Active power distribution network double-layer collaborative voltage regulation and control method based on LSTM neural network

Abstract

The invention provides an active power distribution network double-layer collaborative voltage regulation and control method based on an LSTM neural network. According to the method, self-adaptive partition of the active power distribution network is realized by using a community division algorithm, an active power distribution network double-layer cooperation and photovoltaic power supply descending order control model is established, and rapid regulation and control of off-limit voltage of the power distribution network are realized through a reactive compensation / active power reduction two-stage control mode. The universality of centralized voltage control and the rapidity of cluster voltage control can be considered at the same time, and efficient regulation and control of distributed photovoltaic high-proportion access power distribution network voltage are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of voltage regulation of active distribution networks, and in particular to a double-layer collaborative voltage regulation method for active distribution networks based on LSTM neural networks. Background Art

[0002] With the widespread deployment of renewable energy, especially the high penetration of distributed photovoltaic power generation systems in distribution networks, the voltage regulation of distribution networks faces unprecedented challenges. Traditional centralized voltage control methods, due to the need to coordinate the control variables of the entire network, are increasingly unable to meet the rapidly changing control needs and time scale requirements. In order to meet this challenge, cluster voltage control methods have emerged. By dividing the distribution network into multiple sub-areas, it realizes localized voltage regulation in each area, thereby achieving the purpose of simplified, decentralized and rapid regulation.

[0003] However, the existing cluster voltage control method has the following limitations when dealing with voltage deep over-limit: when the voltage of the target node in the cluster exceeds the limit, and the reactive power adjustable margin of the photovoltaic power source in the cluster is insufficient, it is difficult for the photovoltaic power source to regulate the voltage of the target node. Even if the node voltage in the cluster is normal and the photovoltaic power source has sufficient reactive power margin, the reactive power margin cannot be effectively utilized, resulting in the failure to fully utilize the reactive power adjustable margin of the entire network.

[0004] The above problem is essentially an imbalance between the voltage over-limit level and the adjustable margin of photovoltaic power sources. To this end, the present invention proposes a two-layer collaborative voltage control method for active distribution networks based on LSTM neural networks, which combines the advantages of cluster voltage control and centralized voltage control and solves the problem of limited "autonomy" of over-limit voltage. Summary of the invention

[0005] The present invention proposes a double-layer collaborative voltage control method for an active distribution network based on an LSTM neural network. The method can realize rapid control of over-limit voltage of the distribution network, taking into account the universality of centralized voltage control and the rapidity of cluster voltage control.

[0006] S1: Obtain incomplete measurement data of the historical time section of the distribution network, including: photovoltaic equivalent grid-injected power data, node voltage and voltage-power sensitivity data, and physical topology of the distribution network.

[0007] S2: Establish a voltage-power sensitivity online mapping model based on LSTM neural network, including data preparation and regression learning; the prepared data is obtained by normalizing the historical data set with multi-input-multi-output features, and the regression learning consists of initialization parameters, forward propagation, calculation of loss function, back propagation and iterative training steps.

[0008] S3: A two-layer coordinated voltage control model for active distribution networks based on distributed photovoltaic descending order is established, including inner cluster voltage control and outer centralized voltage control, and voltage control constraints are set based on the Jacobian inverse transformation equation of the power flow.

[0009] S4: When the inner cluster voltage is controlled, the regional node voltage is evaluated in real time and classified in descending order to select the dominant node in the cluster. The voltage-reactive sensitivity vector corresponding to the dominant node cluster is regressively learned and classified in descending order to select the reactive controlled photovoltaic. The actual reactive compensation amount of the controlled photovoltaic is calculated by combining the dominant node voltage over-limit, voltage control constraints and the current controlled photovoltaic reactive adjustable margin, and the dominant node over-limit voltage is regulated.

[0010] S5: Update the voltage of the cluster boundary nodes and the inter-cluster transmission power, and return to S4 to iterate and execute the next round of voltage control. If the voltage of the leading nodes of each cluster returns to normal, the control is terminated; otherwise, if the voltage of the leading node exceeds the limit and is separated from the controlled photovoltaic area, the inner cluster voltage control is exited and the outer centralized voltage control is executed.

[0011] S6: During the outer centralized voltage control, it includes a two-stage control mode of reactive power compensation and active power reduction. The node voltages of the entire network and the corresponding voltage-reactive power sensitivity are arranged in descending order. The dominant node and the controlled photovoltaic in the reactive power compensation stage are selected, the actual reactive power compensation amount of the controlled photovoltaic is calculated, and the over-limit voltage of the dominant node is adjusted sequentially. S6 is repeatedly executed until the adjustable reactive power margin of all controlled photovoltaics is exhausted.

[0012] S7: Set the power factor reduction threshold, determine the voltage over-limit situation according to the current photovoltaic operation power factor, arrange the node voltages and corresponding voltage-active sensitivity of the entire network in descending order, select the leading node and controlled photovoltaic in the active power reduction stage, reduce the active power and release the reactive power regulation margin at the same time, sequentially regulate the over-limit voltage of the leading node, repeat S7 until the voltage of the entire network does not exceed the limit, and output the voltage distribution and power adjustment results.

[0013] In a possible implementation, the voltage-power sensitivity online mapping model based on the LSTM neural network is:

[0014] Prepare the data:

[0015]

[0016] Where: f1, f2 and f3 represent the data training set respectively; V t represents the node voltage matrix at time t; S eint represents the photovoltaic grid-injected power matrix at time t; S ent Represents the voltage-power sensitivity matrix at time t.

[0017] Return to Learning:

[0018] 1) Initialization parameters: set the LSTM neural network weight matrix and bias parameters;

[0019] 2) Training the LSTM neural network model: input the normalized training data set and perform forward propagation calculation;

[0020] Gate calculation:

[0021]

[0022] Where: W i , W l and W o Represents the weight matrices of the input gate, forget gate, and output gate respectively; b i 、b l and b o Represents the bias term of the corresponding gate, I t , L t and O t represents the output of the corresponding gate at time t; x t represents the input at time t, h t-1 represents the hidden state output at time t-1; σ represents the sigmoid function;

[0023] Memory unit calculation:

[0024] c t = l t ec t-1 +i t etanh(W c [h t-1 ,x t ]+b c ) (17)

[0025] Where: W c represents the weight matrix of the memory unit; b c represents the bias term of the memory unit; c t Represents the state of the memory unit at time t; e represents the element-by-element multiplication rule.

[0026] 3) Calculate the loss function: Calculate the global error based on the network output value and the true value of the training set;

[0027]

[0028] Where: E represents the global error of the system; N represents the number of samples; p represents the length of training data; y t Represents the predicted value at time t.

[0029] 4) Back propagation: Use the back propagation algorithm to calculate the gradient loss and update the network parameters;

[0030] 5) Iterative training: monitor the changes in the error function until it converges to the threshold and output the optimal solution of the model.

[0031] In one possible implementation, the voltage control constraint is set based on the power flow Jacobian inverse transformation equation:

[0032] The inverse transformation equation of the power flow Jacobian is:

[0033]

[0034] Where: Δδ and ΔV represent the changes in node voltage phase angle and amplitude; V represents the node voltage amplitude; J represents the Jacobian matrix; X, K, Φ and Ψ represent the phase angle-active power, phase angle-reactive power, voltage-active power and voltage-reactive power sensitivity sub-matrices in the voltage-power sensitivity matrix; ΔP and ΔQ represent the changes in active power and reactive power.

[0035] The voltage regulation constraint is:

[0036]

[0037] Where: ΔV i Indicates that the voltage of the dominant node i exceeds the limit; V i 0 represents the voltage of the dominant node i when photovoltaics are not involved in regulation; Φ ij and ij represents the voltage-active power and voltage-reactive power sensitivity factors of the dominant node i relative to the controlled photovoltaic j; ΔP j and ΔQ j Represents the active and reactive adjustment of the controlled photovoltaic j.

[0038] In a possible implementation, a two-layer coordinated voltage regulation model of an active distribution network integrating distributed photovoltaic descending control is established, and the voltage control of the inner cluster is:

[0039] The regional division modularity function of the inner cluster voltage control is:

[0040]

[0041] Where: A ij represents the weight of the edge between node i and node j, which is 1 when the two nodes are connected and 0 otherwise; k i represents the sum of the weights of the edges connected to node i; m represents the sum of the weights of all edges; c i Represents the cluster assigned to node i. If nodes i and j are in the same cluster, Otherwise it is 0;

[0042] The descending sequence of the dominant nodes of the inner cluster voltage control is:

[0043] δV=[ΔV α ,ΔV β ,…,ΔV γ ,…,ΔV λ ],α,β,γ,λ<n (22)

[0044] Where: α and β represent the node locations where the highest and second highest voltage exceeds the limit.

[0045] Inner cluster voltage control reactive power regulation photovoltaic descending sequence:

[0046]

[0047] Where: and θ represent the locations of the highest and second highest voltage-active power sensitivities corresponding to the PV sources.

[0048] The actual reactive output adjustment of the reactive controlled photovoltaic output of the inner cluster voltage control is:

[0049]

[0050] Where: ΔQ mar,ω and ΔQ dem,ω Represents the reactive power margin of the controlled photovoltaic ω itself and the reactive power demand of the dominant node; S ω , P ω and Q ω It indicates the apparent capacity, active and reactive power output of the controlled photovoltaic ω.

[0051] In a possible implementation, a two-layer coordinated voltage regulation model of the distribution network integrating distributed photovoltaic descending control is established, and the outer layer centralized voltage control is:

[0052] Based on the dominant node voltage evaluation results, the expression of reactive power compensation mode is:

[0053]

[0054] Where: ΔV α are the initial voltage and voltage change of the target node α respectively; ΔQ j is the reactive power change injected into the node by the controlled photovoltaic j; ε represents the number of times the reactive controlled photovoltaic is adjusted.

[0055] The descending sequence of active controlled photovoltaic power under the outer layer centralized voltage control is:

[0056] δΨ=[Ψ α,ω ,Ψ α,ξ ,…,Ψ α,τ ,…,Ψ α,υ ] (26)

[0057] The actual active power reduction adjustment of active power regulation photovoltaic is:

[0058]

[0059] Where: Q' ω Represents the reactive power output after reactive power compensation of the reactive controlled photovoltaic ω; S' ω Represents the usage capacity of reactive controlled photovoltaic ω after reactive compensation; η=Φ αω / Ψ αω Represents the voltage-to-power sensitivity ratio factor.

[0060] Based on the dominant node voltage evaluation result, the expression of active power reduction mode is:

[0061]

[0062] Where: ΔP j , ΔQ j are the changes of active and reactive power injected into the node by the controlled photovoltaic j respectively; χ represents the number of times the active controlled photovoltaic is adjusted. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0064] Figure 1 It is a flow chart of double-layer coordinated voltage control of a distribution network provided by an embodiment of the present invention;

[0065] Figure 2 is a flow chart of a community division algorithm based on a change in modularity function gain provided by an embodiment of the present invention;

[0066] Figure 3 It is a diagram of a double-layer coordinated voltage control structure of a distribution network provided by an embodiment of the present invention;

[0067] Figure 4 It is a schematic diagram of cluster control based on distribution network separation provided by an embodiment of the present invention;

[0068] Figure 5 is a modularity gain variation curve in the distribution network cluster division process provided by an embodiment of the present invention;

[0069] Figure 6 It is a cluster partition binary clustering diagram provided by an embodiment of the present invention;

[0070] Figure 7 It is a voltage distribution diagram of the IEEE 141 node distribution network after reactive power control of the photovoltaic power cluster provided by an embodiment of the present invention;

[0071] Figure 8 It is a voltage distribution diagram of the IEEE 141 node distribution network after the centralized reactive power control of the photovoltaic power source provided by the embodiment of the present invention;

[0072] Fig. 9 It is a voltage distribution diagram of the IEEE 141 node distribution network after the centralized active power control of the photovoltaic power source provided by the embodiment of the present invention;

[0073] Fig.10 This is a schematic diagram of a photovoltaic power source participating in voltage regulation process at 14:15 during the day provided by an embodiment of the present invention;

[0074] Fig.11 It is an IEEE 141-node distribution network topology structure provided by an embodiment of the present invention;

[0075] Fig.12 This is a photovoltaic power output diagram considering regional differences provided by an embodiment of the present invention;

[0076] Fig.13 is a user terminal load power fluctuation curve diagram provided by an embodiment of the present invention;

[0077] Fig.14 It is the voltage distribution of the IEEE 141 node distribution network before photovoltaic participation in regulation provided by an embodiment of the present invention;

[0078] Fig.15 This is the voltage distribution after reactive power compensation of photovoltaic power sources in the IEEE 141-node distribution network provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0079] In order to make the purpose, technical scheme and advantages of the present invention clearer, the following will be described by specific embodiments in conjunction with the accompanying drawings. In the following description, specific details such as specific system structures and technologies are proposed for the purpose of illustration rather than limitation, so as to thoroughly understand the embodiments of the present invention. However, it should be clear to those skilled in the art that the present invention can also be implemented in other embodiments without these specific details.

[0080] In order to make the purpose, technical solutions and advantages of the present invention more clear, specific embodiments will be described below in conjunction with the accompanying drawings.

[0081] Figure 1 The flowchart of the community division algorithm based on the change of modularity function gain provided by the embodiment of the present invention is described in detail as follows:

[0082] 1.1: Initialize the distribution network cluster, treat each node as a separate cluster, and use the formula Calculate the modularity function μ.

[0083] 1.2: For node i, randomly select other nodes j to form a new sub-cluster with it (c i ,c j ), recalculate the modularity function μ, and divide the corresponding nodes with the largest modularity function under all combinations into the same cluster.

[0084] 1.3: Treat the new cluster as an independent node and repeat the second step to form a new cluster division result.

[0085] 1.4: When there are no nodes that can be merged, the clustering process stops and the best clustering result is output.

[0086] Figure 2 The distribution network double-layer coordinated voltage control process provided by the embodiment of the present invention is described in detail as follows:

[0087] In order to maximize the photovoltaic consumption level of the distribution network, based on the control principle of "reactive power compensation first and active power reduction later", reactive power control is divided into cluster reactive power control and centralized reactive power control, and active power reduction control is still carried out in a centralized manner. First, the inner cluster reactive power control is implemented in the target area, and the reactive power adjustable margin of the photovoltaic power source in the area is used to regulate the voltage. When the reactive power adjustable margin in the area is insufficient and the target node voltage is still in an over-limit state, the outer centralized reactive power control is started, and the reactive power adjustable margin of the entire network is used for voltage regulation. If the reactive power adjustable margin of the entire network is insufficient, active power reduction control is performed to release the reactive power regulation margin by reducing the output of the photovoltaic power source, and to regulate the voltage.

[0088] Its advantage is that by giving priority to reactive power compensation, the role of photovoltaic power sources in voltage regulation can be fully utilized. When the reactive power adjustable margin is sufficient, the reduction of active power is avoided, thereby ensuring the stability and economy of power supply. Through hierarchical control, the voltage over-limit problem can be responded to quickly and regulated in time. Through the control strategy combining cluster control and centralized control, the voltage over-limit problem caused by the limited autonomy of the voltage control group in the voltage deep over-limit cluster can be solved.

[0089] In general, this control strategy achieves more efficient and economical voltage regulation by rationally allocating and utilizing power margin resources, while ensuring the stability of the distribution network and the maximum absorption of photovoltaic power output.

[0090] Optionally, cluster reactive power control may include:

[0091] 2.1: Calculate the power flow of the distribution network and determine whether the voltage at the distribution network nodes is normal.

[0092] 2.2: If the voltage is normal, end the regulation. Otherwise, execute step 1.3 to perform inner cluster voltage control.

[0093] 2.3: Select the target node of the adjustable cluster and adjust the reactive power of the controlled PV in the descending order of the voltage-reactive sensitivity within the cluster.

[0094] 2.4: If the node voltage within the cluster returns to normal, the transmission power between clusters and the boundary node voltage are corrected, the reactive power control of the current cluster ends, and the reactive power voltage control of the next cluster is executed; if the node voltage within the cluster still exceeds the limit, it indicates that the reactive power adjustable margin of the current cluster photovoltaic power source is insufficient, and it is listed as an unadjustable cluster.

[0095] 2.5: Repeat steps 2.3 to 2.4 until the adjustable margin of the controlled photovoltaic reactive power of all clusters is limited or the voltage of the entire network returns to normal.

[0096] Optional, see Figure 3 , before calculating the steady-state power flow within the cluster in step 2.1, it also includes:

[0097] The distribution network is separated according to the power flow relationship between boundary nodes. In the figure, a, m and a*, m* are upstream terminal nodes and downstream equivalent virtual balancing nodes respectively; S ab , S mn is the line complex power; S a is the equivalent load power of node a.

[0098] Optionally, centralized reactive power control may include:

[0099] 3.1: Select the node with the highest voltage over-limit level in the distribution network as the target node.

[0100] 3.2: Select the photovoltaic power source with the remaining adjustable reactive power margin in the whole network as the controlled object, arrange the adjustable photovoltaic power sources in descending order of voltage-reactive power sensitivity for the target node, and select the photovoltaic power source with the highest value for voltage regulation.

[0101] 3.3: After the target node voltage is regulated, update the target node and the controlled photovoltaic position. Repeat this step until the voltage of the entire network node returns to normal or the reactive power adjustable margin of all photovoltaic power sources is exhausted.

[0102] 3.4: Similar to the centralized reactive power control steps, in order to improve the photovoltaic consumption level of the distribution network and reduce unnecessary active power reduction, centralized active power control is carried out, and the active power of adjustable photovoltaic power sources is reduced in descending order of voltage-active power sensitivity in the whole network.

[0103] Optional, see Fig.10, the IEEE 141-node distribution network is taken as an example to verify the effectiveness of the proposed control strategy.

[0104] Figure 4 The figure shows the change curve of modularity gain during the distribution network cluster division process. According to the inherent network structure connection of the distribution network, when the number of cluster divisions is 12, the modularity function is the largest and the current cluster division result is the best; Figure 5 The figure shows a binary clustering diagram of cluster partitioning. Based on the optimal number of cluster partitioning, a straight line is drawn to intersect with the binary clustering curve. The node combination below the intersection is the cluster partitioning result. Fig.11 The figure shows the output of photovoltaic power sources taking into account regional differences; Fig.12 Shown is the user terminal load power fluctuation curve. Fig.13 The figure shows the voltage distribution of the IEEE 141 node distribution network before the photovoltaic power source participates in the regulation. When the output of the photovoltaic power source is greater than the load demand, the voltage exceeds the limit problem most of the time throughout the day. Fig.14 The figure shows the voltage distribution after reactive power compensation of photovoltaic power sources in the IEEE 141 node distribution network.

[0105] Figure 6 The figure shows the voltage distribution diagram of the IEEE 141 node distribution network after reactive power control of the photovoltaic power cluster provided by an embodiment of the present invention. When the node voltage is deeply out of limit, the traditional cluster reactive power control can only regulate the over-limit voltage at some moments after cluster reactive power compensation, and cannot effectively regulate the voltage that is seriously out of limit. The over-limit voltage in the cluster cannot achieve effective "autonomy", and the photovoltaic power sources of the remaining clusters cannot actively participate in the voltage regulation of the current cluster. The reactive power regulation margin of the photovoltaic power source is insufficiently utilized, and subsequent voltage regulation cannot be carried out. Figure 7 The figure shows the voltage distribution diagram of the IEEE 141 node distribution network after the centralized reactive power control of photovoltaic power sources provided by the embodiment of the present invention. In view of the limitations of cluster reactive power control, the centralized reactive power control stage uses the reactive power adjustable margin of photovoltaic power sources in the entire network to regulate the voltage of the over-limit node, so as to realize the interactive cooperation of each photovoltaic power source. Most of the time, the node voltage returns to normal, and a few times the node voltage is still in an over-limit state.

[0106] contrast Figure 7 and Fig.14 It can be seen that the double-layer coordinated voltage control effect is very close to the centralized voltage control effect, but the advantage of the proposed control strategy is that cluster control reduces the order of matrix operations and has a faster control speed.

[0107] Figure 8The figure shows the voltage distribution diagram of the IEEE 141 node distribution network after centralized active power control of photovoltaic power sources. Based on the exhaustion of the adjustable reactive power margin of photovoltaic power sources in the entire network, centralized active power reduction control is carried out in this stage, and the reactive power regulation margin is released at the same time to maximize the photovoltaic consumption level of the distribution network and realize rapid voltage regulation.

[0108] like Fig. 9 As shown in the figure, the process of PV power source participating in voltage regulation includes: at 14:15, node 30 is selected as the target node with a voltage of 1.0772pu. Based on the descending order of voltage-power sensitivity, the controlled PV 30# and 25# participate in voltage regulation in turn, and the voltage of node 30 returns to normal; secondly, node 130 is updated as the target node with a voltage of 1.0699pu, and the controlled PV 130# and 124# participate in voltage regulation in turn, and the node voltage returns to normal; finally, node 47 is updated as the target node with a voltage of 1.0661pu. After the controlled PV 50# and 40# participate in voltage regulation, the voltage of node 47 is still in an over-limit state and it is still the target node. At this time, the adjustable reactive power margin of the PV power source in the cluster is exhausted, and the cluster reactive power control cannot complete the effective regulation of the over-limit voltage.

[0109] After the inner cluster voltage control is completed, the outer centralized voltage control is executed, and the photovoltaic power source performs reactive power compensation for the target nodes in turn until the reactive power adjustable margin of the photovoltaic power source of the entire network is exhausted. At this time, the voltage of the target node 47 is 1.0502pu, and the centralized active power reduction control is executed to reduce the active power output of the controlled photovoltaic 50# and 63# in turn, and the voltage of the entire network returns to normal.

[0110] Based on the same voltage regulation standard, the traditional cluster voltage regulation is adjusted to cluster reactive / active power alternating control to ensure that the voltage regulation process can be carried out continuously. At the same time, in order to fully illustrate the superiority of the proposed control strategy, the traditional centralized voltage control (Scheme 1), cluster reactive / active power alternating control (Scheme 2) and the proposed control strategy (Scheme 3) are compared, and the results are shown in Table 1:

[0111] Table 1 Comparison of the results of three control strategies at 14:15 during the day

[0112]

[0113] Scheme 1 follows the global voltage-power sensitivity descending control principle, with the lowest reactive compensation and active power reduction, and is the global optimum; however, due to the need to regress a high-order voltage-power sensitivity matrix, the control speed is relatively slow and the control time is long.

[0114] Scheme 2 follows the descending control principle of voltage-power sensitivity within the cluster and is a local optimum. Based on the voltage depth exceeding the limit, this strategy regulates the voltage by cutting more active power. The amount of active power reduction increases and the corresponding amount of reactive power compensation decreases. This strategy reduces the order of single matrix operations, so the regulation speed is the fastest and the regulation time is the shortest.

[0115] Scheme 3 overcomes the limitation of traditional cluster voltage control that the voltage cannot be effectively "autonomous" when the voltage exceeds the limit by establishing a two-layer collaborative voltage control model of cluster reactive power control and centralized reactive power control. Compared with Scheme 1, the control speed is increased by 20.03%, and the reactive power adjustable margin of photovoltaic power sources in the entire network is fully utilized, which maximizes the photovoltaic absorption level of the distribution network and reduces the active power reduction.

[0116] In summary, the proposed control strategy shows good performance in terms of regulation time and active power reduction, which proves that the scheme has certain advantages.

Claims

1. A double-layer coordinated voltage control method for active distribution network based on LSTM neural network, characterized in that: The method mainly includes the following steps: S1: Obtain incomplete measurement data of the distribution network at historical time sections, including: photovoltaic grid power data, node voltage data, load data, and distribution network physical topology data; S2: Establish a voltage-power sensitivity online mapping model based on LSTM neural network, including data preparation and regression learning; the prepared data is obtained by normalizing the historical data set of multi-input-multi-output features, and the regression learning consists of initialization parameters, forward propagation, calculation of loss function, back propagation and iterative training steps; S3: Establish a two-layer coordinated voltage control model for active distribution network based on distributed photovoltaic descending order, including inner cluster voltage control and outer centralized voltage control, and set voltage control constraints based on the Jacobian inverse transformation equation of the power flow; S4: Inner cluster voltage control: Real-time evaluation of regional node voltages and descending classification to select the dominant node in the area, regression learning of the voltage-reactive sensitivity vector corresponding to the dominant node area and descending classification to select reactive controlled photovoltaics, combined with the dominant node voltage over-limit, voltage regulation constraints and the current controlled photovoltaic reactive adjustable margin, calculate the actual reactive compensation of the controlled photovoltaics, and regulate the dominant node over-limit voltage; S5: Update the voltage of the cluster boundary nodes and the inter-cluster transmission power, and return to S4 to iterate and execute the next round of voltage control; if the voltage of the leading nodes of each cluster returns to normal, the control is terminated; otherwise, if the voltage of the leading node exceeds the limit and is separated from the controlled photovoltaic area, the inner cluster voltage control is exited and the outer centralized voltage control is executed; S6: Outer centralized voltage control: including reactive power compensation and active power reduction two-stage control mode, arranging the node voltages of the whole network and the corresponding voltage-reactive power sensitivity in descending order, selecting the leading node and controlled photovoltaic in the reactive power compensation stage; calculating the actual reactive power compensation amount of the controlled photovoltaic and sequentially regulating the over-limit voltage of the leading node, and repeating S6 until the adjustable reactive power margin of all controlled photovoltaics is exhausted; S7: Set the power factor reduction threshold, determine the voltage over-limit situation according to the current photovoltaic operation power factor, arrange the node voltages of the entire network and the corresponding voltage-active power sensitivity in descending order, select the leading node and the controlled photovoltaic in the active power reduction stage; reduce the active power while releasing the reactive power regulation margin, sequentially regulate the over-limit voltage of the leading node, repeat S7 until the voltage of the entire network does not exceed the limit, and output the voltage distribution and power adjustment results.

2. According to claim 1, a double-layer collaborative voltage control method for active distribution network based on LSTM neural network is characterized in that: The voltage-power sensitivity online mapping model based on LSTM neural network includes: Prepare the data: Where: f1, f2 and f3 represent the data training set respectively; V t represents the distribution network node voltage matrix at time t; S eint represents the photovoltaic grid-injected power matrix at time t; S ent represents the voltage-power sensitivity matrix at time t; Return to Learning: 1) Initialization parameters: set the LSTM neural network weight matrix and bias parameters; 2) Training the LSTM neural network model: input the normalized data training set and perform forward propagation calculation; Gate calculation: Where: W i , W l and W o Represents the weight matrices of the input gate, forget gate, and output gate respectively; b i , b l and b o Respectively represent the bias terms of the corresponding gates; I t , L t and O t They represent the output of the corresponding gate at time t; x t represents the input at time t; h t-1 represents the hidden state output at time t-1; σ represents the sigmoid function; Memory unit calculation: c t =l t ec t-1 +i t etanh(W c [h t-1 ,x t ]+b c ) (3) Where: W c represents the weight matrix of the memory unit; b c represents the bias term of the memory unit; c t represents the state of the memory unit at time t; e represents the element-wise multiplication rule; 3) Calculate the loss function: Calculate the global error based on the network output value and the true value of the training set: Where: E represents the global error of the system; N represents the number of samples; p represents the length of training data; y t represents the predicted value at time t; h t Represents the true value of the training set at time t; 4) Back propagation: Use the back propagation algorithm to calculate the gradient loss and update the network parameters; 5) Iterative training: monitor the changes in the error function until it converges to the threshold and output the optimal solution of the model.

3. According to claim 1, a double-layer collaborative voltage control method for active distribution network based on LSTM neural network is characterized in that: The voltage control constraint is set based on the power flow Jacobian inverse transformation equation: The inverse transformation equation of the power flow Jacobian is: Where: Δδ and ΔV represent the changes in node voltage phase angle and amplitude; V represents the node voltage amplitude; J represents the Jacobian matrix; X, K, Φ and Ψ represent the phase angle-active power, phase angle-reactive power, voltage-active power and voltage-reactive power sensitivity sub-matrices in the voltage-power sensitivity matrix; ΔP and ΔQ represent the changes in active power and reactive power; The voltage regulation constraint is: Where: ΔV i Indicates that the voltage of the dominant node i exceeds the limit; V i 0 represents the voltage of the dominant node i when photovoltaics are not involved in regulation; Φ ij and ij represents the voltage-active power and voltage-reactive power sensitivity factors of the dominant node i relative to the controlled photovoltaic j; ΔP j and ΔQ j Represents the active and reactive adjustment of the controlled PV j.

4. According to claim 1, a double-layer collaborative voltage control method for active distribution network based on LSTM neural network is characterized in that: In the dual-layer coordinated voltage regulation model of the active distribution network based on distributed photovoltaic descending arrangement, the inner cluster voltage control includes: The regional division modularity function of the inner cluster voltage control is: Where: A ij represents the weight of the edge between node i and node j, which is 1 when the two nodes are connected and 0 otherwise; k i represents the sum of the weights of the edges connected to node i; m represents the sum of the weights of all edges; c i Represents the cluster assigned to node i. If nodes i and j are in the same cluster, Otherwise it is 0; The descending sequence of the dominant nodes of the inner cluster voltage control is: δV=[ΔV α ,ΔV β ,…,ΔV γ ,…,ΔV λ ],α,β,γ,λ<n (8) Where: α and β represent the node locations where the highest and second highest voltages exceed the limit; The reactive controlled photovoltaic descending sequence of the inner cluster voltage control is: Where: and θ indicate the locations of the highest and second highest voltage-reactive sensitivities corresponding to the PV sources; The actual reactive output adjustment of the reactive controlled photovoltaic under the inner cluster voltage control is: Where: ΔQ mar,ω and ΔQ dem,ω represents the reactive power margin of the controlled photovoltaic ω and the reactive power demand of the dominant node; S ω , P ω and Q ω It indicates the apparent capacity, active power and reactive power output of the controlled photovoltaic ω.

5. According to claim 1, a double-layer collaborative voltage control method for active distribution network based on LSTM neural network is characterized in that: In the dual-layer coordinated voltage regulation model of the active distribution network based on distributed photovoltaic descending arrangement, the outer layer centralized voltage control includes: Based on the dominant node voltage evaluation results, the expression of the reactive power compensation stage control mode is: Where: ΔV α are the initial voltage and voltage change of the target node α respectively; ΔQ j is the reactive power change injected into the node by the controlled photovoltaic j; ε represents the number of times the reactive controlled photovoltaic is regulated; The descending sequence of active controlled photovoltaic power under the outer layer centralized voltage control is: δΨ=[Ψ α,ω ,P α,ξ ,…,Ψ α,τ ,…,Ψ α,υ ] (12) The actual active power reduction adjustment of the active controlled photovoltaic is: Where: Q' ω Represents the reactive power output after reactive power compensation of the reactive controlled photovoltaic ω; S' ω Represents the usage capacity of reactive controlled photovoltaic ω after reactive compensation; η=Φ αω / Ψ αω represents the voltage-power sensitivity ratio factor; Based on the dominant node voltage evaluation result, the expression of active power reduction mode is: Where: ΔP j , ΔQ j are the changes of active and reactive power injected into the node by the controlled photovoltaic j respectively; χ represents the number of times the active controlled photovoltaic is adjusted.

Images