A reactive power / voltage control method for distribution network with adaptive topology change
Through cloud-edge collaborative architecture and deep neural network, combined with transfer learning algorithm, the response problem of reactive voltage control in distribution network in system topology changes is solved, rapid response and resource optimization are achieved, and computing resources are saved.
Patent Information
- Application Number
- CN202410917739.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-10
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-07-10
AI Technical Summary
The existing reactive voltage control strategies of the distribution network are difficult to respond quickly to system topology changes, and neural network-based methods require a large amount of simulation and computing power support, and lack the ability to quickly solve.
Adopting cloud-edge collaborative architecture, the reactive voltage control tasks of the distribution network are divided into cloud-side global voltage control and edge-side reactive voltage control. Combined with deep neural network and deep reinforcement learning algorithm, the pre-computation and transfer learning of edge networks are used. Algorithm to achieve rapid response and resource optimization.
Effectively allocate computing resources, support fast response of the distribution network, no need to retrain the controller, only small sample data and a small amount of time can be corrected, saving computing resources and computing time.
Smart Images

Figure CN118472964B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a reactive power / voltage control method for a distribution network with adaptive topology changes, and belongs to the technical field of distributed photovoltaic reactive power and voltage control. Background Art
[0002] In recent years, more and more distributed power sources, mainly renewable energy sources, have been connected to distribution networks. Although the popularity of distributed power sources can improve the flexibility and economy of distribution networks, it also poses new challenges to the traditional distribution network reactive power voltage control (VVC) strategy.
[0003] In the traditional reactive power voltage control strategy of distribution network, on-load tap changers and capacitor banks are usually used to solve the voltage over-limit problem caused by load power fluctuations. However, limited by their mechanical characteristics, the regulation capacity of these traditional devices is difficult to meet the reactive power voltage control tasks of the current new distribution network. On the other hand, distributed photovoltaic power sources have the ability to provide fast and flexible reactive power support, and the IEEE 1547.8 standard encourages distributed photovoltaics to participate in distribution network voltage control. Therefore, as an effective and fast-responding reactive power resource, distributed photovoltaic power sources should cooperate with traditional voltage regulation devices to achieve reactive power voltage control of distribution networks at different time scales.
[0004] With the popularization and in-depth research of artificial intelligence algorithms, domestic and foreign scholars have conducted a lot of research on the field of reactive power and voltage control of distribution networks and proposed many new distribution network VVC strategies based on neural networks. However, the current new distribution network VVC strategies based on neural networks do not take into account the changes in system topology and require a lot of simulation and computing power support. There is a lack of ability to quickly solve the VVC results for the changed power system topology. Summary of the invention
[0005] The technical problem to be solved by the present invention is: to provide a reactive power / voltage control method for a distribution network that is adaptive to topology changes. Under the constraints of the cloud-edge-end network architecture, a flexible interconnected distribution network fault recovery optimization algorithm based on edge network pre-calculation and its cloud-edge collaborative control logic are proposed, which can effectively aggregate the computing resources at the edge of the network and support the rapid response of various power supply services in the distribution network.
[0006] The present invention adopts the following technical solutions to solve the above technical problems:
[0007] A reactive power / voltage control method for a distribution network with adaptive topology changes, which is based on a cloud-edge collaborative architecture and divides the distributed photovoltaic reactive power and voltage control method into a distribution network global voltage control strategy based on a cloud-side platform and a distribution network reactive power and voltage control strategy based on an edge computing center;
[0008] The global voltage control strategy of the distribution network based on the cloud-side platform takes the transformer adjustment tap, capacitor switching group and distributed photovoltaic reactive power as the control objects. Combined with the distribution network data collected by the cloud-side platform, the minimum total power loss is taken as the objective function, and the safe and stable operation of the distribution network is taken as the constraint condition. Based on the solution of the reactive voltage control model of the distribution network, the optimal scheduling results within 24 hours of the transformer adjustment tap, capacitor switching group and distributed photovoltaic benchmark reactive power are obtained, and the optimal scheduling results within 24 hours are distributed to the edge computing center.
[0009] The edge computing center uses a deep reinforcement learning algorithm based on a deep neural network to offline train the reactive voltage controller of the distribution network deployed internally. During the offline training process, the controller takes into account the node voltage and distributed photovoltaic reactive output constraints based on the optimal scheduling results of the transformer adjustment tap, capacitor switching group and distributed photovoltaic reactive power issued by the cloud platform, as well as the collected distributed photovoltaic active and reactive output power, and establishes an action value function with the goal of minimizing active power loss. Then, it analyzes and calculates the regression gradient of the action value function relative to the deep neural network structure parameters, and uses the results to update the parameters of the deep neural network. In each iteration, the controller takes action according to the deep neural network and interacts with the simulation environment. The simulation environment receives the controller's action, performs power flow calculations, and updates the next state of the controller.
[0010] After the distribution network topology changes, the edge computing center first conducts offline testing on the original distribution network reactive voltage controller to generate a small sample test data set. It then builds a residual neural network based on the deep neural network structure parameters in the original controller. It then performs transfer learning training on the residual neural network based on the generated small sample test data set, and finally obtains a reactive voltage controller for the distribution network after the topology changes.
[0011] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:
[0012] 1. Based on the cloud-edge collaborative architecture, the present invention divides the reactive voltage optimization task of the distribution network into two stages, thereby transferring part of the originally centralized computing tasks to the edge computing center, which can effectively allocate computing resources and support the rapid response of reactive voltage control of the distribution network.
[0013] 2. The present invention proposes a transfer learning algorithm based on a residual network. After the system topology changes, it avoids retraining the corresponding reactive voltage controller and only requires a small sample of training data and a small amount of training time to correct the original controller, thereby saving a lot of computing resources and computing time. BRIEF DESCRIPTION OF THE DRAWINGS
[0014] Figure 1It is a schematic diagram of the overall architecture of the two-stage voltage control strategy of the distribution network based on the cloud-edge collaborative architecture of the present invention;
[0015] Figure 2 Schematic diagram of the residual neural network structure of the present invention;
[0016] Figure 3 It is a schematic diagram of the transfer learning training process of the present invention. DETAILED DESCRIPTION
[0017] The embodiments of the present invention are described in detail below, and examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be interpreted as limiting the present invention.
[0018] like Figure 1 As shown, under the constraints of the cloud-edge-end network architecture, the present invention adopts a novel reactive voltage control method for distribution network with adaptive topology changes, and divides the reactive voltage optimization task of the distribution network into two stages, as follows:
[0019] (1) In the two-stage voltage control strategy of the distribution network based on the cloud-edge collaborative architecture, the cloud-edge collaborative architecture adopted by the present invention consists of four main components: "cloud, pipe, edge, and end". The basic functions and control modes of each level are as follows:
[0020] 1) Cloud-side platform: From a functional perspective, the cloud-side platform uses OLTC (transformer adjustment taps), CBs (capacitor switching groups) and DPV (distributed photovoltaic) reactive power as the control objects, and uses the situational awareness of the entire network to obtain all-weather meteorological data, load data, electricity data and other data information. It uses the optimal power flow as the objective function and solves the optimal day-ahead dispatch of OLTC, CBs and DPV reactive power, and distributes the solution results to the "edge" computing center through the "pipe" end.
[0021] 2) Pipe-side communication method: Optical fiber, carrier, wireless communication and other methods are used to meet the communication needs of the system, including "cloud"-"edge" communication and "edge"-"end" communication.
[0022] 3) Edge computing center: From the perspective of tasks, the edge computing center mainly includes two tasks: data collection and training the reactive voltage control model of the distribution network. In the data collection task, in addition to receiving the optimal dispatch results sent by the "cloud" end, the edge computing center is also responsible for collecting the real-time operation data of the "end" side equipment and the real-time monitoring data of the distribution network. In the model training task, the edge computing center uses a deep reinforcement learning algorithm based on the massive sample data collected to train the reactive voltage control model of the distribution network offline, and performs real-time control of the distributed photovoltaic power supply on the "end" side based on the real-time monitoring data of the distribution network.
[0023] 4) End-side equipment: The end-side equipment used in the present invention only includes the intelligent inverter of the distributed photovoltaic power source, which has the functions of collecting basic data and executing superior control instructions, and is the final instruction execution basis of the two-stage voltage control strategy of the distribution network.
[0024] (2) Global voltage control strategy of the distribution network deployed on the cloud-side platform:
[0025] The global voltage control strategy of the distribution network deployed on the cloud side takes OLTC, CBs and DPV reactive power as the control object, combines the load forecast data, distributed photovoltaic forecast data and other information collected by the cloud side platform, takes the minimum total power loss as the objective function, takes the safe and stable operation of the distribution network as the constraint condition, and solves the optimal dispatch result of OLTC, CBs and DPV reactive power within 24 hours based on the reactive voltage control model of the distribution network (the dispatch time scale is hourly). Among them, the objective function and constraints of the global voltage control strategy of the distribution network are as follows:
[0026] 1) Objective function:
[0027] (1)
[0028] Where: Indicates the transformer regulation tap (OLTC) at the Adjust gear position at any time; is a [0,1] variable, indicating the access node CB's The capacitor in The switching status at the moment, 0 means not connected, 1 means connected; Indicates access node Distributed photovoltaic Reactive power at the moment; Indicates a 24h scheduling cycle; represents the collection of transmission lines in the distribution network; Indicates that the head end node is The end node is On the transmission line The square of the current amplitude flowing at any moment; Indicates that the head end node is The end node is The resistance of the transmission line.
[0029] 2) Constraints:
[0030] (2)
[0031] (3)
[0032] (4)
[0033] (5)
[0034] (6)
[0035] (7)
[0036] (8)
[0037] Equations (2)-(5) represent the power flow constraints based on second-order cone programming. Represents access nodes The DPV of Active power and reactive power at the moment; Represents access nodes The load of Active power and reactive power at the moment; Indicates access node and nodes The CB in the Reactive power at the moment; Respectively represent Time Node Inflow Node Active power and reactive power; Indicates that the head end node is The end node is The reactance of the transmission line; Indicates Time Node The square of the voltage amplitude; Indicates the minimum and maximum values of the node voltage amplitude; The voltage per unit value at the equilibrium node is ; The adjustment unit of the transformer adjustment tap is ; Indicates access node The capacitor switching group The reactive power of a capacitor, Indicates access node The capacitor switching group The maximum number of times a capacitor's switching state changes within 24 hours.
[0038] Equations (6) and (7) are the operating constraints of OLTCs and CBs, i.e., the number of switching operations per day should be limited considering their mechanical characteristics and lifespan. is an integer variable, indicating the transformer adjustment tap OLTC The maximum value of the gear adjustment at the moment; It indicates the maximum value of OTLC gear adjustment action within 24 hours. is a [0,1] variable used to assist calculations ; is a [0,1] variable, indicating the access node The first Whether the operating state of a capacitor changes at adjacent moments, 0 means the operating state has not changed, and 1 means the operating state has changed.
[0039] Formula (8) is the operating constraint of PV, Indicates the access node Rated apparent power of the DPV.
[0040] (3) Reactive power and voltage control strategy of distribution network deployed in edge computing center:
[0041] The edge computing center uses the deep reinforcement learning (DRL) algorithm to offline train the reactive voltage controller of the distribution network deployed internally based on the deep neural network (DNN). During the offline training process, the controller takes into account the node voltage and DPV reactive output constraints based on the optimal scheduling results of OLTC, CBs and DPV reactive power issued by the cloud-side platform and the DPV active / reactive output power collected by the edge smart inverter, and establishes an action-value function with the goal of minimizing active power loss. Then, the regression gradient of the action-value function relative to the neural network structure parameters is calculated analytically, and the result is used to update the parameters of the deep neural network. This process can be regarded as a closed-loop trial-and-error process in which the controller continuously searches for the optimal DPV reactive output based on the current state of the distribution network. In each iteration, the controller takes action according to its own DNN and interacts with the simulation environment. The simulation environment receives the action value of the controller, thereby performing power flow calculations and updating the next state of the controller.
[0042] The specific training process of the distribution network reactive voltage controller is as follows:
[0043] 1) The current state of the distribution network As the input of the controller, get the output of the controller , and further obtain the actions of the simulation environment .
[0044] Among them, the current status It includes the active output value of each DPV and the dispatch results of OLTC and CBs, that is, The output of DNN represents the reactive output difference of DPV before and after the state, that is, The actions of the simulation environment are generated by the corresponding controller and can be expressed as . are the weights and biases between the connected layers in the DNN. It represents the noise disturbance that the controller receives during the exploration action. The noise follows a normal distribution with a mean of 0 and a standard deviation of .parameter It will decrease at a decay rate during the exploration process.
[0045] 2) Based on the current state of the distribution network , the output of the controller and the next state of the distribution network , get the action value .
[0046] Since the VVC strategy needs to quickly solve the problem of voltage exceeding the limit of the power system while minimizing the active power loss as much as possible, the objective function based on the VVC strategy is constructed as follows:
[0047] (9)
[0048] Among them, the entire objective function consists of three parts: is the active power loss of the power system, is the average voltage violation value, is the value of the non-functional capability violation, and is a maximum value, and the value in the present invention is Specifically, and The calculation formula is as follows:
[0049] (10)
[0050] (11)
[0051] (12)
[0052] In the above formula: and Respectively represent the nodes in the current state and nodes The voltage amplitude of Indicates the node in the current state and nodes The voltage phase angle difference; and They represent the number of nodes exceeding the upper voltage limit and the lower voltage limit in the current state respectively; and Respectively represent the current state access node The upper and lower reactive power limits of the DPV; The first node is The end node is The conductance of the transmission line; is the number of DPVs.
[0053] 3) Since power flow calculation requires a lot of computing resources, the next state of the distribution network is estimated based on voltage sensitivity. The voltage amplitude is repeated from 1) to 3) to generate a certain number of samples. M The training data set.
[0054] Among them, the voltage sensitivity matrix It is expressed as follows:
[0055] (13)
[0056] Where: is the Jacobian matrix.
[0057] The next state of the distribution network The voltage amplitude is calculated as follows:
[0058] (14)
[0059] Where: is the voltage sensitivity matrix Corresponding node and nodes Elements of the relationship between In the action decision of the simulation environment, the corresponding node Reactive power action value.
[0060] 4) A voltage-sensitive deterministic gradient ascent method is used to calculate the gradient of the action-value function and update the parameters of the DNN, so that the action value of the current state continuously approaches the expected action with the maximum reward value. The gradient calculation and update method of the DNN is as follows:
[0061] (15)
[0062] (16)
[0063] (17)
[0064] (18)
[0065] In formula (15), Indicates the learning rate for DNN parameter update, and its value is 0.0001.
[0066] (4) Transfer learning algorithm of reactive power voltage controller of distribution network with original network structure:
[0067] After the distribution network topology changes, the edge computing center first conducts offline testing on the original distribution network reactive power voltage controller to generate a sample with the same number of samples. A small sample test dataset and .in, The test data generated for the controller based on the topology before the change, each sample contains and ;and Test data generated for the controller based on the changed topology. Each sample contains and Although the active / reactive output of PV and the day-ahead dispatch of OLTC and CBs do not change, the node voltage of the system will have a significant shift. Therefore, the states in the two test data sets are and Not the same.
[0068] Secondly, combined with the neural network structure parameters in the original controller, a residual neural network (ResNet) is constructed. The schematic diagram of the constructed residual neural network structure is as follows: Figure 2 As shown, the residual neural network includes sequentially connected convolution layer 1, pooling layer 1, convolution layer 2, pooling layer 2, convolution layer 3, pooling layer 3, connection layer and output layer, and the output end of pooling layer 1 is also connected to the input end of convolution layer 3 through residual link module 1, the output end of pooling layer 1 is also connected to the input end of connection layer through residual link module 3, and the output end of pooling layer 2 is also connected to the input end of connection layer through residual link module 2; each residual link module has the same structure, including sequentially connected network link layer 1 and network link layer 2, the output ends of network link layer 1 and network link layer 2 are both activated by RelU, and network link layer 2 is added to the input of residual link module after RelU activation function as the output of residual link module. ResNet eliminates the neural network training problem caused by too deep network layers by "short-circuit link". The residual module is the main structure of the residual network and the core of the residual network. The residual module retains the features of the shallow network through newly added branches, that is, it performs "identity mapping" processing to prevent the data after multi-layer convolution operations from being over-extracted and losing certain feature connections, thereby ensuring the performance of the model.
[0069] Finally, based on the generated data set, the residual neural network is trained through transfer learning, and finally a distribution network voltage controller suitable for the new topology is obtained. Figure 3 shown.
[0070] The entire transfer learning training process is divided into three stages, and the tasks in each stage are as follows:
[0071] 1) Phase 1:
[0072] In the initial stage of the transfer learning process, a state difference-based migration method is used. The initial target model is pre-trained without labeling the target samples by minimizing the maximum mean difference (Max Mean Discrepancy, MMD) between the states of the new and old topologies in the test dataset. The calculation formula of MMD is as follows:
[0073] (19)
[0074] Secondly, the MMD term is further introduced into the deterministic gradient update strategy to retrain the residual network. The specific gradient update process is as follows:
[0075] (20)
[0076] In the formula, Represents the weight parameter, and its value is 1.
[0077] 2) Phase 2:
[0078] Before the system topology changes, since the distribution network reactive voltage controller has passed the DRL training process for the VVC strategy, In any state, the controller The actions taken can guarantee the corresponding next state The node voltage constraints and reactive power constraints will not be violated.
[0079] For the changed system topology, the controller is based on the state Take Action The corresponding next state , may violate the node voltage constraint or reactive power constraint. At this time, the state ,action and the next state Constitutes 1 tag data.
[0080] At this point, deep reinforcement learning is performed on the residual network based on all the labeled data in the test dataset. In order to ensure that the residual network always remains within the operating constraints during the application of the actual VVC for the temporary system topology, the DRL process based on the labeled data needs to take a sufficiently long step size. , that is, from the initial state Simulate to state .
[0081] In addition, in order to ensure the generalization ability of the residual network for the system state and prevent the neural network from overfitting, a certain number of similar states need to be generated near the target data. , and conduct active generalization learning. The generation process of similar states is as follows:
[0082] (twenty one)
[0083] in, and is the distributed photovoltaic active / reactive power correction value based on normal distribution, It is the node voltage correction value after considering the distributed photovoltaic active / reactive power correction.
[0084] 3) Phase 3:
[0085] Considering that transfer learning and active generalization learning based on the test dataset may cause interference to the overall performance of the controller, in this stage, only active generalization learning is used to train the controller, that is, randomly generating initial states and performing DRL.
[0086] Based on the same inventive concept, an embodiment of the present application provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the aforementioned method for reactive power / voltage control of a distribution network with adaptive topology changes are implemented.
[0087] Based on the same inventive concept, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the steps of the aforementioned distribution network reactive power / voltage control method with adaptive topology changes are implemented.
[0088] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0089] The present invention is described with reference to flowcharts of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process in the flowchart and the combination of processes in the flowchart can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart. Figure 1 A device that specifies functions in a process or multiple processes.
[0090] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A function specified in a process or multiple processes.
[0091] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 The steps of a specified function in a process or multiple processes.
[0092] The above embodiments are only for illustrating the technical idea of the present invention, and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the present invention.
Claims
1. A reactive power / voltage control method for a distribution network with adaptive topology changes, characterized in that: The method is based on the cloud-edge collaborative architecture and divides the distributed photovoltaic reactive voltage control method into a distribution network global voltage control strategy based on the cloud-side platform and a distribution network reactive voltage control strategy based on the edge computing center. The global voltage control strategy of the distribution network based on the cloud-side platform takes the transformer adjustment tap, capacitor switching group and distributed photovoltaic reactive power as the control objects. Combined with the distribution network data collected by the cloud-side platform, the minimum total power loss is taken as the objective function, and the safe and stable operation of the distribution network is taken as the constraint condition. Based on the solution of the reactive voltage control model of the distribution network, the optimal scheduling results within 24 hours of the transformer adjustment tap, capacitor switching group and distributed photovoltaic benchmark reactive power are obtained, and the optimal scheduling results within 24 hours are distributed to the edge computing center. The objective function and constraints of the cloud-side platform-based distribution network global voltage control strategy are as follows: 1.1 Objective function: In the formula, tap t Indicates the adjustment position of the transformer adjustment tap at time t; c im,t represents the switching state of the mth capacitor in the capacitor switching group connected to node i at time t, c im,t is a [0,1] variable; Q PVi,t represents the reactive power of distributed photovoltaic connected to node i at time t; T represents the 24h scheduling period; E represents the set of transmission lines in the distribution network; l ij,t It represents the square of the current amplitude flowing through the transmission line with the first node i and the last node j at the time t; r ij represents the resistance of the transmission line with the first node i and the last node j; 1.2 Constraints: v i,t =v j,t +2(P ij,t r ij +Q ij,t x ij )+l ij,t (r ij 2 +x ij 2 ) Where P PVj,t , Q PVj,t They represent the active power and reactive power of the distributed photovoltaic connected to node j at the tth moment; P Lj,t , Q Lj,t They represent the active power and reactive power of the load at access node j at time t respectively; Q CBi,t , Q CBj,t Respectively represent the reactive power of the capacitor switching group connected to node i and node j at the tth moment; P ij,t , Q ij,t They represent the active power and reactive power flowing from node i to node j at time t respectively; P jk,t , Q jk,t They represent the active power and reactive power flowing from node j to node k at time t respectively; l jk,t It represents the square of the current amplitude flowing through the transmission line with the first node j and the last node k at the time t; r jk represents the resistance of the transmission line with the first node being j and the last node being k; jk represents the reactance of the transmission line with the first node j and the last node k; ij represents the reactance of the transmission line with the first node i and the last node j; v 1,t 、v i,t and v j,t Respectively represent the square of the voltage amplitude of node 1, node i and node j at time t; V min 、V max Respectively represent the minimum and maximum values of the node voltage amplitude; V s Indicates the voltage per unit value at the equilibrium node; ΔV T Indicates the adjustment unit of the transformer adjustment tap; tap t+1 Indicates the adjustment position of the transformer tap at time t+1; β t Indicates the maximum value of the transformer adjustment tap at the t moment; tap max Indicates the maximum value of the transformer adjustment tap position adjustment action within 24 hours; c im,t 、c im,t+1 represents the switching state of the mth capacitor in the capacitor switching group connected to node i at time t and time t+1 respectively; a im,t =c im,t ·c im,t+1 , a im,t is a [0,1] variable; b im,t Indicates whether the operating state of the mth capacitor in the capacitor switching group connected to node i changes at adjacent moments, b im,t is a [0,1] variable; q im,t represents the reactive power of the mth capacitor in the capacitor switching group connected to node i; cap im,max S represents the maximum number of times the switching state of the mth capacitor in the capacitor switching group connected to node i changes within 24 hours; PVi represents the rated apparent power of distributed photovoltaic connected to node i; P PVi,t represents the active power of distributed photovoltaic connected to node i at time t; The edge computing center uses a deep reinforcement learning algorithm to offline train the reactive voltage controller of the distribution network deployed internally based on a deep neural network. During the offline training process, the controller establishes an action value function with the goal of minimizing active power loss based on the optimal dispatch results of transformer adjustment taps, capacitor switching groups and distributed photovoltaic reactive power issued by the cloud platform, as well as the collected distributed photovoltaic active and reactive output power, taking into account the node voltage and distributed photovoltaic reactive output constraints. Then, the regression gradient of the action value function relative to the structural parameters of the deep neural network is calculated analytically, and the result is used to update the parameters of the deep neural network. In each iteration, the controller takes actions based on the deep neural network and interacts with the simulation environment. The simulation environment receives the controller's actions, calculates the power flow, and updates the next state of the controller. The training process of the distribution network reactive voltage controller is as follows: 2.1 The current state of the distribution network x As the input of the controller, get the output of the controller And further obtain the actions of the simulation environment Among them, the current state State x Including the active output value of each distributed photovoltaic and the dispatching results of the transformer adjustment tap and capacitor switching group, that is, V x , tap x They represent the distributed photovoltaic active power vector, node voltage vector, distributed photovoltaic reactive power vector, capacitor switching group reactive power vector, and transformer adjustment tap adjustment position of the current state respectively; the output of the controller represents the reactive output difference of the distributed photovoltaic before and after the state Right now They represent the reactive power of the next state and the current state of the distributed photovoltaic connected to node i respectively; the action of the simulation environment is generated by the corresponding controller, which is expressed as W i ,b i are the weights and biases between the connected layers in a deep neural network, represents the output of the deep neural network, ε(0,σ x ) represents the noise disturbance received by the controller during the exploration action. The noise follows a normal distribution with a mean of 0 and a standard deviation of σ x ; 2.2 Based on the current state of the distribution network x , the output of the controller and the next state of the distribution network x+1 , get the action value R(State x ,a x ); The objective function F of the reactive power / voltage control strategy is constructed as follows: in, is the active power loss of the power system, is the average voltage violation value, is the value of no functional capacity violation, λ1 and λ2 are both maximum values; In the formula, g ij is the conductance of the transmission line with the first node i and the last node j; and Respectively represent the voltage amplitudes of nodes i and j in the current state; Indicates the voltage phase angle difference between node i and node j in the current state; and Respectively represent the number of nodes exceeding the upper voltage limit and the lower voltage limit in the current state; N PV is the number of distributed photovoltaics; and They represent the reactive power upper limit and reactive power lower limit of the distributed photovoltaic connected to node i in the current state, 2.3 Estimating the next state of the distribution network based on voltage sensitivity x+1 The voltage amplitude is , and 2.1-2.3 are repeated to generate the training data set; Among them, the voltage sensitivity matrix S is expressed as follows: In the formula, is the Jacobian matrix, J Pδ , J Qδ , J PV , J QV They are the 4 block matrices corresponding to the Jacobi matrix, P is the node input active power vector, Q is the node input reactive power vector, δ is the node voltage phase angle vector, and V is the node voltage amplitude vector; The next state of the distribution network x+1 The voltage amplitude is calculated as follows: In the formula, represents the voltage amplitude of node i in the next state; S ij is the element in the voltage sensitivity matrix S corresponding to the relationship between node i and node j; is the reactive power action value corresponding to node j in the action decision of the simulation environment; 2.4 The determined gradient ascent method based on voltage sensitivity is used to calculate the gradient of the action value function and update the parameters of the deep neural network so that the action value of the current state approaches the expected action with the maximum reward value; the gradient calculation and update method of the deep neural network is as follows: In the formula, n represents the number of samples used when updating the gradient of the deep neural network, D represents the sample data set, and d represents a sample data, including State x 、output x 、R(State x ,a x ) and State x+1 , represents the correction value of the structural parameters of the deep neural network when it is updated for the rth time, S x 、a x They represent the current state and action value of a sample data in the sample data set D respectively. Represents the gradient of the action value relative to the state value in a sample data; Represents the gradient of the value function relative to the action value in a sample data; θ r ,θ r+1 Respectively represent the deep neural network structure parameters after the deep neural network is updated for the rth and r+1th time, η represents the learning rate of the deep neural network parameter update, W and b represent the weight and bias respectively, N b represents the number of transmission lines in the distribution network, g l represents the conductance of the distribution network transmission line, represents the elements corresponding to node i and node p in the voltage sensitivity matrix; Represents the elements corresponding to node j and node p in the voltage sensitivity matrix, where node i is a node that violates the node voltage upper limit constraint, node j is a node that violates the node voltage lower limit constraint, and node p is a node connected to distributed photovoltaics; Indicates that the deep neural network is based on the current state State x Output value output x The element in the vector corresponding to the distributed photovoltaic connected to node p, Indicates the current state x The reactive power of distributed photovoltaic connected to node p, Q PV,min , Q PV,max They represent the lower and upper limits of the reactive power output of distributed photovoltaics respectively; After the distribution network topology changes, the edge computing center first conducts offline testing on the original distribution network reactive power voltage controller to generate a small sample test data set, and then builds a residual neural network based on the deep neural network structure parameters in the original controller. The residual neural network is then trained through transfer learning based on the generated small sample test data set, and finally a reactive power voltage controller suitable for the distribution network after the topology changes is obtained. The process of transfer learning training is as follows: 3.1 In the initial stage of transfer learning training, a state difference-based migration method is adopted. By minimizing the maximum average difference in the state between the topology before and after the change in the small sample test data set, the initial target model is pre-trained without marking the target sample. The calculation formula of the maximum average difference is as follows: In the formula, They represent the current states in the test data set generated by the controller based on the distribution network topology before and after the change, respectively; The maximum average difference is introduced into the deterministic gradient update strategy to retrain the residual neural network. The specific gradient update process is as follows: In the formula, n′ represents the number of samples used when the residual neural network is updated, D′ represents the sample data set used for transfer learning, D′<<D, d′ represents a sample data in D′, including State x′ 、output x′ , R′(State x′ ,a x′ ) and State x+1′ , represents the gradient calculated based on the maximum average difference, and λ represents the weight parameter; 3.2 For the changed distribution network topology, if the controller is based on the state Take Action The corresponding next state If the node voltage constraint or reactive power constraint is violated, the state at this time action and the next state Constitute 1 tag data; Use all the label data to perform deep reinforcement learning on the residual neural network. Generate similar status nearby In the formula, and They represent the distributed photovoltaic active and reactive correction values based on normal distribution, It is the node voltage correction value after considering the distributed photovoltaic active or reactive power correction; 3.3 Use active generalization learning to train the controller, that is, randomly generate initial states and perform deep reinforcement learning.
2. A computer device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that: When the processor executes the computer program, the steps of the reactive power / voltage control method for a distribution network with adaptive topology change as claimed in claim 1 are implemented.
3. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the reactive power / voltage control method for a power distribution network with adaptive topology change as claimed in claim 1 are implemented.
Citation Information
Patent Citations
Power grid reactive voltage coordination control method based on DDPG algorithm
CN114243718A
High-proportion photovoltaic power distribution network voltage edge control method
CN117239763A
Photovoltaic power distribution network reactive power optimization control method and system based on edge calculation
CN117293853A
