A transformer network stage optimization scheduling method, device and equipment

Abstract

The application provides a transformer network stage optimization scheduling method, device and equipment. It relates to the technical field of electric power. The method comprises the following steps: constructing an optimization scheduling model with the minimum total operation cost of a target power network as the optimization target; solving the optimization scheduling model through a distributed double-layer scheduling algorithm to obtain an optimization scheduling scheme of the target power network; wherein the lower layer learning of the distributed double-layer scheduling algorithm is based on the scheduling experience contained in the historical data of each transformer network and the global coordination variable transmitted from the upper layer scheduling to the current time stage to obtain the scheduling prior knowledge of the current time stage; the upper layer scheduling performs information interaction on the scheduling prior knowledge of each transformer network in the current time stage through a time-varying directed communication graph to obtain the global coordination variable of the current time stage; and the lower layer learning and the upper layer scheduling are mutually optimized through alternating iteration until the preset convergence condition is met.

Description

Technical Field

[0001] This invention relates to the field of power technology, and in particular to a stage-optimized scheduling method, apparatus, and equipment for transformer networks. Background Technology

[0002] An integrated wind-solar-storage transformer is a type of transformer that integrates wind, solar, and energy storage systems, aiming to achieve efficient utilization and stable power supply from renewable energy sources. This transformer integrates wind power generation, photovoltaic power generation, and energy storage technologies into a single system, enabling optimized dispatch and power output based on different energy conditions, thereby improving energy utilization efficiency and grid stability. To ensure the economic efficiency of transformer operation, it is necessary to fully utilize existing technologies and equipment, and select the optimal operating mode to minimize transformer operating costs, while ensuring safe operation and guaranteed power supply.

[0003] Distributed optimization has significant advantages in adapting to the development needs of modern power systems, improving power supply reliability, and reducing energy losses, making it an important direction for future power system development. Its working principle involves decomposing the global optimization problem into multiple sub-problems, which are then solved in parallel by each distributed node based on local information. Each integrated wind-solar-storage transformer acts as an independent decision-making unit, constructing a local optimization model based on local wind and solar power output forecasts, energy storage status, and load demand. By exchanging information with neighboring transformers, it iteratively corrects its own power output plan, ultimately minimizing the overall network operating cost.

[0004] However, environmental changes at different times make wind and solar power generation highly uncertain and intermittent, posing significant challenges to distributed optimization scheduling: during the day-ahead period, prediction errors in wind speed and solar intensity may lead to large deviations between the scheduling plan and the actual output; during the intraday period, sudden weather conditions such as cloud cover and gusts can easily cause power fluctuations; making it difficult for traditional distributed algorithms based on deterministic models to guarantee the feasibility of solutions, thus restricting the full release of the economic operation potential of integrated wind-solar-storage transformers. Summary of the Invention

[0005] Therefore, it is necessary to provide a stage-based optimization scheduling method, apparatus, and equipment for transformer networks to address the aforementioned technical problems. This method can solve the scheduling optimization problem caused by the uncertainty of wind and solar power generation due to environmental changes at different time stages.

[0006] The following technical solution is adopted in this specification: This specification provides a stage-optimized scheduling method for transformer networks, including: An optimization scheduling model is constructed with the goal of minimizing the total operating cost of the target power network over all preset time periods; the target power network includes multiple transformer networks that exchange information through a time-varying directed communication graph. The optimal scheduling model is solved by a distributed two-layer scheduling algorithm to obtain the optimal scheduling scheme for the target power network; the distributed two-layer scheduling algorithm includes lower-layer learning and upper-layer scheduling. The lower layer learns from the scheduling experience contained in the local historical data of each transformer network, as well as the global coordination variables passed from the upper layer scheduling in the previous time stage to the current time stage, to obtain the scheduling prior knowledge for the current time stage. The upper-level scheduler uses a time-varying directed communication graph to exchange information on the scheduling prior knowledge of each transformer network at the current time stage, obtains the global coordination variables at the current time stage, and passes the global coordination variables at the current time stage to the next time stage. The lower-level learning and upper-level scheduling optimize each other through alternating iterations until the preset convergence conditions are met.

[0007] Furthermore, the construction of the optimization scheduling model with the goal of minimizing the total operating cost of the target power network over all preset time periods specifically includes: The optimization objective is to minimize the total operating cost of the target power network over all preset time periods. in, , Indicates the first The transformer network in the first Transformer power at each time period; Indicates the first The transformer network in the first Distribution network trading power at each time period; Indicates the first The transformer network in the first The operating cost function for each time period; Indicates the total number of transformer networks; , and These represent the weights for generation costs, power balance, and electricity trading costs, respectively. This indicates the cost of the transformer; This represents the cost of the imbalance between electricity production and demand; Indicates the transaction costs of the power distribution network; The constraints are transformer power constraints, energy storage system state constraints, ramp rate constraints, and distribution network transaction constraints. in, and They represent the first Minimum and maximum power of a transformer network; Indicates the first The transformer network in the first Battery state of charge at each time period; and They represent Minimum and maximum battery charge of a transformer network; Indicates the first The transformer network in the first Transformer power at each time period; Indicates the first The maximum rate of change of power of transformers in a transformer network.

[0008] Furthermore, before solving the optimized scheduling model using the distributed two-level scheduling algorithm, the optimized scheduling model is transformed into a two-level optimization form, wherein: The lower-level optimization form is as follows: in, Indicates the first The transformer network in the first The time phase, targeting the first A number of similar tasks The optimized parameters obtained after the lower-level training are used as the scheduling experience contained in the local historical data of each transformer network for the lower-level learning. This represents the globally shared initial parameters passed from the upper-level scheduler in the previous time phase to the current time phase, serving as the starting point for training of similar tasks in the lower layer. This refers to historical scheduling tasks sampled from a preset data buffer set that have a similarity to the current scheduling task in terms of load characteristics, renewable energy output characteristics, or network topology characteristics greater than a preset threshold. Indicates similar tasks The training dataset obtained from sampling in the middle layer is used for the lower layer to learn the prior knowledge of this task; Indicates the first The transformer network in the first Nonlinear inequality constraints at each time stage; The upper-level optimization form is as follows: in, Indicates the first The cumulative losses of all transformer networks at each time period are used to evaluate parameters. The advantages and disadvantages; Indicates the first The transformer network in the first The set of tasks observed at each time period; Indicates the first The transformer network in the first Mean squared error loss function for each time period; Indicates similar tasks The test dataset is obtained by sampling from the middle layer and used to evaluate the parameters trained in the lower layer. Performance; Indicates the first Upper-level constraint functions of a transformer network; This represents the model parameters output by the lower-level optimization.

[0009] Furthermore, the step of solving the optimized scheduling model using a distributed two-layer scheduling algorithm specifically includes: I. Initialization steps: Randomly initialize the scheduling parameters and dual parameters, where the scheduling parameters are the power scheduling decision variables of the transformer network, and the dual parameters are the Lagrange multipliers of the constraints in the optimization scheduling model; II. Stage Cycle Steps: According to the preset time stages The algorithm iterates through the time intervals 1, 2, ..., T. For each preset time period, the lower layer learns scheduling prior knowledge based on the historical data of each transformer network and the global coordination variables passed from the upper layer to the current time period. The upper layer then exchanges scheduling prior knowledge for the current time period through a time-varying directed communication graph to obtain the global coordination variables for the current time period. III. Iterative convergence step: Use the updated global coordination variable of each preset time stage as the initial value for scheduling in the next preset time stage, and repeat step II until all preset time stages are scheduled.

[0010] Furthermore, the stage cycle steps specifically include: Real-time scheduling sub-step: For each transformer network, when new data arrives sequentially at the corresponding time stage, a neural network with deployed scheduling parameters is used to perform real-time scheduling based on the current dual parameters; Data update sub-step: After all data for the corresponding time period arrives, add all data to a preset data buffer set; the data buffer set is a queue used to store historical scheduling data; Sampling initialization sub-step: Obtain a training dataset of similar scheduling tasks from the data buffer set. and the test dataset of the current scheduled task And use the current scheduling parameters as the initial model parameters for the next layer of training; Lower-level training sub-step: Based on the training dataset The lower-level training iteratively updates the scheduling parameters and dual parameters, and outputs the scheduling prior knowledge for the current time stage; the scheduling prior knowledge is the optimized scheduling parameters obtained after the lower-level training converges. Upper-level training sub-step: Based on the test dataset The cumulative loss is obtained, and combined with prior scheduling knowledge and local information exchanged between each transformer network through time-varying directed communication graphs, upper-level training is performed to obtain the global coordination variables for the current time stage. The global coordination variables are the scheduling parameters optimized by upper-level training, and the local information is the interaction data of the prior scheduling knowledge of each transformer network between neighboring nodes.

[0011] Furthermore, the training dataset-based The scheduling parameters and dual parameters are updated iteratively during the next training layer using the following formula: in, and They represent the first The transformer network in the first Second and third Scheduling parameters during the next lower-level iteration; This indicates the fixed step size for training in the lower layer; Indicates about The gradient operator; and They represent the first The transformer network in the first Second and third Dual parameters in the next lower iteration; Indicates the first The transformer network in the first Each time phase is for the task The Lagrange function; Indicates about The gradient operator.

[0012] Furthermore, obtaining the global coordination variable for the current time stage is achieved through iteration of the following formula: in, and They represent the first The transformer network in the first Time phase and the Scheduling parameters for each time phase; Represents a closed convex feasible region; Represents a symmetric positive definite matrix; superscript Representation device; Indicates the learning rate; Represents the loss function about The gradient; Represents constraint functions about The gradient; , and Both represent auxiliary variables of the Lagrange function; Indicates the first The transformer network in the first Auxiliary variables for estimating consensus state at different time stages. , In a time-varying directed communication graph, the first... The set of adjacent network nodes of a transformer network node; In a time-varying directed communication graph, the first... The transformer network node and the first The adjacency relationships of nodes in a transformer network.

[0013] Furthermore, the time-varying directed communication graph is a strongly connected balanced graph, which represents the topological structure of information interaction between transformer networks and satisfies local information characteristics, communication topology characteristics, balanced graph characteristics, and strong connectivity characteristics. The local information characteristics include: each transformer network node has information limitations and can only obtain and utilize its own cost information, operating constraints and local historical data; The communication topology characteristics include: defining the correspondence between the set of network nodes and the set of edges by giving a time-varying directed graph, and using the non-zero elements in the adjacency weighting matrix to characterize the existence of information transmission links and link weights between nodes, thereby establishing the neighbor relationship and communication direction between nodes; The characteristics of the balance graph include: for any node in the network, the total weight of information sent to neighboring nodes is equal to the total weight of information received from neighboring nodes, which represents that each transformer network maintains the conservation of input and output weights during information interaction. The strong connectivity characteristic includes: there is a unidirectional reachable directed path between any two nodes in the network, which means that any transformer network node can interact with other nodes through multi-hop relay communication, ensuring that the network has global connectivity.

[0014] This specification provides a stage-optimized scheduling device for transformer networks, comprising: The scheduling model construction module is used to construct an optimized scheduling model with the goal of minimizing the total operating cost of the target power network over all preset time periods; wherein, the target power network includes multiple transformer networks that exchange information through a time-varying directed communication graph; The optimization scheduling module is used to solve the optimization scheduling model through a distributed two-level scheduling algorithm to obtain the optimal scheduling scheme for the target power network; the distributed two-level scheduling algorithm includes lower-level learning and upper-level scheduling. The lower layer learns from the scheduling experience contained in the local historical data of each transformer network, as well as the global coordination variables passed from the upper layer scheduling in the previous time stage to the current time stage, to obtain the scheduling prior knowledge for the current time stage. The upper-level scheduler uses a time-varying directed communication graph to exchange information on the scheduling prior knowledge of each transformer network at the current time stage, obtains the global coordination variables at the current time stage, and passes the global coordination variables at the current time stage to the next time stage. The lower-level learning and upper-level scheduling optimize each other through alternating iterations until the preset convergence conditions are met.

[0015] This specification provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the aforementioned stage-optimized scheduling method for a transformer network.

[0016] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects: This invention employs a two-layer iterative mechanism of "lower-layer learning to mine experience - upper-layer scheduling dynamic collaboration" to accurately address the uncertainties of wind and solar power generation across multiple time scales. For prediction errors at different stages, lower-layer learning utilizes historical data to extract scheduling experience and generate prior knowledge, providing a robust decision-making basis for uncertain environments and effectively compensating for the shortcomings of a single deterministic model. For sudden fluctuations at different stages, upper-layer scheduling leverages time-varying directed graphs to achieve real-time information exchange between nodes and cross-stage transmission of global coordination variables, enabling scheduling strategies to be dynamically corrected and continuously optimized as the environment changes. This organic combination of data-driven learning and distributed collaborative solution breaks through the adaptability bottleneck of traditional static models to source-load fluctuations, ensuring the feasibility of the solution and system stability while fully releasing the economic operation potential of the integrated wind-solar-storage transformer. Attached Figure Description

[0017] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0018] Figure 1 This document provides a flowchart illustrating a stage-optimized scheduling method for a transformer network. Figure 2 This document provides a flowchart of a phased distributed economic scheduling method. Figure 3 This specification provides a time-varying directed communication graph for a 5-node system. Figure 4 This specification provides a dynamic regret graph; Figure 5 This specification provides a constraint violation diagram; Figure 6 This specification provides an error diagram of online scheduling power versus optimal scheduling power. Figure 7 This is a diagram of a microgrid energy structure provided in this specification. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.

[0020] With increasing attention and emphasis on renewable energy, clean energy sources such as wind and solar power are gradually becoming an important part of the power system. However, the unstable and intermittent nature of these renewable energy generation sources makes effective scheduling and management difficult within the power system. Wind-solar-storage integrated transformers combine wind, photovoltaic power generation, and energy storage systems to achieve energy complementarity and rational allocation, thereby improving the utilization rate of renewable energy and the overall system flexibility. The dynamic changes in environment and electricity demand at different stages require intelligent devices to have the ability to quickly adapt to new tasks, and online distributed two-layer scheduling technology can utilize local real-time data and neighbor information for decision-making, effectively improving energy utilization efficiency and system flexibility. This invention focuses on a type of wind-solar-storage integrated transformer network composed of generators, transformers, and energy storage devices, and studies its stage scheduling problem. In this problem, environmental changes at different time stages make wind and solar power generation uncertain, posing difficulties for optimal scheduling. For each time stage, this problem can be described as a distributed optimization problem with nonlinear inequality constraints and convex set constraints, where each transformer network can only access local cost functions, local inequality constraints, local convex set constraints, and local data information, and can only communicate with neighbors through a time-varying directed graph. To address this problem, this invention proposes a novel distributed two-layer scheduling algorithm suitable for online environments. In this algorithm, the lower layer uses a gradient descent-based primal-dual strategy to learn prior knowledge from similar tasks. The upper layer uses an online distributed primal-dual strategy based on auxiliary optimization to summarize experience from similar tasks and quickly achieve optimal scheduling. Finally, the effectiveness of the proposed algorithm is verified using real power grid data provided by the Open Power System Data Platform.

[0021] The following is combined Figures 1-7 The present invention describes a phase-optimized scheduling method for transformer networks.

[0022] Figure 1 This document provides a flowchart illustrating a stage-optimized scheduling method for transformer networks. Figure 1 As shown, the method includes the following: S1. Construct an optimization scheduling model with the goal of minimizing the total operating cost of the target power network over all preset time periods; wherein, the target power network includes multiple transformer networks that exchange information through time-varying directed communication graphs.

[0023] For example, the target power network is... A system consisting of a transformer network is denoted as a set. The transformer networks communicate via a time-varying communication diagram. Local information exchange is performed. Time-varying communication diagram in a transformer network. It is a type A strongly connected balanced graph represents the topological structure of information interaction between transformer networks, satisfying local information characteristics, communication topology characteristics, balanced graph characteristics, and strong connectivity characteristics. Local information characteristics include: each transformer network node has limited information access, only able to obtain and utilize its own cost information, operational constraints, and local historical data. Communication topology characteristics include: defining the correspondence between the set of network nodes and the set of edges using a given time-varying directed graph, and using non-zero elements in the adjacency weighting matrix to represent the existence of information transmission links and link weights between nodes, thereby establishing neighbor relationships and communication directions. Balanced graph characteristics include: for any node in the network, the total weight of information sent to neighboring nodes is equal to the total weight of information received from neighboring nodes, representing the conservation of input and output weights during information interaction in each transformer network. Strong connectivity characteristics include: there exists a unidirectional reachable directed path between any two nodes in the network, representing that any transformer network node can interact with any other node through multi-hop relay communication, ensuring global connectivity of the network. The definition of the graph is as follows:

[0024] Given a time-varying directed graph .in: Represents a set of nodes. Represents the edge set, a nonnegative matrix. It is an adjacency weighted matrix of edges. If ,but ,otherwise If the node Information can be transmitted to nodes Then it is called a node. For nodes Neighbors. Definition For nodes exist The set of all neighbors at any given time. If for all , If all are true, then the graph is called a graph. It is a balance diagram. The Laplacian matrix of the diagram is... Time is defined as ,in ,like , For a fixed directed graph In other words, if a finite sequence of nodes exists Satisfy when hour That is, there exists a node To the node A graph is a directed path that can be found in a graph. It is strongly connected. If there exist two positive real numbers... and For any All , then call It is a Edge. If by edge Definite If it is strongly connected, then it is called The corresponding directed graph for Strongly connected graph.

[0025] For example, based on the above... Time-varying communication diagram composed of transformer networks The goal of the economic dispatch problem is to ensure that each transformer network in the system operates within the specified parameters. The total operating cost is minimized under normal operating conditions during the phase. This invention models the economic scheduling problem as a distributed optimization problem as follows:

[0026] in, , Indicates the first The transformer network in the first Transformer power at each time period; Indicates the first The transformer network in the first Distribution network trading power at each time period; Indicates the first The transformer network in the first The operating cost function for each time period; Indicates the total number of transformer networks; , and These represent the generation cost weight, power balance weight, and electricity trading cost weight, respectively, and are used to determine the relative importance of minimizing generation cost, power balance, and electricity trading cost. This indicates the cost of the transformer; This represents the cost of the imbalance between electricity production and demand; This represents the transaction cost of the power distribution network. and They represent the first Minimum and maximum power of a transformer network; Indicates the first The transformer network in the first Battery state of charge at each time period; and They represent Minimum and maximum battery charge of a transformer network; Indicates the first The transformer network in the first Transformer power at each time period; Indicates the first The maximum rate of change of power of the transformers in a transformer network determines the transformer's... The maximum power output that can vary in the next time period.

[0027] To better address this problem, this invention transforms it into a two-layer optimization form based on a neural network method. In one specific embodiment, consider a problem consisting of... A transformer network consisting of 10 transformers is denoted as a set. Each transformer is connected via a time-varying communication diagram. Local information exchange is performed. Different tasks are stored in different agents, and tasks or data cannot be transferred between agents. Defined as the set of tasks observed by the agent, the distribution of which may change over time. Let... Defined as the first An intelligent agent in time The observed first One task, . and Defined from The sampled datasets contain, respectively Individual and One data point. That is... , For the dataset The loss function evaluated in the middle uses the symbol To minimize the total operating cost under normal conditions, for A two-level optimization problem was designed.

[0028] The lower-level objective is for each agent to learn prior knowledge from local data on similar tasks. The lower-level optimization takes the form of: in, Indicates the first The transformer network in the first The time phase, targeting the first A number of similar tasks The optimized parameters obtained after the lower-level training are used as the scheduling experience contained in the local historical data of each transformer network for the lower-level learning. This represents the globally shared initial parameters passed from the upper-level scheduler in the previous time phase to the current time phase, serving as the starting point for training of similar tasks in the lower layer. This refers to historical scheduling tasks sampled from a preset data buffer set that have a similarity to the current scheduling task in terms of load characteristics, renewable energy output characteristics, or network topology characteristics greater than a preset threshold. Indicates similar tasks The training dataset obtained from sampling in the middle layer is used for the lower layer to learn the prior knowledge of this task; Indicates the first The transformer network in the first Nonlinear inequality constraints at each time stage.

[0029] The upper-layer agent summarizes the prior knowledge learned from the lower-layer agent and the information exchanged between agents to learn a good initial model. The upper-layer optimization form is as follows: in, Indicates the first The cumulative losses of all transformer networks at each time period are used to evaluate parameters. The advantages and disadvantages; Indicates the first The transformer network in the first The set of tasks observed at each time period; Indicates the first The transformer network in the first Mean squared error loss function for each time period; Indicates similar tasks The test dataset is obtained by sampling from the middle layer and used to evaluate the parameters trained in the lower layer. Performance; Indicates the first Upper-level constraint functions of a transformer network; This represents the model parameters output by the lower-level optimization. It is a closed convex set. and Is it a convex function? The function is twice differentiable. The Lagrangian function of the upper-level optimization problem is established as...

[0030] in: Let represent the dual variable and auxiliary variable of the original variable inequality constraint, respectively. Based on the Lagrangian function, Slater's condition, and convex optimization theory, the KKT conditions for the optimization problem are as follows:

[0031] in: , , , , A solution to an optimization problem is a dual variable that exists for inequality constraints if and only if such dual variables exist. This makes the above conditions true. It is a Lagrange multiplier. It is the decay step size. It is an auxiliary variable. , . It is an intelligent agent An estimate of the consensus state. It is a symmetric and positive definite matrix.

[0032] After transforming the optimized scheduling model of step S1 into a two-level optimization form, step S2 can be executed.

[0033] S2. Solve the optimized scheduling model using a distributed two-layer scheduling algorithm to obtain the optimized scheduling scheme for the target power network. The distributed two-layer scheduling algorithm includes lower-layer learning and upper-layer scheduling. Lower-layer learning obtains prior scheduling knowledge for the current time stage based on the scheduling experience contained in the local historical data of each transformer network and the global coordination variables passed from the upper-layer scheduling in the previous time stage. Upper-layer scheduling exchanges the prior scheduling knowledge of each transformer network in the current time stage through a time-varying directed communication graph to obtain the global coordination variables for the current time stage, and then passes these global coordination variables to the next time stage. Lower-layer learning and upper-layer scheduling iteratively optimize each other until a preset convergence condition is met. This may include the following steps:

[0034] I. Initialization steps: Randomly initialize the scheduling parameters and dual parameters, where the scheduling parameters are the power scheduling decision variables of the transformer network, and the dual parameters are the Lagrange multipliers of the constraints in the optimization scheduling model; II. Stage Cycle Steps: According to the preset time stages The algorithm iterates through the time intervals 1, 2, ..., T. For each preset time period, the lower layer learns the scheduling prior knowledge for the current time period based on the scheduling experience contained in the local historical data of each transformer network, as well as the global coordination variables passed from the upper layer scheduling in the previous time period to the current time period. The upper layer scheduling then exchanges the scheduling prior knowledge of each transformer network for the current time period through a time-varying directed communication graph to obtain the global coordination variables for the current time period.

[0035] III. Iterative convergence step: Use the updated global coordination variable of each preset time stage as the initial value for scheduling in the next preset time stage, and repeat step II until all preset time stages are scheduled.

[0036] For example, stage loop step II can be understood as for Based on the two-level optimization method, the following stage optimization scheduling algorithm was designed: the lower-level update is based on the primal-dual method of gradient descent, letting... This indicates the update process of the lower-level steps. Then we have:

[0037] 1. From Sampling ; 2. Order ; ; 3. Output: .

[0038] The upper-level update uses an online distributed primal-dual method based on an auxiliary optimization strategy. The optimization algorithm is as follows: ; in: Is an agent state, . It is an auxiliary variable. , . Is an agent An estimate of the consensus state. It is a non-increasing learning rate, that is . It is a symmetric positive definite matrix. A more detailed description of the phase loop steps can be given as follows:

[0039] 1. Randomly initialize the primitive and dual primitive parameters. ; 2.for ; 3.for ; 4. When When new data arrives in sequence, use To perform real-time scheduling; 5. When all data arrives at time t Conduct training; 6. From the data buffer set Mid-sampling .make , ;

[0040] 7. Use Conduct lower-level training, among which , As the initial model for the lower layer; for : ; endfor; Output 8. Calculate , Used for upper-level training: ; 9. Obtain Decision-making at all times , , as the initial value for scheduling in the next moment; 10.endfor; 11.endfor.

[0041] Step 4 involves applying the initial model obtained from meta-training to the economic scheduling problem after each new data arrival, to obtain the optimal scheduling solution. When all data arrives at time t, meta-training of both inner and outer layers is performed. Step 7 involves inner layer training updating the model and learning prior knowledge under similar tasks. Step 8 involves outer layer training using the prior knowledge obtained from inner layer training under similar tasks to find the optimal initial model. Figure 2 A flowchart of a phased distributed economic scheduling method provided in this specification is shown below. Figure 2 As shown, the above loop can be understood as the following sub-steps:

[0042] Real-time scheduling sub-step: For each transformer network, when new data arrives sequentially at the corresponding time stage, a neural network with deployed scheduling parameters is used to perform real-time scheduling based on the current dual parameters.

[0043] Data update sub-step: When all data for the corresponding time period arrives, add all data to a preset data buffer set; the data buffer set is a queue used to store historical scheduling data.

[0044] Sampling initialization sub-step: Obtain a training dataset of similar scheduling tasks from the data buffer set. and the test dataset of the current scheduled task The current scheduling parameters are used as the initial model parameters for the next layer of training.

[0045] Lower-level training sub-step: Based on the training dataset The lower-level training iteratively updates the scheduling parameters and dual parameters, outputting the scheduling prior knowledge for the current time stage; the scheduling prior knowledge is the optimized scheduling parameters obtained after the lower-level training converges. in, and They represent the first The transformer network in the first Second and third Scheduling parameters during the next lower-level iteration; This indicates the fixed step size for training in the lower layer; Indicates about The gradient operator; and They represent the first The transformer network in the first Second and third Dual parameters in the next lower iteration; Indicates the first The transformer network in the first Each time phase is for the task The Lagrange function; Indicates about The gradient operator.

[0046] Upper-level training sub-step: Based on the test dataset The cumulative loss is obtained, and combined with prior scheduling knowledge and local information from the interaction of time-varying directed communication graphs among the transformer networks, upper-level training is performed to obtain the global coordination variables for the current time stage. The global coordination variables are the scheduling parameters optimized by upper-level training, and the local information is the interaction data of prior scheduling knowledge of each transformer network among neighboring nodes. in, and They represent the first The transformer network in the first Time phase and the Scheduling parameters for each time phase; Represents a closed convex feasible region; Represents a symmetric positive definite matrix; superscript Representation device; Indicates the learning rate; Represents the loss function about The gradient; Represents constraint functions about The gradient; , and Both represent auxiliary variables of the Lagrange function; Indicates the first The transformer network in the first Auxiliary variables for estimating consensus state at different time stages. , In a time-varying directed communication graph, the first... The set of adjacent network nodes of a transformer network node; In a time-varying directed communication graph, the first... The transformer network node and the first The adjacency relationships of nodes in a transformer network.

[0047] This invention aims to design a distributed bi-layer optimization algorithm based on neural network methods, ensuring that the model can continuously learn using limited data to adapt to transformer network scheduling problems at different stages, thereby minimizing the operating cost of the transformer network under normal operating conditions. First, the economic scheduling problem of the transformer network is modeled as a distributed optimization problem. Second, a class of Lagrangian functions is constructed for this optimization problem, and first-order optimality conditions for the optimal point are established based on convex optimization theory. Finally, based on the derived first-order optimality conditions, a continuous-time distributed primal-dual algorithm is designed by combining primal-dual and consensus strategies. When running this algorithm, each transformer makes decisions using only local information. Therefore, this algorithm can protect the privacy of the transformer network, reduce network communication costs, improve network robustness, and ensure efficient network operation. A bi-layer optimization strategy is adopted, with the lower layer learning prior knowledge of similar tasks and the upper layer summarizing prior knowledge of similar tasks to achieve fast scheduling. Existing neural network methods require a large amount of data for training on the target task to achieve good results; this algorithm can accelerate training by utilizing prior knowledge of similar tasks when the target task data is insufficient. Furthermore, this algorithm employs a type of time-varying directed graph for local information exchange, which is less restrictive than the conditions for fixed undirected graphs. Therefore, this algorithm has wider applicability.

[0048] The beneficial effects of this invention are as follows: Compared to traditional centralized optimization methods, distributed optimization of transformer networks has the following advantages: it eliminates the need for long-distance power transmission, reducing energy loss during power transmission; the optimization information is stored independently in each transformer, protecting transformer privacy; it only requires local information exchange with its neighbors, saving communication costs; the failure of one transformer does not affect the entire network, exhibiting strong robustness; and distributed algorithms can be used to perform parallel computing, which helps solve the economic scheduling problem of large-scale transformer networks.

[0049] Compared to traditional data-driven methods, the two-layer scheduling method for transformer networks has the following advantages: when the target transformer data is insufficient for training, prior knowledge can be learned from similar transformer scheduling tasks with insufficient data to compensate for this. General power grid economic scheduling problems only consider coupled equality constraints and convex set constraints; this invention considers a more general problem. In practical applications, ramp rate constraints can often be modeled as nonlinear inequality constraints. Therefore, this invention can be applied to solving economic scheduling problems under more complex constraints such as ramp rate constraints.

[0050] This invention proposes a phased distributed two-layer scheduling algorithm. Agents do not share original training data but can exchange information through a time-varying directed graph for decision-making, thus making the algorithm fully distributed. Furthermore, when the cumulative bias of the minimized sequence grows at a specific rate, the degree of violation of dynamic regret value and coupling inequality constraints both exhibit a sublinear growth trend. Notably, the communication graph required by the algorithm is a time-varying directed graph. Compared to fixed undirected graphs, time-varying directed graphs are more general. Therefore, this algorithm can reduce communication costs and is suitable for situations where communication networks are unstable.

[0051] In one specific embodiment, electricity data from five regions of Italy in 2018 is used, and each agent can only access data from one region. Consider a power network consisting of five transformer networks using a set... This indicates that it includes wind and solar power generation, transformers, and energy storage devices. Each transformer network communicates with its adjacent transformer networks, such as... Figure 3 in Four possible communication diagrams are provided. The switching order is specified as follows. The goal of the economic dispatch problem is to determine the output power of each transformer to minimize the total operating cost while satisfying various power constraints. The distributed economic dispatch problem in a transformer network can be specifically described as follows:

[0052] ; ; ; ; in: , For the first A transformer network in The transformer power and the power exchanged with the distribution network at any given time. It is the maximum rate of change of power of the transformer, which determines the transformer's... The maximum power output that can vary in the next time period. It is a moment The battery state of charge. The weights are set as follows: , , . For the first A transformer network in The operating cost function at time t is derived from the transformer cost function. Costs based on the imbalance between electricity production and demand Transaction costs with the distribution network composition:

[0053] ; in , , These are constants determined based on the actual conditions of the transformer network. and It is the real-time electricity price in the trading market. It is based on wind and solar power generation, load demand, and battery energy storage. It can be written as:

[0054] ; in, Defined as The supply and demand difference at any given moment. Defined as battery in Energy that can be absorbed or released at any time. The expression is as follows:

[0055] ; in, Defined as the maximum charge and discharge rate of the battery. and These represent the active power of wind power and photovoltaic power generation, respectively. For load demand response.

[0056] The system parameters of the transformer network are shown in Table 1.

[0057] Table 1 System parameters of the transformer network To solve the economic dispatch problem in transformer networks, for Based on neural network methods, a phased distributed economic scheduling algorithm was designed for time-varying directed graphs, as shown below: 1. Randomly initialize the primitive and dual primitive parameters. ; 2.for ; 3.for ; 4. When When new data arrives in sequence, use To perform real-time scheduling; 5. When all data arrives at time t Conduct training; 6. From the data buffer set Mid-sampling .make , ;

[0058] 7. Use Conduct lower-level training, among which , As the initial model for the lower layer; for : ; endfor; Output 8. Calculate , Used for upper-level training: ; 9. Obtain Decision-making at all times , , as the initial value for scheduling in the next moment; 10.endfor; 11.endfor.

[0059] Among them, let step length Step length To prevent overcharging and over-discharging, the maximum and minimum capacity constraints for the battery are as follows: and Each agent's neural network model has two hidden layers of size 40 and uses the ReLU activation function. The output layer uses the sigmoid function to restrict the output range to [0,1]. The actual predicted power generation can be obtained by multiplying the output by the power generation, thus ensuring that the predicted value always meets the generator's maximum power constraint.

[0060] Figure 4 and Figure 5 The results of dynamic regret and constraint violation were plotted separately. It can be seen that the output power can quickly adapt to changes in the task environment and asymptotically converge to the optimal power while satisfying the constraints. Figure 6 and Figure 7 Error diagrams of online dispatch power versus optimal dispatch power and microgrid energy structure diagrams were plotted, clearly showing the dispatch values ​​obtained by our method and their comparison with the optimal values. Figure 3-6It can be seen that the model can adapt quickly to changes in the environment, and the scheduling value is close to the optimal scheduling value. Therefore, the method proposed in this invention is effective.

[0061] The stage-optimized scheduling device for transformer networks provided by the present invention will be described below. The stage-optimized scheduling device for transformer networks described below can be referred to in correspondence with the stage-optimized scheduling method for transformer networks described above. The stage-optimized scheduling device for transformer networks may include: The scheduling model construction module is used to construct an optimized scheduling model with the goal of minimizing the total operating cost of the target power network over all preset time periods; wherein the target power network includes multiple transformer networks that exchange information through time-varying directed communication graphs.

[0062] The optimization scheduling module is used to solve the optimization scheduling model through a distributed two-level scheduling algorithm to obtain the optimal scheduling scheme for the target power network; the distributed two-level scheduling algorithm includes lower-level learning and upper-level scheduling. The lower layer learns from the scheduling experience contained in the local historical data of each transformer network, as well as the global coordination variables passed from the upper layer scheduling in the previous time stage to the current time stage, to obtain the scheduling prior knowledge for the current time stage. The upper-level scheduler uses a time-varying directed communication graph to exchange information on the scheduling prior knowledge of each transformer network at the current time stage, obtains the global coordination variables at the current time stage, and passes the global coordination variables at the current time stage to the next time stage. The lower-level learning and upper-level scheduling optimize each other through alternating iterations until the preset convergence conditions are met.

[0063] Specific limitations regarding the stage-optimized scheduling device for transformer networks can be found in the above-mentioned limitations on stage-optimized scheduling of transformer networks, and will not be repeated here. Each module in the aforementioned stage-optimized scheduling device for transformer networks can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware or independently of the processor in a computer device, or stored in software in the memory of a computer device, so that the processor can call and execute the corresponding operations of each module.

[0064] This specification also provides a computer device, which, at the hardware level, includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for the business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then executes it to achieve the above. Figure 1 The provided method is a phase-optimized scheduling method for transformer networks.

[0065] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0066] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A stage-optimized scheduling method for a transformer network, characterized in that, include: An optimization scheduling model is constructed with the goal of minimizing the total operating cost of the target power network over all preset time periods; the target power network includes multiple transformer networks that exchange information through a time-varying directed communication graph. The optimal scheduling model is solved by a distributed two-layer scheduling algorithm to obtain the optimal scheduling scheme for the target power network; the distributed two-layer scheduling algorithm includes lower-layer learning and upper-layer scheduling. The lower layer learns from the scheduling experience contained in the local historical data of each transformer network, as well as the global coordination variables passed from the upper layer scheduling in the previous time stage to the current time stage, to obtain the scheduling prior knowledge for the current time stage. The upper-level scheduler uses a time-varying directed communication graph to exchange information on the scheduling prior knowledge of each transformer network at the current time stage, obtains the global coordination variables at the current time stage, and passes the global coordination variables at the current time stage to the next time stage. The lower-level learning and upper-level scheduling optimize each other through alternating iterations until the preset convergence conditions are met.

2. The stage-optimized scheduling method for transformer networks as described in claim 1, characterized in that, The optimization scheduling model, which aims to minimize the total operating cost of the target power network over all preset time periods, specifically includes: The optimization objective is to minimize the total operating cost of the target power network over all preset time periods. in, , Indicates the first The transformer network in the first Transformer power at each time period; Indicates the first The transformer network in the first Distribution network trading power at each time period; Indicates the first The transformer network in the first The operating cost function for each time period; Indicates the total number of transformer networks; , and These represent the weights for generation costs, power balance, and electricity trading costs, respectively. This indicates the cost of the transformer; This represents the cost of the imbalance between electricity production and demand; Indicates the transaction costs of the power distribution network; The constraints are transformer power constraints, energy storage system state constraints, ramp rate constraints, and distribution network transaction constraints. in, and They represent the first Minimum and maximum power of a transformer network; Indicates the first The transformer network in the first Battery state of charge at each time period; and They represent Minimum and maximum battery charge of a transformer network; Indicates the first The transformer network in the first Transformer power at each time period; Indicates the first The maximum rate of change of power of transformers in a transformer network.

3. The stage-optimized scheduling method for transformer networks as described in claim 2, characterized in that, Before solving the optimized scheduling model using the distributed two-level scheduling algorithm, the optimized scheduling model is transformed into a two-level optimization form, wherein: The lower-level optimization form is as follows: in, Indicates the first The transformer network in the first The time phase, targeting the first A number of similar tasks The optimized parameters obtained after the lower-level training are used as the scheduling experience contained in the local historical data of each transformer network for the lower-level learning. This represents the globally shared initial parameters passed from the upper-level scheduler in the previous time phase to the current time phase, serving as the starting point for training of similar tasks in the lower layer. This refers to historical scheduling tasks sampled from a preset data buffer set that have a similarity to the current scheduling task in terms of load characteristics, renewable energy output characteristics, or network topology characteristics greater than a preset threshold. Indicates similar tasks The training dataset obtained from sampling in the middle layer is used for the lower layer to learn the prior knowledge of this task; Indicates the first The transformer network in the first Nonlinear inequality constraints at each time stage; The upper-level optimization form is as follows: in, Indicates the first The cumulative losses of all transformer networks at each time period are used to evaluate parameters. The advantages and disadvantages; Indicates the first The transformer network in the first The set of tasks observed at each time period; Indicates the first The transformer network in the first Mean squared error loss function for each time period; Indicates similar tasks The test dataset is obtained by sampling from the middle layer and used to evaluate the parameters trained in the lower layer. Performance; Indicates the first Upper-level constraint functions of a transformer network; This represents the model parameters output by the lower-level optimization.

4. The stage-optimized scheduling method for transformer networks as described in claim 3, characterized in that, The solution to the optimized scheduling model using a distributed two-layer scheduling algorithm specifically includes: I. Initialization steps: Randomly initialize the scheduling parameters and dual parameters, where the scheduling parameters are the power scheduling decision variables of the transformer network, and the dual parameters are the Lagrange multipliers of the constraints in the optimization scheduling model; II. Stage Cycle Steps: According to the preset time stages The algorithm iterates through the time intervals 1, 2, ..., T. For each preset time period, the lower layer learns scheduling prior knowledge based on the historical data of each transformer network and the global coordination variables passed from the upper layer to the current time period. The upper layer then exchanges scheduling prior knowledge for the current time period through a time-varying directed communication graph to obtain the global coordination variables for the current time period. III. Iterative convergence step: Use the updated global coordination variable of each preset time stage as the initial value for scheduling in the next preset time stage, and repeat step II until all preset time stages are scheduled.

5. The stage-optimized scheduling method for transformer networks as described in claim 4, characterized in that, The specific steps of the phase cycle include: Real-time scheduling sub-step: For each transformer network, when new data arrives sequentially at the corresponding time stage, a neural network with deployed scheduling parameters is used to perform real-time scheduling based on the current dual parameters; Data update sub-step: After all data for the corresponding time period arrives, add all data to a preset data buffer set; the data buffer set is a queue used to store historical scheduling data; Sampling initialization sub-step: Obtain a training dataset of similar scheduling tasks from the data buffer set. and the test dataset of the current scheduled task And use the current scheduling parameters as the initial model parameters for the next layer of training; Lower-level training sub-step: Based on the training dataset The lower-level training iteratively updates the scheduling parameters and dual parameters, and outputs the scheduling prior knowledge for the current time stage; the scheduling prior knowledge is the optimized scheduling parameters obtained after the lower-level training converges. Upper-level training sub-step: Based on the test dataset The cumulative loss is obtained, and combined with prior scheduling knowledge and local information exchanged between each transformer network through time-varying directed communication graphs, upper-level training is performed to obtain the global coordination variables for the current time stage. The global coordination variables are the scheduling parameters optimized by upper-level training, and the local information is the interaction data of the prior scheduling knowledge of each transformer network between neighboring nodes.

6. The stage-optimized scheduling method for transformer networks as described in claim 5, characterized in that, The training dataset The scheduling parameters and dual parameters are updated iteratively during the next training layer using the following formula: in, and They represent the first The transformer network in the first Second and third Scheduling parameters during the next lower-level iteration; This indicates the fixed step size for training in the lower layer; Indicates about The gradient operator; and They represent the first The transformer network in the first Second and third Dual parameters in the next lower iteration; Indicates the first The transformer network in the first Each time phase is for the task The Lagrange function; Indicates about The gradient operator.

7. The stage-optimized scheduling method for transformer networks as described in claim 6, characterized in that, The global coordination variable for the current time period is obtained through iteration of the following formula: in, and They represent the first The transformer network in the first Time phase and the Scheduling parameters for each time phase; Represents a closed convex feasible region; Represents a symmetric positive definite matrix; superscript Representation device; Indicates the learning rate; Represents the loss function about The gradient; Represents constraint functions about The gradient; , and Both represent auxiliary variables of the Lagrange function; Indicates the first The transformer network in the first Auxiliary variables for estimating consensus state at different time stages. , In a time-varying directed communication graph, the first... The set of adjacent network nodes of a transformer network node; In a time-varying directed communication graph, the first... The transformer network node and the first The adjacency relationships of nodes in a transformer network.

8. The stage-optimized scheduling method for transformer networks as described in claim 1, characterized in that, The time-varying directed communication graph is a strongly connected equilibrium graph, which represents the topological structure of information interaction between transformer networks and satisfies local information characteristics, communication topology characteristics, equilibrium graph characteristics, and strong connectivity characteristics. The local information characteristics include: each transformer network node has information limitations and can only obtain and utilize its own cost information, operating constraints and local historical data; The communication topology characteristics include: defining the correspondence between the set of network nodes and the set of edges by giving a time-varying directed graph, and using the non-zero elements in the adjacency weighting matrix to characterize the existence of information transmission links and link weights between nodes, thereby establishing the neighbor relationship and communication direction between nodes; The characteristics of the balance graph include: for any node in the network, the total weight of information sent to neighboring nodes is equal to the total weight of information received from neighboring nodes, which represents that each transformer network maintains the conservation of input and output weights during information interaction. The strong connectivity characteristic includes: there is a unidirectional reachable directed path between any two nodes in the network, which means that any transformer network node can interact with other nodes through multi-hop relay communication, ensuring that the network has global connectivity.

9. A stage-optimized scheduling device for a transformer network, characterized in that, include: The scheduling model construction module is used to construct an optimized scheduling model with the goal of minimizing the total operating cost of the target power network over all preset time periods; wherein, the target power network includes multiple transformer networks that exchange information through a time-varying directed communication graph; The optimization scheduling module is used to solve the optimization scheduling model through a distributed two-level scheduling algorithm to obtain the optimal scheduling scheme for the target power network; the distributed two-level scheduling algorithm includes lower-level learning and upper-level scheduling. The lower layer learns from the scheduling experience contained in the local historical data of each transformer network, as well as the global coordination variables passed from the upper layer scheduling in the previous time stage to the current time stage, to obtain the scheduling prior knowledge for the current time stage. The upper-level scheduler uses a time-varying directed communication graph to exchange information on the scheduling prior knowledge of each transformer network at the current time stage, obtains the global coordination variables at the current time stage, and passes the global coordination variables at the current time stage to the next time stage. The lower-level learning and upper-level scheduling optimize each other through alternating iterations until the preset convergence conditions are met.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the stage-optimized scheduling method for the transformer network as described in any one of claims 1 to 8.

Images