CN-122000890-A - Distributed resource aggregation method based on input convex neural network and target guidance

CN122000890ACN 122000890 ACN122000890 ACN 122000890ACN-122000890-A

Abstract

The invention discloses a distributed resource aggregation method based on an input convex neural network and target guidance, which comprises the steps of constructing a proxy model based on the input convex neural network to represent the internal operation cost of a distributed resource aggregation main body to obtain an aggregation operation cost model, establishing a multi-period economic dispatch model of a power distribution network, utilizing a dual theory to extract sensitivity information of aggregation parameters in an economic dispatch result to construct a feasible region reconstruction model, dynamically correcting the aggregation parameters based on the multi-period economic dispatch model of the power distribution network and the feasible region reconstruction model, and solving and obtaining an aggregation feasible region and an optimal dispatch scheme of the target guidance. According to the method, the multi-period exchange power and the internal operation cost of the virtual power plant can be fitted with high precision, and meanwhile, the problem that the existing static feasible region is too conservative is solved by the target-oriented reconstruction mechanism, so that the accuracy of multi-period economic dispatching of the power distribution network is greatly improved, and the operation cost of a power grid is remarkably reduced.

Inventors

ZHU JUNPENG
YANG KE
LIU CAN
SHEN HELIAN
Sun Qirun
DONG XIAOXIAO
YUAN YUE

Assignees

河海大学

Dates

Publication Date: 20260508
Application Date: 20260410

Claims (8)

1. The distributed resource aggregation method based on the input convex neural network and the target guidance is characterized by comprising the following steps: S1, constructing a proxy model based on an input convex neural network to represent the internal operation cost of a distributed resource aggregation main body and obtain an aggregation operation cost model; S2, embedding an aggregate operation cost model into an optimization target of the power distribution network, and establishing a multi-period economic dispatch model of the power distribution network by combining an initial aggregate feasible region; S3, extracting sensitivity information of aggregation parameters in economic dispatching results by utilizing a dual theory, and constructing a feasible domain reconstruction model; And S4, dynamically correcting the aggregation parameters based on the multi-period economic dispatching model and the feasible region reconstruction model of the power distribution network, and solving and obtaining an aggregation feasible region and an optimal dispatching scheme of the target guide.
2. The method for aggregating distributed resources based on the input convex neural network and the target direction according to claim 1, wherein the expression of the proxy model based on the input convex neural network in the step S1 is as follows: ; ; ; in the formula (1), the components are as follows, To input a tensor for the convex neural network, In order to obtain the power output information of the internal distributed power supply of the virtual power plant after the deep neural network coding, Is a virtual power plant Exchanging power with an interface of the power distribution network, in (2), Representing a ReLU activation function; Wherein The number of hidden layers for the input convex neural network; And Respectively represent the first An internal transfer weight coefficient matrix and a direct connection entry coefficient matrix of a layer, wherein Is a non-negative value; represents the first The bias of the layers is such that, Is the first A hidden layer, wherein in the formula (3), Representing the minimum running cost of the virtual power plant as an output layer; represents the first Layer output; And Respectively representing a weight coefficient matrix and a direct-connection input item coefficient matrix of an output layer; representing the bias of the output layer and superscript T representing the transpose.
3. The method for aggregating distributed resources based on the input convex neural network and the target direction according to claim 2, wherein the expression of the target function of the multi-period economic dispatch model of the power distribution network in step S2 is as follows: With the minimized operation cost as an objective function, the operation cost comprises transaction cost of a power distribution network and a main network, gas turbine operation cost, distributed generation disuse cost and aggregate operation cost which is embedded through upper mirror diagram constraint and is characterized by an input convex neural network proxy model: ; In the formula (4), the amino acid sequence of the compound, And Representing a cost factor of one and a cost factor of two respectively, Representing the distribution level optimization variables, Representing the exchanged power of the virtual power plant with the distribution network, Representing a node set of the power distribution network, and superscript T represents a transpose.
4. The method for aggregating distributed resources based on the input convex neural network and the target direction according to claim 3, wherein the constraint conditions of the multi-period economic dispatch model of the power distribution network in the step S2 include node active and reactive power balance constraint, branch power flow constraint, node voltage constraint, equipment operation boundary constraint and initial aggregation feasible region constraint, which are specifically expressed as follows: ; ; ; ; ; ; In the formulas (5) and (6), , And Is a node A compact matrix representation parameter subject to a power balance constraint, wherein, Representing nodes Optimization variable of power distribution network Is used for the coefficient matrix of (a), Representing nodes Virtual power plant and distribution network exchange power variable Is used for the coefficient matrix of (a), A constant column vector representing a corresponding power balance constraint; , , And Compact matrix expressing parameters for the remaining network and device operating constraints, wherein, The coefficient matrix is constrained for an equation, The constant column vector is constrained for the equation, The coefficient matrix is constrained for an inequality, Constraint the upper bound column vector for inequality, in equation (7), And Constraint parameters for characterizing an aggregate viable domain of a virtual power plant, wherein, A matrix of constraint coefficients representing a substantially homomorphic polyhedron, Representing the corresponding constraint boundary vector(s), And Representing an aggregate scale factor and a translation factor, respectively, wherein, , , Representative of Equations (8) and (9) are upper mirror constraints of the ReLU activation function, and equation (10) is an upper mirror constraint of the input convex neural network output layer.
5. The method for extracting sensitivity information of aggregation parameters in economic dispatch results by utilizing dual theory in step S3 is characterized by comprising the following steps: A1, constructing a Lagrange function aiming at a multi-period economic dispatch model of the power distribution network: ; In the formula (11), the amino acid sequence of the compound, Representing distribution network equality constraints Is a dual variable of (2); Representing distribution network inequality constraints Is a dual variable of (2); Representative node A dual variable for power balance constraint; Representing a virtual power plant Aggregating the dual variables of the feasible region constraint; Representing a virtual power plant Middle input convex neural network (MIN) first Dual variables of the mirror diagram constraint on the layer; Representing a virtual power plant Middle input convex neural network (MIN) first Layer non-negative constraint of the dual variable; Representing a virtual power plant Inputting a dual variable of a mirror diagram constraint on an output layer of the convex neural network; A2, deriving KKT conditions, and extracting optimal pairs related to the constraint of the aggregation feasible region: ; ; ; ; ; ; ; ; ; ; ; ; ; ; ; ; ; Wherein, superscript Representing an optimal solution for the corresponding variable; representing the variable of the exchange power in the coefficient matrix of the first hidden layer of the input convex neural network A corresponding column block sub-matrix; Representing input convex neural network Variable of exchange power in coefficient matrix of direct connection input item of hidden layer A corresponding column block sub-matrix; Representing the variable of the direct connection input term coefficient matrix of the output layer of the input convex neural network and the exchange power A corresponding column block sub-matrix, wherein the column block sub-matrix is represented by Is extracted from the corresponding coefficient matrix under the block structure of (a) Matching column sub-blocks; A3, calculating partial derivatives of the global objective function relative to aggregation parameters of the aggregation feasible region, and taking the partial derivatives as sensitivity information of modeling of the target-oriented feasible region: ; ; Wherein, the Is that Is the first of (2) A number of variables; to correspond to Is described.
6. The method for aggregating distributed resources based on an input convex neural network and object steering according to claim 5, wherein the expression of the feasible region reconstruction model in step S3 is as follows: ; ; ; ; ; Wherein, the Representing regularization terms for balancing weights between different targets in the optimization model; And Representing the current scaling factor and the shifting factor respectively; And Correction amounts representing the scaling and translation factors, respectively; represents a substantially homomorphic polyhedron, wherein, Representing a power trajectory vector in a substantially homomorphic polyhedron; Represents an accurately viable domain of distributed resources, wherein, Representing a distributed resource power trajectory vector in the precise feasible domain, Representing a matrix of coefficients that constitute a linear constraint of the precise feasible region of the distributed resource, Representing boundary vectors corresponding to the linear constraints, for giving constraint upper bounds of the accurate feasible region of the distributed resource; And Representing the confidence domain of the scaling and translation factors, respectively.
7. The method for aggregating distributed resources based on an input convex neural network and target guidance according to claim 6, wherein in the step S4, a double-layer iterative optimization algorithm is adopted to solve and obtain an aggregate feasible region and an optimal scheduling scheme of the target guidance.
8. The method for aggregating distributed resources based on an input convex neural network and object steering according to claim 7, wherein the solving process of the two-layer iterative optimization algorithm in step S4 includes: b1, solving a feasible domain reconstruction model in the current confidence domain range to obtain a candidate correction quantity of a scaling factor and a shifting factor; Substituting the aggregation parameters updated by the candidate correction amount into a multi-period economic dispatch model of the power distribution network for global trial calculation verification; B3, if the global running cost obtained by trial calculation is reduced, receiving the correction result, if the global running cost is not reduced, rejecting the correction, and reducing the confidence domain range, returning to the step B1, and re-solving; And B4, iterating the process until the convergence condition is met.

Description

Distributed resource aggregation method based on input convex neural network and target guidance Technical Field The invention belongs to the field of power grids, relates to a power distribution network control, operation and optimization technology, and particularly relates to a distributed resource aggregation method based on an input convex neural network and target guidance. Background With the continued reform of 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 power distribution systems is increasing at an unprecedented rate. These resources, while providing flexibility to the system, are highly dispersed in physical location. In order to effectively integrate and manage these decentralized resources, coordinated unified scheduling by intermediaries such as virtual power plants has become a very promising solution. In the current coordination framework, an "aggregate flexibility mode" centered on the distribution network operator is of great interest because of its high friendliness to scheduling. In this mode, each virtual power plant needs to submit its aggregate feasible region and corresponding aggregate operating cost model to the distribution network operator. The distribution network operator then uses this information to perform a global multi-period economic dispatch to optimize the operating efficiency of the overall system. However, existing polymerization methods still have significant limitations in multi-period scenarios. On the one hand, the research on the aggregation operation cost is relatively less, the processing mode is generally and excessively simplified, the time coupling characteristic of heterogeneous flexible resources and the nonlinear characteristic of internal operation cannot be accurately described, so that the cost evaluation accuracy is insufficient, on the other hand, on the construction of an aggregation feasible domain, the existing method is mostly limited to the simple geometric approximation of a static physical boundary, often decision conservation is generated due to the limitation of a fixed boundary, and a target-oriented reconstruction mechanism which is cooperated with a system-level economic dispatching target cannot be considered. Therefore, development of a completely new distributed resource aggregation method is needed. The method not only needs to be capable of efficiently and accurately quantifying the nonlinear and time coupling characteristics of the multi-period aggregation operation cost, but also needs to guide the aggregation feasible region to dynamically reconstruct the system-level scheduling target, and strictly protects the data privacy of the bottom equipment in the virtual power plant in the process, so that the accuracy and the economy of multi-period economic scheduling of the power distribution network are comprehensively improved. Disclosure of Invention The invention aims to overcome the defects in the prior art, and provides a distributed resource aggregation method based on an input convex neural network and target guidance, which can perform high-precision fitting on multi-period exchange power and internal operation cost of a virtual power plant, and meanwhile, a target-guided reconstruction mechanism solves the problem that the existing static feasible region is too conservative, so that the accuracy of multi-period economic scheduling of a power distribution network is greatly improved, the running cost of the power network is obviously reduced, and the method has important significance for large-scale coordination of massive heterogeneous distributed resources and fine and economic operation and planning of the power distribution network. The invention provides a distributed resource aggregation method based on an input convex neural network and target guidance, which comprises the following steps of: S1, constructing a proxy model based on an input convex neural network to represent the internal operation cost of a distributed resource aggregation main body and obtain an aggregation operation cost model; S2, embedding an aggregate operation cost model into an optimization target of the power distribution network, and establishing a multi-period economic dispatch model of the power distribution network by combining an initial aggregate feasible region; S3, extracting sensitivity information of aggregation parameters in economic dispatching results by utilizing a dual theory, and constructing a feasible domain reconstruction model; And S4, dynamically correcting the aggregation parameters based on the multi-period economic dispatching model and the feasible region reconstruction model of the power distribution network, and solving and obtaining an aggregation feasible region and an optimal dispatching scheme of the target guide. Further, the expression of the proxy m