CN-122026533-A - Virtual power plant aggregation operation cost scheduling method based on input convex neural network

CN122026533ACN 122026533 ACN122026533 ACN 122026533ACN-122026533-A

Abstract

The invention discloses a virtual power plant aggregate operation cost scheduling method based on an input convex neural network, which comprises the steps of constructing a feature coding model based on a deep neural network, carrying out feature extraction on distributed power supply output data in a virtual power plant to obtain coding feature data, constructing an equivalent conversion model of the input convex neural network, solving and obtaining a multi-period aggregate operation cost agent model, taking the obtained multi-period aggregate operation cost agent model as an operation cost function of the virtual power plant, constructing a multi-period random economic optimization scheduling model of a power distribution network, solving to obtain an optimal aggregate power scheduling instruction of the virtual power plant, and issuing execution scheduling. The invention ensures the data privacy by utilizing the implicit feature codes, and accurately fits the nonlinear relation and strictly ensures the mathematical convexity by inputting the convex neural network, thereby overcoming the dilemma of non-convex solution and ensuring the multi-period scheduling scheme to have global economic optimality.

Inventors

ZHU JUNPENG
YANG KE
LIU CAN
SHEN HELIAN
Sun Qirun
FU ZHIXIN
YUAN YUE

Assignees

河海大学

Dates

Publication Date: 20260512
Application Date: 20260415

Claims (8)

1. The virtual power plant aggregate operation cost scheduling method based on the input convex neural network is characterized by comprising the following steps of: S1, constructing a feature coding model based on a deep neural network, and performing feature extraction on output data of a distributed power supply in a virtual power plant to obtain coding feature data; S2, fusing the obtained coding characteristic data with the virtual power plant aggregate power, constructing an equivalent transformation model of an input convex neural network, and solving and obtaining a multi-period aggregate operation cost proxy model taking the virtual power plant aggregate power track as input and the predicted aggregate operation cost as output through forward propagation calculation containing non-negative weight constraint; And S3, taking the obtained multi-period aggregation operation cost agent model as an operation cost function of the virtual power plant, constructing a multi-period random economic optimization scheduling model of the power distribution network, solving the model to obtain an optimal aggregation power scheduling instruction of the virtual power plant, and issuing execution scheduling.
2. The method for scheduling the aggregated running cost of the virtual power plant based on the input convex neural network according to claim 1, wherein the expression of the feature coding model in the step S1 is as follows: ; ; ; In the formula (1), the components are as follows, Representing a virtual power plant The internal distributed power output is obtained through Monte Carlo scene generation; Tensors are input for the deep neural network, in equation (2), Representing a ReLU activation function; wherein Hiding the number of layers of the layers for the deep neural network; represents the first Weights of the layers; represents the first The bias of the layers is such that, Is the first A hidden layer, wherein in the formula (3), The output layer represents the internal distributed power supply coding characteristic data of the virtual power plant; representing deep neural network Layer output; Representing the weight of the deep neural network output layer; representing the bias of the deep neural network output layer.
3. The method for scheduling the aggregated running cost of the virtual power plant based on the input convex neural network according to claim 2, wherein the expression of fusing the encoded characteristic data and the aggregated power of the virtual power plant in the step S2 is as follows: ; In the formula (4), the amino acid sequence of the compound, An initial composite input tensor for the input convex neural network; Encoding the characteristic data; the power data is aggregated for the interfaces of the virtual power plant, and the superscript T represents the transpose.
4. The method for scheduling the aggregated running cost of the virtual power plant based on the input convex neural network according to claim 3, wherein the expression of the equivalent transformation model of the input convex neural network in the step S2 is as follows: ; in the formula (5), the amino acid sequence of the compound, For inputting the output of the kth layer of the convex neural network, Wherein The number of hidden layers for the input convex neural network; And Respectively an internal transmission weight matrix and a jump connection weight matrix of the kth layer; Is the bias of the k-th layer.
5. The method for scheduling aggregated operating costs of a virtual power plant based on an input convex neural network according to claim 4, wherein the non-negative weight constraint in step S2 is to constrain an internal transmission weight matrix Is non-negative.
6. The method for scheduling the aggregate operation cost of a virtual power plant based on the input convex neural network according to claim 5, wherein the expression of the multi-period aggregate operation cost proxy model in step S2 is as follows: ; in the formula (6), the amino acid sequence of the compound, For the predicted aggregate running cost output by the output layer of the input convex neural network, adopting a mean square error as a loss function of the input convex neural network; to input the convex neural network Layer output; And Internal transfer weight matrix and jump connection weight matrix respectively representing output layers of input convex neural network, wherein the matrix All coefficients of (2) are non-negative; the bias of the output layer is input to the convex neural network.
7. The method for scheduling the aggregated running cost of the virtual power plant based on the input convex neural network according to claim 6, wherein the multi-period random economic optimization scheduling model of the power distribution network in the step S3 comprises the following objective functions: ; In the formula (7), the superscript T represents the transpose, the subscript s represents the scene index, Representing a set of scenes that are to be combined, Representing a node set of the power distribution network; Representing an optimization variable of the power distribution network; representing the probability of a scene, And Representing a cost factor of one and a cost factor of two, respectively.
8. The method for scheduling the aggregated running cost of the virtual power plant based on the input convex neural network according to claim 7, wherein the multi-period random economic optimization scheduling model of the power distribution network in the step S3 comprises the following constraint conditions: ; ; ; ; ; ; in the formulas (8) to (10), Representing a scene Lower acting power distribution network node Optimization variables Equation constraint coefficient matrix for characterizing distribution network nodes Variable coefficients in the power balance constraint; representing a scene Downward acting on virtual power plant aggregate power variables The equation constraint coefficient matrix of (2) is used for representing the mapping relation of the aggregated power of the virtual power plant when the aggregated power is injected into the power balance equation; Representing a scene Lower node A constant term vector corresponding to the equation constraint for characterizing a known injection amount or a known demand amount at the right end of the node power balance equation; coefficient matrix representing constraints of remaining equations of a power distribution network for characterizing a scene An equality constraint part in the voltage and power flow relation in the lower network; Representing a scene The constant vector corresponding to the equality constraint is arranged below to reflect the external injection parameter or the scene related known quantity; The coefficient matrix representing the inequality constraint of the power distribution network is used for uniformly representing the upper and lower limit constraint of line tide, the upper and lower limit constraint of node voltage, the upper and lower limit constraint of gas turbine output, the climbing constraint of the gas turbine and the upper and lower limit constraint of distributed power supply output; Representing a scene The lower inequality constrains the corresponding upper bound vector; a constraint coefficient matrix representing the aggregate power feasible region of the virtual power plant, The constraint boundary vector represents the constraint boundary vector corresponding to the constraint coefficient matrix, and the constraint boundary vector and the constraint coefficient matrix together form a linear inequality constraint of the virtual power plant aggregate power, which is used for representing the feasible domain range of the virtual power plant aggregate power, wherein the expression (11) -expression (12) is an upper mirror diagram constraint of a ReLU activation function, and the expression (13) is an upper mirror diagram constraint of the predicted aggregate operation cost output by the input convex neural network.

Description

Virtual power plant aggregation operation cost scheduling method based on input convex neural network 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 virtual power plant aggregate operation cost scheduling method based on an input convex neural network. Background With the continued reform of the global power sector, the scale of distributed energy sources such as energy storage systems, electric vehicles, temperature controlled loads, and micro gas turbines in power distribution systems is increasing at an unprecedented rate. In order to effectively integrate and manage these decentralized and massive flexible resources, virtual power plants have emerged as an intermediary aggregate entity. The virtual power plant aggregates a plurality of heterogeneous distributed resources into a whole through internal coordination control, and provides flexibility for an upper-layer power distribution network. In the process, in order to meet the interface aggregate power instruction issued by the power distribution network, the virtual power plant needs to cooperatively schedule various distributed resources in the virtual power plant, so that the generated running loss and physical adjustment cost of the internal equipment form the aggregate running cost of the virtual power plant. Therefore, accurately modeling and quantifying the aggregate operation cost of multiple periods is critical to the upper grid operators to perform global economic dispatch and promote the overall economic operation of the system. However, existing cost modeling methods are often too simplistic and difficult to accurately characterize the complex operating characteristics inside a virtual power plant. For example, some studies use a quadratic equivalent cost function, but such empirical formulas based on historical data fitting are only suitable for cost characterization of homogenized distributed resources, lacking generalization capability in aggregated scenarios involving various heterogeneous resources. For a multi-period scene, considering energy coupling characteristics of energy storage equipment, temperature control load and the like existing across periods, the existing method mostly adopts piecewise linear approximation based on sampling to construct a multi-period cost curve. The external approximation method is easy to sink into dimension disasters, and cannot fully reflect nonlinear marginal cost structures generated by heterogeneous distributed energy due to internal complex physical constraints, so that accuracy of price signals and upper-layer scheduling decisions is affected. In addition, in recent years, research is attempted to introduce a deep neural network to fit high-dimensional complex mapping, but if the traditional deep neural network is directly embedded into power distribution network optimization scheduling as a highly non-convex black box model, the original scheduling planning is converted into non-convex optimization, so that the problem of non-convex optimization is extremely easy to sink into local optimization, and reliable mathematical guarantee cannot be provided for system operation. Meanwhile, detailed parameters and output states of distributed power supplies in the virtual power plant relate to serious data privacy, and the traditional method has a large leakage risk in data interaction. Disclosure of Invention The invention aims to overcome the defects that the prior art is difficult to consider high-precision quantification, global optimal solution and internal data privacy protection of multi-period aggregation cost, and provides a virtual power plant aggregation operation cost scheduling method based on an input convex neural network. The invention provides a virtual power plant aggregation operation cost scheduling method based on an input convex neural network, which comprises the following steps of: S1, constructing a feature coding model based on a deep neural network, and performing feature extraction on output data of a distributed power supply in a virtual power plant to obtain coding feature data; S2, fusing the obtained coding characteristic data with the virtual power plant aggregate power, constructing an equivalent transformation model of an input convex neural network, and solving and obtaining a multi-period aggregate operation cost proxy model taking the virtual power plant aggregate power track as input and the predicted aggregate operation cost as output through forward propagation calculation containing non-negative weight constraint; And S3, taking the obtained multi-period aggregation operation cost agent model as an operation cost function of the virtual power plant, constructing a multi-period random economic optimization scheduling model of the power distribution network, solving the model to obtain an optimal aggregation power scheduling i