Methods, systems, equipment, and storage media for coordinated dispatching and optimization of power main and distribution networks.

By calculating the power transmission sensitivity factor and dynamic weight of distribution network nodes and combining reinforcement learning algorithms to optimize the scheduling of the main distribution network, the problems of uneven weight distribution and security under extreme conditions in the existing technology are solved, and the accuracy and stability of scheduling resources are improved.

CN123121480APending Publication Date: 2026-11-17STATE GRID ZHEJIANG ELECTRIC POWER CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611146268.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-30
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing methods for coordinated scheduling of primary and secondary distribution networks lack a physical mechanism for weight allocation, making it impossible to accurately identify key nodes. Furthermore, reinforcement learning algorithms have insufficient security guarantees under extreme conditions, resulting in uneven allocation of scheduling resources and insufficient system stability.

Method used

By calculating the power transmission sensitivity factor of distribution network nodes, dynamically adjusting node weights, and combining a reinforcement learning algorithm with dual-delay deep deterministic policy gradient, a collaborative scheduling model is constructed, incorporating physical mechanisms and the average dynamic weight of the entire network to optimize scheduling decisions.

Benefits of technology

It enables precise allocation of scheduling resources, improves the security and accuracy of coordinated scheduling of main and distribution networks, and ensures system stability and reliability under extreme operating conditions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123121480A_ABST
    Figure CN123121480A_ABST
Patent Text Reader

Abstract

This invention relates to the field of power grid coordinated dispatching technology, and discloses a method, system, equipment, and storage medium for optimizing power grid coordinated dispatching. The method includes: calculating the physical contribution of power changes at distribution network nodes to the load pressure of key branches in the main grid based on the coordinated topology of the power grid, thus obtaining a power transmission sensitivity factor; calculating the node dynamic weights of distribution network nodes based on the power transmission sensitivity factor and the real-time load of the main grid, thus obtaining the average dynamic weight of the entire network; generating a state feature vector based on the basic operating data of the power grid, node dynamic weights, and the average dynamic weight of the entire network; inputting the vector into a coordinated dispatching model to obtain action vectors; mapping the action vectors to obtain coordinated dispatching instructions; and performing coordinated dispatching of the power grid based on these instructions. This invention introduces the physical laws of the power grid into the dispatching decision-making process, making the dispatching decisions physically interpretable, thereby improving the safety and accuracy of dispatching decisions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of power grid coordinated dispatch technology, and in particular to a power grid coordinated dispatch optimization method, system, equipment and storage medium. Background Technology

[0002] The coordinated operation of the main and distribution networks directly determines the reliability and stability of the power system. However, under new power system operating conditions, coordinated dispatch of the main and distribution networks still faces the following challenges: First, existing dispatch schemes lack a physical mechanism for weight allocation. In actual operation, nodes at different geographical locations and access points have drastically different impacts on power transmission in key sections and weak branches of the main network. However, traditional weight setting methods based on fixed priorities or simple load rate feedback ignore the heterogeneity of distribution network nodes in the physical topology and cannot accurately identify and differentiate the key nodes that contribute the most to system stability. Second, existing reinforcement learning-based dispatch algorithms lack sufficient physical security guarantees under extreme operating conditions. Although deep reinforcement learning has advantages in handling high-dimensional continuous action spaces, existing improvement schemes mostly focus on general hyperparameter optimization or simple L2 regularization constraints. This general regularization mechanism can only prevent overfitting of neural networks but cannot guarantee that dispatch instructions conform to the transient stability laws of the power system. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a method, system, device, and storage medium for optimizing the coordinated dispatch of power main and distribution networks. This addresses the shortcomings of existing coordinated dispatch methods in terms of the accuracy of resource allocation and the reliability under extreme operating conditions, thereby improving the security, accuracy, and rationality of resource allocation in coordinated dispatch of main and distribution networks.

[0004] In a first aspect, the present invention provides a method for optimizing coordinated dispatch of power main and distribution networks, the method comprising: Based on the coordinated topology of the power main and distribution network, the physical contribution of the power change of the distribution network node to the load pressure of the key branch of the main network is calculated, and the power transmission sensitivity factor of each distribution network node is obtained. Based on the power transmission sensitivity factor and the real-time load of the main network, calculate the node dynamic weight of each distribution network node, and obtain the average dynamic weight of the entire network based on the node dynamic weight. Based on the basic operation data of the main power distribution network, the dynamic weight of the nodes, and the average dynamic weight of the entire network, a state feature vector is generated. The state feature vector is input into a preset collaborative scheduling model to obtain an action vector. The action vector is then mapped to an instruction to obtain a collaborative scheduling instruction. The power main and distribution networks are then collaboratively scheduled according to the collaborative scheduling instruction.

[0005] Furthermore, the step of calculating the physical contribution of power changes at distribution network nodes to the load pressure of key branches of the main power grid based on the coordinated topology of the main power distribution network, and obtaining the power transmission sensitivity factor of each distribution network node, includes: Based on the cooperative topology, the power balance equations for the entire network are established, and a Taylor series expansion is performed to obtain the system Jacobian matrix. By performing an inverse mapping on the Jacobian matrix, the sensitivity matrix of node voltage to node injected power is obtained; Based on the sensitivity matrix, the power increment of critical branches in the main network when a distribution network node injects unit power is calculated, and the power increment is normalized to obtain the power transmission sensitivity factor of each distribution network node.

[0006] Further, the step of calculating the node dynamic weight of each distribution network node based on the power transmission sensitivity factor and the real-time load of the main network, and obtaining the average dynamic weight of the entire network based on the node dynamic weight, includes: Based on the power transmission sensitivity factor and the real-time load of the main network, calculate the load component of each distribution network node; Based on the difference between the load component and the preset main network load threshold, a correction coefficient is calculated using an exponential function. The preset base weights of each distribution network node are corrected according to the correction coefficient to obtain the dynamic weights of each distribution network node. The average dynamic weight of the entire network is obtained by averaging the dynamic weights of each distribution network node.

[0007] Furthermore, the step of generating a state feature vector based on the basic operational data of the main power distribution network, the dynamic weights of the nodes, and the average dynamic weight of the entire network includes: Acquire basic operating data of the main power distribution network, including node voltage amplitude, line load rate, real-time load rate of the main network, switching power of main network tie lines, and key branch ID; Based on the real-time load rate of the main network, the current time window identifier is determined. The time window identifier includes a basic time window and a peak time window, and the decision cycles corresponding to the basic time window and the peak time window are different. The basic operational data, the current time window identifier, the node dynamic weight, and the network-wide average dynamic weight are standardized to obtain a state feature vector.

[0008] Furthermore, the cooperative scheduling model is constructed using a reinforcement learning algorithm with dual-delay deep deterministic policy gradient, and the reward function of the cooperative scheduling model consists of a network loss component and a total voltage deviation component. The network loss component is calculated based on the ratio of the total network active power loss to the rated network loss benchmark value, and the total voltage deviation component is calculated by weighted summation of the node dynamic weights of each distribution network node and the voltage deviation term. The collaborative scheduling model uses a noise correction term guided by node dynamic weights to smooth the target action and the target strategy. The noise correction term is calculated by multiplying a weight matrix composed of node dynamic weights with a preset noise coefficient. The evaluation network of the cooperative scheduling model is trained using a loss function based on the average dynamic weight correction of the entire network, the policy network is trained using a delayed update strategy, and the policy gradient is calculated using a gradient calculation formula based on the average dynamic weight correction of the entire network and Lyapunov stability regularization. The policy network is updated according to the calculated policy gradient.

[0009] Furthermore, the steps for constructing the gradient calculation formula include: Based on the voltage potential energy of each distribution network node and the exchange power kinetic energy of the main network interconnection line, the energy of the entire network is calculated using the transient stability theory of the power system. According to Lyapunov stability theory, the gradient of the total network energy is used as a regularization term, and the regularization term is corrected by the average dynamic weight of the entire network to obtain the corrected regularization term. Subtracting the modified regularization term from the original policy gradient yields the gradient calculation formula.

[0010] Furthermore, the cooperative scheduling model employs a manifold mapping experience replay pool mechanism for model training, specifically including the following steps: An experience replay pool is constructed based on historical operational data. The experience replay pool includes a basic time window sample pool and a peak time window sample pool. A preset variational mapping model is used to reconstruct the state features in the sample pool of the basic time window using manifold experience, so as to obtain the mapping sample corresponding to the peak time window. The mapped samples are written into the peak time window sample pool to obtain an updated peak time window sample pool, and random sampling learning is performed based on the updated peak time window sample pool.

[0011] Secondly, the present invention provides a power grid main and distribution network coordinated dispatch optimization system, the system comprising: The weight calculation module is used to calculate the physical contribution of the power change of the distribution network node to the load pressure of the key branch of the main network based on the coordinated topology of the main power distribution network, and to obtain the power transmission sensitivity factor of each distribution network node. Based on the power transmission sensitivity factor and the real-time load of the main network, calculate the node dynamic weight of each distribution network node, and obtain the average dynamic weight of the entire network based on the node dynamic weight. The collaborative scheduling module is used to generate a state feature vector based on the basic operation data of the main power distribution network, the dynamic weight of the nodes, and the average dynamic weight of the entire network. The state feature vector is input into a preset collaborative scheduling model to obtain an action vector. The action vector is then mapped to an instruction to obtain a collaborative scheduling instruction. The power main and distribution networks are then collaboratively scheduled according to the collaborative scheduling instruction.

[0012] Thirdly, embodiments of the present invention also provide 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 computer program to implement the steps of the above-described method.

[0013] Fourthly, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0014] This invention provides a method, system, device, and storage medium for coordinated dispatch optimization of power main and distribution networks. By using power transmission sensitivity factors and node dynamic weights, it achieves deep coupling between distribution network node weights and the physical topology characteristics of the power grid, enabling the differentiated and precise allocation of dispatch resources to key nodes, thus improving the physical accuracy of dispatch resource allocation. By introducing physical mechanism-driven node dynamic weights and the network-wide average dynamic weights into the state feature vector, the dispatch decision-making basis becomes physically interpretable, improving the credibility of dispatch decisions. By co-encoding physical characteristics and basic operational data into a state feature vector and inputting it into the coordinated dispatch model, the decision-making process of the coordinated dispatch model is always guided by the physical operating laws of the power system, thus improving the security and accuracy of dispatch decisions. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the power main and distribution network coordinated scheduling optimization method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the power main and distribution network coordinated dispatch optimization system according to an embodiment of the present invention; Figure 3 This is an internal structural diagram of the computer device in an embodiment of the present invention.

[0016] Figure label: 10. Weight calculation module; 20. Cooperative scheduling module. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 The first embodiment of the present invention proposes a method for optimizing the coordinated dispatch of power main and distribution networks, including steps S10 to S40: Step S10: Based on the coordinated topology of the main power distribution network, calculate the physical contribution of the power change of the distribution network node to the load pressure of the key branch of the main network, and obtain the power transmission sensitivity factor of each distribution network node. Step S20: Calculate the node dynamic weight of each distribution network node based on the power transmission sensitivity factor and the real-time load of the main network, and obtain the average dynamic weight of the entire network based on the node dynamic weight. Step S30: Generate a state feature vector based on the basic operation data of the main power distribution network, the dynamic weight of the nodes, and the average dynamic weight of the entire network; Step S40: Input the state feature vector into a preset collaborative scheduling model to obtain an action vector, perform instruction mapping on the action vector to obtain a collaborative scheduling instruction, and perform collaborative scheduling of the main power grid according to the collaborative scheduling instruction.

[0019] In this embodiment, a power topology consisting of a main grid and multiple distributed distribution networks is first constructed. Then, based on this topology, the node weights of each distribution network node are calculated. Existing weight definition methods typically use fixed priorities or load rate feedback to set weights, but these methods ignore the heterogeneity of distribution network nodes in the physical topology. In actual operation, nodes in different geographical locations and with different access points have drastically different impacts on power transmission in critical and weak branches of the main grid. Existing weight setting methods cannot accurately identify and differentiate the critical nodes that contribute the most to system stability, leading to uneven allocation of scheduling resources under high load conditions and making it difficult to form an effective potential energy barrier to prevent cascading failures caused by local over-limits. To solve this problem, this embodiment provides a dynamic weight calculation method based on a power transmission sensitivity factor. The power transmission sensitivity factor is determined based on the physical contribution of distribution network node power changes to the load pressure of critical branches of the main grid. The specific calculation steps include: Based on the cooperative topology, the power balance equations for the entire network are established, and a Taylor series expansion is performed to obtain the system Jacobian matrix. By performing an inverse mapping on the Jacobian matrix, the sensitivity matrix of node voltage to node injected power is obtained; Based on the sensitivity matrix, the power increment of critical branches in the main network when a distribution network node injects unit power is calculated, and the power increment is normalized to obtain the power transmission sensitivity factor of each distribution network node.

[0020] In this embodiment, based on the coordinated topology of the main and distribution networks, the topology parameters of the entire network are obtained, including the connection relationships between nodes and branches, the impedance parameters of each branch, and the node type. The real-time operating status of the main and distribution networks is also obtained, including the voltage amplitude, phase angle, active / reactive power injection, and active power of each branch at the current moment. Based on the topology parameters and the real-time operating status, a set of power balance equations for all nodes in the network is established in polar coordinates, including active and reactive power balance. At the current operating point, a first-order Taylor expansion of this nonlinear equation set is performed, ignoring higher-order small quantities, and linearized. The Jacobian matrix is ​​obtained according to the linearization formula. Then, the Jacobian matrix is ​​decomposed to obtain the inverse matrix, thereby obtaining the sensitivity matrix of the node voltage vector V to the node active power P and reactive power Q. In the formula, Represents the change in the system's state variables, where, The phase angle change vector of the node voltage determines the direction of active power flow, while The voltage amplitude change vector directly reflects the reactive power balance and voltage stability of the system. (The intermediate term in the formula is...) The inverse of the Jacobian matrix obtained after first-order Taylor expansion and linearization of the nonlinear power balance equations at the current operating point is used as the core physical transformation operator, and its internal block submatrices are... , , and The physical coupling effect of minute disturbances in node-injected active power P and reactive power Q on the voltage phase angle and amplitude was quantified separately. This is the active phase angle sensitivity matrix. This is the active voltage sensitivity matrix. This is the reactive phase angle sensitivity matrix. The reactive voltage sensitivity matrix is ​​represented by the vector on the right side of the formula. This represents the change in node injected power. and These correspond to the active power injection increment and reactive power injection increment caused by load fluctuations or distributed power generation output adjustments, respectively.

[0021] The line with the highest load factor in the main network is defined as the critical branch. Then, using the sensitivity matrix mentioned above, the power increment caused by injecting unit power into distribution network node i on the critical branch is calculated. Specifically, based on power system flow analysis theory, the sensitivity matrix is ​​used to derive the effect of distribution network node i on the critical branch of the main network. Caused power increment The process is as follows: First, the active power flow of the critical branch connecting node m and node n is... Defined as a nonlinear function of node voltage magnitude and phase angle, and employing the existing method of linearizing the partial derivatives of the power flow equations, a linearized expression for the branch power flow variation is obtained by performing a first-order Taylor expansion on the state variables: in, and Let represent the voltage phase angles at nodes m and n, respectively. and Let m and n represent the voltage magnitudes at nodes m and n, respectively. Represents partial derivatives, and Let represent the voltage phase angle change vectors at nodes m and n, respectively. and These represent the voltage magnitude change vectors at nodes m and n, respectively.

[0022] Subsequently, using the aforementioned Jacobian inverse matrix The determined block sensitivity submatrix injects a disturbance of unit active power into node i. Substitute into the sensitivity equation and In this process, the node power disturbance is transformed into the voltage state change of all nodes in the network. Then, using the distribution factor calculation technique, the state change of each node is substituted into the above linearized expression to derive the power increment caused by node i. The expression: in, and These represent the elements in the m-th row and i-th column of the active phase angle sensitivity matrix and the active voltage sensitivity matrix, respectively, quantifying the impact of active power injection at node i on the state variables of node m. and These represent the elements in the nth row and i-th column of the active phase angle sensitivity matrix and the active voltage sensitivity matrix, respectively. This quantifies the impact of active power injection at node i on the state variables of node n. Finally, a per-unit mapping technique is used to normalize the power increment of distribution network node i, yielding the power transmission sensitivity factor of distribution network node i. : Where k is the total number of distribution network nodes, this factor Mapped to the [0,1] interval, it is used to characterize the relative strength of the influence of each node on the key branches of the main network, thereby providing a standardized physical basis for the subsequent construction of a heterogeneous risk potential energy field.

[0023] In this embodiment, the node weight of the distribution network node is calculated based on the power transmission sensitivity factor and the real-time load of the main network. The specific steps include: Based on the power transmission sensitivity factor and the real-time load of the main network, calculate the load component of each distribution network node; Based on the difference between the load component and the preset main network load threshold, a correction coefficient is calculated using an exponential function. The preset base weights of each distribution network node are corrected according to the correction coefficient to obtain the dynamic weights of each distribution network node. The average dynamic weight of the entire network is obtained by averaging the dynamic weights of each distribution network node.

[0024] In this embodiment, basic weights are first preset for distribution network nodes. For example, nodes are pre-classified into important nodes, general nodes, and ordinary nodes according to their importance, with basic weights set to 1.5, 1.2, and 1.0, respectively. However, in actual operation, even important nodes may have different support capabilities for the main network, indicating heterogeneity in the distribution network nodes' physical topology. Therefore, this embodiment combines the power transmission sensitivity factor of each distribution network node with the real-time load of the main network. The product of these two factors represents the load component of each distribution network node. Then, the difference between the load component and the preset main network load threshold is calculated, and a correction coefficient is obtained by mapping using an exponential function. Finally, the basic weights are corrected based on the correction coefficient to obtain the dynamic weight of each distribution network node, expressed as: In the formula, This represents the dynamic weight of distribution network node i at time t. This represents the basic weight of distribution network node i. Indicates the adjustment coefficient, preferred The value is 0.3. This represents the smoothing factor, with a preferred value of 1. This represents the main network load threshold, with a preferred value of 0.85. This represents the power transmission sensitivity factor of distribution network node i. This represents the real-time load on the main network at time t.

[0025] Then, the power transmission sensitivity factors of each distribution network node are added together and the average value is calculated to obtain the average dynamic weight of the entire network. This average value reflects the load status of the entire main and distribution network at the current moment.

[0026] According to the above formula for calculating the dynamic weight of nodes, for distribution network nodes with the same basic weight, the dynamic weight of nodes will differ when their power transmission sensitivity factors are different. For example, for important nodes, under high sensitivity (i.e., a larger power transmission sensitivity factor), their dynamic weight will be significantly increased to ensure that their priority is the highest under high load. Under low sensitivity (i.e., a smaller power transmission sensitivity factor), their dynamic weight will be slightly increased to ensure that the basic priority is not reduced and to avoid excessive resource allocation. For non-important nodes, under low sensitivity, their dynamic weight will hardly be adjusted. The node dynamic weight calculation method provided in this embodiment can differentiate and induce a non-linear surge in node weight under critical operating conditions based on the physical impact of distribution network nodes on the power transmission of key branches of the main network, thereby ensuring that scheduling resources can be accurately allocated to the key nodes that contribute the most to system stability. It should be noted that the dynamic weight of nodes in this embodiment is adapted to the dynamic changes of the critical branches of the main network. The critical branches of the main network will dynamically switch with the load distribution and the output of new energy sources. Some non-critical ordinary nodes may become critical nodes affecting the safety of branches in specific scenarios. Therefore, the critical branches used in this embodiment when calculating the power transmission sensitivity factor are not fixed, but are screened first in each calculation to ensure that the dynamic weight of nodes is adapted to the dynamic changes of the critical branches of the main network and to avoid missing safety risk points.

[0027] In this embodiment, a cooperative scheduling model based on reinforcement learning algorithm is used to optimize the cooperative scheduling of the main and distribution networks. Preferably, a reinforcement learning algorithm with Twin Delayed Deep Deterministic Policy Gradient (TD3) is used to construct the cooperative scheduling model. In this model, the defined state space includes basic operating characteristics and dynamic physical characteristics. The basic operating characteristics include the node voltage amplitude of the distribution network, the load rate of each line, and the switching power of the main network tie lines, etc. The dynamic physical characteristics include the node dynamic weight, the average dynamic weight of the entire network, and the current load rate of the main network, etc. The defined action space is the adjustable resource adjustment amount of each distribution network node, which is mapped to the distributed power output adjustment, controllable load adjustment, and reactive power compensation device switching amount.

[0028] The collaborative scheduling model is built using the backbone network of the basic TD3. The collaborative scheduling model built according to the native TD3 framework includes an Actor network (policy network), dual Critic networks (Q1, Q2, evaluation networks), a target Actor network, and dual target Critic networks (Target Actor, Target Q1, Target Q2). Preferably, the Actor network is a 3-layer fully connected network with the input layer dimension being the state space dimension, including 2 hidden layers, each with 256 neurons, and the activation function being ReLU. The output layer dimension is the action space dimension, with the activation function being Tanh, and the output value range is [-1, 1]. The dual Critic networks are both 3-layer fully connected networks with the input layer dimension equal to the sum of the state dimension and the action dimension, including 2 hidden layers, each with 256 neurons, and the activation function being ReLU. The output layer dimension is 1 (output Q value). The structure of the target Actor network and the dual target Critic network is completely consistent with the corresponding networks mentioned above, and the initial parameters are completely synchronized with the backbone network. This embodiment uses the native TD3 dual evaluation network and delayed update core architecture to ensure the stability of training.

[0029] In the actual operation of a power system, based on the different load levels of the main grid, it can be divided into peak periods and normal periods. To ensure the timeliness of dispatching during peak periods, this embodiment presets time window identifiers and selects different time windows according to different time window identifiers, including basic time windows and peak time windows. Under different time windows, the decision cycle is also different. For example, the decision cycle is 15 minutes under the basic time window and 1 minute under the peak time window. In this embodiment, the steps for constructing the input features of the collaborative dispatching model include: Acquire basic operating data of the main power distribution network, including node voltage amplitude, line load rate, current load rate of the main network, switching power of main network tie lines, and key branch ID; Based on the real-time load rate of the main network, the current time window identifier is determined. The time window identifier includes a basic time window and a peak time window, and the decision cycles corresponding to the basic time window and the peak time window are different. The basic operational data, the current time window identifier, the node dynamic weight, and the network-wide average dynamic weight are standardized to obtain a state feature vector.

[0030] The difference between the state feature vector in this embodiment and the state feature vector in the above embodiment is the addition of a current time window identifier. The current time window identifier is determined based on the degree of change in the real-time load of the main network. When the real-time load of the main network is greater than the main network load threshold, such as 85%, and the load change rate is greater than zero, it indicates that the main network load is continuously rising and the peak period has begun. At this time, the current time window identifier is set as the peak time window, and the decision cycle will switch to the decision cycle of the peak period. The backtracking condition adopts hysteresis logic, that is, the real-time load of the main network is less than the second main network load threshold, such as 75%, and continues for at least N time blocks, such as N=3, before switching to the basic time window. The hysteresis logic can effectively suppress the control oscillation caused by frequent switching.

[0031] In a preferred embodiment, the collaborative scheduling model adopts the model framework of the native TD3 algorithm. Its reward function consists of a network loss component and a total voltage deviation component. Specifically, the network loss component is calculated by the ratio of the total network active power loss to the rated network loss benchmark value. The total network active power loss can be calculated using conventional calculation methods such as power flow calculation or equivalent resistance method. The total voltage deviation component is calculated based on the voltage deviation term of each distribution network node and the node dynamic weight. The voltage deviation term is equal to the square of the difference between the actual voltage value and the rated voltage value of the distribution network node. The voltage deviation term of the distribution network node is multiplied by the node dynamic weight, and the products of each distribution network node are summed to obtain the total voltage deviation component. By directly multiplying the node dynamic weight by the voltage deviation term, the voltage stability of key areas can be forcibly constrained.

[0032] Finally, the network loss component and the total voltage deviation component are standardized and weighted summed, and the negative value is taken to obtain the reward function. In this embodiment, by deeply coupling the node dynamic weights with the voltage deviation term, the reward function can perceive changes in topology risk in real time. In addition, based on the above reward function, an over-limit penalty term can be added. If there is a voltage over-limit or line overload, the over-limit penalty is triggered. The over-limit penalty term is a preset penalty coefficient. When there is a voltage over-limit, the over-limit penalty term is added to the reward function.

[0033] During the training of the cooperative scheduling model, this embodiment employs a noise correction term guided by dynamic node weights to smooth the target action policy. Specifically, weighted noise is introduced into the action output layer of the Actor network: In the formula, Let be the node weight matrix at time t. This matrix is ​​a diagonal matrix composed of the dynamic weights of the distribution network nodes. Nodes with higher weights have larger diagonal elements. The noise figure is a preset value that conforms to a Gaussian distribution. Represents the action space at time t. Represents the policy function. Represents the state space at time t.

[0034] This embodiment introduces weighted noise, which not only smooths the target policy but also guides the model to prioritize exploring the adjustment space of high-risk nodes, avoiding meaningless exploration of low-value regions.

[0035] Furthermore, in this embodiment, a sample set constructed based on historical operational data was used for training the collaborative scheduling model, and the loss function of the Critic network was modified. Specifically, when calculating the target value, the average dynamic weight of the entire network was introduced, and the expression of the loss function at this time is: In the formula, This represents the target value at time t. The Q-value represents the action value estimated by the Critic network. Represents the action space at time t. Represents the state space at time t. Represents the expectation of the state distribution. This represents the loss function of the Q-network (i.e., the Critic network) under parameter θ. This represents the average dynamic weight of the entire network at time t.

[0036] This embodiment introduces a correction coefficient based on the average dynamic weight of the entire network into the loss function of the native TD3 value network. This increases the gradient update magnitude of the neural network during peak load periods (when weights are generally large), thereby accelerating the model's convergence speed under extreme conditions and enabling the model to effectively adapt to the dynamic changes in power operation.

[0037] Furthermore, when updating the Actor network, a conventional delayed update strategy was adopted, such as updating the Actor network once every two updates to the Critic network. A Lyapunov stability regularization term was added to the gradient calculation formula to constrain the action amplitude during high-risk periods and prevent overfitting. Specific steps included: Based on the voltage potential energy of each distribution network node and the exchange power kinetic energy of the main network interconnection line, the energy of the entire network is calculated using the transient stability theory of the power system. According to Lyapunov stability theory, the gradient of the total network energy is used as a regularization term, and the regularization term is corrected by the average dynamic weight of the entire network to obtain the corrected regularization term. Subtracting the modified regularization term from the original policy gradient yields the gradient calculation formula.

[0038] In this embodiment, based on the transient stability theory of power systems, the total network energy is first defined. The total network energy consists of the voltage potential energy of each distribution network node and the exchange power kinetic energy of the main network tie lines, and its expression is: In the formula, The total energy of the network under system state s is represented by k, and the total number of distribution network nodes is represented by C. i V is the equivalent capacitance coefficient of distribution network node i. i V is the real-time voltage of distribution network node i. ref This is the rated reference value for the node voltage. branch road l The transmission power, where L represents the set of main network interconnects. branch road l The phase angle difference between the two endpoints.

[0039] In the above-mentioned total network energy function, the first part is the voltage potential energy term of the distribution network node, and the second part is the transmission power kinetic energy term of the tie line branch. This formula quantifies the degree to which the current grid state deviates from the stable equilibrium point. The higher the energy, the greater the risk of system instability.

[0040] According to Lyapunov stability theory, as long as the system energy decreases monotonically with time, the system will inevitably be asymptotically stable and will not experience voltage instability or oscillations. In reinforcement learning's policy update logic, the scheduling action output by the Actor network must lead to a decrease in system energy at the next time step; that is, the gradient of the energy function with respect to the action must be less than or equal to zero. If the action causes an increase in energy at the next time step, i.e., the gradient of the energy function with respect to the action is greater than zero, then the regularization term will generate a reverse correction gradient, suppressing parameter updates in that direction. After converting this constraint into a gradient update term, we obtain the core gradient of the regularization term. This represents the direction of the influence of Actor parameter adjustment on system energy, where... This represents the trainable parameters of the Actor network. This represents the gradient of the total network energy with respect to the Actor parameters.

[0041] To achieve an adaptive constraint that balances economic efficiency during low-risk periods and prioritizes safety during high-risk periods, the regularization term is coupled with the previously defined dynamic node weights, resulting in a final regularization term embedded in the Actor's gradient. ,in, This represents the weighting coefficient, with a preferred value of 0.5.

[0042] Finally, the aforementioned final regularization term is introduced into the original TD3 Actor basic gradient calculation formula, thus obtaining the improved gradient calculation formula: In the formula, J represents the expected cumulative reward objective function of the strategy. This represents the trainable parameters of the Actor network. This represents the gradient of the objective function with respect to the Actor parameters, where N represents the number of samples in the mini-batch. This represents the output value of the first Q-network in the TD3 dual-Critic network under state s and action a. This represents the output policy of the Actor network in state s. This represents the gradient of the Actor's output action with respect to its own parameters. This represents the gradient of the output of the first Q-network in the Critic network with respect to action a. The ∑ in the formula represents the summation over the sampled data, used to approximate the expectation.

[0043] In this embodiment, the gradient update direction of the model during periods of high grid load is constrained by a regularization term, which forces the scheduling instructions to meet the requirements of system transient stability, prevents overstepping of limits due to the pursuit of economic benefits, and ensures that the scheduling actions generated by reinforcement learning can guide the grid state to evolve towards a stable equilibrium point. Furthermore, by using the average dynamic weight of the entire network, the strength of this physical constraint is linearly increased during periods of high load risk, which forces the Actor network to prioritize system safety rather than simply pursuing economic benefits.

[0044] This embodiment achieves power physics consistency constraints by mathematically reconstructing the loss function of the TD3 framework. By introducing a system energy function regularization term based on Lyapunov stability theory, the gradient update direction of the neural network is forced to conform to the power system transient stability criterion. Combined with a directional action space exploration mechanism guided by dynamic weights, the originally black-box reinforcement learning training process is transformed into an optimization process guided by physical mechanisms. This not only significantly enhances the safety and interpretability of dispatch commands under extreme conditions but also improves the convergence speed and robustness of the model in complex dynamic environments, truly realizing significant progress in response speed and power supply reliability in coordinated dispatching of the main and distribution networks.

[0045] Based on the above embodiments, the cooperative scheduling model first initializes the Actor and Critic networks and their corresponding target networks during training, constructs an experience replay pool for storing interaction data, and then executes a training loop. Specifically, it selects the action space according to the current policy, executes actions to obtain rewards and the next state, and stores the experience in the experience replay pool. It samples small batches of data from the experience pool to calculate the target value, and uses target policy smoothing during target value calculation, adding noise correction terms to the actions of the next state. Then, it updates the Critic network by minimizing the loss function based on the average dynamic weight correction of the entire network, delays the update of the Actor network, and calculates the policy gradient using the gradient calculation formula based on the average dynamic weight correction of the entire network with Lyapunov stability regularization. Finally, it performs soft updates on all target networks, repeating the above steps until convergence, resulting in a trained model.

[0046] In a preferred embodiment, since the number of samples varies across different time windows, this embodiment provides an experience replay pool mechanism for manifold mappings to address the learning efficiency problem of reinforcement learning under sparse, high-risk samples. The specific steps include: An experience replay pool is constructed based on historical operational data. The experience replay pool includes a basic time window sample pool and a peak time window sample pool. A preset variational mapping model is used to reconstruct the state features in the sample pool of the basic time window using manifold experience, so as to obtain the mapping sample corresponding to the peak time window. The mapped samples are written into the peak time window sample pool to obtain an updated peak time window sample pool, and random sampling learning is performed based on the updated peak time window sample pool.

[0047] This embodiment employs a manifold empirical reconstruction mechanism for feature mapping. This mechanism constructs a cross-dimensional feature mapping operator to migrate historical samples within the base time window to the feature manifold space of the peak time window. The detailed reconstruction process is as follows: First, define the variational mapping function. ,in Represents the basic time window The low-frequency state space below, Represents peak time window The high-frequency state space is described. This function employs an encoder-decoder structure, preferably a variational autoencoder model. By minimizing the reconstruction loss and distribution divergence, it extracts the common manifold features of power grid state transitions at different sampling scales. It should be noted that the variational autoencoder uses its conventional structure and a feature training set constructed from historical power grid operating data, and is trained using existing model training methods. Therefore, its model structure and training methods are not limited here.

[0048] Then, meta-samples are extracted from the basic experience pool. ,in, Represents the state space at time t. Represents the action space at time t. This represents the reward at time t. Representing the state space at time t+1, the trained variational mapping function is used to perform feature alignment on the state variables to generate mapped samples. This ensures that its statistical characteristics are consistent with real-time observation data during peak periods, among which, and Let represent the state space at time t and the state space at time t+1 after mapping, respectively. Finally, the mapped samples are written into the peak time window sample pool to obtain the updated peak time window sample pool, and the latest peak time window sample pool is used for learning.

[0049] This embodiment achieves the alignment of empirical data in heterogeneous feature spaces through a manifold experience reconstruction mechanism, effectively eliminating the state transition manifold shift caused by abrupt changes in the decision cycle, solving the learning efficiency problem of reinforcement learning under sparse high-risk samples, and ensuring the convergence stability and policy consistency of the model in the rapid evolution process under extreme conditions.

[0050] By using a trained collaborative scheduling model, action vectors corresponding to the input state vectors can be obtained. Then, the action vectors are mapped to instructions, which are then mapped to the actual adjustment amounts of the adjustable resources of each distribution network node. These are then converted into standardized scheduling instructions and sent to each execution mechanism, thereby realizing the collaborative scheduling of the main power distribution network.

[0051] This embodiment provides a power grid main and distribution network coordinated dispatch optimization method. By using power transmission sensitivity factors and node dynamic weights, it can achieve deep coupling between distribution network node weights and power grid physical topology characteristics, enabling the differentiated and precise allocation of dispatch resources to key nodes, thus improving the physical accuracy of dispatch resource allocation. By introducing physical mechanism-driven node dynamic weights and the network-wide average dynamic weights into the state feature vector, the dispatch decision basis has physical interpretability, thus improving the credibility of dispatch decisions. By co-encoding physical characteristics and basic operating data into a state feature vector and inputting it into the coordinated dispatch model, the decision-making process of the coordinated dispatch model is always guided by the physical operating laws of the power system, thus improving the security and accuracy of dispatch decisions.

[0052] Furthermore, this embodiment achieves empirical data alignment in heterogeneous feature spaces through hysteresis backtracking and variational mapping, effectively eliminating state transition manifold shifts caused by abrupt changes in the decision cycle, and ensuring the convergence stability and policy consistency of the model during rapid evolution under extreme conditions. By reconstructing the underlying mathematical structure of the TD3 framework loss function, the gradient update direction of the neural network is forced to conform to the transient stability criterion of the power system, improving the convergence speed and robustness of the model in complex dynamic environments, thereby effectively improving the response speed and power supply reliability of the main and distribution network coordinated dispatch.

[0053] Please see Figure 2 Based on the same inventive concept, the second embodiment of this invention proposes a power main and distribution network coordinated dispatch optimization system, comprising: The weight calculation module 10 is used to calculate the physical contribution of the power change of the distribution network node to the load pressure of the key branch of the main network according to the cooperative topology of the main power distribution network, and to obtain the power transmission sensitivity factor of each distribution network node. Based on the power transmission sensitivity factor and the real-time load of the main network, calculate the node dynamic weight of each distribution network node, and obtain the average dynamic weight of the entire network based on the node dynamic weight. The collaborative scheduling module 20 is used to generate a state feature vector based on the basic operation data of the main power distribution network, the dynamic weight of the nodes, and the average dynamic weight of the entire network. The state feature vector is input into a preset collaborative scheduling model to obtain an action vector. The action vector is then mapped to an instruction to obtain a collaborative scheduling instruction. The power main and distribution networks are then collaboratively scheduled according to the collaborative scheduling instruction.

[0054] The technical features and effects of the power grid coordinated dispatch optimization system proposed in this embodiment are the same as those of the method proposed in this embodiment, and will not be repeated here. Each module in the above-mentioned power grid coordinated dispatch optimization system can be implemented entirely or partially through software, hardware, or a combination thereof. Each module can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0055] Furthermore, embodiments of the present invention also propose 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 computer program to implement the steps of the above-described method.

[0056] Please see Figure 3The diagram illustrates the internal structure of a computer device in one embodiment. This computer device can specifically be a terminal or a server. The computer device includes a processor, memory, network interface, display, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface of the computer device is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for optimizing the coordinated dispatch of the main power grid and distribution network. The display screen of the computer device can be a liquid crystal display (LCD) or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0057] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computing devices may include more or fewer components than those shown in the figure, or combine certain components, or have the same component arrangement.

[0058] Furthermore, embodiments of the present invention also propose a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described method.

[0059] In summary, the present invention proposes a method, system, device, and storage medium for coordinated scheduling optimization of power distribution networks. The method calculates the physical contribution of power changes at distribution network nodes to the load pressure of critical branches in the main grid based on the coordinated topology of the power distribution network, obtaining the power transmission sensitivity factor of each distribution network node. Based on the power transmission sensitivity factor and the real-time load of the main grid, the method calculates the node dynamic weight of each distribution network node and obtains the average dynamic weight of the entire network based on the node dynamic weight. A state feature vector is generated based on the basic operating data of the power distribution network, the node dynamic weight, and the average dynamic weight of the entire network. The state feature vector is input into a preset coordinated scheduling model to obtain an action vector. The action vector is then mapped to an instruction to obtain a coordinated scheduling instruction, and the power distribution network is coordinated and scheduled according to the coordinated scheduling instruction. This invention achieves deep coupling between distribution network node weights and power grid physical topology characteristics through power transmission sensitivity factors and node dynamic weights. This enables the differentiated and precise allocation of scheduling resources to key nodes, improving the physical accuracy of scheduling resource allocation. By introducing physical mechanism-driven node dynamic weights and the network-wide average dynamic weights into the state feature vector, the scheduling decision-making basis becomes physically interpretable, enhancing the credibility of scheduling decisions. Furthermore, by co-encoding physical characteristics and basic operational data into a state feature vector and inputting it into the collaborative scheduling model, the decision-making process of the collaborative scheduling model is always guided by the physical operating laws of the power system, improving the security and accuracy of scheduling decisions.

[0060] The various embodiments in this specification are described in a progressive manner. For directly identical or similar parts of the embodiments, refer to each other. Each embodiment focuses on its differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0061] The embodiments described above are merely preferred embodiments of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the technical principles of this invention, and these improvements and substitutions should also be considered within the scope of protection of this application. Therefore, the scope of protection of this patent application should be determined by the scope of the claims.

Claims

1. A method for optimizing coordinated dispatch of power main and distribution networks, characterized in that, include: Based on the coordinated topology of the power main and distribution network, the physical contribution of the power change of the distribution network node to the load pressure of the key branch of the main network is calculated, and the power transmission sensitivity factor of each distribution network node is obtained. Based on the power transmission sensitivity factor and the real-time load of the main network, calculate the node dynamic weight of each distribution network node, and obtain the average dynamic weight of the entire network based on the node dynamic weight. Based on the basic operation data of the main power distribution network, the dynamic weight of the nodes, and the average dynamic weight of the entire network, a state feature vector is generated. The state feature vector is input into a preset collaborative scheduling model to obtain an action vector. The action vector is then mapped to an instruction to obtain a collaborative scheduling instruction. The power main and distribution networks are then collaboratively scheduled according to the collaborative scheduling instruction.

2. The power grid main and distribution network coordinated dispatch optimization method according to claim 1, characterized in that, The steps of calculating the physical contribution of power changes at distribution network nodes to the load pressure of key branches in the main power grid based on the coordinated topology of the main power distribution network, and obtaining the power transmission sensitivity factor of each distribution network node, include: Based on the cooperative topology, the power balance equations for the entire network are established, and a Taylor series expansion is performed to obtain the system Jacobian matrix. By performing an inverse mapping on the Jacobian matrix, the sensitivity matrix of node voltage to node injected power is obtained; Based on the sensitivity matrix, the power increment of critical branches in the main network when a distribution network node injects unit power is calculated, and the power increment is normalized to obtain the power transmission sensitivity factor of each distribution network node.

3. The power grid main and distribution network coordinated dispatch optimization method according to claim 2, characterized in that, The steps of calculating the dynamic weight of each distribution network node based on the power transmission sensitivity factor and the real-time load of the main network, and obtaining the average dynamic weight of the entire network based on the dynamic weight of the nodes, include: Based on the power transmission sensitivity factor and the real-time load of the main network, calculate the load component of each distribution network node; Based on the difference between the load component and the preset main network load threshold, a correction coefficient is calculated using an exponential function. The preset base weights of each distribution network node are corrected according to the correction coefficient to obtain the dynamic weights of each distribution network node. The average dynamic weight of the entire network is obtained by averaging the dynamic weights of each distribution network node.

4. The power grid main and distribution network coordinated dispatch optimization method according to claim 1, characterized in that, The step of generating a state feature vector based on the basic operation data of the main power distribution network, the node dynamic weights, and the average dynamic weight of the entire network includes: Acquire basic operating data of the main power distribution network, including node voltage amplitude, line load rate, real-time load rate of the main network, switching power of main network tie lines, and key branch ID; Based on the real-time load rate of the main network, the current time window identifier is determined. The time window identifier includes a basic time window and a peak time window, and the decision cycles corresponding to the basic time window and the peak time window are different. The basic operational data, the current time window identifier, the node dynamic weight, and the network-wide average dynamic weight are standardized to obtain a state feature vector.

5. The power grid main and distribution network coordinated dispatch optimization method according to claim 1, characterized in that, The cooperative scheduling model is constructed using a reinforcement learning algorithm with dual-delay deep deterministic policy gradient. The reward function of the cooperative scheduling model consists of a network loss component and a total voltage deviation component. The network loss component is calculated based on the ratio of the total network active power loss to the rated network loss benchmark. The total voltage deviation component is calculated by weighted summation of the node dynamic weights of each distribution network node and the voltage deviation term. The collaborative scheduling model uses a noise correction term guided by node dynamic weights to smooth the target action and the target strategy. The noise correction term is calculated by multiplying a weight matrix composed of node dynamic weights with a preset noise coefficient. The evaluation network of the cooperative scheduling model is trained using a loss function based on the average dynamic weight correction of the entire network, the policy network is trained using a delayed update strategy, and the policy gradient is calculated using a gradient calculation formula based on the average dynamic weight correction of the entire network and Lyapunov stability regularization. The policy network is updated according to the calculated policy gradient.

6. The power grid main and distribution network coordinated dispatch optimization method according to claim 5, characterized in that, The steps for constructing the gradient calculation formula include: Based on the voltage potential energy of each distribution network node and the exchange power kinetic energy of the main network interconnection line, the energy of the entire network is calculated using the transient stability theory of the power system. According to Lyapunov stability theory, the gradient of the total network energy is used as a regularization term, and the regularization term is corrected by the average dynamic weight of the entire network to obtain the corrected regularization term. Subtracting the modified regularization term from the original policy gradient yields the gradient calculation formula.

7. The power grid main and distribution network coordinated dispatch optimization method according to claim 4, characterized in that, The cooperative scheduling model uses a manifold mapping experience replay pool mechanism for model training, and the specific steps include: An experience replay pool is constructed based on historical operational data. The experience replay pool includes a basic time window sample pool and a peak time window sample pool. A preset variational mapping model is used to reconstruct the state features in the sample pool of the basic time window using manifold experience, so as to obtain the mapping sample corresponding to the peak time window. The mapped samples are written into the peak time window sample pool to obtain an updated peak time window sample pool, and random sampling learning is performed based on the updated peak time window sample pool.

8. A power grid main and distribution network coordinated dispatch optimization system, characterized in that, include: The weight calculation module is used to calculate the physical contribution of the power change of the distribution network node to the load pressure of the key branch of the main network based on the coordinated topology of the main power distribution network, and to obtain the power transmission sensitivity factor of each distribution network node. Based on the power transmission sensitivity factor and the real-time load of the main network, calculate the node dynamic weight of each distribution network node, and obtain the average dynamic weight of the entire network based on the node dynamic weight. The collaborative scheduling module is used to generate a state feature vector based on the basic operation data of the main power distribution network, the dynamic weight of the nodes, and the average dynamic weight of the entire network. The state feature vector is input into a preset collaborative scheduling model to obtain an action vector. The action vector is then mapped to an instruction to obtain a collaborative scheduling instruction. The power main and distribution networks are then collaboratively scheduled according to the collaborative scheduling instruction.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.