Constraint learning based equivalent projection for master-slave coordination intra-day scheduling method and system

Abstract

The application discloses a main-distribution coordination collaborative intra-day dispatching method and system based on constraint learning equivalent projection, and belongs to the technical field of power system dispatching, which comprises the following steps: constructing a distribution network operation model containing a high proportion of distributed resources; initializing an equivalent projection space based on an external cut plane, and generating and labeling boundary-aware samples; learning the feasible region boundary of the distribution network by using a multilayer perception machine, and converting the trained network into a mixed integer linear constraint to analytically generate the equivalent projection of the distribution network; performing effective sub-region screening and redundant constraint elimination on the equivalent projection to obtain a lightweight equivalent projection model; embedding the simplified equivalent projection model into a power transmission network intra-day dispatching model to complete the optimization of tie-line power and the collaborative dispatching of distributed resources in the distribution network. The application can avoid main-distribution multi-round iteration interaction, improve the intra-day dispatching calculation efficiency and scalability while protecting the privacy of the distribution network model, and is suitable for main-distribution collaborative hierarchical regulation and control in a high-proportion distributed resource access scenario.

Description

Technical Field

[0001] This invention belongs to the technical field of power system dispatching, specifically relating to a primary-distributor coordinated intraday dispatching method and system based on constraint learning equivalent projection. Background Technology

[0002] With the large-scale integration of distributed energy sources such as photovoltaics and wind turbines into the power system, the operating characteristics of the distribution network have changed significantly. Transmission and distribution coordinated dispatch has become a key technology for improving the absorption capacity of new energy sources, reducing system operating costs, and ensuring the safety and stability of the power grid. Traditional centralized optimization requires uploading the entire distribution network model and its parameters to the transmission system, which poses a risk of privacy leakage. Moreover, the computational complexity increases sharply with the scale of the system, making it difficult to apply in engineering. Distributed optimization achieves decoupled computation through multiple rounds of iterative information interaction, but it is prone to problems such as convergence oscillation, numerous iterations, and low dispatch efficiency.

[0003] Equivalent projection-based transmission and distribution coordination methods have become the mainstream technology. The core idea is to project the high-dimensional operational constraints of the distribution network onto the power exchange space of tie lines, forming a low-dimensional feasible region that is then fed into the transmission grid. The transmission grid then uses this feasible region to conduct safety-constrained unit combination and economic dispatch, achieving efficient coordination without iteration. However, existing equivalent projection construction methods generally rely on optimization-based iterative searches, including vertex enumeration, Fourier-Motzkin elimination, and internal / external approximation methods. In scenarios with high distributed energy penetration, the dimensionality of distribution network decision variables surges, leading to an exponential increase in the computation time of equivalent projection, making it difficult to meet the real-time requirements of online dispatch. Furthermore, existing data-driven learning methods only use binary classification to determine the feasible region boundary. In high-dimensional convex polyhedral scenarios, the sample demand explodes, resulting in insufficient learning accuracy and an inability to accurately characterize the equivalent projection boundary.

[0004] Faced with the increasing trend of distributed energy and the continuous expansion of the distribution network, how to get rid of the dependence on iterative search, reduce the sensitivity of computational complexity to system scale, achieve high-precision, high-efficiency and scalable equivalent projection construction, and support the coordinated dispatch of transmission and distribution networks has become an urgent technical problem to be solved in the field of power system dispatch. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention generates equivalent projections of the distribution network through constraint learning, replacing the traditional iterative optimization process. This reduces computational load and improves collaborative scheduling efficiency, making computation time less dependent on system scale. This meets the engineering application requirements under high-proportion distributed energy access. The technical solution adopted by this invention is as follows:

[0006] The master-slave collaborative intraday scheduling method based on constraint learning equivalent projection includes the following steps:

[0007] A distribution network operation model with a high proportion of distributed energy is constructed. The feasible domain of tie line power that satisfies all operation constraints of the distribution network is used as the equivalent projection space, so that any tie line power exchange vector in the space corresponds to a set of distribution network operation variables that satisfy all operation constraints.

[0008] The distribution network operation model includes flexible resources such as photovoltaics, wind turbines, micro gas turbines, and controllable loads. The constraints of the distribution network operation model include distributed generation output constraints, controllable load power and energy constraints, linearized power flow equations, node power balance constraints, node voltage and phase angle constraints, and branch power flow constraints.

[0009] The equivalent projection space is initialized based on the external secant plane, and boundary-aware connection line exchange power samples are generated and labeled to obtain a sample dataset with boundary deviation labels.

[0010] Based on the improved multilayer perceptron, the power samples of the tie line exchange are used as input and the labeled labels are used as output. Constraint learning is performed to transform the trained multilayer perceptron into mixed integer linear constraints, and the explicit expression of the equivalent projection of the distribution network is obtained analytically.

[0011] Based on the piecewise linear characteristics of the multilayer perceptron, the parsed equivalent projection is decomposed, effective activation sub-regions are selected and redundant constraints are eliminated to obtain a simplified equivalent projection.

[0012] By uploading the simplified equivalent projection to the power transmission system, a safety-constrained unit combination model with equivalent projection constraints is established for main and distribution coordinated scheduling.

[0013] Furthermore, the equivalent projection space initialization involves setting an initial space that is sufficiently large compared to the equivalent projection space in the tie line switching power space, using the external cutting method to solve the two-layer optimization problem and extract the effective cutting plane, gradually approximating the boundary of the real equivalent projection space, and obtaining an initial space containing the real equivalent projection.

[0014] The external cutting method introduces non-negative auxiliary variables in the initial space to construct an objective function. The inner layer of the objective function minimizes the sum of the non-negative auxiliary variables to characterize the degree of deviation of the sample from the distribution network constraints, while the outer layer maximizes the sample to characterize the degree of infeasible deviation of the sample, which is used to optimize the cutting boundary.

[0015] Furthermore, all operational constraints of the distribution network, in a compact constraint form, are the product of the tie-line exchange power vector and the first coefficient matrix, plus the product of the distribution network operational variables and the second coefficient matrix, which is less than or equal to a set constant vector; the formula is as follows:

[0016]

[0017] in, This represents the power vector exchanged along the tie line. Represents the operating variables of the power distribution network. and These represent two coefficient matrices respectively. This represents a constant vector.

[0018] Furthermore, the inner layer of the objective function is replaced with a dual form, and dual variables and dual variable equality and inequality constraints are introduced to simplify the solution of the original boundary optimization problem. An effective cutting plane is constructed based on the optimal solution of the dual variables to generate an initial space containing a true equivalent projection. The dual form is the product of the constant vector minus the product of the tie line exchange power vector and the first coefficient matrix, and the transposed dual variable. The dual variable equality is that the product of the transposed second coefficient matrix and the dual variable is equal to 0. The dual variable inequality constraint is that the dual variable takes values ​​between -1 and 0. The inequality of the effective cutting plane is that the product of the transposed optimal solution, the tie line exchange power vector, and the first coefficient matrix is ​​greater than or equal to the product of the optimal solution and the constant vector.

[0019] Furthermore, the sample generation and labeling involves randomly generating time-series tie-line exchange power samples within the initialization space, determining whether the sample points are located within the equivalent projection space, calculating the safety margin of internal points and the deviation of external points, and generating a labeled time-series tie-line exchange power sample dataset, thereby achieving precise boundary perception.

[0020] Furthermore, if the objective function is greater than 0, then the sample is not in the equivalent projection space, and the label is set as deviation, with the deviation value being a negative objective function value; if the objective function is equal to 0, then the sample is in the equivalent projection space, and the label is set as safety margin, with the safety margin value being the sum of the squares of the safety margins of the branch power flow and node voltage of the sample. Based on the summation over time steps, a labeled sample dataset with boundary awareness capability is finally obtained that can be used for subsequent neural network training.

[0021] Furthermore, the multilayer perceptron is improved by adopting Leaky ReLU as the activation function and introducing a sign mismatch penalty term and an L2 regularization term into the loss function to improve boundary classification accuracy and suppress overfitting. Labeled samples are input into the multilayer perceptron for training. Through the forward propagation process of the trained multilayer perceptron, the mapping relationship between the link exchange power samples and labels is analyzed. The trained multilayer perceptron is transformed into mixed integer linear constraints using the Big M method. Combined with the non-negative constraint of the equivalent projection (constraining the non-negative condition of the output), the explicit expression of the equivalent projection is finally obtained analytically.

[0022] Furthermore, the simplification of the equivalent projection involves decomposing the equivalent projection into the union of several convex polyhedral sub-regions. Effectively activated sub-regions are selected through optimization problems, and empty regions are eliminated. A parallel multi-constraint synchronous identification method is used to batch eliminate redundant constraints, significantly reducing the number of integer variables and constraints, and improving the computational efficiency of power grid scheduling. Specifically, for n neurons, there exist 2... n There are two activation states, and the entire domain is divided into 2 n There are 10 activation regions, each input x corresponds to a unique activation state, and all activation regions are disjoint. Their union forms the entire space. Input all samples with non-negative labels, perform feasibility optimization, remove invalid sub-regions, and obtain a set of valid sub-regions. Since each sub-region is independent, multiple sub-regions can simultaneously identify redundant constraints. In the end, the equivalent projection significantly reduces the number of binary variables and constraints, and significantly improves scheduling efficiency.

[0023] Furthermore, the coordinated scheduling of the transmission and distribution system aims to minimize the total operating cost of the system. It establishes a unit combination model of the transmission network with equivalent projection constraints, solves for the optimal power command of the tie line, and issues it to the distribution network. The distribution network then performs optimized scheduling of distributed energy resources and controllable loads within the feasible domain of equivalent projection, thereby completing the global coordinated operation.

[0024] The master-partner collaborative intraday scheduling system based on constraint learning equivalent projection includes an equivalent projection construction module, a sample generation module, a multilayer perceptron, an equivalent projection simplification module, and a master-partner collaborative scheduling module. According to the aforementioned master-partner collaborative intraday scheduling method based on constraint learning equivalent projection, the system sequentially executes the construction of equivalent projection, the initialization of equivalent projection, and the generation of sample datasets. The multilayer perceptron is trained using the sample datasets. Based on the piecewise linearity of the multilayer perceptron, the equivalent projection is decomposed to simplify it for use in master-partner collaborative scheduling.

[0025] The advantages and beneficial effects of this invention are as follows:

[0026] This invention generates equivalent projections using a constraint learning framework to replace traditional iterative optimization, ensuring that computational complexity does not increase significantly with system size and distributed energy penetration, thus exhibiting excellent scalability. The invention designs a boundary-aware annotation strategy, enabling accurate learning of equivalent projection boundaries with a small number of samples, addressing the issues of explosive sample demand and insufficient learning accuracy in high-dimensional scenarios. This invention achieves analytical and simplified equivalent projection, significantly reducing the computational burden of transmission network scheduling, balancing accuracy and efficiency, and meeting the needs of online engineering applications. This invention eliminates the need to upload distribution network privacy parameters throughout the process, protecting data privacy while avoiding convergence oscillations in distributed optimization, achieving non-iterative, high-efficiency transmission and distribution coordinated scheduling, demonstrating strong practicality and broad application prospects. Attached Figure Description

[0027] Figure 1 This is a flowchart of the master-slave collaborative intraday scheduling method based on constraint learning equivalent projection in an embodiment of the present invention.

[0028] Figure 2 This is a schematic diagram of the structure of the master-supplier cooperative intraday scheduling system based on constraint learning equivalent projection in an embodiment of the present invention. Detailed Implementation

[0029] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0030] like Figure 1 As shown, this invention proposes a master-slave cooperative intraday scheduling method based on constraint learning equivalent projection, comprising the following steps:

[0031] Step S110: Construct a distribution network operation model with a high proportion of distributed energy resources to form a compact constraint form of the distribution network, providing a basic model for equivalent projection calculation.

[0032] In this embodiment of the invention, the distribution network operation model includes distributed generation output constraints, controllable load power and energy constraints, linearized power flow equations, power balance constraints, node voltage and phase angle constraints, and branch power flow constraints.

[0033] Distributed energy units are wind power and photovoltaic power, and their distributed power output constraints are expressed as follows:

[0034]

[0035]

[0036] in, and They are nodes Photovoltaic power output and wind power output at time t and They are nodes The maximum output of photovoltaic power and the maximum output of wind power at time t.

[0037] The operating constraints of the micro gas turbine are as follows:

[0038]

[0039]

[0040] in, For nodes The active power output of the micro gas turbine at time t and These represent the lower and upper limits of the output of a micro gas turbine, respectively. and These are the upper limits for climbing and descending slopes for micro gas turbines, respectively. For time step.

[0041] The constraints for controllable load operation are expressed as follows:

[0042]

[0043]

[0044]

[0045] in, and They are nodes The controllable load active power and reactive power at time t and These are the lower and upper limits of active power for controllable load, respectively. For controllable load power factor, and These are the lower and upper limits of total energy consumption for controllable loads, respectively.

[0046] In addition to the operational constraints of the above-mentioned flexible resources, the operational constraints of the power distribution system also include linearized power flow equations, power balance constraints, node voltage and phase angle constraints, and branch power flow constraints.

[0047] The power flow equations are expressed as follows:

[0048]

[0049] in, and Branch roads At time t, the active power flow and reactive power flow... and Branch roads The conductivity and susceptance, and They are nodes voltage amplitude, and They are nodes The voltage phase angle.

[0050] The node power balance constraint is expressed as follows:

[0051]

[0052] in, and They are nodes Active and reactive loads at time t For nodes A set of connected nodes.

[0053] The node voltage magnitude and phase angle constraints are expressed as follows:

[0054]

[0055] in, and They are nodes The upper and lower limits of voltage amplitude, and They are nodes The upper and lower limits of the voltage phase angle.

[0056] The power flow constraints of the line are expressed as follows:

[0057]

[0058] in, branch road The upper limit of the active current, branch road The apparent power limit. Therefore, the switching power of the tie line is expressed as:

[0059]

[0060] in, The active power exchanged between the distribution network and the transmission network through tie lines is the core variable of the equivalent projection.

[0061] For ease of analysis, all constraints of the above power distribution system are represented in a compact form, as shown below:

[0062]

[0063] in, To exchange power vectors for tie lines, For distribution network operation variables, and The coefficient matrix, It is a constant vector. Equivalent projection It is the tie-line power feasible region that satisfies all operating constraints of the distribution network, mathematically expressed as:

[0064]

[0065] Therefore, the physical meaning of equivalent projection is: for any There must exist a set of distribution network operation variables. All constraints are satisfied.

[0066] Step S120: Initialize the equivalent projection space based on the external secant plane, and perform boundary-aware sample generation and labeling to obtain a sample dataset with boundary deviation labels.

[0067] In this embodiment of the invention, a sufficiently large space is set in the tie-line power space:

[0068]

[0069] in, and Given the initial spatial coefficient matrix and constant vector, first calculate the following problem:

[0070]

[0071] in, and Let be a pair of non-negative auxiliary variables. If the initial space... If it is large enough and contains a true equivalent projection, then it must exist. Replace the inner form of the above problem with its dual form, where Since the variable is the dual variable, we can derive:

[0072]

[0073] After obtaining the optimal solution Then, an effective cutting plane can be constructed:

[0074]

[0075] Through multiple iterations, an initial space containing the true equivalent projection is obtained to improve the sampling quality of subsequent samples. Time-series samples are then randomly generated within this initial space. Feasibility is determined and labels are calculated by optimizing the problem:

[0076]

[0077] like If the value is negative, it means the sample is not within the equivalent projection, and the label is set to deviation (negative value).

[0078]

[0079] like If the result is positive, it indicates that the sample belongs to the equivalent projection, and the label is set to the safety margin (non-negative value):

[0080]

[0081] in, and These represent the safety margins of the branch power flow and node voltage for sample s at time t, respectively. These labels accurately reflect the distance from the sample to the equivalent projected boundary, enabling boundary-aware learning.

[0082] Step S130: Training the network based on the improved multilayer perceptron for constraint learning. The Big M method is used to transform the trained network into mixed-integer linear constraints, analytically obtaining the explicit expression of the equivalent projection of the distribution network. In this embodiment, a multilayer perceptron with a Leaky ReLU activation function is constructed to... For input, To output the training network, boundary mapping relationships are learned. A sign mismatch penalty term and an L2 regularization term are incorporated into the loss function, as shown below:

[0083]

[0084] in, , and These are the mean squared error, the sign mismatch penalty term, and the L2 regularization term, respectively. , and These are the coefficients for each item. For sample size, For regularization parameters, This represents the L2 paradigm of the network parameters. The forward propagation process of a trained multilayer perceptron can be used to analyze... and Mapping relationship:

[0085]

[0086] in, and These represent the linear mapping of the hidden layer h and the output of the activation function, respectively. The output of the input layer, This is the final value obtained from the output layer. Parameters , and , These represent the weights and biases of the hidden layer and the output layer, respectively. It is usually set to 0.01.

[0087] Since the Leaky ReLU activation function is non-linear, it is transformed into a mixed-integer linear constraint using the Big M method, as shown below:

[0088]

[0089] Finally, combining the non-negativity constraint of the equivalent projection, the analytical form of the trained equivalent projection is obtained as follows:

[0090]

[0091] If the error of the regression model is ignored, then it can be guaranteed that .

[0092] Step S140: Based on the piecewise linearity of the multilayer perceptron, the equivalent projection is decomposed, effective activated sub-regions are selected, and redundant constraints are eliminated to complete the simplification of the equivalent projection. In this embodiment, the piecewise linearity of the multilayer perceptron allows the equivalent projection to be represented as the union of multiple convex polyhedra:

[0093]

[0094] in, In active state, This represents the corresponding activation domain. Its specific representation is as follows:

[0095]

[0096] in, This represents the activation state of the h-th hidden layer. This represents the activation state of the nth neuron in the h-th hidden layer. This indicates that the input to the activated unit is positive. This indicates that the input to the activation unit is negative. This is a constant. In fact, one activation region corresponds to one activation state. If the neural network has n neurons, then there exist 2... n This activation state means that the entire domain can be divided into 2 n Each activation region. Coefficients , and , The derivation is as follows:

[0097]

[0098]

[0099]

[0100] Clearly, each input x corresponds to a unique activation state. All activation domains are disjoint, and their union forms the entire space. Therefore, the learned equivalent projection can be rewritten in the following form:

[0101]

[0102] By eliminating invalid sub-regions through feasibility optimization, a valid activation set is obtained:

[0103]

[0104] in, These are non-negative auxiliary variables. To identify all valid activation regions, input all samples with non-negative labels into the problem to obtain the set of valid sub-regions. Therefore, the equivalent projection can be further written in the following form:

[0105]

[0106] In fact, It is a linear convex polyhedron, whose compact form can be written as Since each subdomain contains a large number of redundant constraints, a multi-constraint synchronous identification model is constructed to batch delete redundant constraints, as shown below:

[0107]

[0108] in, It is a non-negative auxiliary variable. These are 0-1 variables. From the above formula, we can see that if a certain constraint... If the value is 0, it means that the constraint is valid. The optimal value represents identifying as many effective constraints as possible.

[0109] Based on the above analysis, each subdomain is independent of the others, thus multiple subdomains can simultaneously identify redundant constraints. Ultimately, equivalent projection significantly reduces the number of binary variables and constraints, thereby substantially improving scheduling efficiency.

[0110] Step S150: The simplified equivalent projection is uploaded to the transmission system to establish a safety-constrained unit combination model with equivalent projection constraints, thereby achieving coordinated scheduling of the transmission and distribution network. In this embodiment, the objective function of the safety-constrained unit combination model is to minimize the total operating cost, as shown below:

[0111]

[0112] in, and These represent the unit start-up and shutdown costs, and These are the variables for unit start-up and shutdown. and These are the coefficients for the secondary and primary terms of the unit's power generation cost, respectively. Let g be the active power output of unit g at time t.

[0113] Power transmission network operation constraints include unit start-up and shutdown constraints, unit output constraints, node power balance, voltage phase angle constraints, and line power flow constraints, as shown below:

[0114]

[0115] in, This indicates that the switching power of the tie line is within the learned equivalent projected feasible region.

[0116] Solving the above model yields the optimal tie-line power command, which is then sent to the distribution network. The distribution system performs optimized scheduling of controllable resources within the equivalent projection, enabling non-iterative, high-efficiency, and highly scalable transmission and distribution coordinated operation.

[0117] like Figure 2 As shown, this invention proposes a master-partner collaborative intraday scheduling system based on constrained learning equivalent projection. The system includes an equivalent projection construction module, a sample generation module, a multilayer perceptron, an equivalent projection simplification module, and a master-partner collaborative scheduling module. According to the constrained learning-based equivalent projection master-partner collaborative intraday scheduling method, the system sequentially executes the construction of the equivalent projection, the initialization of the equivalent projection, and the generation of the sample dataset. The multilayer perceptron is then trained using the sample dataset. Based on the piecewise linearity of the multilayer perceptron, the equivalent projection is decomposed to simplify it for master-partner collaborative scheduling. Specific examples are the same as described above and will not be repeated here.

[0118] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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 or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A principal-assignment cooperative intraday scheduling method based on constraint learning equivalent projection, characterized in that: The feasible domain of tie line power that satisfies all operating constraints of the distribution network is taken as the equivalent projection space, so that any tie line exchange power vector in the space corresponds to a set of distribution network operating variables that satisfy all operating constraints. The equivalent projection space is initialized based on the external secant plane, and boundary-aware connection line exchange power samples are generated and labeled to obtain a sample dataset with boundary deviation labels. Based on the multilayer perceptron, the power samples of the tie line exchange are used as input and the labeled labels are used as output. Constraint learning is performed to transform the trained multilayer perceptron into mixed integer linear constraints and to obtain the equivalent projection of the distribution network analytically. Based on the piecewise linear characteristics of the multilayer perceptron, the parsed equivalent projection is decomposed, effective activation sub-regions are selected and redundant constraints are eliminated to obtain a simplified equivalent projection. By using simplified equivalent projection, a safety-constrained unit combination model with equivalent projection constraints is established for main and auxiliary coordinated scheduling.

2. The master-distributor cooperative intraday scheduling method based on constraint learning equivalent projection according to claim 1, characterized in that: The equivalent projection space initialization involves setting an initial space that is sufficiently large compared to the equivalent projection space in the tie line switching power space, using the external cutting method to solve the two-layer optimization problem and extract the effective cutting plane, gradually approximating the boundary of the real equivalent projection space, and obtaining an initial space containing the real equivalent projection. The external cutting method introduces non-negative auxiliary variables in the initial space to construct an objective function. The inner layer of the objective function minimizes the sum of the non-negative auxiliary variables to characterize the degree of deviation of the sample from the distribution network constraints, while the outer layer maximizes the sample to characterize the degree of infeasible deviation of the sample, which is used to optimize the cutting boundary.

3. The master-distributor cooperative intraday scheduling method based on constraint learning equivalent projection according to claim 2, characterized in that: All operating constraints of the distribution network are the product of the tie-line exchange power vector and the first coefficient matrix, plus the product of the distribution network operating variables and the second coefficient matrix, which is less than or equal to a set constant vector.

4. The master-distributor cooperative intraday scheduling method based on constraint learning equivalent projection according to claim 3, characterized in that: The inner layer of the objective function is replaced with a dual form, and dual variables and dual variable equality and inequality constraints are introduced. An effective cutting plane is constructed based on the optimal solution of the dual variables to generate an initial space containing the true equivalent projection. The dual form is the product of the constant vector minus the product of the tie-line exchange power vector and the first coefficient matrix, and the transposed dual variable. The dual variable equality is that the product of the transposed second coefficient matrix and the dual variable is equal to 0. The dual variable inequality constraint is that the dual variable takes values ​​between -1 and 0. The inequality of the effective cutting plane is that the product of the transposed optimal solution, the tie-line exchange power vector, and the first coefficient matrix is ​​greater than or equal to the product of the optimal solution and the constant vector.

5. The master-distributor cooperative intraday scheduling method based on constraint learning equivalent projection according to claim 2, characterized in that: The sample generation and labeling process involves randomly generating time-series tie-line exchange power samples within the initialization space, determining whether the sample points are located within the equivalent projection space, calculating the safety margin of internal points and the deviation of external points, and generating a labeled time-series tie-line exchange power sample dataset.

6. The master-distributor cooperative intraday scheduling method based on constraint learning equivalent projection according to claim 5, characterized in that: If the objective function is greater than 0, the sample is not in the equivalent projection space, and the label is set as deviation, with the deviation value being a negative objective function value; if the objective function is equal to 0, the sample is in the equivalent projection space, and the label is set as safety margin, with the safety margin value being the sum of the squares of the safety margins of the branch power flow and node voltage of the sample, calculated based on time steps.

7. The master-distributor cooperative intraday scheduling method based on constraint learning equivalent projection according to claim 1, characterized in that: The multilayer perceptron is improved by using Leaky ReLU as the activation function and introducing a sign mismatch penalty term and an L2 regularization term into the loss function. Labeled samples are input into the multilayer perceptron for training. Through the forward propagation process of the trained multilayer perceptron, the mapping relationship between the link exchange power samples and the labels is analyzed. The trained multilayer perceptron is transformed into a mixed integer linear constraint by the Big M method. Combined with the non-negativity constraint of the equivalent projection, the explicit expression of the equivalent projection is finally obtained analytically.

8. The master-distributor cooperative intraday scheduling method based on constraint learning equivalent projection according to claim 1, characterized in that: The simplification of the equivalent projection involves decomposing it into the union of several convex polyhedral sub-regions. Effective activation sub-regions are selected through an optimization problem, and empty regions are eliminated. A parallel multi-constraint synchronous identification method is used to batch eliminate redundant constraints. Specifically, for n neurons, there exist 2... n There are two activation states, and the entire domain is divided into 2 n There are 10 activation regions, each input corresponds to a unique activation state, and all activation regions are disjoint. Their union forms the entire space. Input all samples with non-negative labels, perform feasibility optimization, remove invalid sub-regions, and obtain a set of valid sub-regions.

9. The master-distributor cooperative intraday scheduling method based on constraint learning equivalent projection according to claim 1, characterized in that: The coordinated scheduling of transmission and distribution systems aims to minimize the total operating cost of the system. It establishes a unit combination model of the transmission network with equivalent projection constraints, solves for the optimal power command of the tie line, and issues it to the distribution network. The distribution network then performs optimal scheduling of distributed energy resources and controllable loads within the feasible domain of equivalent projection.

10. A principal-assignment collaborative intraday scheduling system based on constraint learning equivalent projection, comprising an equivalent projection construction module, a sample generation module, a multilayer perceptron, an equivalent projection simplification module, and a principal-assignment collaborative scheduling module, characterized in that: The master-partner collaborative intraday scheduling method based on constrained learning equivalent projection according to any one of claims 1 to 9 sequentially performs the construction of equivalent projection, the initialization of equivalent projection and the generation of sample dataset, and trains a multilayer perceptron through the sample dataset. Based on the piecewise linearity of the multilayer perceptron, the equivalent projection is decomposed to simplify the equivalent projection for master-partner collaborative scheduling.

Images