A distributed optimal power flow accelerated solution method and system
By dividing the distribution network into regions and using LSTM-VAE networks to accelerate the alternating direction multiplier method, the problems of excessively long solution time and privacy leakage in distributed optimal power flow are solved, and efficient solution of distributed optimal power flow is achieved.
Patent Information
- Application Number
- CN202510235061.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-28
- Publication Date
- 2025-11-21
- Estimated Expiration
- 2045-02-28
AI Technical Summary
In the context of high penetration of renewable energy, the solution time for distributed optimal power flow is too long and centralized architecture poses a risk of privacy leakage. Existing improvement methods have failed to effectively improve the convergence speed of the algorithm.
The community mining theory is used to delineate the power grid area, and the LSTM-VAE network is used to accelerate the alternating direction multiplier method. A two-stage data generation strategy is used to train low-precision and high-precision LSTM-VAE networks to generate a high-precision dataset, reducing the number of iterations and improving the solution efficiency.
While ensuring the quality of the solution, it significantly reduces the solution time for distributed optimal power flow, and improves the convergence speed and privacy protection capabilities of the algorithm.
Smart Images
Figure CN120073742B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to a distributed optimal power flow accelerated solving technology, belonging to the technical field of distribution network dispatching. BACKGROUND
[0002] As the core optimization tool of power system operation, Optimal Power Flow (OPF) faces two fundamental challenges under the background of high penetration of renewable energy: first, with the further increase of new energy unit penetration rate, the system node scale grows exponentially, leading to a sharp increase in the solving time of nonlinear models; second, the centralized architecture requires the dispatch center to obtain the topology parameters and operation data of the whole network, which poses a risk of privacy leakage.
[0003] To break through the above bottlenecks, distributed OPF architecture divides the large power grid into multiple autonomous regions and uses parallel computing framework to improve the solving efficiency of optimal power flow. However, in order to ensure the consistency of the boundary variables of the autonomous regions of the large power grid, a large number of iterations are usually required, which greatly affects the convergence speed of the algorithm. In the current mainstream alternating distributed algorithm, the Alternating Direction Method of Multipliers (ADMM) has attracted widespread attention due to its convex optimization theory support and engineering usability. It balances the optimization accuracy and privacy protection through a decomposition and coordination mechanism, but its hundreds of iterations requirement still cannot meet the minute-level real-time dispatching requirements. The existing ADMM improvement methods have not been able to break through the inherent limitations of the algorithm framework, and it is urgent to mine the spatiotemporal correlation characteristics of historical optimization trajectories through machine learning and build a data-model hybrid driven accelerated solving paradigm to provide real-time optimization capability for new power systems. SUMMARY
[0004] The present application provides a distributed optimal power flow accelerated solving method and system, which can reduce the time required for distributed optimal power flow solving while ensuring the solving quality.
[0005] Technical scheme: In order to achieve the above application purposes, the present application adopts the following technical scheme:
[0006] A distributed optimal power flow accelerated solving method, comprising the following steps:
[0007] Based on the physical characteristics of the distribution network topology, the connectivity between nodes is evaluated using electrical coupling strength, and the distribution network is divided into several regions using community mining theory. The distribution network optimal power flow problem is decomposed into optimal power flow sub-problems in each region according to the divided regions;
[0008] with the first convergence tolerance performing the Alternating Direction Method of Multipliers M lThe first dataset D with low precision is generated. l Using the first dataset D l Training low-precision LSTM-VAE (l) Network; Training-based low-precision LSTM-VAE (l) Network acceleration with second convergence tolerance The alternating direction multiplier method is executed to generate M. h From one high-precision data set, a second high-precision dataset D is obtained. h And using the second dataset D h Training a high-precision LSTM-VAE (h) Network; where the first convergence tolerance Greater than the second convergence tolerance First dataset D l Second dataset D h This includes the values of the global variables z and their corresponding Lagrange multipliers λ used to coordinate various regions of the distribution network during the iterative process of the alternating direction multiplier method, denoted by u = [z, λ], and the convergence values of the global variables and Lagrange multipliers u' = [z', λ'].
[0009] Execute k * The alternating direction multiplier method is used in the next iteration, and the trained high-precision LSTM-VAE is utilized. (h) The network estimates the convergence values of the global variables and Lagrange multipliers of the distributed optimal power flow; the estimated convergence values are used as the hot-start solution of the alternating direction multiplier method, and the alternating direction multiplier method iteration is continued until the algorithm converges.
[0010] Furthermore, based on the physical characteristics of the distribution network topology, the connectivity between nodes is evaluated using electrical coupling strength. Using community mining theory, the distribution network is divided into several regions, including:
[0011] Utilizing the electrical coupling strength e between node i and node j ij To evaluate the connectivity between nodes, community mining theory is used, with the goal of maximizing modularity Q, to divide the distribution network into several regions, represented as follows:
[0012]
[0013] In the formula, E represents element e. ij The symmetric matrix formed by d i Let C represent the sum of the elements in the i-th row of matrix E. i Indicates the region to which node i belongs, when C i =C j At that time, δ(C) i C j ) = 1, conversely, δ(C) = 1. i Cj ) = 0.
[0014] Furthermore, the LSTM-VAE (l) The network is powered by LSTM (l) Modules and VAE (l) The modules are composed of the first dataset D. l Training low-precision LSTM-VAE (l) The network includes: using the first dataset D l Training VAE (l) After the module, VAE (l) The module's parameters are frozen, and its decoder is coupled with LSTM. (l) Modules combined, using LSTM (l) The module extracts temporal features from the input vector and generates a low-dimensional vector. The decoder then reconstructs the low-dimensional vector into high-dimensional global variables and Lagrange multipliers, based on the first dataset D. l Update LSTM (l) Module parameters;
[0015] The LSTM-VAE (h) The network is powered by LSTM (h) Modules and VAE (h) The modules are composed of a second dataset D. h Training a high-precision LSTM-VAE (h) The network includes: using the second dataset D h Training VAE (h) After the module, VAE (h) The module's parameters are frozen, and its decoder is coupled with LSTM. (h) Modules combined, using LSTM (h) The module extracts temporal features from the input vector and generates a low-dimensional vector. The decoder then reconstructs this low-dimensional vector into high-dimensional global variables and Lagrange multipliers, based on the second dataset D. h Update LSTM (h) Module parameters.
[0016] Furthermore, VAE module refers to VAE. (l) Modules and VAE (h) Each module in the VAE module includes an encoder, a normal distribution sampling module, and a decoder. The encoder maps the high-dimensional global variables z and Lagrange multipliers λ in the distributed optimal power flow problem to a low-dimensional latent vector representation l. The decoder reconstructs the low-dimensional latent vector l into high-dimensional global variables z' and Lagrange multipliers λ'. The loss function for training is shown in the following equation:
[0017]
[0018] In the formula uVAE is the true reconstruction value of VAE, u * is the true convergence value of ADMM, d represents the dimension of low-dimensional latent vector l, μ l,(i) and σ l,(i) represent the mean and variance of the i-th dimension of latent vector l, respectively.
[0019] Further, the LSTM module refers to the LSTM (l) module and any of the LSTM (h) modules, the LSTM module includes an input layer, a plurality of hidden layers, a full connection layer and an output layer, the LSTM module extracts the time sequence characteristics of high-dimensional data u, and estimates a low-dimensional latent vector l based on the time sequence characteristics, the latent vector l is reconstructed into a high-dimensional ADMM convergence value u' by the trained VAE module, and the loss function of the LSTM module includes two parts, an accuracy loss and a penalty term loss, as shown in the following formula:
[0020]
[0021] where ω acc and ω pen represent the weights of the accuracy loss and the penalty term loss, respectively, L acc represents the accuracy loss, represents the penalty term loss corresponding to z g , where z g is an element in z, u * represents the accurate convergence value of the distributed optimal power flow global variable and the Lagrange multiplier u * = [z * , λ * ]; Card (u * ) represents the dimension of vector u * , represents an element in λ g ', λ g 'represents a set of all Lagrange multiplier estimation values corresponding to z g , is a normalization factor.
[0022] Further, the alternating direction multiplier method is performed k * times of iterations, and the trained high-precision LSTM-VAE (h) network is used to estimate the convergence value of the distributed optimal power flow global variable and the Lagrange multiplier, including:
[0023] The alternating direction multiplier method is performed k * times of iterations, and the input vector u = [u (1) , u (2) , …, u (k*) ] is collected, where u(k) =[z (k) ,λ (k) ], k = 1, 2, ..., k * And input it into the trained LSTM-VAE (h) In the network, estimate the convergence value u' = [z', λ'] of the global variables and Lagrange multipliers of the distributed optimal power flow.
[0024] A distributed optimal power flow acceleration solution system includes:
[0025] The region division module, based on the physical characteristics of the distribution network topology, uses electrical coupling strength to evaluate the connectivity between nodes and community mining theory to divide the distribution network into several regions. Based on the divided regions, the optimal power flow problem of the distribution network is decomposed into optimal power flow subproblems within each region.
[0026] The offline training module uses the first convergence tolerance. Performing the alternating direction multiplier method M l The first dataset D with low precision is generated. l Using the first dataset D l Training low-precision LSTM-VAE (l) Network; Training-based low-precision LSTM-VAE (l) Network acceleration with second convergence tolerance The alternating direction multiplier method is executed to generate M. h From one high-precision data set, a second high-precision dataset D is obtained. h And using the second dataset D h Training a high-precision LSTM-VAE (h) Network; where the first convergence tolerance Greater than the second convergence tolerance First dataset D l Second dataset D h This includes the values of the global variables z and their corresponding Lagrange multipliers λ used to coordinate various regions of the distribution network during the iterative process of the alternating direction multiplier method, denoted by u = [z, λ], and the convergence values of the global variables and Lagrange multipliers u' = [z', λ'].
[0027] The online solver module executes k * The alternating direction multiplier method is used in the next iteration, and the trained high-precision LSTM-VAE is utilized. (h) The network estimates the convergence values of the global variables and Lagrange multipliers of the distributed optimal power flow; the estimated convergence values are used as the hot-start solution of the alternating direction multiplier method, and the alternating direction multiplier method iteration is continued until the algorithm converges.
[0028] The application further provides an electronic device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the program is implemented when executed by the processor to realize the distributed optimal power flow accelerated solving method as described above.
[0029] The application further provides a computer readable storage medium, which stores a computer program, and the computer program is implemented when executed by the processor to realize the distributed optimal power flow accelerated solving method as described above.
[0030] The application further provides a computer program product, comprising a computer program, and the computer program is implemented when executed by the processor to realize the distributed optimal power flow accelerated solving method as described above.
[0031] Advantages:
[0032] (1) The distributed optimal power flow solving method of the application regards the iteration process of the alternating direction multiplier method ADMM as a time sequence process, estimates the convergence value of the ADMM algorithm through the LSTM-VAE network, and takes the convergence value as a hot start solution of the ADMM algorithm, so that the solving quality is ensured and the time required for solving the distributed optimal power flow is reduced.
[0033] (2) The distributed optimal power flow solving method of the application considers the constraint condition of the Lagrange multiplier in the process of training the LSTM-VAE, so that the estimated convergence value can meet the optimality condition. BRIEF DESCRIPTION OF DRAWINGS
[0034] Figure 1 is a flowchart of the solving method of the application;
[0035] Figure 2 is an improved IEEE123 node distribution network topology partitioning diagram in the embodiment of the application;
[0036] Figure 3 is an offline training process and online solving process schematic diagram of the LSTM-VAE in the embodiment of the application;
[0037] Figure 4 is a network structure diagram of the LSTM-VAE in the embodiment of the application;
[0038] Figure 5 is a convergence curve diagram of the method for solving the distributed optimal power flow proposed by the application. DETAILED DESCRIPTION
[0039] In order to have a clearer understanding of the features and advantages of the technical solutions of the application, the composition and implementation of the specific solutions are described below with reference to the drawings.
[0040] Referring to Figure 1 , the application proposes a distributed optimal power flow accelerated solving method, comprising the following steps:
[0041] Step one, considering the physical characteristics of the distribution network topology, the connectivity between nodes is evaluated by using the electrical coupling strength e ij between node i and node j, and the distribution network is divided into several regions by using community mining theory to maximize the modularity Q.
[0042] In this embodiment, the partition of the improved IEEE-123 node distribution network is as shown in Figure 2 , in which a total of 12 photovoltaic generators and 12 static var compensators are configured in the IEEE-123 node, and the example is divided into five regions. It should be noted that the distribution network partition in this embodiment is made according to the IEEE-123 node distribution network of this embodiment, and when the method of the application is actually implemented, the distribution network region partition can be adaptively adjusted according to the structure of the actual application distribution network.
[0043] The application maximizes the modularity Q as the basis for partitioning the distribution network, and the modularity is in the form of formula (1-1)-(1-2).
[0044]
[0045] In the formula, E is a symmetric matrix composed of elements e ij , d i represents the sum of the i-th row elements of the matrix E, and C i represents the region to which node i belongs. When C i =C j , δ(C i C j )=1, otherwise, δ(C i C j )=0.
[0046] According to the divided regions, the optimal power flow problem of the distribution network is divided into optimal power flow sub-problems in each region, and the distributed optimal power flow realizes power system optimization by solving the optimal power flow sub-problems in each region in parallel. In the alternating direction multiplier method (ADMM) framework, each region solves a local optimal power flow sub-problem containing global variable constraints in turn, and updates the global variables and Lagrange multipliers through multiple iterations until a convergent coordinated solution is generated, and the solving accuracy and speed are improved through the LSTM-VAE network. In the optimization problem, the Lagrange multiplier is a coefficient related to the constraint condition, which is used to introduce the influence of the constraint in the objective function. In ADMM, the Lagrange multiplier is used to balance the relationship between the objective function and the constraint. Specifically, the following steps two to six are included.
[0047] Step two, based on the two-stage data generation strategy, first generate low-precision data, and then use low-precision data to train low-precision LSTM-VAE (l) network, and based on the low-precision LSTM-VAE (l) network, generate a large number of high-precision training data.
[0048] The two-stage data generation strategy of the present application is as follows: first, with a larger first convergence tolerance Perform the traditional ADMM algorithm M l , generate a low-precision first data set D l , and use the low-precision first data set D l to train the low-precision LSTM-VAE (l) network. Subsequently, based on the trained LSTM-VAE (l) network, accelerate the ADMM performed with a smaller second convergence tolerance , generate M h high-precision data, and obtain a high-precision second data set D h . Figure 3 In the offline training phase shown in the figure, the high-precision second training set D h is used to train the high-precision LSTM-VAE (h) network, and the trained LSTM-VAE (h) network is used for Figure 3 the online solution phase shown in the figure, to assist the ADMM to obtain the optimal power flow solution.
[0049] In the context of the present application, the above subscript (l) represents a low-precision network, and the above subscript (h) represents a high-precision network, which are different representations of the same LSTM-VAE network in different stages. In fact, the network structure is unchanged, and the training process of the network is also the same, the difference lies in that the low-precision network is trained with the first data set D l , and the high-precision network is trained with the second data set D h . The LSTM-VAE network is composed of an LSTM module and a VAE module, wherein the LSTM module outputs a low-dimensional hidden vector l by extracting the time sequence features of the high-dimensional data u. The VAE module encodes the high-dimensional data u into a low-dimensional hidden vector through an encoder-decoder structure, and then reconstructs it into a high-dimensional ADMM convergence value u'. After the encoder-decoder training is completed, the decoder of the VAE is used in combination with the LSTM, that is, the LSTM replaces the encoder part, and the low-dimensional vector output by the LSTM is converted into high-dimensional u' through the decoder. Figure 4 The network structure diagram of the LSTM-VAE in the embodiment of the present application is shown.
[0050] Since high-precision networks are directly applied to online solutions, their performance is directly affected by D. h Due to the impact of [the above], and for the sake of convenience, the training process of the high-precision network will be described in detail in steps three and four below. The training process of the low-precision network is the same and will not be repeated here. Furthermore, low-precision LSTM-VAE [is also mentioned]. (l) Network-accelerated ADMM for generating high-precision second dataset D h The process is similar to the high-precision LSTM-VAE described in steps five and six below. (h) The process of network-assisted ADMM is the same, the only difference is that a low-precision LSTM-VAE is used. (l) The network, i.e., the second dataset D h Through low-precision LSTM-VAE (l) The optimal solution for distributed optimal power flow generated by network-assisted ADMM will not be repeated here.
[0051] The dataset of this invention specifically includes the top k * In the next iteration, the values of the global variables z and their corresponding Lagrange multipliers λ are used to coordinate the values of the global variables z and λ in each region of the distribution network. The value u is denoted by u = [z, λ] and the convergence value u' of the global variables and Lagrange multipliers.
[0052] In this embodiment M l =100, M h =300, with load fluctuation range of [90%, 110%]. Table 1 shows the number of iterations required to generate training data.
[0053] Table 1 shows the number of iterations required to generate training data.
[0054]
[0055] Step 3: Based on the high-precision training dataset D h Training high-precision VAEs (h) .
[0056] Specifically, VAE (h) The module includes an encoder, a normal sampling method, and a decoder. The encoder maps the high-dimensional global variable z and the Lagrange multiplier λ in the distributed optimal power flow problem to a vector representation l in a low-dimensional space. The normal sampling method introduces randomness into the output of the encoder. The decoder is used to decode and reconstruct the low-dimensional vector l obtained by random sampling into the high-dimensional global variable z' and the Lagrange multiplier λ'.
[0057] Training VAE (h) The process is as follows: First, the training set D... h The global variable and the target value u of the Lagrange multiplier* The input is fed into the encoder, which produces a low-dimensional vector l. Subsequently, the decoder decodes l to obtain the reconstructed target value u. VAE The VAE is completed by minimizing its loss function. (h) The loss function for training the network is as follows:
[0058]
[0059] Where d is the dimension of l, σ l,(i) and μ l,(i) Let be the mean and variance of the values in the i-th dimension, respectively.
[0060] Step 4: VAE (h) Parameters are frozen, and its decoder module is connected to LSTM. (h) Combined, by LSTM (h) Temporal features of the input vector are extracted and a low-dimensional vector is generated, which is then processed by a VAE. (h) The LSTM module parameters are then reconstructed into high-dimensional global variables and Lagrange multipliers, and further updated based on high-precision training data. That is, the output of the LSTM is used as the VAE. (h) The decoder's input updates only the LSTM parameters during this process. Its loss function consists of two parts: accuracy loss and penalty term loss, as shown in equations (1-4)-(1-7).
[0061]
[0062]
[0063] Where ω acc and ω pen L represents the weights of the accuracy loss and the penalty term loss, respectively. acc Indicates a loss of accuracy. Indicate z g The corresponding penalty term loss, where z g For an element in z, u * The exact convergence value u of the global variable and Lagrange multiplier of the distributed optimal power flow * =[z * ,λ * Card(u) * ) represents vector u * Dimensions λ g The element in ', λ g 'represents z g The set of all corresponding Lagrange multiplier estimates, for The normalization factor.
[0064] trained LSTM-VAE (h) structure is LSTM (h) + VAE (h) decoder. The training process of steps two, three, and four is an offline training process, as shown in Figure 3
[0065] step five, using the trained LSTM-VAE (h) estimate ADMM convergence value: perform k * times ADMM iterations, collect input vectors u = [u (1) , u (2) , …, u (k*) ], where u (k) = [z (k) , λ (k) ], k = 1, 2, …, k * , and input them into the trained LSTM-VAE (h) , the vector u is first input into the LSTM (h) module to generate a low-dimensional vector representation of the estimate l, and then l is decoded by the decoder module of the VAE (h) to output the estimated convergence value of the distributed optimal power flow global variables and Lagrange multipliers u' = [z', λ']. In this example, k * = 10.
[0066] step six, use the estimated convergence value u' as the warm start solution of the ADMM algorithm, and continue to perform ADMM iterations until the algorithm converges, at which time the system obtains the optimal solution of the distributed optimal power flow, i.e., the optimal power flow solution of each sub-region and the final convergence value of the global variables and Lagrange multipliers.
[0067] To further demonstrate the superiority of the method proposed in the present application, the number of iterations required by the ADMM based on LSTM-VAE to solve the distributed optimal power flow is tested, and Table 2 summarizes the number of iterations required by the traditional ADMM algorithm and the method proposed in the present application to solve the distributed optimal power flow. The proposed method reduces the number of iterations from 87 to 60, achieving a speedup of 1.45 times. The method proposed in the present application performs a total of 60 iterations to solve the distributed OPF problem of the modified IEEE-123-bus system. The first 10 iterations are dedicated to generating input vectors u for the LSTM-VAE model. The LSTM-VAE then uses u to generate a warm start solution for the ADMM. Based on the warm start solution, the ADMM performs an additional 50 iterations to reach convergence.
[0068] Table 2 Number of iterations required to solve the distributed optimal power flow
[0069]
[0070] To further demonstrate the superiority of the model, the convergence curves of different methods were compared, such as... Figure 5 As shown, in the 11th iteration, LSTM-VAE estimates the convergence values of the global variables and Lagrange multipliers. The large difference between the full estimate and the value in the 10th iteration leads to a sudden increase in the residuals, resulting in an overall increase in the residuals in subsequent iterations. Nevertheless, the method proposed in this invention still converges faster than the traditional ADMM. This is because the residuals decrease more rapidly compared to the traditional ADMM based on the estimates.
[0071] Based on the same technical concept as the method embodiments, the present invention also provides a distributed optimal power flow acceleration solution system, comprising:
[0072] A distributed optimal power flow acceleration solution system includes:
[0073] The region division module, based on the physical characteristics of the distribution network topology, uses electrical coupling strength to evaluate the connectivity between nodes and community mining theory to divide the distribution network into several regions. Based on the divided regions, the optimal power flow problem of the distribution network is decomposed into optimal power flow subproblems within each region.
[0074] The offline training module uses the first convergence tolerance. Performing the alternating direction multiplier method M l The first dataset D with low precision is generated. l Using the first dataset D l Training low-precision LSTM-VAE (l) Network; Training-based low-precision LSTM-VAE (l) Network acceleration with second convergence tolerance The alternating direction multiplier method is executed to generate M. h From one high-precision data set, a second high-precision dataset D is obtained. h And using the second dataset D h Training a high-precision LSTM-VAE (h) Network; where the first convergence tolerance Greater than the second convergence tolerance First dataset D l Second dataset D h This includes the values of the global variables z and their corresponding Lagrange multipliers λ used to coordinate various regions of the distribution network during the iterative process of the alternating direction multiplier method, denoted by u = [z, λ], and the convergence values of the global variables and Lagrange multipliers u' = [z', λ'].
[0075] The online solver module executes k *Alternating Direction Method of Multipliers (ADMM) with a few iterations, and a well-trained high-precision LSTM-VAE (h) The network estimates the convergence values of the global variables and the Lagrange multipliers of the distributed optimal power flow, and takes the estimated convergence values as the warm start solution of the Alternating Direction Method of Multipliers (ADMM) to continue the iteration of the ADMM until the algorithm converges.
[0076] It should be understood that the distributed optimal power flow accelerated solving system in the embodiments of the present application can implement all the technical solutions in the above method embodiments, and the functions of each functional module can be implemented according to the methods in the above method embodiments, and the specific implementation process can be referred to the related description in the above embodiments, which will not be described here.
[0077] The present application also provides an electronic device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the program is executed by the processor to implement the steps of the distributed optimal power flow accelerated solving method as described above.
[0078] The present application also provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the distributed optimal power flow accelerated solving method as described above.
[0079] It should be understood that the distributed optimal power flow accelerated solving system in the embodiments of the present application can implement all the technical solutions in the above method embodiments, and the functions of each functional module can be implemented according to the methods in the above method embodiments, and the specific implementation process can be referred to the related description in the above embodiments, which will not be described here.
[0080] The present application also provides an electronic device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the program is executed by the processor to implement the steps of the distributed optimal power flow accelerated solving method as described above.
[0081] The present application also provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the distributed optimal power flow accelerated solving method as described above.
[0082] The present application also provides a computer program product comprising a computer program, wherein the computer program is executed by a processor to implement the distributed optimal power flow accelerated solving method as described above.
[0083] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus (systems), computer devices, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied 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.
[0084] This invention is described with reference to a flowchart of a method according to embodiments of the invention. It should be understood that each step in the flowchart and combinations thereof can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the process. Figure 1 A device for a function specified in one or more processes.
[0085] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 The function specified in one or more processes.
[0086] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 Steps of a specified function in one or more processes.
Claims
1. A method for accelerating the solution of distributed optimal power flow, characterized in that, Includes the following steps: Based on the physical characteristics of the distribution network topology, the connectivity between nodes is evaluated using electrical coupling strength. Using community mining theory, the distribution network is divided into several regions, and the optimal power flow problem of the distribution network is decomposed into optimal power flow subproblems within each region according to the divided regions. With the first convergence tolerance Performing the alternating direction multiplier method M l The first dataset D with low precision is generated. l Using the first dataset D l Training low-precision LSTM-VAE (l) Network; Training-based low-precision LSTM-VAE (l) Network acceleration with second convergence tolerance The alternating direction multiplier method is executed to generate M. h From one high-precision data set, a second high-precision dataset D is obtained. h And using the second dataset D h Training a high-precision LSTM-VAE (h) Network; where the first convergence tolerance Greater than the second convergence tolerance First dataset D l Second dataset D h This includes the values of the global variables z and their corresponding Lagrange multipliers λ used to coordinate various regions of the distribution network during the iterative process of the alternating direction multiplier method, represented by high-dimensional data u = [z, λ], and the convergence values of the global variables and Lagrange multipliers u' = [z', λ']. Execute k * The alternating direction multiplier method is used in the next iteration, and the trained high-precision LSTM-VAE is utilized. (h) The network estimates the convergence values of the global variables and Lagrange multipliers of the distributed optimal power flow; the estimated convergence values are used as the hot-start solution of the alternating direction multiplier method, and the alternating direction multiplier method iteration is continued until the algorithm converges.
2. The method according to claim 1, characterized in that, Based on the physical characteristics of the distribution network topology, the connectivity between nodes is evaluated using electrical coupling strength. Using community mining theory, the distribution network is divided into several regions, including: Utilizing the electrical coupling strength e between node i and node j ij To evaluate the connectivity between nodes, community mining theory is used, with the goal of maximizing modularity Q, to divide the distribution network into several regions, represented as follows: In the formula, E represents element e. ij The symmetric matrix formed by d i Let C represent the sum of the elements in the i-th row of matrix E. i Indicates the region to which node i belongs, when C i =C j At that time, δ(C) i C j ) = 1, conversely, δ(C) = 1. i C j ) = 0.
3. The method according to claim 1, characterized in that, The LSTM-VAE (l) The network is powered by LSTM (l) Modules and VAE (l) The modules are composed of the first dataset D. l Training low-precision LSTM-VAE (l) The network includes: using the first dataset D l Training VAE (l) After the module, VAE (l) The module's parameters are frozen, and its decoder is coupled with LSTM. (l) Modules combined, using LSTM (l) The module extracts temporal features from the input vector and generates a low-dimensional vector. The decoder then reconstructs the low-dimensional vector into high-dimensional global variables and Lagrange multipliers, based on the first dataset D. l Update LSTM (l) Module parameters; The LSTM-VAE (h) The network is powered by LSTM (h) Modules and VAE (h) The modules are composed of a second dataset D. h Training a high-precision LSTM-VAE (h) The network includes: using the second dataset D h Training VAE (h) After the module, VAE (h) The module's parameters are frozen, and its decoder is coupled with LSTM. (h) Modules combined, using LSTM (h) The module extracts temporal features from the input vector and generates a low-dimensional vector. The decoder then reconstructs this low-dimensional vector into high-dimensional global variables and Lagrange multipliers, based on the second dataset D. h Update LSTM (h) Module parameters.
4. The method according to claim 3, characterized in that, VAE module refers to VAE (l) Modules and VAE (h) Each module in the VAE module includes an encoder, a normal distribution sampling module, and a decoder. The encoder maps the high-dimensional global variables z and Lagrange multipliers λ in the distributed optimal power flow problem to a low-dimensional latent vector representation l. The decoder reconstructs the low-dimensional latent vector l into high-dimensional global variables z' and Lagrange multipliers λ'. The loss function for training is shown in the following equation: In the formula u VAE u represents the actual reconstructed value of the VAE. * Let d represent the true convergence value of the alternating direction multiplier method, where d represents the dimension of the low-dimensional latent vector l, and μ represents the true convergence value. l,(i) and σ l,(i) Let represent the mean and variance of the i-th dimension of the latent vector l, respectively.
5. The method according to claim 3, characterized in that, LSTM module refers to LSTM (l) Modules and LSTM (h) Each module in the module, the LSTM module includes an input layer, several hidden layers, a fully connected layer, and an output layer. The LSTM module extracts the temporal features of high-dimensional data u and estimates the low-dimensional latent vector l based on this. The trained VAE module reconstructs the latent vector l into the high-dimensional ADMM convergence value u'. Its loss function includes two parts: accuracy loss and penalty term loss, as shown in the following formula: Where ω acc and ω pen L represents the weights of the accuracy loss and the penalty term loss, respectively. acc Indicates a loss of accuracy. Indicate z g The corresponding penalty term loss, where z g For an element in z, u * The exact convergence value u of the global variable and Lagrange multiplier of the distributed optimal power flow * =[z * ,λ * ];Card(u * ) represents vector u * Dimensions λ g The element in ', λ g 'represents z g The set of all corresponding Lagrange multiplier estimates, for The normalization factor.
6. The method according to claim 1, characterized in that, Execute k * The alternating direction multiplier method is used in the next iteration, and the trained high-precision LSTM-VAE is utilized. (h) Network estimation of the convergence values of global variables and Lagrange multipliers for distributed optimal power flow includes: Execute k * The alternating direction multiplier method in the next iteration collects the input vector u = [u (1) ,u (2) ,…,u (k*) ], where u (k) =[z (k) ,λ (k) ], k = 1, 2, ..., k * And input it into the trained LSTM-VAE (h) In the network, estimate the convergence value u' = [z', λ'] of the global variables and Lagrange multipliers of the distributed optimal power flow.
7. A distributed optimal power flow acceleration solution system, characterized in that, include: The region division module, based on the physical characteristics of the distribution network topology, uses electrical coupling strength to evaluate the connectivity between nodes and community mining theory to divide the distribution network into several regions. Based on the divided regions, the optimal power flow problem of the distribution network is decomposed into optimal power flow subproblems within each region. The offline training module uses the first convergence tolerance. Performing the alternating direction multiplier method M l The first dataset D with low precision is generated. l Using the first dataset D l Training low-precision LSTM-VAE (l) Network; Training-based low-precision LSTM-VAE (l) Network acceleration with second convergence tolerance The alternating direction multiplier method is executed to generate M. h From one high-precision data set, a second high-precision dataset D is obtained. h And using the second dataset D h Training a high-precision LSTM-VAE (h) Network; where the first convergence tolerance Greater than the second convergence tolerance First dataset D l Second dataset D h This includes the values of the global variables z and their corresponding Lagrange multipliers λ used to coordinate various regions of the distribution network during the iterative process of the alternating direction multiplier method, represented by high-dimensional data u = [z, λ], and the convergence values of the global variables and Lagrange multipliers u' = [z', λ']. The online solver module executes k * The alternating direction multiplier method is used in the next iteration, and the trained high-precision LSTM-VAE is utilized. (h) The network estimates the convergence values of the global variables and Lagrange multipliers of the distributed optimal power flow; the estimated convergence values are used as the hot-start solution of the alternating direction multiplier method, and the alternating direction multiplier method iteration is continued until the algorithm converges.
8. An electronic device, characterized in that, include: One or more processors; Memory; And one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, wherein when the programs are executed by the processors, they implement the distributed optimal power flow acceleration solution method as described in any one of claims 1-6.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the distributed optimal power flow acceleration solution method as described in any one of claims 1-6.
10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the distributed optimal power flow acceleration solution method as described in any one of claims 1-6.