Virtual power plant aggregation operation cost scheduling method based on input convex neural network

Abstract

The invention discloses a virtual power plant aggregation operation cost scheduling method based on an input convex neural network, and the method comprises the steps: constructing a feature coding model based on a deep neural network, carrying out the feature extraction of the output data of a distributed power supply in a virtual power plant, and obtaining the coding feature data; constructing an equivalent transformation model of the input convex neural network, and solving to obtain a multi-period aggregation operation cost agent model; and 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, obtaining an optimal aggregation power scheduling instruction of the virtual power plant, and issuing the optimal aggregation power scheduling instruction to execute scheduling. According to the method, implicit feature coding is utilized to guarantee data privacy, a nonlinear relation is accurately fitted by inputting the convex neural network, mathematical convexity is strictly guaranteed, the difficulty of non-convex solution is overcome, and it is guaranteed that a multi-period scheduling scheme has global economic optimality.

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 virtual power plant aggregation operation cost scheduling method based on an input convex neural network. Background Technology

[0002] With the ongoing reforms in 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 growing at an unprecedented rate. To effectively integrate and manage these dispersed and massive flexible resources, virtual power plants (VPS) have emerged as an intermediary aggregation entity. VPS aggregate numerous heterogeneous distributed resources into a unified whole through internal coordination and control, providing flexibility to the upper-level distribution network. In this process, to meet the interface aggregation power commands issued by the distribution network, the VPS needs to coordinate the scheduling of its various internal distributed resources. The resulting internal equipment operating losses and physical regulation costs constitute the "aggregated operating cost" of the VPS. Therefore, accurately modeling and quantifying the aggregated operating cost across multiple time periods is crucial for upper-level grid operators to execute global economic dispatch and promote the overall economic operation of the system.

[0003] However, existing cost modeling methods are often oversimplified and fail to accurately depict the complex operational characteristics of virtual power plants. For example, some studies employ quadratic equivalent cost functions, but these empirical formulas, based on historical data fitting, are only suitable for depicting the costs of homogeneous distributed resources and lack generalization ability in aggregation scenarios involving various heterogeneous resources. For multi-period scenarios, considering the cross-period energy coupling characteristics of energy storage devices and temperature-controlled loads, existing methods often use piecewise linear approximations based on sampling to construct multi-period cost curves. This external approximation method is prone to the curse of dimensionality and cannot fully reflect the nonlinear marginal cost structure of heterogeneous distributed energy sources due to their complex internal physical constraints, thus affecting the accuracy of price signals and upper-level dispatch decisions.

[0004] Furthermore, although recent studies have attempted to introduce deep neural networks to fit high-dimensional complex mappings, traditional deep neural networks, as highly non-convex black-box models, can transform the original scheduling plan into a non-convex optimization problem if directly embedded into distribution network optimization and scheduling. This problem is prone to getting trapped in local optima and cannot provide reliable mathematical guarantees for system operation. At the same time, the detailed parameters and output status of distributed power sources within virtual power plants involve serious data privacy concerns, and traditional methods pose a significant risk of data leakage during data interaction. Summary of the Invention

[0005] Purpose of the invention: To overcome the shortcomings of existing technologies in simultaneously achieving high-precision quantification of multi-time period aggregated costs, global optimal solution, and internal data privacy protection, this invention provides a virtual power plant aggregated operation cost scheduling method based on an input convex neural network. It utilizes implicit feature encoding to ensure data privacy, and accurately fits nonlinear relationships through an input convex neural network while strictly guaranteeing mathematical convexity, overcoming the difficulties of non-convex solutions and ensuring that the multi-time period scheduling scheme possesses global economic optimality. This is of great significance for the efficient, safe, and economical operation of the power distribution network.

[0006] Technical Solution: To achieve the above objectives, this invention provides a virtual power plant aggregated operating cost scheduling method based on an input convex neural network, comprising the following steps:

[0007] S1: Based on a deep neural network, a feature encoding model is constructed to extract features from the output data of distributed power sources inside the virtual power plant, and coded feature data is obtained.

[0008] S2: The acquired encoded feature data is fused with the aggregated power of the virtual power plant to construct an equivalent transformation model of the input convex neural network. Through forward propagation calculation with non-negative weight constraints, a multi-period aggregated operating cost proxy model is obtained with the aggregated power trajectory of the virtual power plant as input and the predicted aggregated operating cost as output.

[0009] S3: The acquired multi-period aggregated operating cost proxy model is used as the operating cost function of the virtual power plant. A multi-period stochastic economic optimization scheduling model of the distribution network is constructed and solved to obtain the optimal aggregated power scheduling instruction of the virtual power plant, and then the scheduling is executed.

[0010] Furthermore, the expression for the feature encoding model in step S1 is as follows:

[0011]

[0012]

[0013]

[0014] In equation (1), Represents a virtual power plant The output of the internal distributed power source is generated through the Monte Carlo scene; The input tensor for the deep neural network; in equation (2), Represents the ReLU activation function; ,in The number of hidden layers in a deep neural network; Representing the Layer weights; Representing the Layer bias, For the first Hidden layers; in equation (3), This is the output layer, representing the coded characteristic data of distributed power sources within the virtual power plant; Representing the deep neural network's first Layer output; These represent the weights of the output layer of a deep neural network. This represents the bias of the output layer of a deep neural network.

[0015] Furthermore, the expression for fusing the encoded feature data with the aggregated power of the virtual power plant in step S2 is as follows:

[0016]

[0017] In equation (4), This is the initial composite input tensor for the input convex neural network; Encode feature data; This is the interface for aggregating power data for virtual power plants; the superscript T represents transpose.

[0018] Furthermore, the expression for the equivalent transformation model of the input convex neural network in step S2 is as follows:

[0019]

[0020] In equation (5), The output of the k-th layer of the input convex neural network, ,in The number of hidden layers in the input convex neural network; and These are the internal transit weight matrix and the skip connection weight matrix of the k-th layer, respectively; This is the bias of the k-th layer.

[0021] Furthermore, the non-negative weight constraint in step S2 is: the internal transitive weight matrix of the constraint. All coefficients are non-negative.

[0022] Furthermore, the expression for the multi-time-period aggregated operating cost proxy model in step S2 is as follows:

[0023]

[0024] In equation (6), The mean squared error is used as the loss function of the input convex neural network to calculate the predicted aggregate running cost of the output layer of the input convex neural network. For the input convex neural network, the first Layer output; and These represent the internal transit weight matrix and the skip connection weight matrix of the output layer of the input convex neural network, respectively. All coefficients are non-negative. This is the bias of the input convex neural network output layer.

[0025] Furthermore, the multi-time-period stochastic economic optimization scheduling model for the distribution network in step S3 includes the following objective function:

[0026]

[0027] In equation (7), the superscript T represents transpose, and the subscript s represents the scene index. Representative scene set, Represents the set of nodes in the distribution network; Represents variables for distribution network optimization; Represents the probability of a scenario. and These represent cost coefficient one and cost coefficient two, respectively.

[0028] Furthermore, the multi-time-period stochastic economic optimization scheduling model for the distribution network in step S3 includes the following constraints:

[0029]

[0030]

[0031]

[0032]

[0033]

[0034]

[0035] In equations (8) to (10), Representative scenarios Downward action on distribution network nodes Optimization variables The equality constraint coefficient matrix is ​​used to characterize the distribution network nodes. Variable coefficients in power balance constraints; Representative scenarios The effect of the aggregated power variable of the virtual power plant The equality constraint coefficient matrix is ​​used to characterize the mapping relationship when the virtual power plant aggregated power injection power balance equation is used. Representing a scene Next node The constant term vector corresponding to the equality constraint is used to characterize the known injection amount or known demand amount on the right-hand side of the nodal power balance equation; The coefficient matrix represents the remaining equality constraints of the distribution network, used to characterize the scenario. The equality constraint part in the voltage and power flow relationships in the lower network; Representing a scene The constant vector corresponding to the equality constraint reflects the externally injected parameters or known quantities related to the scene; The coefficient matrix represents the inequality constraints of the distribution network, which is used to uniformly characterize the upper and lower limits of line power flow constraints, node voltage constraints, gas turbine output constraints, gas turbine ramping constraints, and distributed power output constraints. Representing a scene The upper bound vector corresponding to the lower inequality constraint; The constraint coefficient matrix represents the feasible region of aggregated power from the virtual power plant. The constraint boundary vectors corresponding to the constraint coefficient matrix together constitute the linear inequality constraint of the aggregated power of the virtual power plant, which is used to characterize the feasible domain range of the aggregated power of the virtual power plant; Equations (11)-(12) are the mirror image constraints of the ReLU activation function; Equation (13) is the mirror image constraint of the predicted aggregated operating cost of the input convex neural network output.

[0036] Beneficial Effects: Compared with existing technologies, this invention employs a data-driven proxy model based on an input convex neural network to accurately fit the complex nonlinear relationship between aggregated power and internal operating costs across multiple time periods. This effectively overcomes the difficulty in characterizing the time coupling of flexible resources, thus providing a solid basis for issuing high-precision, refined scheduling instructions for the distribution network. Simultaneously, by applying non-negative constraints to specific hidden layer weights, this invention strictly guarantees the mathematical convexity of the cost proxy model, breaking the non-convexity solution dilemma faced by traditional deep neural networks when integrated into the scheduling framework. This ensures that the final multi-time period scheduling scheme possesses global economic optimality. Furthermore, this invention utilizes a front-end cascaded deep neural network to perform high-dimensional implicit feature encoding on the output status of underlying devices, avoiding direct interaction of raw physical data at the input structure level. This effectively protects data privacy and internal information confidentiality when virtual power plants participate in distribution network scheduling interactions. Attached Figure Description

[0037] Figure 1 This is a flowchart of the method of the present invention;

[0038] Figure 2 This is a structural diagram of the input convex neural network;

[0039] Figure 3 This is a structural diagram of the power distribution network system used in Example 2. Detailed Implementation

[0040] 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.

[0041] Example 1:

[0042] like Figure 1 As shown, this embodiment provides a virtual power plant aggregated operating cost scheduling method based on an input convex neural network, including the following steps:

[0043] S1: Based on a deep neural network, a feature encoding model is constructed to extract features from the output data of distributed power sources inside the virtual power plant, and coded feature data is obtained.

[0044] like Figure 2 As shown in the left box diagram, the expression for the feature encoding model is as follows:

[0045]

[0046]

[0047]

[0048] In equation (1), Represents a virtual power plant The output of the internal distributed power source is generated through the Monte Carlo scene; The input tensor for the deep neural network; in equation (2), Represents the ReLU activation function; ,in The number of hidden layers in a deep neural network; Representing the Layer weights; Representing the Layer bias, For the first Hidden layers; in equation (3), This is the output layer, representing the coded characteristic data of distributed power sources within the virtual power plant; Representing the deep neural network's first Layer output; These represent the weights of the output layer of a deep neural network. This represents the bias of the output layer of a deep neural network.

[0049] S2: The acquired encoded feature data is fused with the aggregated power of the virtual power plant to construct an equivalent transformation model of the input convex neural network. Through forward propagation calculation with non-negative weight constraints, a multi-period aggregated operating cost proxy model is obtained with the aggregated power trajectory of the virtual power plant as input and the predicted aggregated operating cost as output.

[0050] like Figure 2 As shown in the right-hand diagram, the expression for fusing coded feature data with the aggregated power of the virtual power plant is as follows:

[0051]

[0052] In equation (4), This is the initial composite input tensor for the input convex neural network; Encode feature data; This is the interface for aggregating power data for virtual power plants; the superscript T represents transpose.

[0053] The equivalent transformation model of the input convex neural network is expressed as follows:

[0054]

[0055] In equation (5), The output of the k-th layer of the input convex neural network, ,in The number of hidden layers in the input convex neural network; and These are the internal transit weight matrix and the skip connection weight matrix of the k-th layer, respectively; This is the bias of the k-th layer.

[0056] Non-negative weight constraints are defined as: the transitive weight matrix within the constraint. All coefficients are non-negative.

[0057] The expression for the multi-period aggregated operating cost proxy model is as follows:

[0058]

[0059] In equation (6), The mean squared error is used as the loss function of the input convex neural network to calculate the predicted aggregate running cost of the output layer of the input convex neural network. For the input convex neural network, the first Layer output; and These represent the internal transit weight matrix and the skip connection weight matrix of the output layer of the input convex neural network, respectively. All coefficients are non-negative. This is the bias of the input convex neural network output layer.

[0060] S3: Use the acquired multi-period aggregated operating cost proxy model as the operating cost function of the virtual power plant, construct and solve the multi-period stochastic economic optimization scheduling model of the distribution network, obtain the optimal aggregated power scheduling instruction of the virtual power plant, and issue the scheduling to be executed.

[0061] The multi-period stochastic economic optimization scheduling model for distribution networks includes the following objective function:

[0062]

[0063] In equation (7), the superscript T represents transpose, and the subscript s represents the scene index. Representative scene set, Represents the set of nodes in the distribution network; These represent variables for distribution network optimization, including gas turbine output, distributed generation output, line power flow, and load power. Represents the probability of a scenario. and These represent cost coefficient one and cost coefficient two, respectively.

[0064] The multi-period stochastic economic optimization scheduling model for the distribution network includes the following constraints:

[0065]

[0066]

[0067]

[0068]

[0069]

[0070]

[0071] In equations (8) to (10), Representative scenarios Downward action on distribution network nodes Optimization variables The equality constraint coefficient matrix is ​​used to characterize the distribution network nodes. Variable coefficients in power balance constraints; Representative scenarios The effect of the aggregated power variable of the virtual power plant The equality constraint coefficient matrix is ​​used to characterize the mapping relationship when the virtual power plant aggregated power injection power balance equation is used. Representing a scene Next node The constant term vector corresponding to the equality constraint is used to characterize the known injection amount or known demand amount on the right-hand side of the nodal power balance equation; The coefficient matrix represents the remaining equality constraints of the distribution network, used to characterize the scenario. The equality constraint part in the voltage and power flow relationships in the lower network; Representing a scene The constant vector corresponding to the equality constraint reflects the externally injected parameters or known quantities related to the scene; The coefficient matrix represents the inequality constraints of the distribution network, which is used to uniformly characterize the upper and lower limits of line power flow constraints, node voltage constraints, gas turbine output constraints, gas turbine ramping constraints, and distributed power output constraints. Representing a scene The upper bound vector corresponding to the lower inequality constraint; The constraint coefficient matrix represents the feasible region of aggregated power from the virtual power plant. The constraint boundary vectors corresponding to the constraint coefficient matrix together constitute the linear inequality constraint of the aggregated power of the virtual power plant, which is used to characterize the feasible domain range of the aggregated power of the virtual power plant; Equations (11)-(12) are the mirror image constraints of the ReLU activation function; Equation (13) is the mirror image constraint of the predicted aggregated operating cost of the input convex neural network output.

[0072] In this embodiment, the multi-period stochastic economic optimization scheduling model for the distribution network is a linear programming model, which is solved using the simplex method.

[0073] The method of this invention should not only be able to accurately characterize the complex nonlinear mapping relationship between the aggregated power trajectory of the virtual power plant and the internal minimum operating cost, but also overcome the non-convex solution dilemma brought about by the integration of traditional neural networks into distribution network scheduling in terms of mathematical mechanism, so as to provide the system with a scheduling decision basis with global optimality guarantee, and effectively protect the internal data privacy of the virtual power plant.

[0074] Example 2:

[0075] To verify the effectiveness of the method of the present invention, this embodiment applies the method in a practical manner, selecting modified real distributed energy device data. The parameters of the distributed energy resources are shown in Tables 1, 2, 3, and 4.

[0076] Table 1 Energy Storage Battery Parameters

[0077]

[0078] Table 2 Delayed Load Parameters

[0079]

[0080] Table 3 Temperature control load parameters

[0081]

[0082] Table 4 Gas Turbine Parameters

[0083]

[0084] The parameter settings for the input convex neural network are shown in Table 5.

[0085] Table 5 Input Parameters of Convex Neural Network

[0086]

[0087] In this embodiment, the test results of two virtual power plants are shown. The average percentage error between the cost value obtained by inputting the convex neural network aggregated operating cost proxy model and the actual cost is shown in Table 6.

[0088] Table 6. Average Error Percentage

[0089]

[0090] As shown in Table 6, the average error of the input convex neural network surrogate model constructed in this invention is extremely low on the training set for different virtual power plants, and the average error is still controlled within 0.85% when facing unknown test set data. The test set error curve closely follows the training set error curve, and the difference between the two is minimal. This indicates that the trained cost surrogate model can not only approximate the complex nonlinear mapping relationship between the multi-time period aggregated power trajectory and the internal minimum operating cost with extremely high accuracy, but also has strong generalization ability, avoiding the overfitting phenomenon that is prone to occur in traditional neural networks.

[0091] To verify the effectiveness of this invention in actual power distribution network dispatching, this embodiment underwent simulation testing on a modified power distribution network system, such as... Figure 3As shown, the test system includes 33 nodes, numbered 1 to 33. The final comparison results of power grid operation costs are shown in Table 7.

[0092] Table 7 Operating Cost Results

[0093]

[0094] As shown in Table 7, the distribution network operating cost obtained using the input convex neural network surrogate model proposed in this invention is very close to the actual cost calculated by the accurate model. This indicates that the scheduling method of this invention can accurately approximate the actual operating state of the system and has excellent practical application effects in multi-time-period economic scheduling of distribution networks.

[0095] As can be seen from this embodiment, in scenarios where virtual power plants containing various heterogeneous distributed energy sources are connected to the distribution network, the multi-period aggregated operating cost scheduling method based on input convex neural networks provided by this invention can accurately quantify aggregated operating costs and ensure the mathematical convexity of the model while strictly protecting internal data privacy. This enables grid dispatchers to efficiently formulate economic dispatch plans with global optimality, providing significant practical engineering support for improving the accuracy, solution efficiency, and overall operational economy of multi-period economic dispatching in distribution networks.

Claims

1. A virtual power plant aggregated operation cost scheduling method based on an input convex neural network, characterized in that, Includes the following steps: S1: Based on a deep neural network, a feature encoding model is constructed to extract features from the output data of distributed power sources inside the virtual power plant, and coded feature data is obtained. S2: The acquired encoded feature data is fused with the aggregated power of the virtual power plant to construct an equivalent transformation model of the input convex neural network. Through forward propagation calculation with non-negative weight constraints, a multi-period aggregated operating cost proxy model is obtained with the aggregated power trajectory of the virtual power plant as input and the predicted aggregated operating cost as output. S3: The acquired multi-period aggregated operating cost proxy model is used as the operating cost function of the virtual power plant. A multi-period stochastic economic optimization scheduling model of the distribution network is constructed and solved to obtain the optimal aggregated power scheduling instruction of the virtual power plant, and then the scheduling is executed.

2. The virtual power plant aggregated operation cost scheduling method based on an input convex neural network according to claim 1, characterized in that, The expression for the feature encoding model in step S1 is as follows: ； ； ； In equation (1), Represents a virtual power plant The output of the internal distributed power source is generated through the Monte Carlo scene; The input tensor for the deep neural network; in equation (2), Represents the ReLU activation function; ,in The number of hidden layers in a deep neural network; Representing the Layer weights; Representing the Layer bias, For the first Hidden layers; in equation (3), This is the output layer, representing the coded characteristic data of distributed power sources within the virtual power plant; Representing the deep neural network's first Layer output; These represent the weights of the output layer of a deep neural network. This represents the bias of the output layer of a deep neural network.

3. The virtual power plant aggregated operation cost scheduling method based on an input convex neural network according to claim 2, characterized in that, The expression for fusing the encoded feature data with the aggregated power of the virtual power plant in step S2 is as follows: ； In equation (4), This is the initial composite input tensor for the input convex neural network; Encode feature data; This is the interface for aggregating power data for virtual power plants; the superscript T represents transpose.

4. The virtual power plant aggregated operation cost scheduling method based on an input convex neural network according to claim 3, characterized in that, The expression for the equivalent transformation model of the input convex neural network in step S2 is as follows: ； In equation (5), The output of the k-th layer of the input convex neural network, ,in The number of hidden layers in the input convex neural network; and These are the internal transit weight matrix and the skip connection weight matrix of the k-th layer, respectively; This is the bias of the k-th layer.

5. The virtual power plant aggregated operation cost scheduling method based on an input convex neural network according to claim 4, characterized in that, The non-negative weight constraint in step S2 is: the weight transfer matrix inside the constraint. All coefficients are non-negative.

6. The virtual power plant aggregated operation cost scheduling method based on an input convex neural network according to claim 5, characterized in that, The expression for the multi-time period aggregated operating cost proxy model in step S2 is as follows: ； In equation (6), The mean squared error is used as the loss function for the predicted aggregation operation cost of the output layer of the input convex neural network. For the input convex neural network, the first Layer output; and These represent the internal transit weight matrix and the skip connection weight matrix of the output layer of the input convex neural network, respectively. All coefficients are non-negative; This is the bias of the input convex neural network output layer.

7. The virtual power plant aggregated operation cost scheduling method based on an input convex neural network according to claim 6, characterized in that, The multi-period stochastic economic optimization scheduling model for the distribution network in step S3 includes the following objective function: ； In equation (7), the superscript T represents transpose, and the subscript s represents the scene index. Representative scene set, Represents the set of nodes in the distribution network; Represents variables for distribution network optimization; Represents the probability of a scenario. and These represent cost coefficient one and cost coefficient two, respectively.

8. The virtual power plant aggregated operation cost scheduling method based on an input convex neural network according to claim 7, characterized in that, The multi-period stochastic economic optimization scheduling model for the distribution network in step S3 includes the following constraints: ； ； ； ； ； ； In equations (8) to (10), Representative scenarios Downward action on distribution network nodes Optimization variables The equality constraint coefficient matrix is ​​used to characterize the distribution network nodes. Variable coefficients in power balance constraints; Representative scenarios The effect of the aggregated power variable of the virtual power plant The equality constraint coefficient matrix is ​​used to characterize the mapping relationship when the virtual power plant aggregated power injection power balance equation is used. Representing a scene Next node The constant term vector corresponding to the equality constraint is used to characterize the known injection amount or known demand amount on the right-hand side of the nodal power balance equation; The coefficient matrix represents the remaining equality constraints of the distribution network, used to characterize the scenario. The equality constraint part in the voltage and power flow relationships in the lower network; Representing a scene The constant vector corresponding to the equality constraint reflects the externally injected parameters or known quantities related to the scene; The coefficient matrix represents the inequality constraints of the distribution network, which is used to uniformly characterize the upper and lower limits of line power flow constraints, node voltage constraints, gas turbine output constraints, gas turbine ramping constraints, and distributed power output constraints. Representing a scene The upper bound vector corresponding to the lower inequality constraint; The constraint coefficient matrix represents the feasible region of aggregated power from the virtual power plant. The constraint boundary vectors corresponding to the constraint coefficient matrix together constitute the linear inequality constraint of the aggregated power of the virtual power plant, which is used to characterize the feasible domain range of the aggregated power of the virtual power plant; Equations (11)-(12) are the mirror image constraints of the ReLU activation function; Equation (13) is the mirror image constraint of the predicted aggregated operating cost of the input convex neural network output.

Images