DPVs fast active power regulation method based on optimal dynamic droop coefficient

Abstract

The invention relates to a DPVs fast active power adjustment method based on an optimal dynamic droop coefficient, and the method comprises the steps: selecting adjustable capacity and response time as clustering features, and carrying out the clustering division of DPVs through employing a U-k-means algorithm; constructing a frequency deviation-optimal droop coefficient model considering adjustment performance and network loss based on the divided DPVs cluster body; gAT is applied to rapidly solve an optimal droop coefficient adjustment strategy under various frequency deviation conditions in advance, active response of DPVs cluster optimal droop coefficient adjustment when a frequency event occurs is ensured, and rapid adjustment of DPVs active power output is realized according to a power-frequency relationship. The method has the advantages of accuracy and high efficiency in the aspect of realizing rapid active power regulation based on the droop coefficient by using massive DPVs.

Description

Technical Field

[0001] The invention relates to a DPVs fast active power regulation method based on an optimal dynamic droop coefficient, and belongs to the technical field of smart grids. Background Art

[0002] The deployment of renewable energy in the power system is the only way to achieve carbon neutrality. With the advantages of low maintenance cost, high energy conversion rate and long life, distributed photovoltaics (DPVs) have become a major component of renewable energy, and their penetration rate in the power grid continues to rise. Connected to the distribution network. On the one hand, massive DPVs replace traditional generators, resulting in insufficient frequency support capacity of the power system; on the other hand, DPVs based on inverter interaction have fast active power regulation response performance. Therefore, it is imperative to develop and utilize the fast active power regulation potential of massive distributed photovoltaics.

[0003] However, directly controlling a large number of dispersed distributed resources will cause dimensionality disasters and huge computational complexity. To this end, the present invention introduces clustering technology to classify DPVs with similar characteristics into the same group, and then treats them as an entity to participate in the rapid active power regulation of the power system, which can greatly improve the computational efficiency. At the same time, more and more studies have realized that it is of great significance to realize the rapid active power regulation of distributed resources based on the droop coefficient strategy. However, there are few researches on the development of DPVs droop coefficient regulation at present, and the existing research ignores the influence of network flow in the process of designing the droop coefficient, which not only leads to inaccurate response to the regulation instruction, but also increases the risk of voltage over-limit. In addition, the existing droop coefficient regulation strategy must be calculated by the control center every time a frequency event occurs before being issued to DPVs. Under the timeliness requirement of the rapid regulation service, when frequency events occur frequently, this instruction issuance process has extremely high requirements on communication resources, and the communication delay problem cannot be avoided. Moreover, such a response process is passive, which does not match the active response mechanism based on droop coefficient regulation. To make up for the above shortcomings, the present invention constructs a frequency deviation-optimal droop coefficient model to achieve accurate and reliable response to regulation instructions from the perspective of power flow optimization. Taking into account the initiative and timeliness requirements of the rapid active power regulation response based on the droop coefficient, the present invention uses a graph attention network (GAT) that introduces an attention mechanism to aggregate neighbor information in the topological space according to its importance to quickly solve the droop coefficient optimization model under various frequency deviations in advance, and obtain the optimal droop coefficient adjustment strategy for distributed photovoltaics. Finally, the active power output regulation of DPVs is realized according to the power-frequency relationship. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method for actively regulating and controlling the active power of massive distributed photovoltaic aggregates based on the optimal dynamic droop coefficient. Taking into account that directly regulating massive distributed photovoltaics will lead to dimensionality disasters and severe computational burdens, clustering of distributed photovoltaics is achieved based on the Uk-means algorithm. At the same time, in order to accurately and safely meet the regulation instructions issued by the balancing mechanism, a frequency deviation-optimal droop coefficient model that takes into account the regulation performance and transmission loss is constructed from the perspective of power flow optimization, and the optimal droop coefficient adjustment strategy under various frequency deviation conditions is quickly calculated in advance through the GAT algorithm to ensure active response when frequency events occur.

[0005] To achieve the above object, the present invention adopts the following technical solution: a DPVs fast active power regulation method based on optimal dynamic droop coefficient, comprising the following steps: Step 1: Directly controlling massive DPVs will cause huge optimization dimensions and severe computing burden. Based on the business needs of fast active power regulation, we select adjustable capacity and response time as clustering features and apply the Uk-means algorithm to cluster massive DPVs. Step 2: In order to accurately and safely meet the regulation tasks issued by the balancing agency, a frequency deviation-optimal droop coefficient model considering the regulation performance of the DPVs cluster and the active power transmission loss is constructed from the perspective of power flow optimization; Step 3: Considering the high efficiency of graph neural networks in solving power flow related problems, and GAT can give greater weight to important features in neighbor information by introducing the attention mechanism, and has stronger expression performance, GAT is used to quickly solve the optimal droop coefficient adjustment strategy of DPVs under various frequency deviations in advance.

[0006] Furthermore, step 1.1. Establish the clustering characteristics of massive distributed photovoltaics. In view of the business needs of fast active power regulation, the adjustable capacity of DPVs is selected from the perspective of the active power adjustable margin and rapidity of regulation of DPVs. and response time As clustering features, , Indicates the initial active power output value of distributed photovoltaics, Step 1.2. Apply the Uk-means algorithm to cluster massive distributed photovoltaics. (1) Input a busbar Clustering feature data of distributed photovoltaic , ; and, due to the difference in the order of magnitude of the characteristic data, the range method is used to normalize the data. (1), (2) Deviation threshold for updating given cluster center , maximum number of iterations And the maximum number of clusters ; Set the initial value of the number of iterations respectively , the initial value of the number of clusters , the initial value of the cluster center , the data belongs to the cluster The initial value of the probability and the initial values ​​of the learning parameters , (3) Calculate data according to formula (2) Belongs to a cluster , (2), (4) Update the learning parameters according to formula (3) , (3), (5) Update the data attribution probability according to formula (4) , (4), (6) Update the learning parameters through formula (5) , (5), (7) Update the number of clusters according to formula (6) , and adjust them by equations (7) and (8) respectively and ,if and ,Adjustment , (6), (7), (8), (8) Update the cluster center according to formula (8) , if the maximum updated cluster center is less than the cluster center deviation threshold And the updated number of clusters is less than the given maximum number of clusters , then the final clustering result is obtained, otherwise , repeat steps (3)-(8).

[0007] Furthermore, in step 2, based on the clustered DPVs, in order to accurately and safely meet the regulation tasks issued by the balancing mechanism, a frequency deviation-optimal droop coefficient model considering the regulation performance of the DPVs cluster and the active power transmission loss is constructed from the perspective of power flow optimization, as follows: Step 2.1. Targeting For a power grid with 1 busbars, the comprehensive optimization objective function is established as: (9), Among them, the decision variable is the droop coefficient of the DPVs cluster , For bus Department The number of distributed photovoltaic clusters at the moment, To adjust the performance index, the adjustable capacity of the cluster and response time Decide, , Indicates busbar DPV clusters exist The active power output value at each moment, and Respectively represent busbar and busbar Department The voltage vector at time , and Respectively represent busbar , The resistance and reactance between represents the vector modulus, and is the weight coefficient, Step 2.2. For the above optimization model, the corresponding constraints are: (10), (11), (12), (13), (14), (15), (16), (17), (18), Among them, formula (10) is the droop constraint, , They represent the upper and lower bounds of the droop coefficient adjustment, respectively. Equations (11)-(13) are the operating constraints of the distributed photovoltaic cluster. for Moment bus DPV clusters The active power output value of Indicates busbar DPV clusters The initial output active power, , for The system frequency at the moment, is the frequency regulation dead zone, is the nominal frequency of the system, For bus Cluster The rated active power value of The rated active power values ​​of distributed photovoltaics are summed up to get: For distributed photovoltaic The rated active power value of For bus DPV clusters exist The upper limit of active power output at the moment, For bus DPV clusters exist The reactive power output value at the moment, is the power factor value of the cluster, equations (14) and (15) are the power flow constraints, , They are Moment bus The active power and reactive power of the load at , for Moment bus ， The voltage value at , Respectively represent busbar , The conductance and susceptance between for Moment bus ， The phase angle difference between them is given by equation (16), which is the bus voltage constraint. and are the upper and lower limits of the bus voltage amplitude, equations (17) and (18) are the power balance constraints, Indicates the active power value of the root node interacting with the upper grid, that is, the regulation demand issued by the balancing agency The corresponding active power value is changed from the initial value and change value Decide.

[0008] Furthermore, in step 3, considering the high efficiency of graph neural networks in solving power flow related problems, GAT can give greater weight to important features in neighbor information by introducing the attention mechanism, and has stronger expression performance. GAT is used to quickly solve the droop coefficient optimization model under various frequency deviation conditions in advance, and the optimal droop coefficient adjustment strategy of DPVs is obtained, as follows: Introducing a node feature , and connection relationship The diagram is used to describe the power grid information, where and Respectively represent the number and characteristic dimension of buses in the power grid, the initial node characteristics , Contains the active power and reactive power of the load, the upper and lower bounds of the adjustable droop coefficient of the cluster and the corresponding regulation performance indicators, is a binary matrix. When the busbars are directly connected, the matrix element is 1, otherwise it is 0. Through the graph attention layer, the input features Advanced Features , The conversion process is as follows. First, in order to ensure that GAT has sufficient expression performance, a learnable linear transformation is introduced. , ; Then, a dynamic masked self-attention strategy is applied , Get the attention coefficient between nodes , calculated by the following formula (19), in, Representation Node Characteristics for busbars The importance of the node, according to Only for busbar The node's neighbors perform the calculation of the attention coefficient. LeakyReLU is a nonlinear activation function. represents transpose, Indicates a connection operation. Apply the softmax function to normalize the attention coefficient: (20), in, express The set of all neighbor nodes, Expanding the above formula, the attention coefficient based on the attention mechanism is expressed as: (twenty one), Then, each node linearly combines the features of its neighbors according to the obtained attention coefficient and passes a nonlinear activation function , complete the update of node features, as shown below (twenty two) , Since the multi-head attention mechanism helps to stabilize the learning process of the self-attention mechanism, we set An independent attention mechanism is used, and these features are averaged to finally obtain the following graph attention network output feature representation: (twenty three), in, Indicated by The normalized attention coefficient calculated by the attention mechanism, is the corresponding weight matrix for linear transformation, and then the calculation results of all individual attention mechanisms are averaged. Therefore, the output feature dimension of each node in the final graph attention layer is .

[0009] Finally, the constructed graph attention network learns the mapping relationship between system operation status, grid topology and optimal droop coefficient, and pre-calculates the optimal droop coefficient adjustment strategy of distributed photovoltaic under any frequency deviation. After that, the power-frequency relationship Realize DPVs active power output regulation.

[0010] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the program, the DPVs fast active power regulation method based on the optimal dynamic droop coefficient is implemented.

[0011] A computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implements the DPVs fast active power regulation method based on the optimal dynamic droop coefficient Compared with the prior art, the advantages of the present invention are as follows: The present invention takes into account that directly regulating a large number of distributed photovoltaics will cause huge optimization dimensions and serious computational burdens. Starting from the needs of fast active power regulation services, the adjustable capacity and response speed of distributed photovoltaics are selected as clustering features, and the Uk-means algorithm is applied to achieve clustering of massive distributed photovoltaics, which greatly reduces the optimization dimension and improves the computational efficiency. Different from the unsafe and inaccurate power regulation caused by usually ignoring the power flow constraints in the prior art, the present invention considers the influence of network power flow in the process of designing the droop coefficient, see formulas (14)-(18), and considers the power flow constraints in the process of establishing the frequency deviation-optimal droop coefficient model, which can not only ensure the accurate response to the issued regulation instructions, but also ensure the bus voltage safety during the response process. Furthermore, the present invention realizes the efficient solution of the optimal droop coefficient regulation strategy of DPVs through the GAT method. GAT can aggregate information according to the importance of neighbor features in the topological space, and is more advanced and intelligent than the existing data-driven technology in solving optimization problems related to power flow. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a flow chart of a DPVs fast active power regulation method based on an optimal dynamic droop coefficient provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0013] In order to deepen the understanding of the present invention, the present embodiment is described in detail below with reference to the accompanying drawings.

[0014] Example: Figure 1 As shown, the purpose of the present invention is to provide a DPVs fast active power regulation method based on the optimal dynamic droop coefficient. In order to avoid the dimensionality disaster and heavy computational burden problems caused by directly regulating massive distributed photovoltaics, the present invention implements clustering of distributed photovoltaics based on the Uk-means algorithm. In order to accurately and safely achieve the active power regulation requirements of massive distributed photovoltaics specified by the balancing mechanism, based on the clustering results of distributed photovoltaics, from the perspective of power flow optimization, the present invention constructs a frequency deviation-optimal droop coefficient model that takes into account the active power regulation performance and transmission loss of DPVs clusters, and quickly solves through GAT, pre-calculates the optimal droop coefficient regulation strategy of distributed photovoltaics under various frequency deviations, ensures active response when frequency events occur, and realizes the output regulation of DPVs active power according to the power-frequency relationship.

[0015] The invention discloses a DPVs fast active power regulation method based on an optimal dynamic droop coefficient, comprising the following steps: Step 1: In order to realize the clustering of massive distributed photovoltaics, the adjustable capacity and response time are selected as clustering features, and the Uk-means algorithm is applied to cluster the massive DPVs; (1) Input a busbar Clustering feature data of distributed photovoltaic , ; and, due to the difference in the order of magnitude of the characteristic data, the range method is used to normalize the data. (1), (2) Deviation threshold for updating given cluster center , maximum number of iterations And the maximum number of clusters ; Set the initial value of the number of iterations respectively , the initial value of the number of clusters , the initial value of the cluster center , the data belongs to the cluster The initial value of the probability and the initial values ​​of the learning parameters , (3) Calculate data according to formula (2) Belongs to a cluster , (2), (4) Update the learning parameters according to formula (3) , (3), (5) Update the data attribution probability according to formula (4) , (4), (6) Update the learning parameters through formula (5) , (5), (7) Update the number of clusters according to formula (6) , and adjust them by equations (7) and (8) respectively and ,if and ,Adjustment , (6), (7), (8), (8) Update the cluster center according to formula (8) , if the maximum updated cluster center is less than the cluster center deviation threshold And the updated number of clusters is less than the given maximum number of clusters , then the final clustering result is obtained, otherwise , repeat steps (3)-(8).

[0016] Step 2: Based on the clustered DPVs, in order to accurately and safely meet the regulation tasks issued by the balancing agency, a frequency deviation-optimal droop coefficient model considering the regulation performance of the DPVs cluster and the active power transmission loss is constructed from the perspective of power flow optimization: Step 2.1. Targeting For a power grid with 1 busbars, the comprehensive optimization objective function is established as: (9), Among them, the decision variable is the droop coefficient of the DPVs cluster , For bus Department The number of distributed photovoltaic clusters at the moment, To adjust the performance index, the adjustable capacity of the cluster and response time Decide, , Indicates busbar DPV clusters exist The active power output value at each moment, and Respectively represent busbar and busbar Department The voltage vector at time , and Respectively represent busbar , The resistance and reactance between represents the vector modulus, and is the weight coefficient, Step 2.2. For the above optimization model, the corresponding constraints are: (10), (11), (12), (13), (14), (15), (16), (17), (18), Among them, formula (10) is the droop constraint, , They represent the upper and lower bounds of the droop coefficient adjustment, respectively. Equations (11)-(13) are the operating constraints of the distributed photovoltaic cluster. for Moment bus DPV clusters The active power output value of Indicates busbar DPV clusters The initial output active power, , for The system frequency at the moment, is the frequency regulation dead zone, is the nominal frequency of the system, For bus Cluster The rated active power value of The rated active power values ​​of distributed photovoltaics are summed up to get: For distributed photovoltaic The rated active power value of For bus DPV clusters exist The upper limit of active power output at the moment, For bus DPV clusters exist The reactive power output value at the moment, is the power factor value of the cluster, equations (14) and (15) are the power flow constraints, , They are Moment bus The active power and reactive power of the load at , for Moment bus ， The voltage value at , Respectively represent busbar , The conductance and susceptance between for Moment bus ， The phase angle difference between them is given by equation (16), which is the bus voltage constraint. and are the upper and lower limits of the bus voltage amplitude, equations (17) and (18) are the power balance constraints, Indicates the active power value of the root node interacting with the upper grid, that is, the regulation demand issued by the balancing agency The corresponding active power value is changed from the initial value and change value Decide.

[0017] Step 3: Considering the high efficiency of graph neural networks in solving power flow related problems, GAT can give greater weight to important features in neighbor information by introducing the attention mechanism, and has stronger expression performance. GAT is used to quickly solve the droop coefficient optimization model under various frequency deviation conditions in advance, and the optimal droop coefficient adjustment strategy of DPVs is obtained; Introducing a node feature , and connection relationship The diagram is used to describe the power grid information, where and Respectively represent the number and characteristic dimension of buses in the power grid, the initial node characteristics , Contains the active power and reactive power of the load, the upper and lower bounds of the adjustable droop coefficient of the cluster and the corresponding regulation performance indicators, is a binary matrix. When the busbars are directly connected, the matrix element is 1, otherwise it is 0. (2) Through the graph attention layer, the input features Advanced Features , The conversion process is as follows. First, in order to ensure that GAT has sufficient expression performance, a learnable linear transformation is introduced. , ; Then, a dynamic masked self-attention strategy is applied , Get the attention coefficient between nodes , calculated by the following formula (19), in, Representation Node Characteristics for busbars The importance of the node, according to Only for busbar The node's neighbors perform the calculation of the attention coefficient. LeakyReLU is a nonlinear activation function. represents transpose, Indicates a connection operation. Furthermore, in order to avoid the adverse effects of coefficients of different orders of magnitude between nodes on feature aggregation, the softmax function is applied to normalize the attention coefficients: (20), in, express The set of all neighbor nodes, Expanding the above formula, the attention coefficient based on the attention mechanism is expressed as: (twenty one), Then, each node linearly combines the features of its neighbors according to the obtained attention coefficient and passes a nonlinear activation function , complete the update of node features, as shown below (twenty two) , Since the multi-head attention mechanism helps to stabilize the learning process of the self-attention mechanism, we set An independent attention mechanism is used, and these features are averaged to finally obtain the following graph attention network output feature representation: (twenty three), in, Indicated by The normalized attention coefficient calculated by the attention mechanism, is the corresponding weight matrix for linear transformation, and then the calculation results of all individual attention mechanisms are averaged. Therefore, the output feature dimension of each node in the final graph attention layer is .

[0018] Finally, the constructed graph attention network learns the mapping relationship between system operation status, grid topology and optimal droop coefficient, and pre-calculates the optimal droop coefficient adjustment strategy of distributed photovoltaic under any frequency deviation. After that, the power-frequency relationship Realize DPVs active power output regulation.

[0019] (3) In order to verify the effectiveness of GAT, the mean absolute error (MAE), mean absolute percentage error (MAPE) and root mean square error (RMSE) performance indicators were used for testing. (twenty four), (25), (26), in, represents the number of samples in the test set, is the output data dimension of GAT, is the sample label, is the output value of GAT.

[0020] It should be noted that the above embodiments are not intended to limit the protection scope of the present invention, and equivalent changes or substitutions made on the basis of the above technical solutions all fall within the protection scope of the claims of the present invention.

Claims

1. A DPVs fast active power regulation method based on an optimal dynamic droop coefficient, characterized in that: The method comprises the following steps: Step 1: Directly controlling massive DPVs will cause huge optimization dimensions and severe computing burden. Based on the business needs of fast active power regulation, we select adjustable capacity and response time as clustering features and apply the Uk-means algorithm to cluster massive DPVs. Step 2: Construct a frequency deviation-optimal droop coefficient model considering the regulation performance of DPVs cluster and active power transmission loss; Step 3: Use GAT to quickly solve the optimal droop coefficient adjustment strategy of DPVs under various frequency deviation conditions in advance.

2. The DPVs fast active power regulation method based on the optimal dynamic droop coefficient according to claim 1 is characterized in that: In step 1, the Uk-means algorithm is used to cluster massive DPVs, as follows: Step 1.

1. Establish clustering characteristics of massive distributed photovoltaics. In view of the business needs of fast active power regulation, the adjustable capacity of DPVs is selected from the perspective of the active power adjustable margin and rapidity of regulation of DPVs. and response time As clustering features, , Indicates the initial active power output value of distributed photovoltaics, Step 1.

2. Apply the Uk-means algorithm to cluster massive distributed photovoltaics. (1) Input a busbar Clustering feature data of distributed photovoltaic , ; Use the range method to normalize the data, （1）， (2) Deviation threshold for updating given cluster center , maximum number of iterations And the maximum number of clusters ; Set the initial value of the number of iterations respectively , the initial value of the number of clusters , the initial value of the cluster center , the data belongs to the cluster The initial value of the probability and the initial values ​​of the learning parameters , (3) Calculate data according to formula (2) Belongs to a cluster , （2）， (4) Update the learning parameters according to formula (3) , （3）， (5) Update the data attribution probability according to formula (4) , （4）， (6) Update the learning parameters through formula (5) , （5）， (7) Update the number of clusters according to formula (6) , and adjust them by equations (7) and (8) respectively and ,if and ,Adjustment , （6）， （7）， （8）， (8) Update the cluster center according to formula (8) , if the maximum updated cluster center is less than the cluster center deviation threshold And the updated number of clusters is less than the given maximum number of clusters , then the final clustering result is obtained, otherwise , repeat steps (3)-(8).

3. The DPVs fast active power regulation method based on the optimal dynamic droop coefficient according to claim 1 is characterized in that: Step 2: Based on the clustered DPVs, a frequency deviation-optimal droop coefficient model is constructed that takes into account the regulation performance of the DPVs cluster and the active power transmission loss. The details are as follows: Step 2.

1. For a power grid with 1 busbars, the comprehensive optimization objective function is established as: （9）， Among them, the decision variable is the droop coefficient of the DPVs cluster , For bus Department The number of distributed photovoltaic clusters at the moment, To adjust the performance index, the adjustable capacity of the cluster and response time Decide, , Indicates busbar DPV clusters exist The active power output value at each moment, and Respectively represent busbar and busbar Department The voltage vector at time , and Respectively represent busbar , The resistance and reactance between represents the vector modulus, and is the weight coefficient, Step 2.

2. For the above optimization model, the corresponding constraints are: （10）， （11）， （12）， （13）， （14）， （15）， （16）， （17）， （18）， Among them, formula (10) is the droop constraint, , They represent the upper and lower bounds of the droop coefficient adjustment, respectively. Equations (11)-(13) are the operating constraints of the distributed photovoltaic cluster. for Moment bus DPV clusters The active power output value of Indicates busbar DPV clusters The initial output active power, , for The system frequency at the moment, is the frequency regulation dead zone, is the nominal frequency of the system, For bus Cluster The rated active power value of The rated active power values ​​of distributed photovoltaics are summed up to get: For distributed photovoltaic The rated active power value of For bus DPV clusters exist The upper limit of active power output at the moment, For bus DPV clusters exist The reactive power output value at the moment, is the power factor value of the cluster, equations (14) and (15) are the power flow constraints, , They are Moment bus The active power and reactive power of the load at , for Moment bus ， The voltage value at , Respectively represent busbar , The conductance and susceptance between for Moment bus ， The phase angle difference between them is given by equation (16), which is the bus voltage constraint. and are the upper and lower limits of the bus voltage amplitude, equations (17) and (18) are the power balance constraints, Indicates the active power value of the root node interacting with the upper grid, that is, the regulation demand issued by the balancing agency The corresponding active power value is changed from the initial value and change value Decide.

4. The DPVs fast active power regulation method based on the optimal dynamic droop coefficient according to claim 1 is characterized in that: In step 3, GAT is used to quickly solve the droop coefficient optimization model under various frequency deviation conditions in advance to obtain the optimal droop coefficient adjustment strategy for DPVs, as follows: Introducing a node feature , and connection relationship The diagram is used to describe the power grid information, where and Respectively represent the number and characteristic dimension of buses in the power grid, the initial node characteristics ， Contains the active power and reactive power of the load, the upper and lower bounds of the adjustable droop coefficient of the cluster and the corresponding regulation performance indicators, is a binary matrix. When the busbars are directly connected, the matrix element is 1, otherwise it is 0. Through the graph attention layer, the input features Advanced Features , The conversion process is as follows. First, in order to ensure that GAT has sufficient expression performance, a learnable linear transformation is introduced. , ; Then, a dynamic masked self-attention strategy is applied , Get the attention coefficient between nodes , calculated by the following formula （19）， in, Representation Node Characteristics for busbars The importance of the node, according to Only for busbar The node's neighbors perform the calculation of the attention coefficient. LeakyReLU is a nonlinear activation function. represents transpose, Indicates a connection operation. Apply the softmax function to normalize the attention coefficient: （20）， in, express The set of all neighbor nodes, Then, each node linearly combines the features of its neighbors according to the obtained attention coefficient and passes a nonlinear activation function , complete the update of node features, as shown below （21） ， Since the multi-head attention mechanism helps to stabilize the learning process of the self-attention mechanism, we set An independent attention mechanism is used, and these features are averaged to finally obtain the following graph attention network output feature representation: （22）， in, Indicated by The normalized attention coefficient calculated by the attention mechanism, is the corresponding weight matrix for linear transformation. Then, the calculation results of all individual attention mechanisms are averaged. Therefore, the output feature dimension of each node in the final graph attention layer is , Finally, the mapping relationship between system operation status, grid topology and optimal droop coefficient is learned through the constructed graph attention network, and the optimal droop coefficient adjustment strategy of distributed photovoltaic under any frequency fluctuation range is pre-calculated. Then, through the power-frequency relationship Realize DPVs active power output regulation.

5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the DPVs fast active power regulation method based on the optimal dynamic droop coefficient as described in any one of claims 1 to 4 above is implemented.

6. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the computer instruction is executed by the processor, the DPVs fast active power regulation method based on the optimal dynamic droop coefficient as described in any one of claims 1-4 is implemented.

Images