ensure grid safety. Furthermore, if a more stringent static power limit is used to constrain the tie-line transmission power, i.e., Model D, using 77.5% of the thermal rating instead of the FNN congestion safety classification constraint, such as... Figure 5 As shown in (d), the control strategy output by the model effectively kept the tie-line power below the blocking limit. However, because the same relatively strict static power limit was used for all operating scenarios, the tie-line power under the control of Model D was more likely to reach the limit boundary, such as... Figure 5 (d) As shown on the left, this leads to over-regulation at many times. While this approach is beneficial for ensuring the system's congestion safety, overly conservative threshold settings can result in underutilization of the channel's transmission capacity, thus increasing unnecessary regulation costs. In this example, the operating cost of Model D is $189,310.57. In contrast, Model A can better utilize the available margin of the transmission channel by tracking the actual operation of the system based on the dynamic feedback of the surrogate neural network. Figure 5 As shown in (a), the difference between tie-line power and blocking limit under Model A control is smaller, which can reduce unnecessary control expenses. Its corresponding operating cost is $186,537.12, which is $2,773.45 lower than that of Model D.

[0066] The above results indicate that while static safety constraints can simplify the complexity of online analysis and decision-making, they are difficult to adapt to dynamic system operating scenarios. Relatively lenient threshold settings may lead to under-regulation, making it difficult to guarantee the satisfaction of safety criteria, while relatively strict threshold settings may lead to over-regulation, increasing unnecessary regulation costs. Based on the data-driven agent modeling method proposed in this invention, when a highly accurate congestion safety classification agent model is embedded, the resulting neural network agent constraints can account for changing operating conditions, constituting dynamic congestion safety constraints. This allows for the full utilization of the channel's transmission capacity while ensuring the output strategy meets safety criteria, thereby reducing unnecessary regulation costs.

[0067] Furthermore, a comprehensive safety verification of the control strategies output by models A through D was conducted based on multiple blocking event cases to test the average control performance of the proposed method. Specifically, 30 blocking event cases were selected in the experiment, and corresponding active control strategies were formulated using models A through D. Each case contained 16 time segments, so each model output control strategies for a total of 480 time segments. Subsequently, the control strategies were verified for AC power flow, N-1 static safety, and transient power angle stability to test the overall safety criterion satisfaction rate of the output strategies of each model. The experimental results are shown in Table 1, where the operating cost refers to 16 time segments.

[0068] Table 1. Overall security verification results of output strategies for different blocking security constraint models

[0069] As shown in Table 1, Model A, using the FNN congestion safety classification proxy constraint proposed in this invention, still achieved the highest safety verification pass rate. Its N-1 static safety and transient power angle stability verification pass rates reached 98.33% and 97.29%, respectively, verifying the effectiveness of the FNN congestion safety classification constraint in ensuring the congestion safety of the output strategy. For Model B, one case failed the AC power flow verification, mainly due to the superposition of errors in the AC power flow linearization procedure and the single-layer hidden layer FNN model. For Models C and D, their safety verification pass rates were not ideal, especially in terms of transient power angle stability. This was mainly because the congestion safety constraints they set did not incorporate transient stability considerations. In particular, Model D failed to track system changes in some periods, leading to violations of the congestion safety criterion. In other periods, it failed to fully utilize the transmission capacity of the channel, increasing unnecessary control expenses and ultimately resulting in a relatively high average operating cost.

[0070] In summary, the dynamic constraints formed by the FNN blocking safety classification agent model allow the regulation and optimization model to take into account the changing operating conditions of the system, thereby maintaining the system within a safe range at a reasonable operating cost.

[0071] c. Solution efficiency analysis of embedded FNN surrogate model Since embedding FNN blocking safety classification surrogate constraints introduces integer decision variables into the regulation optimization model, this section will analyze the solution efficiency of the proposed method. Specifically, as mentioned in the previous section, the number of 0 / 1 auxiliary variables introduced in the neural network hybrid integer linear reconstruction process is the same as the number of neurons in the network's hidden layers. If the surrogate network used has The nth hidden layer, where the nth Each hidden layer has If there are 10 neurons, then combined with the number of new energy random scenarios considered in this paper... and the number of control periods The number of integer variables introduced can be calculated as follows: If the FNN surrogate model contains a large number of neurons, it will introduce a large number of integer decision variables, affecting the model's computational efficiency. To improve the timeliness of model decisions, as mentioned earlier, L1 regularization can be used to improve the efficiency of the weight matrix. The sparsity of the structure and the auxiliary boundaries introduced during the mixed-integer linear reconstruction of neural networks. and The following section tests the effectiveness of the two acceleration methods using an FNN proxy network with three hidden layers, each containing 48 neurons. Experiments were conducted with and without L1 regularization and with and without boundary compaction to solve the active regulation optimization model, thus statistically analyzing the model's solution time. The number of uncertain new energy scenarios considered remained at 5, and the number of regulation optimization periods remained at 16. The model was solved 10 times for each case to report the average results, as shown in Table 2.

[0072] Table 2. Average solution time of the model under different acceleration settings

[0073] When the regulation optimization model does not include the FNN blocking safety classification surrogate constraint, the optimization problem transforms into a linear programming problem, with an average solution time of only 1.275 seconds. Adding the FNN blocking safety surrogate constraint significantly increases the solution time. Specifically, if L1 regularization is not added during FNN network training, solving the regulation optimization model becomes extremely time-consuming, failing to meet the timeliness requirements of online regulation. This is mainly due to the large number of 0 / 1 auxiliary variables introduced during the linear reconstruction process of the neural network. In this example, for an FNN surrogate network with the structure {48, 48, 48}, the multi-time-period stochastic regulation optimization model contains 11520 integer variables, drastically increasing the difficulty of solving the model. By using L1 regularization during FNN network training, the solution efficiency of the corresponding regulation optimization model can be significantly improved. This is mainly because L1 regularization forces some network weight parameters to be 0, thereby promoting matrix... The sparsity of L1 regularization is used to reduce the number of integer variables actually involved in decision-making in the optimization model. For example, when the L1 regularization parameter is set to 0.01, the number of integer variables that ultimately take effect is only 935, which can significantly reduce the solution time of the optimization problem. It is worth noting that increasing the strength of L1 regularization can also reduce the classification accuracy of the model, which may affect the accuracy of subsequent model decisions. Therefore, the strength of L1 regularization should not be too high and needs to be set within a reasonable range; 0.01 is a commonly used parameter setting. In addition, as shown in Table 2, for auxiliary boundaries... and Boundary compaction can also effectively improve the model's solution efficiency, reducing the average solution time by approximately 24% in this experiment. For a trained FNN network, the boundary compaction process only needs to be executed offline once, and the corresponding parameter results can be saved for subsequent optimization. Therefore, the boundary compaction process does not affect the model's online computation efficiency. Based on L1 regularization and... , The boundary contraction strategy, model number 6, has an average solution time of 35.614 seconds, which can well meet the timeliness requirements of online active control.

[0074] In summary, this invention constructs a congestion safety classification neural network proxy model based on power grid operation scenario data, and then analytically embeds it into a multi-time-period congestion active control optimization model through mixed-integer linear reconstruction. This achieves an analytical expression and rapid solution for dynamic congestion safety constraints. Using this method, considering the uncertainty of renewable energy output, AC power flow constraints, and operational constraints of various regulatory resources, it can output active control strategies that meet congestion safety requirements, including conventional unit output, energy storage charging and discharging, renewable energy curtailment, and load reduction. The results show that the proposed method can improve the solution efficiency of congestion active control optimization while avoiding complex online repetitive safety checks and time-domain simulations. It also enhances the adaptability of control strategies to different operating conditions and reduces unnecessary regulation costs while ensuring power grid congestion safety. This has positive significance for the real-time identification, early elimination, and active control of power grid transmission congestion risks.

[0075] Compared with existing technologies, this invention applies a classification neural network surrogate modeling method to the optimization process of dynamic congestion control in power grids. It utilizes a classification neural network to learn the nonlinear mapping relationship between the power grid operating state and the congestion safety state, transforming the congestion safety determination process, which originally relied on complex safety verification and time-domain simulation, into a rapidly computeable classification surrogate model. Furthermore, through mixed-integer linear reconstruction, the neural network classification boundary is analytically embedded into the active control optimization model, achieving unified modeling and rapid solution of congestion safety constraints and multi-time-period active control optimization. This invention has the following advantages: 1. It fully considers the relationship between the power grid operating state and the dynamic congestion safety boundary. By learning the congestion safety characteristics under different operating scenarios through a classification neural network, it can better adapt to changes in power grid operating conditions compared to the fixed transmission limit method, improving the dynamics and accuracy of the congestion safety constraint expression. 2. Introducing a regularization term during the classification neural network training process and employing interval operations and boundary contraction strategies during mixed-integer linear reconstruction improves network sparsity, reduces the number of binary auxiliary variables, narrows the optimization range of the mixed-integer linear programming model, and improves online solution efficiency. 3. By combining a data-driven congestion safety classification model with a mixed-integer linear programming optimization model, the complex dynamic congestion safety control problem can be transformed into a fast-solvable active control optimization problem, providing effective support for real-time identification, early elimination, and online active control of power grid transmission congestion risks. 4. Proactive regulation triggered in advance based on congestion risk perception results can reserve more time for conventional unit ramp-up and energy storage charging / discharging, enhancing the initiative of real-time congestion prevention and control, and reducing the use of high-cost control measures such as emergency load shedding.

[0076] The following describes the dynamic blocking active control system based on the analytical embedding of a classification neural network provided by the present invention. The dynamic blocking active control system based on the analytical embedding of a classification neural network described below can be referred to in correspondence with the dynamic blocking active control method based on the analytical embedding of a classification neural network described above.

[0077] This application provides a dynamic blocking active control system based on a classification neural network with embedded analytical methods, including: The benchmark model establishment and linearization module is used to establish a benchmark model for blocking regulation optimization that includes multiple types of operational constraints, and to linearize the nonlinear constraints in the benchmark model for blocking regulation optimization. The data-driven proxy model building module is used to build a congestion safety classification neural network proxy model based on historical power grid operation data and corresponding congestion safety status judgment results. The mixed-integer linear reconstruction module is used to perform mixed-integer linear reconstruction on the blocking safe classification neural network agent model to determine the blocking safe classification agent constraints. The embedded solution and strategy output module is used to embed the blocking safety classification agent constraints into the linearized blocking control optimization benchmark model to construct and solve a mixed integer linear programming model, which is used to output multi-time period active control strategies.

[0078] Figure 6 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 6 As shown, the electronic device may include a processor 610, a communications interface 620, a memory 630, and a communication bus 640. The processor 610, communications interface 620, and memory 630 communicate with each other via the communication bus 640. The processor 610 can call logical instructions from the memory 630 to execute a dynamic blocking active control method based on a classification neural network parsing embedded within it.

[0079] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0080] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0081] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dynamic blocking active control method based on the parsing embedding of a classification neural network, characterized in that, include: A benchmark model for blockage control optimization, which includes multiple types of operational constraints, is established, and the nonlinear constraints in the benchmark model for blockage control optimization are linearized. Based on historical power grid operation data and corresponding congestion safety status determination results, a congestion safety classification neural network proxy model is constructed. The blocking security classification neural network agent model is reconstructed using mixed integer linear reconstruction to determine the blocking security classification agent constraints; The congestion safety classification agent constraint is embedded into the linearized congestion control optimization benchmark model to construct and solve a mixed-integer linear programming model, which is used to output a multi-period active control strategy.

2. The dynamic blocking active control method based on the parsing embedded classification neural network according to claim 1, characterized in that, The step of establishing a baseline model for congestion control optimization that includes multiple types of operational constraints specifically includes: To minimize the overall cost based on conventional unit power generation cost, energy storage system charging and discharging cost, renewable energy curtailment cost, and load reduction cost within the control cycle, an objective function is established for the congestion control optimization benchmark model. The various types of operational constraints in the congestion control optimization benchmark model include at least AC power flow constraints, generator operation constraints, operation safety constraints, energy storage operation constraints, and new energy and load regulation constraints.

3. The dynamic blocking active control method based on the parsing of a classification neural network as described in claim 2, characterized in that, The nonlinear constraints include the quadratic term of the conventional unit generation cost in the objective function, the apparent power constraint of the branch in the AC power flow constraint and the operational safety constraint.

4. The dynamic blocking active control method based on the parsing embedding of a classification neural network according to claim 1, characterized in that, The step of constructing a congestion safety classification neural network proxy model based on historical power grid operation data and corresponding congestion safety status determination results specifically includes: Historical power grid operation data is acquired to extract the operation status characteristics at each historical moment. The operation status characteristics include at least node injected active power, node injected reactive power, and the square of generator node voltage amplitude. For each historical moment, the operation status characteristics are used to determine the blocking safety and generate the corresponding status label; Using the operating state features as input and the state labels as output, a feedforward classification neural network is constructed and trained to obtain a congestion safety classification neural network surrogate model that characterizes the mapping relationship between the power grid operating state and the congestion safety state.

5. The dynamic blocking active control method based on the parsing of a classification neural network embedded according to claim 4, characterized in that, During the training of the feedforward classification neural network, an L1 regularization term is added to the loss function to promote the sparsity of the network weight matrix.

6. The dynamic blocking active control method based on the parsing of a classification neural network embedded according to claim 5, characterized in that, The step of performing mixed-integer linear reconstruction on the blocking security classification neural network proxy model specifically includes: For the ReLU activation function in the feedforward classification neural network, 0 / 1 auxiliary variables are introduced, and the Big-M strategy is used to transform the nonlinear logic of the ReLU activation function into linearity; A boundary contraction strategy is adopted to calculate the upper and lower boundaries of the variables before activation of each hidden layer neuron in the feedforward classification neural network and implement boundary contraction to narrow the range of Big-M values. For neurons whose activation state can be determined according to the upper and lower boundaries, their activation state is fixed and the corresponding 0 / 1 auxiliary variables are omitted.

7. The dynamic blocking active control method based on the parsing embedded classification neural network according to claim 1, characterized in that, The expression for the blocking security classification agent constraint is: ； In the formula, For random scene indexing, For scheduling period index, For hidden layer index, This represents the total number of hidden layers. In the scene Scheduling period The input layer below receives information. In the scene Scheduling period The input feature vector, In the scene Scheduling period The next The input vector before activation of the hidden layer. In the scene Scheduling period The next The output vector after activation of the hidden layer. Here is the weight matrix of the FNN model. For bias vectors, , These are the upper and lower boundary vectors, respectively. In the scene Scheduling period Hidden layer below The corresponding 0 / 1 auxiliary variable vector, In the scene Scheduling period The following network logarithm In the scene Scheduling period The number of network segments corresponding to the normal class under the following conditions. In the scene Scheduling period The number of network pairs corresponding to the blocking class below. , These represent the weight matrix and bias vector of the output layer, respectively. For category tags.

8. The dynamic blocking active control method based on the parsing embedding of a classification neural network according to claim 1, characterized in that, After outputting the multi-period active control strategy, a verification operation is performed on the multi-period active control strategy; The verification operations include: performing AC power flow verification, N-1 static security verification, and transient power angle stability verification on the multi-time period active control strategy.

9. A dynamic blocking active control system based on a classification neural network with analytical embedding, characterized in that, include: The benchmark model establishment and linearization module is used to establish a blocking control optimization benchmark model containing multiple types of operational constraints, and to linearize the nonlinear constraints in the blocking control optimization benchmark model. The data-driven proxy model building module is used to build a congestion safety classification neural network proxy model based on historical power grid operation data and corresponding congestion safety status judgment results. The mixed-integer linear reconstruction module is used to perform mixed-integer linear reconstruction on the blocking security classification neural network agent model to determine the blocking security classification agent constraints. An embedded solution and strategy output module is used to embed the blocking safety classification agent constraint into the linearized blocking control optimization benchmark model to construct and solve a mixed integer linear programming model, which is used to output a multi-period active control strategy.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the dynamic blocking active control method based on the parsing of a classification neural network as described in any one of claims 1 to 8.

Images