Power distribution substation low voltage treatment method and system based on spatio-temporal graph neural network and improved sparrow search

Abstract

The application discloses a power distribution area low-voltage treatment method based on a space-time graph neural network and an improved sparrow search, and comprises the following steps: constructing a graph structure of a power distribution area; collecting multi-source operation data and constructing a node feature vector; inputting the feature sequence of all nodes in a continuous sampling period and the adjacency matrix thereof into a space-time graph neural network model, outputting voltage prediction values and node hidden states of each node, and generating a low-voltage cause vector; constructing and online correcting a voltage sensitivity matrix, and determining candidate treatment resources; constructing a multi-resource collaborative treatment optimization model, and solving the optimization model by using an improved sparrow search algorithm to obtain a collaborative treatment strategy; issuing a collaborative treatment instruction to corresponding treatment resources according to the collaborative treatment strategy, and correcting the voltage sensitivity matrix and the space-time graph neural network model according to the measured voltage response after the treatment action. The method improves the accuracy, real-time performance, economy and adaptability of low-voltage treatment.

Description

Technical Field

[0001] This invention relates to the field of power quality management technology for distribution networks, specifically to a method, system, and equipment for managing low voltage in distribution substations based on spatiotemporal graph neural networks and improved sparrow search. Background Technology

[0002] With the large-scale integration of distributed photovoltaic (PV) systems, electric vehicle charging facilities, user-side energy storage, air source heat pumps, agricultural irrigation loads, and high-power residential appliances into low-voltage distribution substations, the voltage operation characteristics of these substations have gradually shifted from the traditional unidirectional, low-fluctuation, and weakly coupled characteristics to bidirectional interaction between source and load, multi-point fluctuations, and multi-factor coupling. Low voltage issues may not only occur at end users with long power supply radii, but may also occur during specific periods such as evening peak hours, peak summer and winter demand, concentrated charging, or rapid declines in PV output.

[0003] Current methods for managing low voltage typically include increasing the capacity of distribution transformers, upgrading transmission lines, reactive power compensation, boosting voltage with voltage regulators, and using end-point voltage management devices or energy storage discharge. While these methods can improve voltage quality in certain scenarios, they still have the following shortcomings:

[0004] First, existing methods typically use whether the voltage is below a threshold as the trigger condition for governance, making it difficult to accurately distinguish whether the low voltage is caused by heavy transformer load, line voltage drop, three-phase imbalance, insufficient reactive power, photovoltaic output fluctuations, concentrated charging load access, or seasonal load growth, resulting in a mismatch between governance measures and the actual causes.

[0005] Second, existing governance strategies mostly target individual devices for localized control, such as switching capacitors on and off separately, controlling energy storage discharge separately, or adjusting voltage regulator settings separately. They lack synergistic optimization among voltage regulators, reactive power compensation devices, energy storage devices, commutation devices, photovoltaic inverters, and adjustable loads.

[0006] Third, most existing methods are passive responses after voltage exceeds the limit, and cannot make short-term predictions based on information such as load, photovoltaic, charging load and weather before low voltage occurs, which can easily lead to a long duration of low voltage.

[0007] Fourth, the parameters of low-voltage lines, user load structure, distributed power supply access points, and transformer area topology in distribution substations have strong dynamic change characteristics. Existing control models are difficult to correct online in a timely manner based on the actual governance effect, which can easily lead to insufficient regulation, over-regulation, or frequent equipment operation. Summary of the Invention

[0008] The purpose of this invention is to provide a method, system, and device for low voltage management in distribution substations based on spatiotemporal graph neural networks and improved sparrow search, in order to solve the problems of inaccurate cause identification, insufficient coordination of management equipment, delayed management actions, and difficulty in adaptive correction of control models in existing low voltage management.

[0009] This invention is achieved through the following technical solution:

[0010] In a first aspect, the first embodiment of the present invention provides a method for managing low voltage in distribution substations based on a spatiotemporal graph neural network and an improved sparrow search, comprising:

[0011] The distribution radio area is constructed as a graph structure containing nodes, edges, and an adjacency matrix;

[0012] Collect multi-source operational data from each node in the distribution area and construct node feature vectors;

[0013] The feature sequences and adjacency matrices of all nodes within a continuous sampling period are input into the spatiotemporal graph neural network model, which outputs the voltage prediction values ​​and hidden states of each node in the future prediction time domain, and generates a low voltage cause vector based on the hidden states of the nodes.

[0014] The voltage sensitivity matrix is ​​constructed and corrected online based on the node voltage changes before and after the governance action, and candidate governance resources are determined based on the low voltage cause vector and the voltage sensitivity matrix.

[0015] A multi-resource collaborative governance optimization model is constructed based on the low voltage cause vector, voltage sensitivity matrix and equipment constraints. An improved sparrow search algorithm is used to solve the multi-resource collaborative governance optimization model to obtain the collaborative governance strategy.

[0016] Based on the collaborative governance strategy, collaborative governance instructions are issued to the corresponding governance resources, and the voltage sensitivity matrix and spatiotemporal neural network model are corrected based on the measured voltage response after the governance action.

[0017] Furthermore, the construction of the distribution radio area into a graph structure containing nodes, edges, and an adjacency matrix specifically includes:

[0018] The low-voltage side of the transformer, the beginning and end of the branch, the meter box, the user terminal, the photovoltaic inverter access point, the energy storage access point, the reactive power compensation equipment access point, and the voltage regulation equipment access point are all regarded as nodes, and the low-voltage lines between the nodes are regarded as edges. The electrical connection relationship between the nodes is represented by the adjacency matrix A.

[0019] If node With nodes If there is a direct electrical connection between adjacent matrix elements, then... If there is no direct electrical connection, then Alternatively, the adjacency matrix elements are determined by a weighted average of line impedance, line length, or node electrical distance, calculated using the following formula:

[0020] ;

[0021] in, Represents a node With nodes The equivalent line impedance between them This represents the attenuation coefficient.

[0022] Furthermore, the spatiotemporal graph neural network model includes a graph convolutional layer, a time series feature extraction layer, and an output layer. The graph convolutional layer is used to extract the spatial correlation features between nodes in the transformer area. The time series feature extraction layer is used to extract the temporal variation features of load, photovoltaic output, charging load, and temperature. The output layer outputs the predicted voltage values ​​of each node within the sampling period.

[0023] Furthermore, the low voltage cause vector Represented as:

[0024] ;

[0025] in, This indicates the quantity causing the heavy load on the distribution transformer. This indicates the causes of line voltage drop. This indicates the quantity that causes the three-phase imbalance. Indicates the quantity of cause of insufficient work. Indicates the quantity that causes source load fluctuations;

[0026] The formula for generating the low voltage cause vector based on the hidden state of the node is as follows:

[0027] ;

[0028] in, The output of the spatiotemporal graph neural network represents the first... The hidden state of each node This represents the causal diagnosis weight matrix. Represents the causal diagnosis bias vector;

[0029] After the step of generating the low voltage cause vector based on the implicit state of the node, the method further includes: correcting the low voltage cause vector using regular cause factors, which include: line voltage drop cause factor, three-phase imbalance cause factor, reactive power deficiency cause factor, transformer overload cause factor and source load fluctuation cause factor.

[0030] Furthermore, the step of constructing and online correcting the voltage sensitivity matrix based on the node voltage changes before and after the governance action specifically includes:

[0031] A voltage sensitivity matrix is ​​established based on line parameters, historical control actions, and real-time management action feedback.

[0032] ;

[0033] in, Indicates the first The voltage change of each node before and after the governance action. Indicates the first Changes in the amount of governance resource actions;

[0034] The pressure sensitivity matrix is ​​corrected using the following online correction formula:

[0035] ;

[0036] in, Indicates the sensitivity learning rate. This indicates a small positive quantity that prevents the denominator from being zero. and These represent the node voltages before and after the treatment action. and These represent the amount of governance resources used before and after the governance action.

[0037] Furthermore, determining candidate governance resources based on the low voltage cause vector and voltage sensitivity matrix specifically includes:

[0038] When the cause of distribution transformer overload is the largest, priority should be given to energy storage discharge, delayed start of charging load, adjustable load reduction and distribution transformer voltage regulation measures.

[0039] When the cause of line voltage drop is the largest, priority should be given to branch voltage regulation equipment, end voltage management equipment, end energy storage or distributed reactive power compensation equipment that are closer to low voltage nodes.

[0040] When the three-phase imbalance is at its maximum, priority should be given to using commutation equipment or interphase power transfer equipment to transfer part of the load on the heavily loaded phase to the lightly loaded phase.

[0041] When the cause of insufficient reactive power is the largest, SVG, smart capacitors, or photovoltaic inverters with reactive power regulation capabilities should be given priority.

[0042] When the source-load fluctuation causes are the greatest, priority should be given to energy storage for early charging and discharging, photovoltaic inverter reactive power support, and adjustable load transfer.

[0043] Furthermore, the construction of the multi-resource collaborative governance optimization model based on the low-voltage cause vector, voltage sensitivity matrix, and device constraints specifically includes:

[0044] Construct a governance resource action vector, the formula is:

[0045]

[0046] in, This indicates the amount of change in the voltage adjustment level. This indicates the reactive power output of the SVG. Indicates the energy storage charging and discharging power. Indicates the switching capacity of the capacitor. Represents the commutation action variable. Indicates the adjustable load adjustment amount;

[0047] The formula for the multi-resource collaborative governance optimization model is:

[0048]

[0049] in, Indicates the voltage deviation target. This indicates the target for low-voltage over-limit penalty. Indicates the target cost of equipment operation. Indicates the target number of equipment actions. This indicates that the energy storage state of charge deviates from the target. Indicates a three-phase unbalanced target. , , , , and These correspond to the weighting coefficients for voltage deviation target, low voltage over-limit penalty target, equipment operation cost target, equipment operation frequency target, energy storage state of charge deviation target, and three-phase imbalance target, respectively.

[0050] Furthermore, the specific method for obtaining the collaborative governance strategy by solving the multi-resource collaborative governance optimization model using the improved sparrow search algorithm includes:

[0051] By mixing and encoding the voltage regulation level, reactive power compensation capacity, energy storage charging and discharging power, commutation action variables, photovoltaic inverter reactive power output and adjustable load regulation, individual candidate governance strategies are obtained.

[0052] The sparrow population is initialized in a targeted manner based on the low voltage cause vector, so that the population is preferentially distributed in the governance resource action space corresponding to the main low voltage cause.

[0053] Calculate the fitness value of each individual sparrow and divide the sparrow population into producers, followers, and vigilants based on the fitness value;

[0054] During the iteration process, the positions of producers, followers and vigilant are updated respectively, and dynamic safety threshold, elite reverse learning and constraint penalty mechanism are introduced to perturb the search for the current optimal governance strategy.

[0055] Sigmoid discrete mapping is performed on the discrete variables corresponding to voltage regulation level, capacitor switching and commutation action;

[0056] When the maximum number of iterations is reached or the change in fitness is less than a set threshold, the optimal collaborative governance strategy is output.

[0057] Secondly, another embodiment of the present invention provides a low-voltage management system for distribution substations based on a spatiotemporal graph neural network and improved sparrow search, used to implement the method described in the first embodiment above, including:

[0058] The graph structure building module is used to construct the distribution radio area into a graph structure containing nodes, edges, and adjacency matrices;

[0059] The data acquisition module is used to collect multi-source operational data from each node in the distribution area and construct node feature vectors;

[0060] The low voltage prediction module is used to input the feature sequences and their adjacency matrices of all nodes within a continuous sampling period into the spatiotemporal graph neural network model, output the voltage prediction value and hidden state of each node in the future prediction time domain, and generate a low voltage cause vector based on the hidden state of the nodes.

[0061] The sensitivity calculation module is used to construct and correct the voltage sensitivity matrix online based on the node voltage change before and after the governance action, and to determine candidate governance resources based on the low voltage cause vector and the voltage sensitivity matrix.

[0062] The optimization model building module constructs a multi-resource collaborative governance optimization model based on the low voltage cause vector, voltage sensitivity matrix, and equipment constraints.

[0063] The optimization model solving module uses an improved sparrow search algorithm to solve the multi-resource collaborative governance optimization model to obtain a collaborative governance strategy.

[0064] The instruction execution module is used to issue collaborative governance instructions to the corresponding governance resources according to the collaborative governance strategy;

[0065] The correction module is used to correct the voltage sensitivity matrix and the spatiotemporal neural network model based on the measured voltage response after the governance action.

[0066] Thirdly, another embodiment of the present invention provides an electronic device comprising: a processor, an input device, an output device, and a memory, wherein the processor, the input device, the output device, and the memory are interconnected, the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to invoke the program instructions to execute the method described in the first embodiment above.

[0067] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0068] This invention provides a method, system, and device for managing low voltage in distribution substations based on a spatiotemporal graph neural network and improved sparrow search. The distribution substation is abstracted as a graph structure, and the spatiotemporal graph neural network simultaneously extracts the topological spatial correlation features and load temporal variation features of the substation, enabling prediction of future short-term low voltage risks and effectively shortening the duration of low voltage. Low voltage causes are quantified into multiple causal quantities such as transformer overload, line voltage drop, three-phase imbalance, insufficient reactive power, and source-load fluctuations through a low voltage cause vector. Combined with a voltage sensitivity matrix, the improvement effect of different management resources on each node is evaluated, accurately identifying the dominant causes of low voltage and prioritizing targeted management resources, avoiding blind voltage regulation based solely on voltage exceeding limits. The voltage sensitivity matrix assesses the impact of different management resources on the voltage of different nodes, prioritizing the management resources that maximize the improvement effect on the target low-voltage node while minimizing disturbance to other nodes. A multi-objective optimization model is used to collaboratively control voltage regulation, reactive power, energy storage, commutation, photovoltaic inverters, and adjustable loads, reducing the risk of excessive and frequent operation of individual devices. Employing an improved sparrow search algorithm, this method incorporates low-voltage cause vector-oriented initialization, dynamic safety thresholds, elite back-learning, and discrete mapping mechanisms. It efficiently handles mixed-integer optimization scenarios where discrete variables such as voltage regulation levels, capacitor switching, and commutation actions coexist with continuous variables such as energy storage power and reactive power output. This results in fast convergence and high-quality feasible solutions. It is suitable for mixed-integer optimization scenarios in low-voltage management of distribution substations. Furthermore, it can online correct the voltage sensitivity matrix and prediction model based on the measured voltage response after management actions, enabling the management strategy to adapt to dynamic changes in substation topology, line parameters, and load structure. Attached Figure Description

[0069] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0070] Figure 1A flowchart of a low-voltage management method for distribution substations based on spatiotemporal graph neural networks and improved sparrow search provided in an embodiment of the present invention;

[0071] Figure 2 The convergence curve of the improved sparrow search algorithm in this embodiment of the invention is shown.

[0072] Figure 3 This is a comparison diagram of node voltage before and after treatment according to an embodiment of the present invention;

[0073] Figure 4 This is a bar chart showing the effect of low-voltage node mitigation in an embodiment of the present invention.

[0074] Figure 5 The diagram below shows a structural block diagram of a low-voltage management system for a distribution substation based on a spatiotemporal graph neural network and an improved sparrow search, as provided in another embodiment of the present invention. Detailed Implementation

[0075] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0076] like Figure 1 As shown, the first embodiment of the present invention provides a method for low voltage management of distribution substations based on spatiotemporal graph neural networks and improved sparrow search, applicable to a low voltage management system for distribution substations based on spatiotemporal graph neural networks and improved sparrow search provided in this embodiment of the present invention. The method includes:

[0077] S1. Construct the distribution radio area into a graph structure containing nodes, edges, and an adjacency matrix.

[0078] In some embodiments, the specific process of step S1 is as follows:

[0079] S11. First, abstract the distribution radio station area into a graph structure. In this system, the low-voltage side of the transformer, the beginning and end of the branch, the meter box, the user terminal, the photovoltaic inverter access point, the energy storage access point, the reactive power compensation equipment access point, and the voltage regulation equipment access point are all considered as nodes; the low-voltage lines between nodes are considered as edges; whether there is an electrical connection between nodes is determined by the adjacency matrix. express.

[0080] S12. In some embodiments, if node With nodes If there is a direct electrical connection between adjacent matrix elements, then... If there is no direct electrical connection, then In other embodiments, the adjacency matrix elements may also be determined by weighting line impedance, line length, or node electrical distance, for example:

[0081] ,

[0082] in, Represents a node With nodes The equivalent line impedance between them This represents the attenuation coefficient.

[0083] S2. Collect multi-source operational data from each node in the distribution area and construct node feature vectors.

[0084] The specific process of step S2 is as follows:

[0085] S21, Regarding the first Each node, the system in Collect and construct node feature vectors at all times:

[0086]

[0087] in, , , Let be the three-phase voltages at time t. Let be the current at time t. Let be the active power and reactive power at time t, respectively. Let be the power factor at time t. Let t be the state of charge of the energy storage. For the photovoltaic output at time t, Let be the charging load at time t. Let be the temperature at time t. The date type is at time t;

[0088] S22. For nodes lacking a certain type of data, the system supplements the data using zero values, estimated values ​​from adjacent nodes, or historical average values ​​for the same period. For example, for nodes not connected to energy storage, their... It can be set to 0; for nodes not connected to photovoltaics, its It can be set to 0.

[0089] S23. Perform outlier removal and normalization on the collected data. For continuous variables, the following normalization method can be used:

[0090] ,

[0091] in, Represents the original feature values. and These represent the minimum and maximum values ​​of the corresponding features in the sample set, respectively.

[0092] S3. Input the feature sequences and adjacency matrices of all nodes within the continuous sampling period into the spatiotemporal graph neural network model, output the voltage prediction values ​​and hidden states of each node in the future prediction time domain, and generate a low voltage cause vector based on the hidden states of the nodes.

[0093] The specific process of step S3 is as follows:

[0094] S31, will continue The input matrix is ​​composed of the feature sequences of all nodes within each sampling period. and its adjacency matrix The inputs are fed into the spatiotemporal graph neural network model. The spatiotemporal graph neural network model includes graph convolutional layers, time-series feature extraction layers, and an output layer. The graph convolutional layers are used to extract spatial correlation features between nodes in the transformer area; their calculation process is as follows:

[0095] ,

[0096] in, , It is the identity matrix. for The corresponding degree matrix, Indicates the first Layer node features Indicates the first Layer training weights, This represents the activation function.

[0097] S32. The time series feature extraction layer can employ a GRU, LSTM, or Transformer structure to extract the time-series variation features of load, photovoltaic output, charging load, and temperature. This embodiment uses a GRU, and its implicit state update process can be represented as:

[0098]

[0099]

[0100]

[0101]

[0102] in, Indicates an update to the door. This indicates that the door is being reset. express The state is implied at all times. This indicates element-wise multiplication.

[0103] S33, Spatiotemporal graph neural network model outputs the future Predicted voltage values ​​of each node within each sampling period:

[0104]

[0105] in, Indicates the prediction time domain, Indicates the duration of historical input. This represents the trained spatiotemporal graph neural network model. Indicates model parameters, Indicates continuity Node feature sequence within a sampling period.

[0106] During the training phase, the model's loss function includes voltage prediction error, low-voltage risk classification loss, and cause diagnosis loss:

[0107]

[0108] in, Indicates the number of nodes. Indicates the prediction time domain, and These are the weighting coefficients. Losses are classified as low-voltage risk. To diagnose the cause of the loss, This represents the voltage prediction error.

[0109] S34, Regarding the Future Time of the first Each node calculates the low-voltage risk probability based on the predicted voltage. The calculation formula is:

[0110] ,

[0111] in, Indicates the lower limit of voltage. Represents the smoothing coefficient, when Greater than the preset risk threshold At that time, the judgment of the first Each node in the future There is always a risk of low voltage. If multiple adjacent nodes are at risk of low voltage simultaneously, the risk area and risk propagation path are further determined based on the transformer substation topology.

[0112] S35, Low Voltage Cause Vector Represented as:

[0113] ;

[0114] in, This indicates the quantity causing the heavy load on the distribution transformer. This indicates the causes of line voltage drop. This indicates the quantity that causes the three-phase imbalance. Indicates the quantity of cause of insufficient work. This represents the quantity that causes source load fluctuations.

[0115] Hidden states of nodes output by spatiotemporal graph neural network model Generate the low voltage cause vector:

[0116] ,

[0117] in, The output of the spatiotemporal graph neural network represents the first... The hidden state of each node This represents the causal diagnosis weight matrix. This represents the causal diagnosis bias vector.

[0118] S36. To improve the interpretability of diagnostic results, the output of the artificial intelligence model is further modified using rule-based causal factors. The modified causal vector... The calculation formula is:

[0119] ,

[0120] in, Indicates the fusion weight. This represents the rule-causing vector composed of rule-causing factors.

[0121] The regular causal factors include: line voltage drop causal factors, three-phase imbalance causal factors, reactive power insufficiency causal factors, transformer overload causal factors, and source load fluctuation causal factors.

[0122] Factors causing line voltage drop The calculation formula is:

[0123] ,

[0124] Triphase imbalance causative factors The calculation formula is:

[0125] ,

[0126] Factors contributing to insufficient work The calculation formula is:

[0127] ,

[0128] Factors contributing to overloading of distribution transformers The calculation formula is:

[0129] ,

[0130] Source load fluctuation cause factor The calculation formula is:

[0131] ,

[0132] in, Indicates the first The voltage of the upstream node of each node, Indicates the rated voltage. Indicates the rated current. This indicates the real-time capacity load of the distribution transformer. Indicates the rated capacity of the distribution transformer. This indicates the low-voltage side voltage of the transformer. Indicates the sampling interval.

[0133] S4. Construct and correct the voltage sensitivity matrix online based on the node voltage changes before and after the governance action, and determine candidate governance resources based on the low voltage cause vector and the voltage sensitivity matrix.

[0134] The specific process of step S4 is as follows:

[0135] S41. Establish a voltage sensitivity matrix based on line parameters, historical control actions, and real-time management action feedback:

[0136] ,

[0137] in, Indicates the first The voltage change of each node before and after the governance action. Indicates the first Changes in the amount of governance resources involved.

[0138] S42. In actual operation, to avoid calculation fluctuations caused by excessively small action amounts, the following online correction formula is adopted:

[0139] ,

[0140] in, Indicates the sensitivity learning rate. This indicates a small positive quantity that prevents the denominator from being zero. and These represent the node voltages before and after the treatment action. and These represent the amount of governance resources used before and after the governance action.

[0141] S43. Determine candidate governance resources based on the low voltage cause vector and voltage sensitivity matrix. When the distribution transformer overload cause is the largest, prioritize governance measures such as energy storage discharge, delayed charging load start-up, adjustable load reduction, and distribution transformer voltage regulation; when the line voltage drop cause is the largest, prioritize branch voltage regulation equipment, end-point voltage governance equipment, end-point energy storage, or distributed reactive power compensation equipment located close to the low voltage node; when the three-phase imbalance cause is the largest, prioritize commutation equipment or inter-phase power transfer equipment to transfer part of the load on the heavily loaded phase to the lightly loaded phase; when the reactive power deficiency cause is the largest, prioritize SVG, smart capacitors, or photovoltaic inverters with reactive power regulation capabilities; when the source-load fluctuation cause is the largest, prioritize early charging and discharging of energy storage, reactive power support from photovoltaic inverters, and adjustable load transfer.

[0142] S5. Based on the low voltage cause vector, voltage sensitivity matrix and equipment constraints, a multi-resource collaborative governance optimization model is constructed. An improved sparrow search algorithm is used to solve the multi-resource collaborative governance optimization model to obtain the collaborative governance strategy.

[0143] The specific process of step S5 is as follows:

[0144] S51. Construct governance resource action vectors:

[0145] ,

[0146] in, This indicates the amount of change in the voltage adjustment level. This indicates the reactive power output of the SVG. Indicates the energy storage charging and discharging power. Indicates the switching capacity of the capacitor. Represents the commutation action variable. This indicates the adjustable load adjustment amount.

[0147] In the prediction time domain, the node voltage response can be expressed as:

[0148] ,

[0149] in, This represents the disturbance term caused by load fluctuations, photovoltaic fluctuations, and forecast errors. Represents the voltage sensitivity matrix. This indicates the amount of governance resources used.

[0150] S52. Construct a multi-objective optimization function:

[0151]

[0152] in, , , , , and These correspond to the weighting coefficients for voltage deviation target, low voltage over-limit penalty target, equipment action cost target, equipment action number target, energy storage state of charge deviation target, and three-phase imbalance target, respectively. Indicates the voltage deviation target. This indicates the target for low voltage over-limit penalty. Indicates the target cost of equipment operation. Indicates the target number of equipment actions. This indicates that the energy storage state of charge deviates from the target. This indicates a three-phase unbalanced target.

[0153] Voltage deviation target The calculation formula is:

[0154] ,

[0155] Low voltage over-limit penalty target The calculation formula is:

[0156] ,

[0157] Equipment operating cost target The calculation formula is:

[0158] ,

[0159] Equipment action count target The calculation formula is:

[0160] ,

[0161] Energy storage state of charge deviates from target The calculation formula is:

[0162] ,

[0163] Three-phase imbalance target The calculation formula is:

[0164] .

[0165] The multi-resource collaborative governance optimization model satisfies the following constraints:

[0166]

[0167]

[0168]

[0169]

[0170]

[0171] in, Indicates the upper limit of voltage. and These represent the lower and upper limits of governance resource actions, respectively. and These represent the lower and upper limits of the energy storage state of charge, respectively. Indicates the energy storage discharge power. Indicates the energy storage charging power. Indicates discharge efficiency. Indicates charging efficiency. Indicates the rated capacity of energy storage. Indicates governance resources during the statistical period Number of actions within, This indicates the maximum number of actions allowed.

[0172] Through the above objective function, the system can ensure that the low voltage node recovers to the qualified range while reducing equipment operation costs, number of operations, and energy storage life loss.

[0173] The multi-resource collaborative governance optimization model also satisfies voltage upper and lower limit constraints, equipment action boundary constraints, energy storage state of charge constraints, equipment action frequency constraints, and three-phase imbalance constraints.

[0174] S53. Since the multi-resource collaborative governance optimization model includes both continuous variables such as energy storage power and SVG reactive power output, as well as discrete variables such as voltage regulation level, capacitor switching and commutation actions, and the constraints include voltage upper and lower limits, energy storage state of charge, number of equipment actions and three-phase imbalance, this embodiment adopts an improved sparrow search algorithm for solving the problem. This algorithm introduces low voltage cause vector oriented initialization, dynamic safety threshold, elite back learning, mixed integer discrete mapping and constraint penalty mechanism on the basis of traditional sparrow search algorithm to improve the search efficiency and feasible solution quality in complex low voltage governance scenarios.

[0175] Each individual sparrow represents a candidate governance strategy:

[0176] ,

[0177] in, This indicates the individual sparrow's ID number. Indicates the first One candidate collaborative governance strategy.

[0178] The population matrix is ​​represented as:

[0179] ,

[0180] in, Indicates population size, This indicates optimization of variable dimensions.

[0181] The fitness function is expressed as:

[0182] ,

[0183] in, This indicates the penalty for exceeding voltage limits. This indicates the penalty for exceeding the state of charge limit of energy storage. This indicates penalties for frequent device actions. This indicates the penalty for exceeding the limit due to three-phase imbalance. , , and These correspond to the penalty coefficients for voltage over-limit penalty items, energy storage state of charge over-limit penalty items, equipment frequent operation penalty items, and three-phase imbalance over-limit penalty items.

[0184] S54. To make the initial solution closer to the feasible governance region, this embodiment performs targeted initialization of the population based on the low voltage cause vector. The initial position is represented as:

[0185] ,

[0186] in, Indicates the first The first individual sparrow The initial position of the dimensional variable. and They represent the first Upper and lower bounds of a dimensional variable A random number between 0 and 1 Indicates the directional initial strength coefficient. The first term obtained by mapping low voltage cause vectors represents the... Maintenance resource bias. Through this setting, when the cause of line voltage drop is large, the population will preferentially expand towards branch voltage regulation, end energy storage and local reactive power compensation variables; when the cause of three-phase imbalance is large, the population will preferentially expand towards commutation variables; when the cause of reactive power deficiency is large, the population will preferentially expand towards SVG and capacitor reactive power output variables.

[0187] S55. In each iteration, the sparrow population is divided into producers, followers, and vigilant individuals based on their fitness values. To balance global search and local exploitation, a dynamic safety threshold is set:

[0188] ,

[0189] in, Indicates the first The safety threshold at the next iteration and These represent the lower and upper limits of the safety threshold, respectively. The producer position is updated as follows:

[0190] ,

[0191] in, Indicates the current iteration number. Indicates the maximum number of iterations. It is a random number. Indicates the warning value. Indicates the dynamic security threshold. Represents a random number that follows a normal distribution. This represents a vector whose elements are all 1s.

[0192] Follower position updated to:

[0193] ,

[0194] in, Indicates the first The th iteration of the globally optimal individual Dimensional position, Indicates the first The worst individual in the next iteration Dimensional position, Representation matrix The generalized inverse.

[0195] The location of the vigilant has been updated to:

[0196] ,

[0197] in, Indicates the first The fitness value of an individual sparrow. This represents the globally optimal fitness value. This represents the worst global fitness value. For random numbers that follow a normal distribution, A random number in the range [-1, 1]. To avoid constants with a denominator of zero. For example... Figure 2 As shown, the convergence curve of the improved sparrow search algorithm is illustrated.

[0198] S56. To improve local development capabilities, this embodiment performs reverse learning perturbation on the current elite individual:

[0199] ,

[0200] in, This represents the first [unit / item] after reverse learning by elite individuals. Dimensional position, Indicates the current elite individual's... Dimensional position, The disturbance coefficient is in the range [-0.5, 0.5].

[0201] S57. For discrete variables such as voltage regulation level, capacitor switching, and commutation operation, the system uses the Sigmoid function for discrete mapping:

[0202] ,

[0203] in, Indicates the first A discrete action variable, A random number between 0 and 1.

[0204] Through the aforementioned targeted initialization, producer / follower / vigilant collaborative search, elite reverse learning, and discrete mapping, the embodiments of the present invention can quickly obtain a collaborative governance strategy that satisfies voltage, safety, and device constraints in mixed integer variable scenarios.

[0205] S6. Issue collaborative governance instructions to the corresponding governance resources according to the collaborative governance strategy, and correct the voltage sensitivity matrix and spatiotemporal neural network model according to the measured voltage response after the governance action.

[0206] The specific process of step S6 is as follows:

[0207] S61. After obtaining the optimal governance strategy, issue coordinated governance instructions to the corresponding governance resources. These instructions include one or more of the following: voltage regulator speed adjustment, SVG reactive power output, energy storage charging / discharging power, capacitor switching capacity, commutation operation instructions, photovoltaic inverter reactive power regulation, and adjustable load regulation. For example... Figure 3 As shown, a voltage comparison diagram of the node before and after the treatment is presented. From Figure 3 It can be seen that after implementing the above collaborative governance strategy, the voltage of each low-voltage node was effectively raised, so that all nodes were restored to the qualified range after governance, and no voltage exceeded the upper limit.

[0208] S62. After the governance action is executed, the actual voltage response of each node is collected, and the prediction error is calculated. :

[0209] ,

[0210] in, This represents the actual voltage of node i at time t. Represents the predicted voltage of node i at time t, when When the error exceeds the set error threshold, the operational data before and after this treatment are added to the training sample library, and the spatiotemporal graph neural network model is incrementally trained according to the set period. Incremental training loss function. Represented as:

[0211] ,

[0212] in, This represents the weight coefficient of the newly added samples. This is the loss function without incremental training.

[0213] Figure 4 This is a bar chart showing the effect of the low-voltage node mitigation method of the present invention. From... Figure 4 As can be seen, after the collaborative governance strategy described in this embodiment was implemented, the voltage of each low-voltage node was restored to the qualified range, and the governance effect was significant.

[0214] The above methods can continuously improve the accuracy of low voltage risk prediction, cause diagnosis, and collaborative governance strategy generation as field operation data accumulates.

[0215] This invention provides a method for managing low voltage in distribution transformer areas based on a spatiotemporal graph neural network and an improved sparrow search. The method abstracts the distribution transformer area into a graph structure and utilizes the spatiotemporal graph neural network to simultaneously extract the topological spatial correlation features and load temporal variation features of the transformer area, enabling prediction of future short-term low voltage risks. By using a low voltage cause vector, the causes of low voltage are quantified into multiple causal quantities such as transformer overload, line voltage drop, three-phase imbalance, insufficient reactive power, and source-load fluctuations, avoiding blind voltage regulation based solely on voltage exceeding limits. A voltage sensitivity matrix is ​​used to evaluate the impact of different management resources on the voltage of different nodes, allowing priority to be given to management resources that maximize the improvement effect on the target low-voltage node while minimizing disturbance to other nodes. A multi-objective optimization model is employed to collaboratively control voltage regulation, reactive power, energy storage, commutation, photovoltaic inverters, and adjustable loads, reducing the risk of excessive and frequent operation of individual equipment. An improved sparrow search algorithm is employed to handle discrete variables such as voltage regulation levels, capacitor switching, and commutation actions, as well as continuous variables such as energy storage power and SVG reactive power output. By employing low-voltage cause vector-oriented initialization, dynamic safety thresholds, elite back-learning, and constraint penalty mechanisms, the algorithm's convergence speed and feasible solution search capability are enhanced, making it suitable for mixed-integer optimization scenarios in low-voltage management of distribution substations. It can online correct the voltage sensitivity matrix and prediction model based on the measured voltage response after management actions, enabling the management strategy to adapt to dynamic changes in substation topology, line parameters, and load structure.

[0216] like Figure 5 According to another embodiment of the present invention, a low-voltage intelligent management system for distribution substations based on spatiotemporal graph neural networks and improved sparrow search is provided, comprising:

[0217] The topology modeling module is used to construct the electrical topology diagram of the transformer substation based on the connection relationships between each monitoring node and line within the substation.

[0218] The data acquisition module is used to collect multi-source operational data from the low-voltage side of the transformer, branch lines, meter boxes, user terminals, distributed photovoltaic systems, energy storage devices, reactive power compensation devices, voltage regulation devices, commutation devices, and charging loads.

[0219] The low voltage prediction module is used to predict the voltage value and low voltage risk probability of each node in the future prediction time domain based on the node feature sequence and the transformer area adjacency matrix using a spatiotemporal graph neural network model.

[0220] The cause diagnosis module is used to generate low voltage cause vectors based on the hidden state of nodes and regular cause factors;

[0221] The sensitivity calculation module is used to construct and correct the voltage sensitivity matrix based on the changes in node voltage before and after the governance action;

[0222] The collaborative optimization module is used to construct a multi-resource collaborative governance optimization model based on the low voltage cause vector, voltage sensitivity matrix, and equipment constraints.

[0223] The algorithm solution module is used to solve the multi-resource collaborative governance optimization model using an improved sparrow search algorithm;

[0224] The instruction execution module is used to send collaborative management instructions to voltage regulation equipment, reactive power compensation equipment, energy storage equipment, commutation equipment, photovoltaic inverters, and adjustable load control equipment.

[0225] The online correction module is used to correct the voltage sensitivity matrix and the spatiotemporal neural network model based on the measured voltage response after the governance action.

[0226] The execution process of each module can be carried out according to the steps of the low voltage management method for distribution substations based on spatiotemporal graph neural network and improved sparrow search provided in the first embodiment, and will not be described in detail in this embodiment.

[0227] The modules described above can be integrated into the smart terminal of the distribution area, or deployed on a computing terminal, distribution automation master station, or collaborative control platform. Each module can be implemented through software, hardware, or a combination of both.

[0228] For example, a rural power distribution area includes a 400kVA distribution transformer with three branch lines on the low-voltage side. The third branch line has a longer power supply radius, connecting multiple residential users, electric vehicle charging stations, and distributed photovoltaic systems at its end. The area is equipped with a 100kW / 200kWh energy storage device, an SVG reactive power compensation device, several smart capacitors, and smart phase-switching switches.

[0229] During the summer evening peak hours, the system predicted that multiple users on phase A at the end of the third branch would experience a low voltage risk within the next 20 minutes. After obtaining the low-voltage side voltage of the transformer, the system found that the transformer's output voltage was basically normal. However, after obtaining the voltages at the beginning and end of the third branch, it was found that the voltage of phase A at the end was significantly lower than that at the beginning, while the current of phase A was significantly higher than that of phases B and C, and the power factor of the transformer area was lower than the set threshold.

[0230] The feature sequences and adjacency matrices of all nodes within a continuous sampling period are input into a spatiotemporal graph neural network model. The model outputs the predicted voltage values ​​and hidden states of each node in the future prediction time domain. A low-voltage cause vector is generated based on the hidden states, indicating that the low voltage is primarily caused by line voltage drop and three-phase imbalance, with insufficient reactive power being a secondary factor. Based on the voltage sensitivity matrix, energy storage discharge near the end of the third branch, SVG reactive power output, and intelligent commutation actions are found to have good effects on improving the voltage of the target node. However, overall voltage regulation of the transformer may lead to excessively high voltage at the first-end users, therefore these are not considered priority mitigation measures.

[0231] An improved sparrow search algorithm is used to solve the multi-resource collaborative governance optimization model, resulting in the following collaborative governance strategy: control the intelligent phase-switching switch to transfer part of the load from phase A to phase B; control the reactive power output of the SVG; control the energy storage to discharge at a set power during the evening peak period; if the terminal voltage is still lower than the lower voltage limit, control the branch voltage regulating equipment to perform a slight voltage boost.

[0232] After implementing the above governance strategy, the voltage of phase A at the end of the third branch returned to the acceptable range, while the voltages of phases B and C did not exceed the upper voltage limit, and the three-phase imbalance was reduced. The system records the voltage response before and after this action and updates the voltage sensitivity coefficients of the energy storage, SVG, and commutation devices to the end node of the third branch online.

[0233] The low voltage management system for distribution substations based on spatiotemporal graph neural networks and improved sparrow search provided in this invention embodiment and the low voltage management method for distribution substations based on spatiotemporal graph neural networks and improved sparrow search are based on the same inventive concept and have the same beneficial effects, and will not be described again here.

[0234] Another embodiment of the present invention provides an electronic device, which includes a processor, an input device, an output device, and a memory. The processor, the input device, the output device, and the memory are interconnected. The memory is used to store a computer program, which includes program instructions. The processor is configured to call the program instructions to execute the method described in the first embodiment above.

[0235] It should be understood that, in the embodiments of the present invention, the processor may be a Central Processing Unit (CPU), but it may also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0236] Input devices may include touchpads, microphones, etc., while output devices may include displays (LCDs, etc.), speakers, etc.

[0237] The memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of the memory may also include non-volatile random access memory. For example, the memory may also store information about the device type.

[0238] In specific implementations, the processor, input device, and output device described in the embodiments of the present invention can execute the implementation of the method embodiments described in the embodiments of the present invention, or they can execute the implementation of the system embodiments described in the embodiments of the present invention, which will not be repeated here.

[0239] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A method for managing low voltage in distribution substations based on spatiotemporal graph neural networks and improved sparrow search, characterized in that, include: The distribution radio area is constructed as a graph structure containing nodes, edges, and an adjacency matrix; Collect multi-source operational data from each node in the distribution area and construct node feature vectors; The feature sequences and adjacency matrices of all nodes within a continuous sampling period are input into the spatiotemporal graph neural network model, which outputs the voltage prediction values ​​and hidden states of each node in the future prediction time domain, and generates a low voltage cause vector based on the hidden states of the nodes. The voltage sensitivity matrix is ​​constructed and corrected online based on the node voltage changes before and after the governance action, and candidate governance resources are determined based on the low voltage cause vector and the voltage sensitivity matrix. A multi-resource collaborative governance optimization model is constructed based on the low voltage cause vector, voltage sensitivity matrix and equipment constraints. An improved sparrow search algorithm is used to solve the multi-resource collaborative governance optimization model to obtain the collaborative governance strategy. Based on the collaborative governance strategy, collaborative governance instructions are issued to the corresponding governance resources, and the voltage sensitivity matrix and spatiotemporal neural network model are corrected based on the measured voltage response after the governance action.

2. The method according to claim 1, characterized in that, The construction of the distribution radio area into a graph structure containing nodes, edges, and an adjacency matrix specifically includes: The low-voltage side of the transformer, the beginning and end of the branch, the meter box, the user terminal, the photovoltaic inverter access point, the energy storage access point, the reactive power compensation equipment access point, and the voltage regulation equipment access point are all regarded as nodes, and the low-voltage lines between the nodes are regarded as edges. The electrical connection relationship between the nodes is represented by the adjacency matrix A. If node With nodes If there is a direct electrical connection between adjacent matrix elements, then... If there is no direct electrical connection, then Alternatively, the adjacency matrix elements are determined by a weighted average of line impedance, line length, or node electrical distance, calculated using the following formula: ； in, Represents a node With nodes The equivalent line impedance between them This represents the attenuation coefficient.

3. The method according to claim 1, characterized in that, The spatiotemporal graph neural network model includes a graph convolutional layer, a time series feature extraction layer, and an output layer. The graph convolutional layer is used to extract the spatial correlation features between nodes in the power distribution area. The time series feature extraction layer is used to extract the temporal variation features of load, photovoltaic output, charging load, and temperature. The output layer outputs the predicted voltage values ​​of each node within the sampling period.

4. The method according to claim 1, characterized in that, The low voltage cause vector Represented as: ； in, This indicates the quantity causing the heavy load on the distribution transformer. This indicates the causes of line voltage drop. This indicates the quantity that causes the three-phase imbalance. Indicates the quantity of cause of insufficient work. Indicates the quantity that causes source load fluctuations; The formula for generating the low voltage cause vector based on the hidden state of the node is as follows: ； in, The output of the spatiotemporal graph neural network represents the first... The hidden state of each node This represents the causal diagnosis weight matrix. Represents the causal diagnosis bias vector; After the step of generating the low voltage cause vector based on the implicit state of the node, the method further includes: correcting the low voltage cause vector using regular cause factors, which include: line voltage drop cause factor, three-phase imbalance cause factor, reactive power deficiency cause factor, transformer overload cause factor and source load fluctuation cause factor.

5. The method according to claim 4, characterized in that, The specific steps of constructing and online correcting the voltage sensitivity matrix based on the node voltage changes before and after the governance action include: A voltage sensitivity matrix is ​​established based on line parameters, historical control actions, and real-time management action feedback. ； in, Indicates the first The voltage change of each node before and after the governance action. Indicates the first Changes in the amount of governance resource actions; The pressure sensitivity matrix is ​​corrected using the following online correction formula: ； in, Indicates the sensitivity learning rate. This indicates a small positive quantity that prevents the denominator from being zero. and These represent the node voltages before and after the treatment action. and These represent the amount of governance resources used before and after the governance action.

6. The method according to claim 5, characterized in that, Determining candidate governance resources based on the low voltage cause vector and voltage sensitivity matrix specifically includes: When the cause of distribution transformer overload is the largest, priority should be given to energy storage discharge, delayed start of charging load, adjustable load reduction and distribution transformer voltage regulation measures. When the cause of line voltage drop is the largest, priority should be given to branch voltage regulation equipment, end voltage management equipment, end energy storage or distributed reactive power compensation equipment that are closer to low voltage nodes. When the three-phase imbalance is at its maximum, priority should be given to using commutation equipment or interphase power transfer equipment to transfer part of the load on the heavily loaded phase to the lightly loaded phase. When the cause of insufficient reactive power is the largest, SVG, smart capacitors, or photovoltaic inverters with reactive power regulation capabilities should be given priority. When the source-load fluctuation causes are the greatest, priority should be given to energy storage for early charging and discharging, photovoltaic inverter reactive power support, and adjustable load transfer.

7. The method according to claim 1, characterized in that, The construction of the multi-resource collaborative governance optimization model based on the low voltage cause vector, voltage sensitivity matrix, and equipment constraints specifically includes: Construct a governance resource action vector, the formula is: ； in, This indicates the amount of change in the voltage adjustment level. This indicates the reactive power output of the SVG. Indicates the energy storage charging and discharging power. Indicates the switching capacity of the capacitor. Represents the commutation action variable. Indicates the adjustable load adjustment amount; The formula for the multi-resource collaborative governance optimization model is: ； in, Indicates the voltage deviation target. This indicates the target for low voltage over-limit penalty. Indicates the target cost of equipment operation. Indicates the target number of equipment actions. This indicates that the energy storage state of charge deviates from the target. Indicates a three-phase unbalanced target. , , , , and These correspond to the weighting coefficients for voltage deviation target, low voltage over-limit penalty target, equipment operation cost target, equipment operation frequency target, energy storage state of charge deviation target, and three-phase imbalance target, respectively.

8. The method according to claim 7, characterized in that, The specific method for obtaining the collaborative governance strategy by solving the multi-resource collaborative governance optimization model using the improved sparrow search algorithm includes: By mixing and encoding the voltage regulation level, reactive power compensation capacity, energy storage charging and discharging power, commutation action variables, photovoltaic inverter reactive power output and adjustable load regulation, individual candidate governance strategies are obtained. The sparrow population is initialized in a targeted manner based on the low voltage cause vector, so that the population is preferentially distributed in the governance resource action space corresponding to the main low voltage cause. Calculate the fitness value of each individual sparrow and divide the sparrow population into producers, followers, and vigilants based on the fitness value; During the iteration process, the positions of producers, followers and vigilant are updated respectively, and dynamic safety threshold, elite reverse learning and constraint penalty mechanism are introduced to perturb the search for the current optimal governance strategy. Sigmoid discrete mapping is performed on the discrete variables corresponding to voltage regulation level, capacitor switching and commutation action; When the maximum number of iterations is reached or the change in fitness is less than a set threshold, the optimal collaborative governance strategy is output.

9. A low-voltage management system for distribution substations based on spatiotemporal graph neural networks and improved sparrow search, characterized in that, For implementing the method as described in any one of claims 1-8, comprising: The graph structure building module is used to construct the distribution radio area into a graph structure containing nodes, edges, and adjacency matrices; The data acquisition module is used to collect multi-source operational data from each node in the distribution area and construct node feature vectors; The low voltage prediction module is used to input the feature sequences and their adjacency matrices of all nodes within a continuous sampling period into the spatiotemporal graph neural network model, output the voltage prediction value and hidden state of each node in the future prediction time domain, and generate a low voltage cause vector based on the hidden state of the nodes. The sensitivity calculation module is used to construct and correct the voltage sensitivity matrix online based on the node voltage change before and after the governance action, and to determine candidate governance resources based on the low voltage cause vector and the voltage sensitivity matrix. The optimization model building module constructs a multi-resource collaborative governance optimization model based on the low voltage cause vector, voltage sensitivity matrix, and equipment constraints. The optimization model solving module uses an improved sparrow search algorithm to solve the multi-resource collaborative governance optimization model to obtain a collaborative governance strategy. The instruction execution module is used to issue collaborative governance instructions to the corresponding governance resources according to the collaborative governance strategy; The correction module is used to correct the voltage sensitivity matrix and the spatiotemporal neural network model based on the measured voltage response after the governance action.

10. An electronic device, characterized in that, include: The processor, input device, output device, and memory are interconnected, the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to invoke the program instructions to execute the method described in any one of claims 1-8.

Images