Distributed resource aggregation method based on input convex neural network and target orientation

Abstract

The invention discloses a distributed resource aggregation method based on an input convex neural network and target orientation, and the method comprises the steps: constructing an agent model based on the input convex neural network, so as to represent the internal operation cost of a distributed resource aggregation main body, and obtaining an aggregation operation cost model; establishing a power distribution network multi-period economic dispatching model; sensitivity information of aggregation parameters in an economic dispatching result is extracted by using a duality theory, and a feasible region reconstruction model is constructed; and dynamically correcting aggregation parameters based on the power distribution network multi-period economic dispatching model and the feasible region reconstruction model, and solving to obtain a target-oriented aggregation feasible region and an optimal dispatching scheme. According to the method, the multi-period exchange power and the internal operation cost of the virtual power plant can be subjected to high-precision fitting, and meanwhile, a target-oriented reconstruction mechanism solves the problem that an existing static feasible region is too conservative, so that the accuracy of multi-period economic dispatching of the power distribution network is greatly improved, and the operation cost of the power grid is remarkably reduced.

Description

Technical Field

[0001] This invention belongs to the field of power grids and relates to distribution network control, operation and optimization technologies, specifically to a distributed resource aggregation method based on input convex neural networks and goal-oriented approaches. Background Technology

[0002] With the ongoing reforms in the global power industry, the penetration rate of distributed resources such as energy storage systems, electric vehicles, temperature-controlled loads, photovoltaics, wind turbines, and micro gas turbines in distribution systems is growing at an unprecedented rate. While these resources provide flexibility to the system, they are highly geographically dispersed. To effectively integrate and manage these dispersed resources, coordinated and unified dispatch through intermediaries such as virtual power plants has become a promising solution.

[0003] Within the current coordination framework, the "aggregated flexibility model" centered on distribution network operators has garnered significant attention due to its high dispatchability. In this model, each virtual power plant submits its aggregated feasible region and corresponding aggregated operating cost model to the distribution network operator. The operator then utilizes this information to perform global multi-period economic dispatch, thereby optimizing the overall system's operational efficiency. However, existing aggregation methods still have significant limitations in multi-period scenarios. On one hand, research on aggregated operating costs is relatively limited, and the processing methods are generally overly simplistic, often failing to accurately characterize the temporal coupling characteristics and internal nonlinear features of heterogeneous flexible resources, resulting in insufficient cost assessment accuracy. On the other hand, in constructing the aggregated feasible region, existing methods are mostly limited to simple geometric approximations of static physical boundaries, often leading to conservative decision-making due to fixed boundary constraints, and failing to consider goal-oriented reconfiguration mechanisms that align with system-level economic dispatch objectives.

[0004] Therefore, there is an urgent need to develop a novel distributed resource aggregation method. This method not only needs to be able to efficiently and accurately quantify the nonlinear and time-coupled characteristics of multi-period aggregation operation costs, but also needs to guide the aggregation feasible region to be dynamically reconstructed towards the system-level scheduling target, while strictly protecting the data privacy of the underlying equipment inside the virtual power plant during this process, thereby comprehensively improving the accuracy and economy of multi-period economic scheduling of the distribution network. Summary of the Invention

[0005] Purpose of the invention: To overcome the shortcomings of existing technologies, this invention provides a distributed resource aggregation method based on input convex neural networks and goal orientation. This method can accurately fit the power exchanged in multiple time periods with the internal operating costs of virtual power plants. At the same time, the goal orientation reconstruction mechanism solves the problem of the overly conservative static feasible region in existing methods. This significantly improves the accuracy of economic dispatching of distribution networks in multiple time periods and significantly reduces the operating costs of the power grid. This method is of great significance for the large-scale collaborative use of massive heterogeneous distributed resources and the refined and economical operation and planning of distribution networks.

[0006] Technical Solution: To achieve the above objectives, this invention provides a distributed resource aggregation method based on an input convex neural network and goal-oriented approach, comprising the following steps:

[0007] S1: Construct an agent model based on an input convex neural network to characterize the internal operating cost of the distributed resource aggregation entity, and obtain the aggregation operating cost model;

[0008] S2: Embed the aggregated operating cost model into the optimization objective of the distribution network, and combine it with the initial aggregated feasible region to establish a multi-period economic dispatch model for the distribution network;

[0009] S3: Utilize duality theory to extract the sensitivity information of the aggregation parameters in the economic scheduling results and construct a feasible region reconstruction model;

[0010] S4: Based on the multi-period economic dispatch model and feasible region reconfiguration model of distribution network, the aggregation parameters are dynamically corrected to obtain the target-oriented aggregated feasible region and the optimal dispatch scheme.

[0011] Furthermore, the surrogate model based on the input convex neural network in step S1 is expressed as follows:

[0012]

[0013]

[0014]

[0015] In equation (1), The input tensor is used to input the convex neural network. This refers to the output information of distributed power sources within a virtual power plant, encoded using a deep neural network. For virtual power plants Power exchanged with the distribution network interface; in equation (2), Represents the ReLU activation function; ,in The number of hidden layers in the input convex neural network; and Representing the first The layer's internal transitive weight coefficient matrix and direct input term coefficient matrix, where The coefficient is non-negative; Representing the Layer bias, For the first Hidden layers; in equation (3), This is the output layer, representing the minimum operating cost of the virtual power plant; Representing the Layer output; and These represent the weight coefficient matrix of the output layer and the coefficient matrix of the directly connected input terms, respectively. This represents the bias of the output layer.

[0016] Furthermore, the objective function of the multi-period economic dispatch model for the distribution network in step S2 is expressed as follows:

[0017] With the objective function of minimizing operating costs, these costs include transaction costs between the distribution network and the main grid, gas turbine operating costs, distributed generation curtailment costs, and aggregate operating costs represented by an input convex neural network surrogate model embedded through upper-view graph constraints.

[0018]

[0019] In equation (4), and These represent cost coefficient one and cost coefficient two, respectively. Represents the optimization variables of the power distribution layer. This represents the exchange power between the virtual power plant and the distribution network. It represents the set of nodes in the distribution network.

[0020] Furthermore, the constraints of the multi-period economic dispatch model of the distribution network in step S2 include node active and reactive power balance constraints, branch power flow constraints, node voltage constraints, equipment operating boundary constraints, and initial aggregate feasible region constraints, which are specifically expressed as follows:

[0021]

[0022]

[0023]

[0024]

[0025]

[0026]

[0027] In equations (5) and (6), , and For nodes The parameters are expressed in a compact matrix form for the power balance constraint, where, Represents a node Distribution network optimization variables The coefficient matrix, Represents a node Virtual power plants exchange power variables with distribution networks The coefficient matrix, This represents the constant column vector corresponding to the power balance constraint; , , and The parameters are expressed as a compact matrix for the remaining network and device operational constraints, where, This is the equation constraint coefficient matrix. This is a column vector of constants constrained by equality. This is the inequality constraint coefficient matrix. The column vector is the upper bound of the inequality constraint; in equation (7), and The constraint parameters used to characterize the feasible region for virtual power plant aggregation are as follows: The constraint coefficient matrix represents the basic homomorphic polyhedron. This represents the corresponding constraint boundary vector. and These represent the aggregation scaling factor and the translation factor, respectively. , , represent The dimension; Equations (8) and (9) are the mirror image constraints of the ReLU activation function; Equation (10) is the mirror image constraint of the output layer of the input convex neural network.

[0028] Furthermore, the method for extracting the sensitivity information of the aggregation parameters in the economic scheduling results using duality theory in step S3 includes:

[0029] A1: Constructing the Lagrangian function for the multi-period economic dispatch model of the distribution network:

[0030]

[0031] In equation (11), Represents the equality constraints of the distribution network The dual variable; Represents the inequality constraints of the distribution network The dual variable; Representative node The dual variable of the power balance constraint; Represents a virtual power plant Aggregate the dual variables of the feasible region constraint; Represents a virtual power plant The first input convex neural network Dual variables of the layer mirror diagram constraint; Represents a virtual power plant The first input convex neural network The dual variables of the layer nonnegativity constraint; Represents a virtual power plant The dual variables of the mirror image constraint on the output layer of the input convex neural network;

[0032] A2: Derive its KKT conditions and extract the optimal dual multipliers related to the aggregate feasible region constraint:

[0033]

[0034]

[0035]

[0036]

[0037]

[0038]

[0039]

[0040]

[0041]

[0042]

[0043]

[0044]

[0045]

[0046]

[0047]

[0048]

[0049]

[0050] Among them, superscript This represents the optimal solution for the corresponding variable; This represents the exchange power variable in the coefficient matrix of the first hidden layer of the input convex neural network. The corresponding column block submatrix; Indicates the input to the convex neural network. The coefficient matrix of the directly connected input terms in the hidden layer is related to the exchange power variable. The corresponding column block submatrix; This represents the exchange power variable in the coefficient matrix of the directly connected input terms of the output layer of a convex neural network. The corresponding column block submatrix; where, the column block submatrix refers to the submatrix of the corresponding column blocks. Under the block structure, extract the corresponding coefficient matrix and... Matching sub-blocks;

[0051] A3: Calculate the partial derivatives of the global objective function with respect to the aggregated feasible region parameters, and use them as sensitivity information for objective-oriented feasible region modeling:

[0052]

[0053]

[0054] in, for The One variable; For corresponding The block matrix.

[0055] Furthermore, the feasible region reconstruction model in step S3 is expressed as follows:

[0056]

[0057]

[0058]

[0059]

[0060]

[0061] in, This represents a regularization term, used to balance the weights between different objectives in the optimization model; and These represent the current scaling factor and translation factor, respectively. and These represent the correction amounts for the scaling and translation factors, respectively. Represents a basic homomorphic polyhedron, in which, This represents a power trajectory vector in a fundamental homomorphic polyhedron; Represents the exact feasible region of the distributed resource, where, This represents a distributed resource power trajectory vector within the exact feasible region. This represents the coefficient matrix that constitutes the linear constraints of the exact feasible region of distributed resources. This represents the boundary vector corresponding to the linear constraint, used to give the upper bound of the constraint on the exact feasible region of the distributed resource; and These represent the confidence regions of the scaling and translation factors, respectively, to ensure the feasibility of the correction.

[0062] Furthermore, in step S4, a two-layer iterative optimization algorithm is used to solve for the target-oriented aggregated feasible region and the optimal scheduling scheme.

[0063] Furthermore, the solution process of the two-layer iterative optimization algorithm in step S4 includes:

[0064] B1: Solve the feasible region reconstruction model within the current confidence region to obtain candidate corrections for scaling factors and translation factors;

[0065] B2: Substitute the aggregated parameters updated with candidate corrections into the multi-period economic dispatch model of the distribution network for global trial calculation and verification;

[0066] B3: If the calculated global operating cost decreases, accept the result of this correction; if it does not decrease, reject the correction and narrow the confidence region, then return to step B1 to solve again.

[0067] B4: Iterate the above process until the convergence condition is met.

[0068] Beneficial Effects: Compared with existing technologies, this invention employs implicit feature encoding and input convex neural network proxy aggregation of operating costs. While avoiding direct interaction of physical data to protect internal privacy, it achieves high-precision and strictly convex fitting of the nonlinear mapping between multi-period exchange power and internal operating costs. This overcomes the difficulties in accurately characterizing the temporal coupling of heterogeneous resources and the non-convexity of traditional neural network solutions, providing a solid model foundation for obtaining the globally optimal solution of economic scheduling using mature algorithms. This invention replaces the traditional static geometric aggregation method with a goal-oriented feasible region reconstruction mechanism. It uses dual sensitivity information to guide the dynamic correction of the aggregated feasible region, solving the problem that the existing static aggregation boundary is too conservative and excludes high-economic-value operating points. This significantly improves the accuracy and global economy of multi-period economic scheduling of distribution networks, effectively bridging the gap between the mathematical feasibility and economic optimality of pre-scheduling of distribution networks. It is of great significance for the large-scale collaborative calling of massive heterogeneous distributed resources and the refined and economical operation and planning of distribution networks. Attached Figure Description

[0069] Figure 1 This is a flowchart of the method of the present invention;

[0070] Figure 2 This is the power distribution network structure diagram used in Example 2;

[0071] Figure 3 This is a diagram of the iterative solution process. Detailed Implementation

[0072] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0073] Example 1:

[0074] like Figure 1 As shown, this embodiment provides a distributed resource aggregation method based on an input convex neural network and a goal-oriented approach, including the following steps:

[0075] S1: Construct an agent model based on an input convex neural network to characterize the internal operating cost of the distributed resource aggregation entity, and obtain the aggregation operating cost model;

[0076] The surrogate model based on the input convex neural network is expressed as follows:

[0077]

[0078]

[0079]

[0080] In equation (1), The input tensor is used to input the convex neural network. This refers to the output information of distributed power sources within a virtual power plant, encoded using a deep neural network. For virtual power plants Power exchanged with the distribution network interface; in equation (2), Represents the ReLU activation function; ,in The number of hidden layers in the input convex neural network; and Representing the first The layer's internal transitive weight coefficient matrix and direct input term coefficient matrix, where The coefficient is non-negative, while It can take positive or negative values ​​without affecting convexity; Representing the Layer bias, For the first Hidden layers; in equation (3), This is the output layer, representing the minimum operating cost of the virtual power plant; Representing the Layer output; and These represent the weight coefficient matrix of the output layer and the coefficient matrix of the directly connected input terms, respectively. This represents the bias of the output layer.

[0081] S2: Embed the aggregated operating cost model into the optimization objective of the distribution network, and combine it with the initial aggregated feasible region to establish a multi-period economic dispatch model for the distribution network;

[0082] The objective function of the multi-period economic dispatch model for the distribution network is expressed as follows:

[0083] With the objective function of minimizing operating costs, these costs include transaction costs between the distribution network and the main grid, gas turbine operating costs, distributed generation curtailment costs, and aggregate operating costs represented by an input convex neural network surrogate model embedded through upper-view graph constraints.

[0084]

[0085] In equation (4), and These represent cost coefficient one and cost coefficient two, respectively. Represents the optimization variables of the power distribution layer. This represents the exchange power between the virtual power plant and the distribution network. It represents the set of nodes in the distribution network.

[0086] The constraints of the multi-period economic dispatch model for distribution networks include active and reactive power balance constraints at nodes, branch power flow constraints, node voltage constraints, equipment operating boundary constraints, and initial aggregate feasible region constraints, which are expressed as follows:

[0087]

[0088]

[0089]

[0090]

[0091]

[0092]

[0093] In equations (5) and (6), , and For nodes The parameters are expressed in a compact matrix form for the power balance constraint, where, Represents a node Distribution network optimization variables The coefficient matrix, Represents a node Virtual power plants exchange power variables with distribution networks The coefficient matrix, This represents the constant column vector corresponding to the power balance constraint; , , and The parameters are expressed as a compact matrix for the remaining network and device operational constraints, where, This is the equation constraint coefficient matrix. This is a column vector of constants constrained by equality. This is the inequality constraint coefficient matrix. is the column vector of the upper bound of the inequality constraint; the other network and equipment operation constraints include reactive power balance constraints, node voltage constraints, line active and reactive power flow constraints, upper and lower limits of gas turbine output constraints, ramping constraints, distributed power generation output constraints, and load demand constraints; in equation (7), and The constraint parameters used to characterize the feasible region for virtual power plant aggregation are as follows: The constraint coefficient matrix represents the basic homomorphic polyhedron. This represents the corresponding constraint boundary vector. and These represent the aggregation scaling factor and the translation factor, respectively. , , represent The dimension; Equations (8) and (9) are the mirror image constraints of the ReLU activation function; Equation (10) is the mirror image constraint of the output layer of the input convex neural network.

[0094] S3: Utilize duality theory to extract the sensitivity information of the aggregation parameters in the economic scheduling results and construct a feasible region reconstruction model;

[0095] Methods for extracting sensitivity information of aggregation parameters from economic scheduling results using duality theory include:

[0096] A1: Constructing the Lagrangian function for the multi-period economic dispatch model of the distribution network:

[0097]

[0098] In equation (11), Represents the equality constraints of the distribution network The dual variable; Represents the inequality constraints of the distribution network The dual variable; Representative node The dual variable of the power balance constraint; Represents a virtual power plant Aggregate the dual variables of the feasible region constraint; Represents a virtual power plant The first input convex neural network Dual variables of the layered mirror diagram constraint; Represents a virtual power plant The first input convex neural network The dual variables of the layer nonnegativity constraint; Represents a virtual power plant The dual variables of the mirror image constraint on the output layer of the input convex neural network;

[0099] A2: Derive its KKT conditions and extract the optimal dual multipliers related to the aggregate feasible region constraint:

[0100]

[0101]

[0102]

[0103]

[0104]

[0105]

[0106]

[0107]

[0108]

[0109]

[0110]

[0111]

[0112]

[0113]

[0114]

[0115]

[0116]

[0117] Among them, superscript This represents the optimal solution for the corresponding variable; This represents the exchange power variable in the coefficient matrix of the first hidden layer of the input convex neural network. The corresponding column block submatrix; Indicates the input to the convex neural network. The coefficient matrix of the directly connected input terms in the hidden layer is related to the exchange power variable. The corresponding column block submatrix; This represents the exchange power variable in the coefficient matrix of the directly connected input terms of the output layer of a convex neural network. The corresponding column block submatrix; where, the column block submatrix refers to the submatrix of the corresponding column blocks. Under the block structure, extract the corresponding coefficient matrix and... Matching sub-blocks;

[0118] A3: Calculate the partial derivatives of the global objective function with respect to the aggregated feasible region parameters, and use them as sensitivity information for objective-oriented feasible region modeling:

[0119]

[0120]

[0121] in, for The One variable; For corresponding The block matrix.

[0122] The feasible region reconstruction model is expressed as follows:

[0123]

[0124]

[0125]

[0126]

[0127]

[0128] in, This represents a regularization term, used to balance the weights between different objectives in the optimization model; and These represent the current scaling factor and translation factor, respectively. and These represent the correction amounts for the scaling and translation factors, respectively. Represents a basic homomorphic polyhedron, in which, This represents a power trajectory vector in a fundamental homomorphic polyhedron; Represents the exact feasible region of the distributed resource, where, This represents a distributed resource power trajectory vector within the exact feasible region. This represents the coefficient matrix that constitutes the linear constraints of the exact feasible region of distributed resources. This represents the boundary vector corresponding to the linear constraint, used to give the upper bound of the constraint on the exact feasible region of the distributed resource; and These represent the confidence regions of the scaling and translation factors, respectively, to ensure the feasibility of the correction.

[0129] S4: Based on the multi-period economic dispatch model and feasible region reconfiguration model of distribution network, the aggregation parameters are dynamically corrected, and a two-level iterative optimization algorithm is used to solve for the goal-oriented aggregated feasible region and the optimal dispatch scheme.

[0130] The solution process of the two-level iterative optimization algorithm includes:

[0131] B1: Solve the feasible region reconstruction model within the current confidence region to obtain candidate corrections for scaling factors and translation factors;

[0132] B2: Substitute the aggregated parameters updated with candidate corrections into the multi-period economic dispatch model of the distribution network for global trial calculation and verification;

[0133] B3: If the calculated global operating cost decreases, accept the result of this correction; if it does not decrease, reject the correction and narrow the confidence region, then return to step B1 to solve again.

[0134] B4: Iterate the above process until the convergence condition is met.

[0135] This invention provides a highly accurate and convex-guaranteed model of the multi-period aggregated operating cost, taking into account the multi-period coupling and nonlinear operating characteristics of heterogeneous distributed flexible resources. It also guides the static aggregated feasible region to be dynamically reconstructed towards the system-level scheduling objective. This invention effectively bridges the gap between the mathematical feasibility and economic optimality of pre-scheduling in distribution networks, and is of great significance for improving the accuracy of multi-period economic scheduling in distribution networks, reducing overall system operating costs, and promoting efficient collaboration of distributed resources.

[0136] Example 2:

[0137] To verify the effectiveness and impact of the method of the present invention, this embodiment applies the method in a practical manner, selecting a modified distribution network example, and connecting distributed power sources, gas turbines, and virtual power plants to the original test system. For example... Figure 2As shown, the test system includes 33 nodes, numbered 1 to 33. The locations of the distributed power source, gas turbine, and virtual power plant are as follows. Figure 2 As shown in Tables 1 and 2, the line parameters and load parameters of the distribution network are shown in Table 1 and Table 2.

[0138] Table 1 Line Parameters

[0139]

[0140] Table 2 Load Parameters

[0141]

[0142] The distributed power supply parameter settings are shown in Table 3.

[0143] Table 3 Renewable Distributed Power Generation Parameters

[0144]

[0145] To verify the effectiveness of the method of this invention in multi-time-period economic dispatching of distribution networks, comparative tests were conducted in a distribution network system in this embodiment. Two different resource aggregation scenarios were set up for the tests: one was a "goal-free" aggregation model using traditional static geometric approximation, which has fixed boundaries and exhibits a certain degree of conservatism; the other was a goal-oriented reconstruction model using dynamic boundary correction based on dual sensitivity information. The operating cost results are shown in Table 4, comparing the impact of these two different boundary handling methods on the global dispatching economy of the system.

[0146] Table 4 Operating Cost Results

[0147]

[0148] Figure 3 The graph shows the iterative process of solving the method of this invention. It can be seen that as the number of iterations increases, the system operating cost gradually decreases and tends to stabilize after several iterations. This indicates that the proposed goal-oriented dynamic correction method for feasible regions can continuously improve the scheduling results during the iterative solution process and eventually achieve convergence.

[0149] As shown in Table 4, on the one hand, when using the traditional aggregation model without target orientation, the system has a high operating cost due to the conservatism of the fixed geometric boundary; on the other hand, after introducing the target-oriented reconstruction mechanism, the aggregation feasible region can be dynamically modified based on the global scheduling requirements, which effectively releases the flexibility of the underlying resources and significantly reduces the total operating cost, thus verifying the effectiveness of the method of the present invention in overcoming the conservatism of the static aggregation model and improving the global scheduling economy of the distribution network.

[0150] This embodiment demonstrates that, in scenarios involving massive heterogeneous distributed resources in distribution network scheduling, the method of the present invention can not only provide a high-precision and strictly convex proxy model that takes into account both the temporal coupling characteristics of heterogeneous resources and the nonlinear operating costs, but also guide the dynamic correction of the aggregation feasible region through global scheduling sensitivity information. This enables distribution network dispatchers to effectively overcome the decision-making conservatism brought about by traditional static geometric aggregation, and significantly reduce the overall operating cost of the system while strictly protecting the data privacy of underlying devices. It plays an important supporting role in the large-scale collaborative invocation and refined, economical operation and planning of distribution networks with high penetration of distributed resources.

Claims

1. A distributed resource aggregation method based on input convex neural networks and goal-oriented approaches, characterized in that, Includes the following steps: S1: Construct an agent model based on an input convex neural network to characterize the internal operating cost of the distributed resource aggregation entity, and obtain the aggregation operating cost model; S2: Embed the aggregated operating cost model into the optimization objective of the distribution network, and combine it with the initial aggregated feasible region to establish a multi-period economic dispatch model for the distribution network; S3: Utilize duality theory to extract the sensitivity information of the aggregation parameters in the economic scheduling results and construct a feasible region reconstruction model; S4: Based on the multi-period economic dispatch model and feasible region reconfiguration model of distribution network, the aggregation parameters are dynamically corrected to obtain the target-oriented aggregated feasible region and the optimal dispatch scheme.

2. The distributed resource aggregation method based on input convex neural network and goal orientation as described in claim 1, characterized in that, The surrogate model based on the input convex neural network in step S1 is expressed as follows: ； ； ； In equation (1), The input tensor is used to input the convex neural network. This refers to the output information of distributed power sources within a virtual power plant, encoded using a deep neural network. For virtual power plants Power exchanged with the distribution network interface; in equation (2), Represents the ReLU activation function; ,in The number of hidden layers in the input convex neural network; and Representing the first The layer's internal transitive weight coefficient matrix and direct input term coefficient matrix, where The coefficient is non-negative; Representing the Layer bias, For the first Hidden layers; in equation (3), This is the output layer, representing the minimum operating cost of the virtual power plant; Representing the Layer output; and These represent the weight coefficient matrix of the output layer and the coefficient matrix of the directly connected input terms, respectively. Represents the output layer bias; the superscript T represents transpose.

3. The distributed resource aggregation method based on input convex neural network and goal orientation as described in claim 2, characterized in that, The objective function of the multi-period economic dispatch model for the distribution network in step S2 is expressed as follows: With the objective function of minimizing operating costs, these costs include transaction costs between the distribution network and the main grid, gas turbine operating costs, distributed generation curtailment costs, and aggregate operating costs represented by an input convex neural network surrogate model embedded through upper-view graph constraints. ； In equation (4), and These represent cost coefficient one and cost coefficient two, respectively. Represents the optimization variables of the power distribution layer. This represents the exchange power between the virtual power plant and the distribution network. This represents the set of nodes in the distribution network, and the superscript T indicates transpose.

4. The distributed resource aggregation method based on input convex neural network and goal orientation as described in claim 3, characterized in that, The constraints of the multi-time period economic dispatch model of the distribution network in step S2 include node active and reactive power balance constraints, branch power flow constraints, node voltage constraints, equipment operation boundary constraints, and initial aggregate feasible region constraints, which are specifically expressed as follows: ； ； ； ； ； ； In equations (5) and (6), , and For nodes The parameters are expressed in a compact matrix form for the power balance constraint, where, Represents a node Distribution network optimization variables The coefficient matrix, Represents a node Virtual power plants exchange power variables with distribution networks The coefficient matrix, This represents the constant column vector corresponding to the power balance constraint; , , and The parameters are expressed as a compact matrix for the remaining network and device operational constraints, where, This is the equation constraint coefficient matrix. This is a column vector of constants constrained by equality. This is the inequality constraint coefficient matrix. The column vector is the upper bound of the inequality constraint; in equation (7), and The constraint parameters used to characterize the feasible region for virtual power plant aggregation are as follows: The constraint coefficient matrix represents the basic homomorphic polyhedron. This represents the corresponding constraint boundary vector. and These represent the aggregation scaling factor and the translation factor, respectively. , , represent The dimension; Equations (8) and (9) are the mirror image constraints of the ReLU activation function; Equation (10) is the mirror image constraint of the output layer of the input convex neural network.

5. The distributed resource aggregation method based on input convex neural network and goal orientation according to claim 4, characterized in that, The method for extracting sensitivity information of aggregation parameters from economic scheduling results using duality theory in step S3 includes: A1: Constructing the Lagrangian function for the multi-period economic dispatch model of the distribution network: ； In equation (11), Represents the equality constraints of the distribution network The dual variable; Represents the inequality constraints of the distribution network The dual variable; Representative node The dual variable of the power balance constraint; Represents a virtual power plant Aggregate the dual variables of the feasible region constraint; Represents a virtual power plant The first input convex neural network Dual variables of the layered mirror diagram constraint; Represents a virtual power plant The first input convex neural network The dual variables of the layer nonnegativity constraint; Represents a virtual power plant The dual variables of the mirror image constraint on the output layer of the input convex neural network; A2: Derive its KKT conditions and extract the optimal dual multipliers related to the aggregate feasible region constraint: ； ； ； ； ； ； ； ； ； ； ； ； ； ； ； ； ； Among them, superscript This represents the optimal solution for the corresponding variable; This represents the exchange power variable in the coefficient matrix of the first hidden layer of the input convex neural network. The corresponding column block submatrix; Indicates the input to the convex neural network. The coefficient matrix of the directly connected input terms in the hidden layer is related to the exchange power variable. The corresponding column block submatrix; This represents the exchange power variable in the coefficient matrix of the directly connected input terms of the output layer of a convex neural network. The corresponding column block submatrix; where, the column block submatrix refers to the submatrix of the corresponding column blocks. Under the block structure, extract the corresponding coefficient matrix and... Matching sub-blocks; A3: Calculate the partial derivatives of the global objective function with respect to the aggregated feasible region parameters, and use them as sensitivity information for objective-oriented feasible region modeling: ； ； in, for The One variable; For corresponding The block matrix.

6. The distributed resource aggregation method based on input convex neural network and goal orientation according to claim 5, characterized in that, The feasible region reconstruction model in step S3 is expressed as follows: ； ； ； ； ； in, This represents a regularization term, used to balance the weights between different objectives in the optimization model; and These represent the current scaling factor and translation factor, respectively. and These represent the correction amounts for the scaling and translation factors, respectively. Represents a basic homomorphic polyhedron, in which, This represents a power trajectory vector in a fundamental homomorphic polyhedron; Represents the exact feasible region of the distributed resource, where, This represents a distributed resource power trajectory vector within the exact feasible region. This represents the coefficient matrix that constitutes the linear constraints of the exact feasible region of distributed resources. This represents the boundary vector corresponding to the linear constraint, used to give the upper bound of the constraint on the exact feasible region of the distributed resource; and These represent the confidence regions of the scaling and translation factors, respectively.

7. The distributed resource aggregation method based on input convex neural network and goal orientation as described in claim 6, characterized in that, In step S4, a two-layer iterative optimization algorithm is used to solve for the target-oriented aggregate feasible region and the optimal scheduling scheme.

8. The distributed resource aggregation method based on input convex neural network and goal orientation according to claim 7, characterized in that, The solution process of the two-level iterative optimization algorithm in step S4 includes: B1: Solve the feasible region reconstruction model within the current confidence region to obtain candidate corrections for scaling factors and translation factors; B2: Substitute the aggregated parameters updated with candidate corrections into the multi-period economic dispatch model of the distribution network for global trial calculation and verification; B3: If the calculated global operating cost decreases, accept the result of this correction; if it does not decrease, reject the correction and narrow the confidence region, then return to step B1 to solve again. B4: Iterate the above process until the convergence condition is met.

Images