A flexible power distribution network distribution robust rolling control method and system based on optimization guided learning

Abstract

The application discloses a flexible power distribution network distribution robust rolling control method and system based on optimized guided learning. The method acquires the topology of the flexible power distribution network, line parameters, node operation quantity and source load prediction deviation, and constructs a multi-attribute state matrix under a rolling time window; a probability distribution fuzzy set is constructed according to historical samples of the source load prediction deviation, and a distribution robust optimization model subject to physical constraints of the power distribution network is established; the model is solved to generate training samples composed of a current state, an optimized control quantity and a subsequent state, and a learning auxiliary model is trained; at a current scheduling moment, the learning auxiliary model outputs a control variable initial value according to the current state, the distribution robust optimization model solves a control quantity sequence based on the initial value, executes the current control quantity, and updates the multi-attribute state matrix according to the node operation quantity after the execution. The application can improve the solving efficiency of the rolling optimization control on the basis of reserving the physical constraints of the power distribution network and checking the source load uncertainty.

Description

Technical Field

[0001] This invention belongs to the field of distribution network optimization control, specifically relating to a flexible distribution network sub-brush rolling control method and system based on optimization-guided learning. Background Technology

[0002] With the integration of distributed photovoltaic (PV) power, energy storage devices, and flexible interconnection devices into the distribution network, the distribution network is gradually transforming from a unidirectional power supply network into a flexible distribution network that coordinates the operation of power sources, grids, and loads. PV power output and load demand are volatile, and deviations exist between predicted results and actual operating conditions, which can easily lead to node voltage exceeding limits, changes in line power flow, and increased network losses. Therefore, it is necessary to dynamically control flexible regulation equipment based on the real-time operating status of the distribution network.

[0003] Existing distribution network optimization control methods typically establish deterministic optimization models based on predicted photovoltaic output and load values, or pre-assuming that prediction deviations follow a specific probability distribution. When the actual prediction deviations differ from the preset distribution, the resulting control scheme may fail to meet the operational constraints of the distribution network. The bibliometric optimization method can construct a set of probability distributions using historical prediction deviations and solve for control schemes under conditions where the probability distributions are not entirely certain. However, in the rolling control process, it requires repeatedly solving the optimization model, which includes uncertainties and distribution network physical constraints, as the operating state changes. This results in a large computational load, and the solution efficiency is easily affected by the initial values ​​of the optimization variables and the problem size.

[0004] Furthermore, while methods that directly generate control quantities based on historical operating data can reduce online computation, the generated control quantities may not meet the physical constraints and uncertainty constraints of the current distribution network. Therefore, existing technologies still struggle to simultaneously address source-load uncertainty, improve the efficiency of rolling optimization, and ensure the physical feasibility of the final control quantities. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide a rolling control method for the distributed bar in flexible distribution networks based on optimization-guided learning, so as to solve the problems of repeated solving of the distributed bar optimization model, low efficiency of online solving, and difficulty in guaranteeing the physical feasibility of data-driven control results in the rolling control process of flexible distribution networks.

[0006] Technical solution: A flexible distribution network sub-bar rolling control method based on optimization-guided learning, the method comprising:

[0007] The topology, line parameters, node voltage, node power, and source-load prediction deviation of the flexible distribution network are obtained. The electrical distance relationship between nodes is determined based on the topology and line parameters. The electrical coupling relationship between nodes is determined based on the line parameters and node power transmission relationship. Based on the electrical distance relationship and electrical coupling relationship between nodes, the node voltage, node active power, and node reactive power are organized into a multi-attribute state matrix according to the rolling time window.

[0008] An empirical distribution is established based on historical samples of source-load prediction deviations. A probability distribution fuzzy set is constructed with the empirical distribution as the center. The probability distribution fuzzy set is used to characterize the probability distribution of source-load prediction deviations. The feasible domain of control quantities and distribution network operation status is limited by the physical constraints of the distribution network. A sub-Bruker optimization model in the prediction time domain is established.

[0009] The sub-Bruker optimization model is solved by using different multi-attribute state matrices as the current state to obtain the corresponding optimized control quantity and the subsequent multi-attribute state matrix after the optimized control quantity is applied. The current multi-attribute state matrix, the optimized control quantity and the subsequent multi-attribute state matrix form a training sample set, and the auxiliary model is trained using the training sample set.

[0010] At the current scheduling moment, the current multi-attribute state matrix is ​​input into the trained learning auxiliary model to obtain the initial values ​​of the control variables. The control variables of the split-bar optimization model are initialized with the initial values ​​of the control variables, and the control quantity sequence in the prediction time domain is obtained by solving under the probability distribution fuzzy set and the physical constraints of the distribution network.

[0011] The system issues control quantities corresponding to the current scheduling time in the control quantity sequence to the flexible regulation equipment, collects the node voltage, node active power and node reactive power after the control quantity is executed, updates the multi-attribute state matrix and enters the next scheduling time.

[0012] A flexible distribution network distributed bar rolling control system based on optimization-guided learning includes:

[0013] The state characterization module is used to obtain the topology, line parameters, node voltage, node power and source-load prediction deviation of the flexible distribution network. Based on the topology and line parameters, it determines the electrical distance relationship between nodes. Based on the line parameters and node power transmission relationship, it determines the electrical coupling relationship between nodes. Based on the electrical distance relationship and electrical coupling relationship between nodes, it organizes the node voltage, node active power and node reactive power into a multi-attribute state matrix according to the rolling time window.

[0014] The robust optimization module is used to establish an empirical distribution based on historical samples of source load prediction deviations, construct a probability distribution fuzzy set centered on the empirical distribution, characterize the probability distribution of source load prediction deviations with the probability distribution fuzzy set, limit the feasible domain of control quantities and distribution network operating status with distribution network physical constraints, and establish a partial robust optimization model in the prediction time domain.

[0015] The model training module is used to solve the split-bar optimization model with different multi-attribute state matrices as the current state, obtain the corresponding optimized control quantity, and obtain the subsequent multi-attribute state matrix after the optimized control quantity is applied. The current multi-attribute state matrix, the optimized control quantity, and the subsequent multi-attribute state matrix form a training sample set, which is used to train and learn the auxiliary model.

[0016] The online control module is used to input the current multi-attribute state matrix into the trained learning auxiliary model at the current scheduling time to obtain the initial values ​​of the control variables. The control variables of the split-bar optimization model are initialized with the initial values ​​of the control variables, and the control quantity sequence in the prediction time domain is obtained by solving under the probability distribution fuzzy set and the physical constraints of the distribution network.

[0017] The feedback update module is used to send the control quantity corresponding to the current scheduling moment in the control quantity sequence to the flexible regulation equipment, collect the node voltage, node active power and node reactive power after the execution of the control quantity, update the multi-attribute state matrix and enter the next scheduling moment.

[0018] The present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and executable by the processor, wherein the computer program, when executed by the processor, implements the above-described flexible distribution network split-bar rolling control method based on optimized guided learning.

[0019] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the above-described flexible distribution network spool rolling control method based on optimization-guided learning.

[0020] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described flexible distribution network split-bar rolling control method based on optimization-guided learning.

[0021] Beneficial effects:

[0022] (1) This invention improves the ability to characterize complex operating states by constructing a continuous state space that integrates the distribution network topology, line parameters and node voltage information; by introducing a sub-Blu-ray bar optimization model based on Wasserstein distance, it effectively characterizes the uncertainty caused by distributed power output and load fluctuations, and improves the robustness of optimization decisions; by using a rolling time-domain optimization framework, it realizes dynamic prediction and real-time correction of the distribution network operating state, which improves the online response capability of the control strategy; and by introducing a learning-assisted model, it provides initial values ​​of control variables based on the current multi-attribute state matrix at each scheduling time, so that the sub-Blu-ray bar optimization model can reuse the optimization solution results obtained under different operating states, reduce the computational overhead required to re-search for control variables during each rolling solution, thereby alleviating the computational burden caused by repeated solutions of the sub-Blu-ray bar optimization model in the rolling control process and improving the efficiency of online solution.

[0023] (2) This invention uses the control quantity prediction results output by the learning-aid model as the initial values ​​of the control variables in the sub-Blu-ray bar optimization model, instead of directly using them as control commands for the flexible regulation equipment. The sub-Blu-ray bar optimization model uses a probability distribution fuzzy set to represent the probability distribution of the source-load prediction deviation, and re-solves the control quantity sequence under the physical constraints of the distribution network. This ensures that the final issued control quantity undergoes unified verification by the source-load uncertainty and the physical constraints of the distribution network, thereby avoiding the direct application of data-driven control results to the flexible regulation equipment without physical constraint verification, and improving the physical feasibility of the control results. Compared with existing methods, this invention can achieve rapid response and stable regulation of source-load fluctuations in high-penetration distributed photovoltaic scenarios, significantly improving the operational reliability and comprehensive carrying capacity of the distribution network. Attached Figure Description

[0024] Figure 1 A schematic diagram of a three-layer model structure of a flexible distribution network split-blower rolling control system provided in an embodiment of the present invention;

[0025] Figure 2 The flowchart of the flexible distribution network split-blower rolling control method provided in the embodiments of the present invention;

[0026] Figure 3 This is a schematic diagram illustrating the voltage change trend of nodes after power distribution network reconfiguration, provided in an embodiment of the present invention.

[0027] Figure 4 This is a schematic diagram of the spatiotemporal distribution changes of voltage at each node after the power distribution network is reconfigured, provided in an embodiment of the present invention. Detailed Implementation

[0028] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0029] This invention provides a method for flexible distribution network sub-bulb rolling control based on optimized guided learning. Figure 1 The rolling control system is implemented using a three-layer model structure consisting of a prediction perception layer, an optimization decision-making layer, and a learning execution layer. This three-layer model structure achieves a closed-loop process of flexible power distribution network operation status perception, control strategy optimization, and control result correction through the synergy between data perception, uncertainty optimization decision-making, and learning-assisted execution.

[0030] The predictive perception layer is used to collect and process multi-source heterogeneous information during the operation of the flexible distribution network, including distribution network topology data, line parameter data, node voltage data, node active power data, node reactive power data, and distributed generation output and load prediction deviation data. This layer first preprocesses and extracts features from the acquired data to construct a multi-attribute state matrix that reflects the current operating status. It then combines historical operating information to perceive the source-load change trends in future periods, providing state input for subsequent optimization decisions.

[0031] The optimization decision layer is used to establish a sub-Blu-ray optimization model based on the system state information output by the prediction sensing layer, and to generate an optimized control strategy that satisfies the operating constraints of the distribution network. This layer first constructs a probability distribution fuzzy set based on historical source-load prediction deviation samples to describe the possible range of source-load uncertainty. Then, combining node power balance constraints, line power flow constraints, node voltage constraints, and flexible regulation equipment operating constraints, it establishes a sub-Blu-ray optimization model under a rolling time window, and obtains the control quantity sequence within the prediction time domain by solving the model. The initial values ​​of the control variables output by the learning auxiliary model are used as the initial reference for solving the optimization model, thereby reducing the search range in the rolling optimization process and improving online solution efficiency.

[0032] The learning execution layer is used to assist in the generation and safe execution of optimized control strategies. This layer uses state-control samples generated by the optimization decision layer to train the learning auxiliary model, enabling it to learn the mapping relationship between the system's operating state and the optimized control variables. During actual operation, the learning auxiliary model quickly outputs initial values ​​of control variables based on the current multi-attribute state matrix, and performs feasibility corrections on the control results through action projection and safety constraint correction mechanisms to ensure that the final executed control strategy meets the physical constraints and operational boundary requirements of the distribution network.

[0033] Through the coordinated operation of the aforementioned prediction perception layer, optimization decision layer, and learning execution layer, this invention forms a rolling control closed loop of "state perception - optimization decision - learning-assisted execution - state update," which improves the computational efficiency and response capability of flexible distribution networks in the online control process while ensuring the reliability of distributed bar optimization control.

[0034] The method for acquiring data in a flexible distribution network system includes: collecting voltage amplitude and voltage phase angle, line power flow data, active and reactive power injection data of each node in the distribution network, and operating parameters of each flexible regulating device, wherein the flexible regulating device includes soft switching interconnection device, energy storage device and reactive power compensation device, and acquiring forecast information of distributed power output and load.

[0035] In this embodiment of the invention, distributed photovoltaic (PV) grid integration and distribution network collaborative optimization are taken as the research object, and the IEEE-33 node system is selected as the benchmark case for simulation verification. In this system, distributed PV power sources are connected to multiple typical nodes to simulate the uncertainty caused by fluctuations in PV output. Simultaneously, multi-source information such as historical load data, electricity prices, and operating conditions is combined to improve the model's ability to characterize actual operating characteristics. PV output data is modeled using typical daily irradiance curves, and random disturbances and prediction errors are superimposed on this model to reflect the uncertainty characteristics in actual operation. Relevant parameter settings are based on actual distribution network operating experience and typical operating conditions to ensure the rationality and representativeness of the simulation environment.

[0036] Based on the constructed system model data, the specific implementation steps of the method of the present invention are described below. The flowchart of the method is as follows: Figure 2 As shown.

[0037] S1: Obtain the topology, line parameters, node voltage, node power, and source-load prediction deviation of the flexible distribution network. Determine the electrical distance relationship between nodes based on the topology and line parameters. Determine the electrical coupling relationship between nodes based on the line parameters and node power transmission relationship. Based on the electrical distance relationship and electrical coupling relationship between nodes, organize the node voltage, node active power, and node reactive power into a multi-attribute state matrix according to the rolling time window.

[0038] First, basic parameters and operational status data of the system are acquired through power measurement devices and state estimation equipment configured in the distribution network. Data for the distributed photovoltaic (PV) distribution network system includes operating parameters of reactive power compensation devices, energy storage devices, and transformers configured at each node. Node voltage information includes voltage amplitude and phase angle at each node. Simultaneously, the active and reactive power injected into nodes after distributed PV integration is collected. In the IEEE-33 node system, multiple operable tie lines are configured to adjust the network structure, thereby changing the power flow distribution. During optimization, the operating mode is switched by adjusting network connections without altering the original line parameters.

[0039] Then, an optimized operating environment for the distribution network is constructed based on the data. To improve the state representation capability, a continuous state space is used to uniformly model the system operating state, fusing topological information, node voltage information, and power injection information. Traditional topologies are usually described using node adjacency matrices, which exhibit discrete changes when the network structure changes, hindering stable modeling and optimization solutions in subsequent iterations. Therefore, this invention constructs the system state as a continuous two-dimensional mapping in a single time section, fusing multi-source information such as system topology, node voltage, active and reactive power, to achieve a unified representation of complex operating states.

[0040] Specifically, the flexible distribution network is abstracted as a weighted network with nodes as vertices and lines as edges, and the basic electrical distance between adjacent nodes is defined using line impedance parameters. For nodes with direct connections... and nodes Its corresponding line complex impedance Represented as:

[0041]

[0042] In the formula, and They represent the lines respectively. The resistance and reactance. Indicates directly connected nodes With nodes The basic electrical distance between the two nodes. When there is no direct connection between the two nodes, the feasible route connecting the two nodes is determined according to the distribution network topology, and the basic electrical distances of each line on the route are summed. This indicates that for a node and nodes The equivalent electrical distance, Indicates the connection node and nodes Any path in the set of feasible route paths; when there are multiple connection paths, the path with the smallest cumulative line impedance modulus is taken as the shortest electrical path; when the distribution network is a radial structure, there is only one connection path between nodes, and the equivalent electrical distance is the sum of the basic electrical distances of each line on that path. An electrical distance matrix is ​​constructed based on the equivalent electrical distances between all node pairs. This matrix describes the electrical proximity relationships between different nodes in a flexible distribution network, formed based on topology and line impedance. In this invention, the electrical distance serves as an auxiliary feature of node association; it does not directly represent the operating state of node voltage, active power, or reactive power, nor does it replace the electrical coupling relationships between nodes described by the power flow equations.

[0043] The electrical distance relationship is mainly used to supplement the topology association information of the distribution network. It does not change the row and column structure of the node voltage, active power, and reactive power attribute matrices, nor does it independently complete the conversion from discrete topology to continuous state space. The continuous state representation of the flexible distribution network is mainly composed of normalized continuous operating variables such as node voltage, active power, and reactive power. The electrical distance relationship, as node association information, is input into the subsequent learning auxiliary model along with the multi-attribute state matrix to help the state feature extraction module identify the topology association features between different nodes. Combined with the distribution network power flow calculation model, the electrical coupling relationship between nodes is established based on the physical constraints between node power injection, line impedance, and node voltage. The electrical distance relationship focuses on describing the degree of topology association between nodes, reflecting the network structure characteristics; the electrical coupling relationship describes the physical interaction between node operating states through power flow equations, reflecting electrical operating characteristics. Together with continuous operating variables such as node voltage, active power, and reactive power, they constitute a complete representation of the operating state of the flexible distribution network.

[0044] For any given node, its operating state is influenced not only by its own power injection but also by the combined effects of the operating states of its connected lines and neighboring nodes. Therefore, the electrical coupling relationships between nodes are described using line power flow constraint equations:

[0045]

[0046] in, and They represent the lines respectively. Active power and reactive power, and This indicates the active and reactive power injected into the node. and These represent the line resistance and reactance, respectively. Indicates the line current. and These represent the node voltage amplitudes, and Indicates the node number in the distribution network. Represents nodes Adjacent downstream nodes, Indicates the load node. This represents a distributed power node. and Representing nodes respectively The active and reactive power injected at the load side, and Representing nodes respectively The active and reactive power output of the distributed power source.

[0047] The aforementioned power flow model describes the electrical coupling relationships between distribution network nodes through constraints between line impedance parameters, node power injection, and node voltage. The line power transmission process is jointly determined by line impedance parameters and node power states, while node voltage changes are influenced by the power changes of adjacent nodes and line parameters. Therefore, this model establishes a correlation mapping relationship between node operating states. Based on this, this invention does not directly treat discrete nodes as independent state variables, but rather utilizes the distribution network topology and line parameters to establish electrical correlation relationships between nodes, mapping operating information such as node voltage, active power, and reactive power to a unified state space. By normalizing the state variables of each node and arranging them according to node and time dimensions, a continuous multi-attribute state matrix is ​​formed. Therefore, the continuous state space refers to expressing the originally discretely distributed operating variables with node numbers through correlation based on electrical distance relationships, electrical coupling relationships, and time series information, enabling the operating states of different nodes at different time scales to be described within a unified mathematical space. This continuous state space serves as the state input for subsequent learning auxiliary models and sub-Bruker optimization models, characterizing the dynamic changes in the operating state of the flexible distribution network.

[0048] A reference node is selected as the voltage reference node, and the system state variables are shifted and normalized to unify the scale of state expression. The expression is as follows:

[0049]

[0050] in, Represents a node voltage amplitude, This represents the reference node voltage, used to standardize the voltage reference. This represents the normalized node voltage deviation; and These represent the active power and reactive power injected into the node, respectively. and These represent the baseline active and reactive power capacities, respectively. and These represent the normalized active and reactive power, respectively. Indicates time The system-standardized state vector is used to characterize the operating state of the flexible distribution network at a uniform scale.

[0051] Next, based on the aforementioned continuous state-space representation, the operational information such as node voltage and power injection is organized temporally within a rolling time window to construct a multi-attribute state matrix. To ensure that the environmental state has a unified expression form under different operating stages, the voltage, active power, and reactive power data of each node on a continuous time scale are arranged and combined to form a unified-dimensional state input structure.

[0052] Specifically, within a given time window Inside, at the current moment Starting from this point, we first obtain the time of each node. to The operational data includes information such as node voltage, active power, and reactive power. Since the operating state of a distribution network is determined by multiple nodes, the operational data of a single node cannot fully reflect the overall network operating characteristics. Therefore, this invention combines the aforementioned electrical distance and electrical coupling relationships to uniformly express the correlation characteristics between different nodes. The electrical distance relationship between nodes describes the spatial correlation between different nodes in the network topology, while the power flow model describes the electrical coupling relationship between node voltage, node power, and line parameters. This maps the original discretely distributed operational data according to node numbers to a continuous state space with node correlation characteristics. In this state space, the operational variables of each node are no longer independent data points, but rather form a state representation with spatial correlation and temporal evolution characteristics through the physical connection relationship of the distribution network. After completing the above continuous state expression, various operational variables are arranged and combined according to the node dimension and the time dimension to construct voltage attribute matrices respectively. Active power attribute matrix and reactive power attribute matrix In this matrix, rows represent node numbers, and columns represent the corresponding time series within a rolling time window. This method transforms discrete node operational data into a unified, multi-attribute state matrix, used to describe the dynamic operation of the flexible distribution network over time. The attribute matrix construction process is as follows:

[0053] ;

[0054] in, Represents a node At any moment The voltage amplitude; and Representing nodes respectively At any moment Active power and reactive power; Indicates the number of system nodes; Indicates the length of the scrolling time window; , and These represent the voltage attribute matrix, active power attribute matrix, and reactive power attribute matrix, respectively, and are used to characterize the evolution of the operating state of the flexible distribution network over time.

[0055] In this invention, the electrical coupling relationship is used to characterize the correlation between the corresponding operating quantities of different nodes in the multi-attribute state matrix. The electrical coupling relationship is determined based on the distribution network topology, line impedance parameters, and node power transmission relationships, and is used to reflect the impact of a node's operating state change on the operating states of adjacent and related nodes. Specifically, the voltage attribute matrix, active power attribute matrix, and reactive power attribute matrix are used to record the time-series data of node voltage, active power, and reactive power within a rolling time window, respectively. The electrical coupling relationship does not change the original row and column structure of each attribute matrix, but rather serves as inter-node correlation information, used together with the attribute matrices for system state characterization. In the subsequent learning-assisted model, the state feature extraction module combines the multi-attribute state matrix and the inter-node electrical coupling relationship to extract the operating state characteristics of each node and the correlation characteristics between different nodes. This allows the learning-assisted model to simultaneously utilize node time-series operating information and the distribution network physical connection relationship, providing state input for subsequent state transition prediction and initial value generation of control variables.

[0056] S2. Based on historical samples of source-load prediction deviation, establish an empirical distribution. Construct a probability distribution fuzzy set centered on the empirical distribution. Use the probability distribution fuzzy set to characterize the probability distribution of source-load prediction deviation. Use the physical constraints of the distribution network to limit the feasible domain of control quantities and distribution network operating status, and establish a sub-Blu-shaped optimization model in the prediction time domain.

[0057] This invention constructs an uncertainty modeling process based on the attribute state matrix, and uses the sub-Bruker optimization method to model source load fluctuations. For finite sample data, it generates a continuous uncertainty description by constructing a probability distribution fuzzy set. The sub-Bruker modeling process is as follows:

[0058]

[0059] in, This represents a vector of uncertain variables, including distributed generation output deviation and load forecast deviation. Indicates the first A historical sample; Indicates the number of samples; This represents an empirical distribution constructed based on historical samples; Indicates at sample points Dirac distribution at location; Represents the candidate probability distribution; Represents the Wasserstein distance; Indicates the radius of the distributed fuzzy set; Represents a set of uncertain distributions; These represent the optimization decision variables, including node voltage, power flow, and SOP control variables. This represents the upper limit of the constraint violation probability, used to describe the level of risk in system operation. The last constraint in the modeling expression is used as an opportunity constraint to limit the probability of violation of the physical constraints of the distribution network under the candidate probability distribution.

[0060] Furthermore, centering on the empirical distribution formed by historical samples, a fuzzy set of probability distributions capable of covering the actual range of prediction error variations is constructed by limiting the Wasserstein distance range between the candidate probability distribution and the empirical distribution. This avoids pre-assuming that the source load bias follows a fixed probability distribution, thereby improving the ability to describe uncertainty. The decision variables of the sub-Blu-ray bar optimization model include the node voltage state, line power flow state, and flexible regulation equipment control quantities at each time point in the prediction time domain. The flexible regulation equipment control quantities include SOP active power regulation quantity, reactive power regulation quantity, and other adjustable equipment control variables. Indicates an uncertain scenario The system operation cost function is used to comprehensively evaluate the economic efficiency, safety, and control smoothness of the distribution network. Its specific forms include network loss cost, voltage deviation cost, and control quantity change cost at adjacent scheduling times.

[0061]

[0062] in, This indicates the system's active power loss. Indicates the node voltage deviation index. Indicates the change index of the control quantity; These represent the corresponding weight coefficients. The function represents the power distribution network operation constraint function, which is used to limit the feasible domain of the optimization results, i.e., the power distribution network physical constraints, including node power balance constraints, line capacity constraints, node voltage constraints, and flexible regulation equipment operation constraints. Specifically, it includes: (1) node power balance constraints, which are used to ensure that the injected power of each node and the transmitted power of the line meet the power flow balance relationship; (2) line capacity constraints, which are used to limit the transmitted power of the line to not exceed the allowable capacity; (3) node voltage constraints, which are used to ensure that the node voltage amplitude is within the safe operating range; (4) flexible regulation equipment operation constraints, which are used to limit the active and reactive power regulation range of equipment such as SOP.

[0063] By solving the above-mentioned bicycle bar optimization model, the optimal decision variables in the prediction time domain can be obtained, including the control quantities of flexible regulating equipment, node operating states, and line power flow states corresponding to each scheduling time. Subsequently, only the control quantities corresponding to the current scheduling time are executed, and the state matrix and optimization model are updated again in the next rolling time, realizing the rolling optimization control of the flexible distribution network.

[0064] S3. Solve the sub-Bruker optimization model using different multi-attribute state matrices as the current state to obtain the corresponding optimized control quantity and the subsequent multi-attribute state matrix after the optimized control quantity is applied. The current multi-attribute state matrix, the optimized control quantity, and the subsequent multi-attribute state matrix form a training sample set. Use the training sample set to train the learning auxiliary model to learn the state transition relationship from the current multi-attribute state matrix and the optimized control quantity to the subsequent multi-attribute state matrix, as well as the correspondence between the current multi-attribute state matrix and the optimized control quantity.

[0065] This invention constructs a state-control sample set using historical operating data and optimization solution results. Specifically, within a rolling time window, based on the system state matrix at each moment, and combined with the optimized control quantity output by the split-bar optimization model, a sequence of "state-control-next state" data is formed. The next state is the subsequent multi-attribute state matrix after the optimized control quantity is applied, which is used to characterize the dynamic evolution process of the distribution network's operating state, thereby providing training samples for the learning model.

[0066] Based on the aforementioned sample data, a learning-aid model is constructed to model the system state transition relationships. This learning-aid model employs a neural network-based state mapping model, which is essentially a data-driven model that uses historical optimization results to learn the nonlinear mapping relationship between the flexible distribution network's operating state and control variables, rather than directly replacing the sub-Brønsted bar optimization model. The learning-aid model includes a state feature extraction module, a state transition prediction module, and a control variable prediction module. Specifically, the state feature extraction module encodes the input multi-attribute state matrix, extracting node voltage, active power, reactive power, and inter-node correlation information; the state transition prediction module learns the dynamic change pattern of the current operating state under control and outputs the state prediction result for the next moment; the control variable prediction module generates corresponding approximate control variable results based on the current state characteristics and uses these results as the initial values ​​of the control variables for the sub-Brønsted bar optimization model. The input to the learning-aid model includes the current system state matrix. and the control variables obtained from the Bruker bar optimization model Output the state matrix at the next time step. The prediction results are obtained. By comparing the model's predicted state with the actual operating state, a state transition error is constructed. This error, combined with the control variable approximation error, is used to iteratively update the model parameters, enabling the learning model to gradually learn the changing patterns of the distribution network's operating state and the corresponding relationships between optimized control strategies. Among these, the state... A multi-attribute state matrix representation, consisting of node voltage, active power, and reactive power, can reflect the comprehensive state characteristics of flexible distribution networks under different operating conditions; control variables This represents the control variables of flexible regulating equipment obtained from the sub-Blu-ray bar optimization model, including SOP active power regulation, reactive power regulation, and other adjustable equipment control variables. During actual operation, the learning-aid model quickly predicts the initial values ​​of the control variables based on the current state matrix and inputs them into the sub-Blu-ray bar optimization model. Subsequently, the sub-Blu-ray bar optimization model is re-solved under the probabilistic fuzzy set and distribution network physical constraints to obtain the final execution control variables. Therefore, the learning-aid model is only used to improve the online solution efficiency of the optimization model and does not directly replace the optimization decision-making process, thus ensuring that the control results meet the physical feasibility requirements of the distribution network.

[0067] To improve the learning model's ability to fit the system's dynamic characteristics and enhance the feasibility and stability of the control strategy, the following joint loss function is constructed to train the learning model parameters:

[0068]

[0069] in, Represents the parameters of the learning model; and Representing time respectively and The system state vector, which consists of node voltage, active power and reactive power; This represents the optimized control quantity obtained from the split-bar optimization model; This represents the initial control values ​​generated by the learning model; This represents the state transition mapping relationship constructed by the learning model; Represents the L2 norm; The weighting coefficients are used to balance the state prediction error and control deviation. The loss function characterizes the learning model's ability to fit the dynamic evolution of the system and its approximation of the optimal control strategy. The joint loss function consists of two parts: the first part measures the learning model's predicted state... Compared with actual operating status The differences between them enable the learning model to learn the dynamic evolution of the distribution network under control; the second part is used to measure the initial values ​​of the control variables output by the learning model. Solution results of the Bruker bar optimization model The deviation between the two allows the learning model to learn the mapping relationship between the optimized control strategy and the system operating state. Through the above joint training method, the learning auxiliary model can simultaneously learn the correspondence between "state-state transition" and "state-control variable". In actual operation, it can quickly generate initial values ​​of control variables close to the optimal solution based on the current multi-attribute state matrix. Then, the sub-Bruker optimization model further combines the probability distribution fuzzy set and the physical constraints of the distribution network to solve again, ensuring the feasibility of the final control strategy.

[0070] S4. At the current scheduling moment, input the current multi-attribute state matrix into the learning auxiliary model to obtain the initial values ​​of the control variables. Use the initial values ​​of the control variables to initialize the control variables of the sub-Blu-ray optimization model. Solve the control quantity sequence in the prediction time domain under the probability distribution fuzzy set and the physical constraints of the distribution network.

[0071] In the actual decision-making process, based on the current system state matrix By combining the operational information in the prediction time domain, the initial values ​​of the control variables are generated using the learning-assisted model and used as the input of the rolling optimization model. Under the conditions of satisfying the node power balance constraints, line capacity constraints and voltage constraints, the optimal control decision vector at the current moment is solved, thereby realizing the real-time optimization and adjustment of the distribution network operation status.

[0072] For any node, the constraint form is:

[0073]

[0074] in, and These represent the active power and reactive power output by the distributed power source, respectively. and These represent the active and reactive power regulation quantities of the flexible interconnection device, respectively. and These represent the active power and reactive power consumed by the node load, respectively. This indicates the reactive power provided by the reactive power compensation equipment; Represents nodes A set of connected lines. Indicates the line The maximum allowable transmission capacity is used to ensure the safe operation of the line. and These represent the lower and upper limits of the node voltage, respectively. This represents the upper limit of the SOP equipment capacity. Based on the above constraints, the sub-Blu-ray optimization model can further optimize and correct the initial values ​​of the control variables output by the learning-aid model, while satisfying the physical operating boundaries of the distribution network, to obtain the optimal control quantity that satisfies both source-load uncertainty constraints and equipment operating constraints.

[0075] S5: Send the control quantity corresponding to the current scheduling moment in the control quantity sequence to the flexible regulation equipment, collect the node voltage, node active power and node reactive power after executing the control quantity, update the multi-attribute state matrix and enter the next scheduling moment.

[0076] Furthermore, the learning model evaluates the system state and control strategy data generated during real-time operation, and filters and updates the strategies based on comprehensive performance indicators to improve the adaptability and decision stability of the learning model. The comprehensive performance indicators are obtained by evaluating the system's operating state in the prediction time domain, and their calculation formula is as follows:

[0077]

[0078] in, This represents the total operating cost within the predicted time domain; This indicates a rolling optimization prediction time window; Indicates time The system has active power loss; Represents a node At any moment The voltage amplitude; Indicates the reference voltage; Indicates the number of system nodes; and These represent the control decision vectors at adjacent time points; Represents the L2 norm; , , These are the network loss weighting coefficient, voltage deviation weighting coefficient, and control smoothing weighting coefficient, respectively, used to characterize the system's operating economy, voltage safety, and control stability.

[0079] Furthermore, to improve the adaptability of the learning-assisted model to different operating scenarios, this invention uses the aforementioned comprehensive operating cost index to evaluate the initial control values ​​generated by the learning model, and updates the learning model parameters based on the evaluation results. At each rolling optimization time step, the initial control variable values ​​generated by the learning-assisted model based on the current multi-attribute state matrix are... Input the sub-Bruker optimization model, and then re-solve the sub-Bruker optimization model under the conditions of satisfying the source load uncertainty constraint and the distribution network physical constraint to obtain the final control variables. Subsequently, based on the system's operating state after the control variable takes effect, the comprehensive operating cost within the corresponding prediction time window is calculated. When the initial values ​​of the control variables generated by the learning-aid model enable the sub-Blu-ray optimization model to achieve a low overall operating cost, it indicates that the initial values ​​of the control variables have a good matching relationship with the current operating state. In this case, the corresponding state-control sample is retained and used to update the learning model parameters. When the calculated overall operating cost is high (e.g., comparable to a preset threshold, the previous scheduling time, or the historical average), it indicates that the current learning model output deviates significantly from the actual operating requirements. In this case, the optimal control result obtained by resolving the sub-Blu-ray optimization model is used to replace the original initial control values, forming new training samples to correct the learning model. Specifically, during the update process, the current system state, the multi-attribute state matrix, the initial values ​​of the control variables output by the learning model, and the optimal control variables corrected by the optimization model are collectively used to form the feedback sample. The learning model is then incrementally trained using this sample, gradually approximating the control decision-making patterns of the sub-Blu-rod optimization model under different operating conditions. This screening and updating mechanism based on comprehensive operating costs avoids the learning model directly outputting unverified control results. Simultaneously, the high-quality optimization solutions provided by the sub-Blu-rod optimization model are used to continuously refine the learning model, improving the control accuracy and generalization ability of the learning-aid model under different source load fluctuation scenarios.

[0080] Figure 3 This paper demonstrates the voltage variation trend of system nodes after implementing the distribution network distribution bar rolling control method of this invention. Overall, the maximum system voltage remains relatively stable with small fluctuations; the average voltage remains near its rated value for most of the time with relatively gentle changes; however, the minimum voltage shows a significant downward trend during periods of high load, especially exhibiting large fluctuations at night. This is mainly because the higher photovoltaic output during the day effectively supports the node voltage level, keeping the overall system voltage stable; while at night, the photovoltaic output weakens or even disappears, and load demand increases, leading to voltage drops at some nodes. Optimizing and adjusting the system operating state using the proposed method can suppress voltage fluctuations to a certain extent, improve low-voltage problems, and make the voltage distribution more balanced. Although some voltage drop still exists during periods of extreme load, the overall voltage level remains within the safe operating range, indicating that the proposed method can effectively improve the stability and safety of distribution network voltage operation.

[0081] Under conditions of high-penetration distributed photovoltaic access, uneven voltage distribution and fluctuations become key challenges for optimizing the operation of distribution networks. Figure 4The three-dimensional distribution of system voltage across different time periods and node numbers is presented. Overall, the system voltage is mainly distributed within the range of 0.95 to 1.04 pu. During the initial operation phase, the voltage distribution at each node exhibits a certain degree of dispersion, with some nodes showing significant voltage drops during periods of high load. As operation progresses, during periods of reduced photovoltaic output and increased load, the voltage at some end nodes shows a more significant downward trend, demonstrating strong spatiotemporal coupling characteristics.

[0082] After optimizing and adjusting the system's operating status using the proposed method, the voltage distribution at each node tends to be more balanced, low-voltage areas are effectively alleviated, and the overall voltage level is closer to the rated range. Especially during peak load periods and periods of fluctuating photovoltaic output, the voltage drop is significantly suppressed, thereby reducing the risk of voltage exceeding limits and improving system stability and power supply quality. In summary, the method of this invention, through dynamic optimization and coordinated adjustment of the system's operating status, effectively improves the distribution network's adaptability to distributed photovoltaic fluctuations and enhances the system's safety and reliability.

[0083] Based on the same technical concept as the method embodiments, the present invention also provides a flexible distribution network distribution bar rolling control system based on optimization-guided learning, comprising:

[0084] The state characterization module is used to obtain the topology, line parameters, node voltage, node power and source-load prediction deviation of the flexible distribution network. Based on the topology and line parameters, it determines the electrical distance relationship between nodes. Based on the line parameters and node power transmission relationship, it determines the electrical coupling relationship between nodes. Based on the electrical distance relationship and electrical coupling relationship between nodes, it organizes the node voltage, node active power and node reactive power into a multi-attribute state matrix according to the rolling time window.

[0085] The robust optimization module is used to establish an empirical distribution based on historical samples of source load prediction deviations, construct a probability distribution fuzzy set centered on the empirical distribution, characterize the probability distribution of source load prediction deviations with the probability distribution fuzzy set, limit the feasible domain of control quantities and distribution network operating status with distribution network physical constraints, and establish a partial robust optimization model in the prediction time domain.

[0086] The model training module is used to solve the split-bar optimization model with different multi-attribute state matrices as the current state, obtain the corresponding optimized control quantity, and obtain the subsequent multi-attribute state matrix after the optimized control quantity is applied. The current multi-attribute state matrix, the optimized control quantity, and the subsequent multi-attribute state matrix form a training sample set, which is used to train and learn the auxiliary model.

[0087] The online control module is used to input the current multi-attribute state matrix into the trained learning auxiliary model at the current scheduling time to obtain the initial values ​​of the control variables. The control variables of the split-bar optimization model are initialized with the initial values ​​of the control variables, and the control quantity sequence in the prediction time domain is obtained by solving under the probability distribution fuzzy set and the physical constraints of the distribution network.

[0088] The feedback update module is used to send the control quantity corresponding to the current scheduling moment in the control quantity sequence to the flexible regulation equipment, collect the node voltage, node active power and node reactive power after the execution of the control quantity, update the multi-attribute state matrix and enter the next scheduling moment.

[0089] The specific implementation process of each module can be referred to the relevant descriptions in the above method embodiments, and will not be repeated here.

[0090] The present invention also provides an electronic device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, wherein when the programs are executed by the processors, they implement the flexible distribution network split-bar rolling control method based on optimized guided learning as described above.

[0091] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the above-described method for flexible distribution network branch-brush rolling control based on optimization-guided learning.

[0092] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method for flexible distribution network branch-bar rolling control based on optimized guided learning.

Claims

1. A method for flexible distribution network sub-bar rolling control based on optimization-guided learning, characterized in that, The method includes: The topology, line parameters, node voltage, node power, and source-load prediction deviation of the flexible distribution network are obtained. The electrical distance relationship between nodes is determined based on the topology and line parameters. The electrical coupling relationship between nodes is determined based on the line parameters and node power transmission relationship. Based on the electrical distance relationship and electrical coupling relationship between nodes, the node voltage, node active power, and node reactive power are organized into a multi-attribute state matrix according to the rolling time window. An empirical distribution is established based on historical samples of source-load prediction deviations. A probability distribution fuzzy set is constructed with the empirical distribution as the center. The probability distribution fuzzy set is used to characterize the probability distribution of source-load prediction deviations. The feasible domain of control quantities and distribution network operation status is limited by the physical constraints of the distribution network. A sub-Bruker optimization model in the prediction time domain is established. The sub-Bruker optimization model is solved by using different multi-attribute state matrices as the current state to obtain the corresponding optimized control quantity and the subsequent multi-attribute state matrix after the optimized control quantity is applied. The current multi-attribute state matrix, the optimized control quantity and the subsequent multi-attribute state matrix form a training sample set, and the auxiliary model is trained using the training sample set. At the current scheduling moment, the current multi-attribute state matrix is ​​input into the trained learning auxiliary model to obtain the initial values ​​of the control variables. The control variables of the split-bar optimization model are initialized with the initial values ​​of the control variables, and the control quantity sequence in the prediction time domain is obtained by solving under the probability distribution fuzzy set and the physical constraints of the distribution network. The system issues control quantities corresponding to the current scheduling time in the control quantity sequence to the flexible regulation equipment, collects the node voltage, node active power and node reactive power after the control quantity is executed, updates the multi-attribute state matrix and enters the next scheduling time.

2. The method according to claim 1, characterized in that, The determination of the electrical distance relationship and the electrical coupling relationship between nodes includes: For directly connected nodes and nodes The modulus of the corresponding line complex impedance is determined as the basic electrical distance. For nodes that are not directly connected and nodes Connect nodes and nodes The minimum sum of the basic electrical distances along each line path is determined as the equivalent electrical distance. The equivalent electrical distances between each pair of nodes form a node electrical distance matrix. : The electrical coupling relationship between nodes is determined based on the power flow constraints between the active and reactive power injected at the nodes, the line impedance, and the node voltage. The electrical distance relationship and electrical coupling relationship between nodes are used to help map node voltage, active power and reactive power to a unified state space. Neither of them changes the row and column structure of the multi-attribute state matrix.

3. The method according to claim 1, characterized in that, The method for constructing a multi-attribute state matrix is ​​as follows: Let the set of nodes of the flexible distribution network be denoted as The set of times corresponding to the rolling time window is denoted as ,according to The node voltage of node i Active power at nodes and node reactive power Normalize, Reference node voltage, and These are the active power baseline capacity and the reactive power baseline capacity, respectively. From the normalized node voltage Active power at nodes and node reactive power Form voltage attribute matrices respectively Active power attribute matrix and reactive power attribute matrix And satisfy In this context, the rows of each attribute matrix correspond to the nodes of the flexible distribution network, and the columns of each attribute matrix correspond to the sampling times within the rolling time window. The voltage attribute matrix, the active power attribute matrix, and the reactive power attribute matrix constitute the multi-attribute state matrix.

4. The method according to claim 1, characterized in that, The method for constructing fuzzy sets of probability distributions is as follows: The deviations between the predicted and actual photovoltaic power output, as well as the deviations between the predicted and actual load, are combined to form a source-load prediction deviation sample. An empirical distribution is then formed from the historical samples of these source-load prediction deviations. , For historical samples of source load prediction bias, Let H be the Dirac distribution corresponding to the historical samples, and H be the number of historical samples. This represents the empirical distribution; Candidate probability distributions whose Wasserstein distances satisfy the distance constraint with the empirical distribution are grouped into the fuzzy set of probability distributions. ,in, This represents the fuzzy set of the probability distribution. Represents the candidate probability distribution. Represents the Wasserstein distance. This represents the radius of the fuzzy set of the distribution determined based on historical samples of the source load prediction bias.

5. The method according to claim 4, characterized in that, The method for establishing the Bruker bar optimization model is as follows: The optimization problem of the sub-Bruker optimization model, with the worst expected running cost under the candidate probability distribution within the fuzzy probability set as the optimization objective, satisfies... ,in, The optimization decision variables include the node voltage state, line power flow state, and flexible regulation equipment control quantities at each scheduling moment in the prediction time domain. This represents the uncertain variable constituted by the source load prediction bias. , and These represent the network loss cost, node voltage deviation cost, and control quantity change cost at adjacent scheduling times, respectively. , and These are the corresponding weighting coefficients; The distribution network physical constraints of the sub-Bruker optimization model include: node power balance constraints, line capacity constraints, node voltage constraints, and flexible regulation equipment operation constraints. Furthermore, the probability of violating these distribution network physical constraints under the candidate probability distribution is limited by chance constraints. in, For the physical constraint functions of the distribution network, To constrain the upper limit of the probability of violation.

6. The method according to claim 1, characterized in that, Training the learning-assisted model includes: By inputting different multi-attribute state matrices as the current multi-attribute state matrix into the Bruker optimization model, a control quantity sequence in the prediction time domain is obtained. The control quantity corresponding to the current scheduling time is then extracted from the control quantity sequence as the optimized control quantity. The subsequent multi-attribute state matrix is ​​formed based on the node voltage, node active power, and node reactive power after the optimization control is applied. ; the corresponding current multi-attribute state matrices Optimize control quantity and subsequent multi-attribute state matrix Forming training samples; The current multi-attribute state matrix and optimize control quantity Input the learning auxiliary model to obtain the subsequent state prediction result; then input the current multi-attribute state matrix. Input the learning-assisted model to obtain the control quantity prediction results. ; Based on the state transition prediction error between the subsequent state prediction result and the subsequent multi-attribute state matrix, and the control quantity prediction error between the control quantity prediction result and the optimized control quantity, a joint loss function is constructed, and the parameters of the learning auxiliary model are updated according to the joint loss function: in, For the joint loss function, To learn the parameters of the auxiliary model, These are the weighting coefficients. To learn from the successor state predictions generated by the auxiliary model, This represents the L2 norm.

7. The method according to claim 1, characterized in that, The update of the multi-attribute state matrix includes: The node voltage, node active power, and node reactive power collected after executing the control quantity corresponding to the current scheduling moment are added to the corresponding attribute matrix, and the node voltage, node active power, and node reactive power corresponding to the earliest sampling moment in the rolling time window are removed. The current multi-attribute state matrix before executing the control quantity, the initial values ​​of the control variables output by the learning auxiliary model, the optimized control quantity corresponding to the current scheduling moment obtained by resolving the Bruker optimization model, and the subsequent multi-attribute state matrix formed after executing the control quantity are used to form a feedback sample, and the learning auxiliary model is incrementally trained using the feedback sample.

8. A flexible distribution network distributed rod rolling control system based on optimization-guided learning, characterized in that, include: The state characterization module is used to obtain the topology, line parameters, node voltage, node power and source-load prediction deviation of the flexible distribution network. Based on the topology and line parameters, it determines the electrical distance relationship between nodes. Based on the line parameters and node power transmission relationship, it determines the electrical coupling relationship between nodes. Based on the electrical distance relationship and electrical coupling relationship between nodes, it organizes the node voltage, node active power and node reactive power into a multi-attribute state matrix according to the rolling time window. The robust optimization module is used to establish an empirical distribution based on historical samples of source load prediction deviations, construct a probability distribution fuzzy set centered on the empirical distribution, characterize the probability distribution of source load prediction deviations with the probability distribution fuzzy set, limit the feasible domain of control quantities and distribution network operating status with distribution network physical constraints, and establish a partial robust optimization model in the prediction time domain. The model training module is used to solve the split-bar optimization model with different multi-attribute state matrices as the current state, obtain the corresponding optimized control quantity, and obtain the subsequent multi-attribute state matrix after the optimized control quantity is applied. The current multi-attribute state matrix, the optimized control quantity, and the subsequent multi-attribute state matrix form a training sample set, which is used to train and learn the auxiliary model. The online control module is used to input the current multi-attribute state matrix into the trained learning auxiliary model at the current scheduling time to obtain the initial values ​​of the control variables. The control variables of the split-bar optimization model are initialized with the initial values ​​of the control variables, and the control quantity sequence in the prediction time domain is obtained by solving under the probability distribution fuzzy set and the physical constraints of the distribution network. The feedback update module is used to send the control quantity corresponding to the current scheduling moment in the control quantity sequence to the flexible regulation equipment, collect the node voltage, node active power and node reactive power after the execution of the control quantity, update the multi-attribute state matrix and enter the next scheduling moment.

9. An electronic device, characterized in that, The system includes a processor, a memory, and a computer program stored in the memory and executable by the processor, wherein the computer program, when executed by the processor, implements the flexible distribution network branch-bar rolling control method based on optimization-guided learning as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores a computer program, which, when executed by a processor, implements the flexible distribution network split-bar rolling control method based on optimization-guided learning as described in any one of claims 1 to 7.

Images