Method and system for dynamic balance optimization of ring main unit load

Abstract

This invention discloses a method and system for dynamic load balancing optimization of ring main units, belonging to the field of smart grid technology. It aims to solve the problem of reduced load dispatch accuracy caused by timing misalignment errors resulting from different data transmission delays at various electrical nodes. The method includes: constructing an admittance matrix and obtaining a topology matrix through singular value decomposition; obtaining the original load current vector and timing difference vector; inputting historical feature vectors into an improved graph neural network to obtain a delay factor vector; constructing a confidence matrix based on the delay factor vector and timing difference vector, and reconstructing the load current vector by filtering with the topology matrix; when the reconstruction confidence level meets a threshold, sending the reconstructed load current vector to a cloud optimizer to generate and distribute tie switch state change strategies. This invention improves the accuracy and response efficiency of load dispatch under asynchronous measurement data.

Description

Technical Field

[0001] This invention belongs to the field of smart grid technology, specifically relating to a method and system for dynamic load balance optimization of ring network boxes. Background Technology

[0002] In medium-voltage distribution networks, dynamic load balancing refers to the selective alteration of the opening and closing states of tie switches and sectionalizing switches within the ring main unit (RMU) when load imbalances occur on feeders. This allows some load to be transferred from heavily loaded feeders to lightly loaded feeders, thereby maintaining the load rate of each feeder within a reasonable range. With the development of distribution automation, existing technologies synchronously collect analog and status variables such as voltage, current, and switch positions through substation terminals or feeder terminals installed within the RMU, and transmit this data to the distribution automation master station via a wireless network. The master station uses a network topology model for state estimation and online power flow calculation. When it detects that the load rate of a specific feeder exceeds the limit or the imbalance between feeders exceeds a set threshold, the master station automatically triggers a load balancing optimization strategy, generating an optimal switch action sequence to achieve dynamic load balancing across the entire network.

[0003] However, due to differences in the wireless communication environments of each terminal, there are variations in transmission delays when measurement data is sent to the master station, resulting in asynchronous timing errors in the data received by the master station from each electrical node. If the master station directly treats the asynchronously arriving data as arriving at the same time for state estimation and power flow calculation, it will introduce severe time-alignment measurement distortion, causing the calculation results to deviate from the actual power grid operating state. Furthermore, relying entirely on the master station for centralized topology analysis and power flow calculation results in heavy communication and computational burdens and large response delays, making it difficult to meet the high timeliness requirements of dynamic balance control under scenarios of severe load fluctuations. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a method and system for dynamic load balancing optimization of ring network boxes, thereby resolving the issues present in the background art.

[0005] To achieve the aforementioned objectives, this invention proposes a method for dynamic load balancing optimization of ring main units, comprising: Based on the electrical connection relationship and resistance between electrical nodes in the ring network, an admittance matrix is ​​constructed. Singular value decomposition is performed on the admittance matrix, and the left singular vectors corresponding to the first K largest singular values ​​are extracted to form a topology matrix. Obtain the load current of each electrical node in the ring network and construct the original load current vector for the current sampling period. Obtain the receiving time of the most recent measurement data of each electrical node, calculate the time difference between the current sampling period and the receiving time, and construct the current time difference vector based on the time difference. Obtain the historical feature vector of the previous sampling period, input the historical feature vector into the improved graph neural network, and calculate the delay factor vector of the current sampling period; Based on the time delay factor vector and the current time series difference vector, the confidence matrix is ​​calculated. Based on the confidence matrix and the topology matrix, the original load current vector is filtered to calculate the reconstructed load current vector. The reconstruction confidence level is calculated based on the reconstructed load current vector. When the reconstruction confidence level is greater than the preset confidence threshold, the reconstructed load current vector is sent to the cloud optimizer so that the cloud optimizer can calculate and issue the state change strategy of the interconnection switch based on the reconstructed load current vector.

[0006] The present invention also provides a ring network box load dynamic balancing optimization system, which is used to implement the above-described method, and the system includes: The topology construction module constructs an admittance matrix based on the electrical connection relationship and resistance between electrical nodes in the ring network. It then performs singular value decomposition on the admittance matrix and extracts the left singular vectors corresponding to the top K largest singular values ​​to form the topology matrix. The timing difference module obtains the load current of each electrical node in the ring network and constructs the original load current vector for the current sampling period. It obtains the receiving time of the most recent measurement data of each electrical node, calculates the time difference between the current sampling period and the receiving time, and constructs the current timing difference vector based on the time difference. The delay estimation module obtains the historical feature vector of the previous sampling period, inputs the historical feature vector into the improved graph neural network, and calculates the delay factor vector of the current sampling period. The load reconfiguration module calculates a reliability matrix based on the delay factor vector and the current timing difference vector. Based on the reliability matrix and the topology matrix, it filters the original load current vector and calculates the reconfigured load current vector. The reconfiguration distribution module calculates the reconfiguration confidence level based on the reconfiguration load current vector. When the reconfiguration confidence level is greater than a preset confidence threshold, the reconfiguration load current vector is sent to the cloud optimizer so that the cloud optimizer can calculate and distribute the state change strategy of the interconnection switch based on the reconfiguration load current vector.

[0007] The beneficial effects of this invention are as follows: This invention first captures the K dominant distribution patterns with the strongest electrical coupling in the ring network physical topology by performing singular value decomposition on the ring network admittance matrix and extracting the left singular vectors corresponding to the top K largest singular values ​​to form a topology matrix. Then, historical feature vectors are input into an improved graph neural network to estimate the delay factor of each node and fuse it with the current time-series difference vector to generate a confidence matrix. This assigns differentiated trust weights to asynchronously arriving measurement data, effectively suppressing time-alignment measurement distortion caused by communication delay differences. Subsequently, the original load current vector is filtered and reconstructed based on the confidence matrix and the topology matrix. The weighted data is projected onto the topology pattern space and non-physical components are filtered out to obtain a reconstructed load current vector that eliminates asynchronous errors. Finally, the data quality is evaluated by the reconstruction confidence level. The reconstructed vector is sent to the cloud optimizer only when the confidence level meets the standard to generate a tie switch state change strategy, thereby avoiding erroneous control based on distorted data. This invention completes the time alignment and reliable reconstruction of asynchronous data at the edge, improving the accuracy of dynamic load balancing. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating the steps of a dynamic load balancing optimization method for a ring network box according to the present invention. Figure 2 This is a comparison chart of the load imbalance between the present invention and the prior art; Figure 3 This is a schematic diagram of the structure of a ring network box load dynamic balance optimization system according to the present invention. Detailed Implementation

[0009] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0010] like Figure 1 As shown, a method for dynamic load balancing optimization of a ring main unit includes: S1: Construct an admittance matrix based on the electrical connection relationship and resistance between electrical nodes in the ring network. Perform singular value decomposition on the admittance matrix and extract the left singular vectors corresponding to the first K largest singular values ​​to form a topology matrix.

[0011] Before the ring main unit leaves the factory, a topology matrix is ​​pre-calculated and stored in the edge controller based on the physical ring network structure to which the ring main unit is connected. Specifically, when calculating the topology matrix, the impedance parameters (including resistance and reactance) of each line constituting the physical ring network and the equivalent impedance of the transformer are first obtained. A node admittance matrix is ​​constructed based on the electrical connection relationship between each electrical node. The element in the i-th row and j-th column of the admittance matrix represents the mutual admittance between electrical node i and electrical node j, and the diagonal elements represent the self-admittance of the electrical node. The larger the admittance value, the easier the current transmission between the two electrical nodes and the stronger the electrical coupling. For distributed parameters such as ground capacitance in actual distribution networks, their impact is relatively small at the voltage level of the ring main unit. To meet the real-time calculation requirements of the edge controller, this embodiment adopts a lumped parameter model for simplification.

[0012] Singular value decomposition (SVD) is performed on the admittance matrix, denoted as the left singular vector matrix, the singular value diagonal matrix, and the right singular vector matrix. The method for performing SVD on the matrix is ​​existing technology and will not be described here. The first K columns of the left singular vector matrix corresponding to the first K largest singular values ​​are taken, and these K column vectors are concatenated sequentially to obtain the topology matrix. Each column vector in the topology matrix represents the electrical distribution pattern of the physical ring network corresponding to the k-th order singular value. The value of the i-th element in the column vector represents the load current of electrical node i participating in the weighting under the k-th distribution pattern.

[0013] In mathematics, the square of the singular value represents the proportion of that pattern in the total energy of the admittance matrix. Since the elements in the admittance matrix physically represent the admittance values ​​between electrical nodes, the energy proportion reflects the physical electrical characteristics. The larger the singular value, the more the distribution pattern is composed of electrical branches with high admittance in the network. According to circuit principles, current transmission tends to choose the path of least impedance. High admittance branches represent the channels with the strongest electrical coupling and the least resistance to current transmission in the network. Therefore, the larger the singular value, the stronger the electrical coupling and the smaller the transmission impedance corresponding to that pattern, which is the most easily excited current distribution pattern in the network physically.

[0014] By retaining the column vectors corresponding to the top K largest singular values, the K most dominant electrical linkage patterns in the ring network physical topology are extracted. When the actual load current distribution of each electrical node in the ring network can be represented as a linear combination of the K coupling modes, that is, when the physical constraint that an increase in the load current of an electrical node must be accompanied by a change in the load current of its neighboring electrical nodes in a specific proportion is satisfied, the load current distribution is considered to conform to the electrical transmission law of the ring network topology.

[0015] S2: Obtain the load current of each electrical node in the ring network and construct the original load current vector for the current sampling period. Obtain the receiving time of the most recent measurement data of each electrical node, calculate the time difference between the current sampling period and the receiving time, and construct the current time difference vector based on the time difference.

[0016] During the sampling period, the load current of each electrical node is collected through its current transformer. The load current reflects the actual power consumption of that electrical node. The load currents of the N electrical nodes are arranged into an original load current vector. Since the measurement data of each electrical node is transmitted to the edge controller via wireless communication, the communication environments of different electrical nodes vary, resulting in different arrival times of the data at the edge controller. Treating the measurement data collected at different times by each electrical node as if they were at the same time would introduce time misalignment errors. Therefore, the receiving time of the most recent measurement data of the i-th electrical node at the edge controller is recorded. The difference between the current sampling period time t and the receiving time is used as the time difference value of the electrical node's data. For example, if the current sampling period time t is 10:00:00, the data of electrical node A arrives at the edge controller at 09:59:55, and the data of electrical node B arrives at 09:59:50, then the time difference value for electrical node A is 5 seconds, and for electrical node B it is 10 seconds. The larger the time difference, the further back in time the measurement data currently being used by that electrical node is from its actual acquisition time, and the greater the deviation of the data from the current true load state. The time differences of the N electrical nodes are arranged into a time-series difference vector.

[0017] S3: Obtain the historical feature vector of the previous sampling period, input the historical feature vector into the improved graph neural network, and calculate the delay factor vector of the current sampling period.

[0018] In this embodiment, the historical feature vector includes the time-series difference vector, historical load current change rate, historical dominance participation, number of node topologies, historical prediction error, and historical load current standard deviation. The historical dominance participation is the product of the row vector of the corresponding electrical node in the topology matrix and the filter coefficient vector of the previous sampling period. The number of node topologies is the total number of other electrical nodes connected to the electrical node in the ring network.

[0019] The historical feature vector is the feature vector of the previous sampling period. This vector includes the time-series difference vector, historical load current change rate, historical dominance participation, number of node topologies, historical prediction error, and historical load current standard deviation. The historical load current change rate is the historical load current change rate between two adjacent sampling periods; that is, the load current change rate of each electrical node in the previous sampling period t and the period before that t-1. It is calculated as the ratio of the load current difference between adjacent sampling periods to the time interval. A larger load current change rate indicates a greater impact of time delay on the measured value of that electrical node.

[0020] The historical dominant participation degree is the historical dominant participation degree of the previous sampling period, which is the product of the row vector of the corresponding electrical node in the topology matrix and the filter coefficient vector of the previous sampling period. The filter coefficient vector contains K coefficients, where the k-th coefficient represents the projection weight of the k-th electrical distribution mode in the previous sampling period. The projection weight represents the proportion of the current actual load current distribution that conforms to the k-th electrical distribution mode. The specific calculation method of the filter coefficients will be introduced later. The row vector of the i-th row of the topology matrix represents the participation weight of electrical node i in the K electrical distribution modes. The historical dominant participation degree obtained by multiplying it with the filter coefficients represents the participation intensity of electrical node i in the main electrical distribution mode of the previous sampling period. The higher the participation intensity, the more important the role of the electrical node in the ring network load current distribution, and the greater the impact of the accuracy of its measurement data on the overall balance optimization.

[0021] The node topology number represents the total number of other electrical nodes connected to the i-th electrical node in the ring network. It reflects the density of connections of this electrical node within the ring network topology. A higher topology number indicates that the electrical node has more electrical connections with its neighbors, and the wider the range of influence of the electrical node's measurement error propagating to other electrical nodes through electrical coupling.

[0022] Historical prediction error is obtained by averaging the absolute values ​​of the differences between the reconstructed value of the (k-1)th sampling period and the actual measured value of the kth sampling period over the past M periods. Historical prediction error reflects the average degree to which the predicted estimate for electrical node i deviates from its true value over historical periods. A larger historical prediction error indicates that the load current of that electrical node has been more difficult to predict accurately historically, and the reliability of its current measurement should be reduced.

[0023] The historical load current standard deviation is obtained by calculating the standard deviation of the load current measurement of the i-th electrical node within a preset historical time period. The standard deviation measures the dispersion of the load current of that electrical node over the historical time period. The larger the standard deviation, the more drastic the fluctuation of the load current of that electrical node, and the higher the risk of measurement distortion introduced by time delay.

[0024] In this embodiment, the historical feature vector is input into the improved graph neural network, including: The admittance magnitude matrix is ​​obtained by obtaining the absolute value of each element in the admittance matrix. The admittance magnitude matrix is ​​normalized to obtain the weight matrix. The historical feature vector is input into the improved graph neural network. Each neuron in the hidden layer of the improved graph neural network performs a weighted summation of the features of its electrically connected electrical nodes based on the weight matrix. The output value of the hidden layer is calculated by the activation function. The output layer calculates the original output evaluation value of each electrical node based on the hidden layer output value. The original output evaluation value is input into the constrained activation function for mapping and outputs the time delay factor of each electrical node. The time delay factors of all electrical nodes constitute the time delay factor vector.

[0025] Before inputting the historical feature vectors into the improved graph neural network, the improved graph neural network needs to be trained offline. First, multiple historical time period data with synchronized electrical node data are obtained from the historical operation database of the ring network, which are used as the reference for the actual load current. In the historical time period data, asynchronous timing scenarios are simulated by simulating or randomly sampling the arrival data of different electrical nodes under different communication delays, and the corresponding historical feature vectors are calculated as training input samples.

[0026] The training input samples are fed into the improved graph neural network, which outputs the delay factor vector for the current sampling period. Based on this delay factor vector and the current time-series difference vector, the confidence matrix is ​​calculated according to the aforementioned formula. Subsequently, the reconstructed load current vector is calculated using the confidence matrix, the topology matrix, and the asynchronous original load current vector.

[0027] Construct a loss function, defined as the mean square error between the reconstructed load current vector and the true load current vector: Where N is the total number of electrical nodes. and These represent the load current values ​​of the i-th electrical node in the reconstructed load current vector and the actual load current vector, respectively. Using an adaptive moment estimation optimization algorithm, the gradient of the loss function with respect to trainable parameters such as the linear mapping weights of the output layer in the improved graph neural network is calculated via backpropagation, and the parameters are iteratively updated until the loss function converges. It should be noted that the connection weights in the hidden layers of the improved graph neural network, constructed based on the admittance magnitude matrix, are fixed values ​​and do not participate in training updates.

[0028] Through the above training process, the improved graph neural network can autonomously learn the intrinsic mapping relationship between historical feature vectors and time delay decay rate. When historical features indicate that the load of a certain electrical node changes drastically and the historical prediction error is large, the network will automatically output a larger time delay factor, so that the old data of the node is given a smaller confidence weight in subsequent calculations, thereby ensuring the accuracy of the final reconstructed load current vector at the end-to-end level.

[0029] The historical feature vector from the previous sampling period is used as input data for the improved graph neural network. The hidden layer of the improved graph neural network contains N neurons, and each of the N neurons corresponds one-to-one with N electrical nodes. For the i-th neuron in the hidden layer, the set of electrical nodes in the ring network that are electrically connected to the i-th electrical node is retrieved. Based on the hidden layer connection weight matrix, the connection weights corresponding to each electrical node j in the set of electrical nodes are extracted from the i-th neuron. The historical feature vector of electrical node j is multiplied by the corresponding connection weight and then accumulated to obtain the aggregated feature of the i-th neuron. The aggregated feature is input into the activation function for nonlinear mapping to calculate the hidden layer output value of the i-th neuron. This process is repeated for all neurons in the hidden layer to calculate the hidden layer output values ​​of all neurons.

[0030] The improved graph neural network's output layer receives the hidden layer output values ​​from all neurons. A linear mapping is performed on these hidden layer output values ​​to calculate the raw output evaluation value for each electrical node. This raw output evaluation value is then input into the aforementioned constrained activation function for mapping, and the mapping result is used as the delay factor for each electrical node. The delay factors of all electrical nodes are arranged according to their electrical node numbers to form the delay factor vector for the current sampling period.

[0031] To improve the hidden layer connection weights of a graph neural network, during construction, the absolute values ​​of each element in the admittance matrix are obtained to obtain the admittance magnitude matrix. This matrix is ​​then normalized to its maximum value to generate the hidden layer weight matrix. The element in the i-th row and j-th column of the hidden layer weight matrix represents the connection weight between the neuron corresponding to electrical node i and the neuron corresponding to electrical node j in the hidden layer. A larger admittance magnitude indicates stronger electrical coupling between the two electrical nodes, resulting in a larger connection weight.

[0032] The construction of the hidden layer weight matrix ensures that the internal information flow path of the improved graph neural network aligns with the current transmission path in the physical topology of a ring network. In a physical ring network, a large admittance value between two adjacent electrical nodes indicates low resistance to current transmission between them, meaning that load changes in one node will strongly affect the other through this low-impedance path. Correspondingly, in the hidden layer of the improved graph neural network, neurons corresponding to two electrical nodes with large admittance values ​​have larger connection weights, resulting in stronger information transmission. Conversely, neurons corresponding to two electrical nodes with small or zero admittance values ​​have smaller connection weights or no connection, resulting in weaker or no information transmission.

[0033] The output value of the i-th neuron in the hidden layer is obtained by weighting and summing the features of all electrical nodes with physical electrical connections and then inputting it into the activation function. This structure makes the information propagation path of the hidden layer follow the physical topology connection relationship of the ring network. The information received by each electrical node neuron comes only from its direct neighbor electrical nodes in the physical ring network, and the receiving weight is proportional to the electrical coupling strength between the two electrical nodes.

[0034] The output layer calculates the original output evaluation value of the i-th electrical node based on the output value of the hidden layer, and maps it to the delay factor of the current sampling period through a constrained activation function, where the constrained activation function is: ,in, The preset minimum delay factor, This is the preset maximum delay factor, whose value is greater than 1. This is the original output evaluation value calculated for electrical node i by the output layer. The function in the constraint activation function. The original output evaluation value is mapped to a value between 0 and 1, and then transformed into the maximum and minimum attenuation range through a linear transformation preceding the function. The reason for this quadratic transformation is that if the delay factor only takes values ​​between 0 and 1, its confidence weight might be too large when calculating timing differences later, failing to suppress older data. Therefore, the preset maximum delay factor is set to a value greater than 1, allowing for the calculation of a smaller confidence weight and thus suppressing the measurement value of that electrical node.

[0035] Arrange the delay factors of N electrical nodes according to their electrical node numbers to form the delay factor vector for the current sampling period, which will be used to construct the reliability matrix later.

[0036] S4: Based on the time delay factor vector and the current time series difference vector, the confidence matrix is ​​calculated. Based on the confidence matrix and the topology matrix, the original load current vector is filtered to calculate the reconstructed load current vector.

[0037] Based on the delay factor vector and the current time-series difference vector of each electrical node in the current sampling period, a confidence weight is calculated for each electrical node. The confidence weight quantifies the degree to which the current measurement data of that electrical node participates in the subsequent filtering process. The confidence weight of the i-th electrical node is... Calculated based on the following formula ,in, Let be the time delay factor of the i-th electrical node in the current sampling period, which is a dimensionless value output by the improved graph neural network. It represents the time difference between the current sampling time and the receiving time for the i-th electrical node after normalization.

[0038] In the above formula, when the time difference of an electrical node is small, it indicates that the measurement data of that electrical node is close to the acquisition time, and the calculated confidence weight is close to 1. The measurement data of that electrical node maintains a high degree of participation in subsequent filtering. When the time difference of an electrical node is large, the calculated confidence weight approaches 0. The larger the time delay factor, the faster the confidence weight decreases, and the stronger the suppression of old data. Since the time delay factor is calculated by the improved graph neural network based on multi-dimensional characteristics such as the load current change rate, topology connection density, and historical prediction accuracy of each electrical node, the confidence weight of different electrical nodes varies under the same time difference. For example, electrical nodes connected to impactful industrial loads have a large time delay factor, and even if the time difference is small, the confidence weight decreases faster, reflecting that the measurement data of this type of electrical node is more sensitive to time delay. Electrical nodes connected to residential lighting loads have a small time delay factor, and even if the time difference is large, the confidence weight is still high, reflecting that the measurement data of this type of electrical node can still be trusted within a certain time delay range.

[0039] Arrange the trust weights of N electrical nodes in the diagonal position according to the electrical node number to form a trust matrix. In the trust matrix, the diagonal element in the i-th row and i-th column represents the degree to which the measurement data of the i-th electrical node is trusted by the subsequent calculation process in the current sampling period.

[0040] In this embodiment, filtering of the original load current vector includes: The original load current vector is weighted based on the credibility matrix, and the weighted original load current vector is projected onto the electrical distribution pattern space represented by the topology matrix. The projection result is normalized and corrected by combining the credibility matrix and the topology matrix to obtain the filter coefficient vector. The original load current vector is reconstructed based on the filter coefficient and the topology matrix.

[0041] The purpose of filtering the original load current vector is to reduce the measurement distortion caused by time misalignment within the original load current vector under the constraints of the ring network physical topology, thereby obtaining a reconstructed load current vector that conforms to the electrical transmission laws. In the specific calculation, the original load current of each electrical node is first scaled according to its confidence weight. The measured values ​​corresponding to electrical nodes with small confidence weights are compressed to near zero, while the measured values ​​corresponding to electrical nodes with large confidence weights are kept near their original values. Then, the weighted load current vector is projected onto the space of K electrical distribution patterns using a topology matrix to obtain the preliminary projection weights for each distribution pattern. Finally, the preliminary projection weights are corrected using an intermediate matrix to obtain the final filtering coefficient vector.

[0042] The intermediate matrix is ​​calculated based on the following formula: ,in, It is a topological matrix. This is the confidence matrix for the current sampling period. In a physical sense, This represents the correlation between topology patterns, taking into account the differences in the reliability of measurement data from each electrical node. When the measurement data from each electrical node are time-aligned and have the same reliability, the reliability matrix degenerates into an identity matrix. Become Since the column vectors of the topological matrix are derived from the left singular vectors of the singular value decomposition, they satisfy the orthogonality. Since it is an identity matrix, its inverse matrix is ​​also an identity matrix. When the measurement data of each electrical node are not aligned in time and have different levels of reliability, the diagonal elements corresponding to different electrical nodes in the reliability matrix will be different, leading to... It is no longer an identity matrix; an intermediate matrix needs to be obtained by calculating its inverse matrix.

[0043] Filter coefficient vector Calculated based on the following formula: ,in, This is the original load current vector for the current sampling period. That is, scaling the original load current based on the credibility matrix. The projection weights represent the projection weights obtained by projecting the weighted original load current vector onto the electrical distribution pattern space represented by the topological matrix. The filter coefficient vector is a K-dimensional vector, where the k-th element represents the component of the k-th electrical distribution pattern in the current sampling period. Distribution patterns with larger projection weights have a stronger effect on the current load current distribution.

[0044] Finally, the reconstructed load current vector is calculated based on the following formula: This formula linearly combines K electrical distribution patterns according to their projected weights to reconstruct a load current distribution that conforms to the physical topology constraints of the ring network. The i-th element in the reconstructed load current vector represents the estimated load current of electrical node i after topology space smoothing filtering.

[0045] During the reconstruction process, components in the original load current vector that violate the electrical transmission laws of the ring network are filtered out. These violations mainly originate from measurement distortion introduced by time misalignment, manifested as the measured values ​​of some electrical nodes deviating from the normal correlation ratio of their neighboring electrical nodes. Since these violations cannot be represented by a linear combination of the K dominant electrical distribution patterns, they are discarded after being projected onto the space spanned by the topology matrix. The reconstructed load current vector retains only the components that can be represented by a linear combination of the K dominant electrical distribution patterns. These components satisfy the physical constraints of electrical linkage between electrical nodes in the ring network and reflect the coordinated change law of load current at each electrical node according to the electrical coupling relationship when the load current changes at a certain electrical node.

[0046] S5: Calculate the reconstruction confidence level based on the reconstructed load current vector. When the reconstruction confidence level is greater than the preset confidence threshold, send the reconstructed load current vector to the cloud optimizer so that the cloud optimizer can calculate and issue the state change strategy of the interconnection switch based on the reconstructed load current vector.

[0047] In this embodiment, when calculating the reconstruction confidence level, the difference vector between the original load current vector and the reconstructed load current vector is first calculated, and the sum of the squares of each element of this difference vector is calculated as the first result. Then, the sum of the squares of each element of the original load current vector is calculated as the second result. The reconstruction confidence level is obtained by subtracting the ratio of the first result to the second result from 1. The reconstruction confidence level represents the degree to which the reconstructed load current vector fits the original measurement data in terms of direction and amplitude. When the reconstruction confidence level is close to 1, it indicates that most of the energy in the original load current vector can be represented by a linear combination of K electrical distribution patterns represented by the topology matrix, and the measurement distortion introduced by time misalignment is relatively small. When the reconstruction confidence level is much less than 1, it indicates that the original measurement data deviates significantly from the physical topology constraints of the ring network, and there is a large amount of distortion.

[0048] When the reconstruction confidence level is greater than the preset confidence threshold, it is determined that the time alignment and confidence level of the measurement data of each electrical node in the current sampling period meet the requirements of subsequent load balancing calculation. The reconstructed load current vector is then sent to the cloud optimizer, which calculates the state change strategy of the interconnection switch based on the reconstructed load current vector.

[0049] In this embodiment, the cloud optimizer calculates the state change strategy of the tie switch based on the reconstructed load current vector, including: Electrical nodes are divided into heavy-load electrical nodes and light-load electrical nodes. The overload of heavy-load electrical nodes and the load margin of light-load electrical nodes are calculated. The smaller value between the overload and the load margin is used as the transfer amount. Based on the transfer amount, the target state of the electrical path connection switch between heavy-load and light-load electrical nodes is determined, and a state change strategy table including the connection switch number, current state and target state is generated.

[0050] When the reconstruction confidence level is less than or equal to the preset confidence threshold, it is determined that the measurement data quality of the current sampling period does not meet the load balancing calculation requirements. The edge controller does not send the reconstructed load current vector to the cloud optimizer and waits for the next sampling period to reacquire the measurement data of each electrical node and repeat the above steps until the reconstruction confidence level meets the requirements.

[0051] This judgment mechanism avoids the cloud optimizer from calculating the contact switch action strategy based on severely distorted data, reducing the risk of erroneous control due to insufficient data quality. The preset threshold is set according to the complexity of the ring network's physical topology and the required control accuracy. Lower thresholds can be set in scenarios with simple ring network topologies and lower control accuracy requirements, while higher thresholds are required in scenarios with complex ring network topologies and higher control accuracy requirements.

[0052] In this embodiment, electrical nodes are divided into heavy-load electrical nodes and light-load electrical nodes, including: When the reconfigured load current value of an electrical node is greater than the corresponding heavy load threshold, the electrical node is determined to be a heavy load electrical node and recorded in the heavy load electrical node set. In the heavy load electrical node set, the electrical nodes adjacent to the heavy load electrical node in the ring network through the tie switch are searched. When the reconfigured load current value of the adjacent electrical node is less than the corresponding light load threshold, the adjacent electrical node is determined to be a light load electrical node and recorded in the light load electrical node set.

[0053] The cloud optimizer receives the reconfigured load current vector reported by the edge controller, iterates through the N elements in the reconfigured load current vector, and compares the value of the i-th element with the overload threshold of the electrical node. The overload threshold is determined by the rated capacity and safety margin factor of the distribution transformer connected to the electrical node, representing the maximum allowable load current of the electrical node. When the value of the i-th element is greater than the overload threshold of the electrical node, the electrical node i is determined to be an overloaded electrical node, and the electrical node is recorded in the set of overloaded electrical nodes.

[0054] For each heavily loaded electrical node in the set of heavily loaded electrical nodes, retrieve its adjacent electrical nodes connected via tie switches in the ring network physical topology. For each adjacent electrical node j, obtain its reconfigured load current value and compare it with the light load threshold of that electrical node. The light load threshold is determined by the rated capacity and minimum load rate coefficient of the distribution transformer connected to the electrical node, representing the lower limit of the load current that the electrical node is allowed to receive load transfer while ensuring power supply efficiency. When the reconfigured load current value is less than the light load threshold, electrical node j is determined to be a light load electrical node and recorded in the set of light load electrical nodes.

[0055] In this embodiment, the difference between the reconstructed load current value of the heavy-load electrical node and the heavy-load threshold is taken as the load overload, and the difference between the heavy-load threshold and the reconstructed load current value of the light-load electrical node is taken as the load margin.

[0056] For each heavily loaded electrical node i in the set of heavily loaded electrical nodes, the difference between the reconfigured load current value and the heavy-load threshold is taken as the load overload. The load overload represents the portion of the current load of electrical node i that exceeds its maximum allowable value, which needs to be reduced to a safe range through load transfer. For each lightly loaded electrical node j in the set of lightly loaded electrical nodes, the difference between the heavy-load threshold and the reconfigured load current value is taken as the load margin. The load margin represents the remaining space between the current load current of electrical node j and its maximum allowable value, indicating the upper limit of the load transfer that electrical node can accept.

[0057] Based on the physical connection relationship between heavily loaded and lightly loaded electrical nodes, load transfer paths are constructed. For example, for heavily loaded electrical node i and lightly loaded electrical node j, if the two electrical nodes are directly connected through a tie switch that is in an open state, this electrical node pair is recorded as a candidate transfer path. For each candidate transfer path, the transfer amount is calculated, and the transfer amount is the smaller value between the load overload and the load margin. Under the standard radial operation constraint of the distribution network, i.e., based on the transfer amount as demand, the target state of the tie switch between heavily loaded electrical node i and lightly loaded electrical node j is set to closed. At the same time, each section switch that is in a closed state upstream of the heavily loaded electrical node i is retrieved and used as a candidate open switch. Combined with the reconstructed load current vector, a simulated power flow calculation is performed to evaluate the actual load transfer amount generated after opening each candidate section switch. The section switch that makes the actual load transfer amount closest to and not greater than the target transfer amount demand is selected, and its target state is set to open. Thus, while maintaining the radial topology of the network, part of the load is transferred from the heavily loaded side to the lightly loaded side. Record all contact switches and sectionalizing switches that need to change their status and their target status in the status change strategy table. Each item in the table includes the switch number, current status, and target status.

[0058] Based on the state change strategy table and the reconstructed load current vector, ring network power flow calculation is performed to obtain the expected load current vector after the tie switch is closed and the sectionalizing switch is opened, and the actual load transfer amount is calculated. When the expected load current values ​​of all electrical nodes in the expected load current vector meet the preset conditions, and the actual load transfer amount meets the target transfer amount requirement, the state change strategy is deemed feasible, and an action command is generated and issued based on the state change strategy table.

[0059] Based on the reconstructed load current vector and the state change strategy table, the expected load current vector after the state change of the tie switch and sectionalizing switch is calculated. Specifically, the cloud optimizer first modifies the admittance value of the corresponding switch branch in the admittance matrix according to the target state in the state change strategy table. For example, if the switch is closed, the corresponding line admittance is included; if the switch is open, the corresponding admittance is set to 0, thus obtaining the updated admittance matrix. The reconstructed load current vector is used as the node injection quantity. Combined with the updated admittance matrix and the preset node voltage reference value, the solution is obtained using the forward back substitution method or the Newton-Raphson method. This calculates the expected load current vector of each electrical node in steady state after the switch action, and the actual load transfer amount of the heavily loaded electrical node is calculated based on this expected load current vector.

[0060] Iterate through the N elements in the expected load current vector and compare the value of the i-th element with the overload threshold of the electrical node. If the value is greater than the overload threshold, it is determined that electrical node i will be overloaded after the switch state changes, and the electrical node is recorded in the overload electrical node set.

[0061] If the set of overloaded electrical nodes is empty, and the actual load transfer amount meets the target transfer amount requirement, it means that the switching actions in the state change strategy table will not cause any electrical nodes to overload and can alleviate the load on heavily loaded nodes, making the strategy feasible. The cloud-based optimizer generates switching action instructions based on the state change strategy table. These instructions include the numbers of the tie switches and sectionalizing switches that need to be activated, as well as their target states. The instructions are then sent to the corresponding ring network box's actuators via the communication link. Upon receiving the instructions, the actuators control the operating mechanisms of the relevant tie switches and sectionalizing switches to perform physical switching actions, completing the loop closing and then opening operation, thus achieving load balancing optimization of the ring network box.

[0062] If the set of overloaded electrical nodes is not empty, or the actual load transfer cannot meet the target transfer requirement, it indicates that the switching action in the state change strategy table will cause the electrical nodes to overload or fail to achieve the expected balance effect, and the strategy is not feasible. The cloud optimizer intercepts this action command and does not send any switching action command to the ring network box, maintaining the current switching state. At the same time, the cloud optimizer generates grid congestion alarm information, which includes the number of the electrical nodes in the overloaded electrical node set, the expected load current value, and the overload threshold. The alarm information is sent to the grid dispatch center, where dispatchers take manual intervention measures according to the actual situation.

[0063] like Figure 2As shown in the figure, this invention is a comparison of the system load imbalance between the present invention and the prior art. When the load change occurs in the third sampling period, causing the imbalance to spike, the present invention, because it has completed the time alignment and reliable reconstruction of asynchronous data in advance at the edge side, eliminates the measurement distortion introduced by the communication delay difference. Therefore, it can quickly and accurately generate a state change strategy and rapidly reduce the load imbalance to a safe and stable range of less than 10% in the following periods.

[0064] like Figure 3 As shown, the present invention also provides a ring network box load dynamic balancing optimization system, which is used to implement the above-described method. The system includes: The topology construction module constructs an admittance matrix based on the electrical connection relationships and resistances between electrical nodes in the ring network. It then performs singular value decomposition on the admittance matrix and extracts the left singular vectors corresponding to the top K largest singular values ​​to form the topology matrix.

[0065] The timing difference module obtains the load current of each electrical node in the ring network and constructs the original load current vector for the current sampling period. It obtains the receiving time of the most recent measurement data of each electrical node, calculates the time difference between the current sampling period and the receiving time, and constructs the current timing difference vector based on the time difference.

[0066] The delay estimation module obtains the historical feature vector of the previous sampling period, inputs the historical feature vector into the improved graph neural network, and calculates the delay factor vector of the current sampling period.

[0067] The load reconfiguration module calculates a reliability matrix based on the delay factor vector and the current timing difference vector. Based on the reliability matrix and the topology matrix, it filters the original load current vector and calculates the reconfigured load current vector.

[0068] The reconfiguration distribution module calculates the reconfiguration confidence level based on the reconfiguration load current vector. When the reconfiguration confidence level is greater than a preset confidence threshold, the reconfiguration load current vector is sent to the cloud optimizer so that the cloud optimizer can calculate and distribute the state change strategy of the interconnection switch based on the reconfiguration load current vector.

[0069] Finally, it should be noted that the various thresholds, algorithm parameters, and specific calculation methods for normalization processing involved in the above specific embodiments can all be determined by those skilled in the art through conventional experiments or experience based on the physical topology, communication environment characteristics, and historical operating data of the actual ring network. These are conventional technical means in the field and are not intended to limit the scope of protection of this invention.

Claims

1. A method for dynamic load balancing optimization of a ring main unit, characterized in that, include: Based on the electrical connection relationship and resistance between electrical nodes in the ring network, an admittance matrix is ​​constructed. Singular value decomposition is performed on the admittance matrix, and the left singular vectors corresponding to the first K largest singular values ​​are extracted to form a topology matrix. Obtain the load current of each electrical node in the ring network and construct the original load current vector for the current sampling period. Obtain the receiving time of the most recent measurement data of each electrical node, calculate the time difference between the current sampling period and the receiving time, and construct the current time difference vector based on the time difference. Obtain the historical feature vector of the previous sampling period, input the historical feature vector into the improved graph neural network, and calculate the delay factor vector of the current sampling period; Based on the time delay factor vector and the current time series difference vector, the confidence matrix is ​​calculated. Based on the confidence matrix and the topology matrix, the original load current vector is filtered to calculate the reconstructed load current vector. The reconstruction confidence level is calculated based on the reconstructed load current vector. When the reconstruction confidence level is greater than the preset confidence threshold, the reconstructed load current vector is sent to the cloud optimizer so that the cloud optimizer can calculate and issue the state change strategy of the interconnection switch based on the reconstructed load current vector.

2. The method according to claim 1, characterized in that, The historical feature vectors are input into the improved graph neural network, including: The admittance magnitude matrix is ​​obtained by obtaining the absolute value of each element in the admittance matrix. The admittance magnitude matrix is ​​normalized to obtain the weight matrix. The historical feature vector is input into the improved graph neural network. Each neuron in the hidden layer of the improved graph neural network performs a weighted summation of the features of its electrically connected electrical nodes based on the weight matrix. The output value of the hidden layer is calculated by the activation function. The output layer calculates the original output evaluation value of each electrical node based on the hidden layer output value. The original output evaluation value is input into the constrained activation function for mapping and outputs the time delay factor of each electrical node. The time delay factors of all electrical nodes constitute the time delay factor vector.

3. The method according to claim 1, characterized in that, Filtering the original load current vector includes: The original load current vector is weighted based on the credibility matrix, and the weighted original load current vector is projected onto the electrical distribution pattern space represented by the topology matrix. The projection result is normalized and corrected by combining the credibility matrix and the topology matrix to obtain the filter coefficient vector. The original load current vector is reconstructed based on the filter coefficient and the topology matrix.

4. The method according to claim 1, characterized in that, The cloud-based optimizer calculates the state change strategy for tie switches based on the reconstructed load current vector, including: The electrical nodes are divided into heavy-load electrical nodes and light-load electrical nodes. The overload of the heavy-load electrical nodes and the load margin of the light-load electrical nodes are calculated. The smaller value between the overload and the load margin is used as the transfer amount. Based on the transfer amount, the target state of the electrical path connection switch between the heavy-load and light-load electrical nodes is determined, and a state change strategy table including the connection switch number, current state and target state is generated. Based on the state change strategy table and the reconstructed load current vector, the ring network power flow is calculated to obtain the expected load current vector after the tie switch is closed and the sectionalizing switch is opened. The actual load transfer amount is calculated. When the expected load current values ​​of all electrical nodes in the expected load current vector meet the preset conditions and the actual load transfer amount meets the target transfer amount requirement, the state change strategy is deemed feasible. Based on the state change strategy table, an action command is generated and issued.

5. The method according to claim 4, characterized in that, Electrical nodes are classified into heavy-load electrical nodes and light-load electrical nodes, including: When the reconfigured load current value of an electrical node is greater than the corresponding heavy load threshold, the electrical node is determined to be a heavy load electrical node and recorded in the heavy load electrical node set. In the heavy load electrical node set, the electrical nodes adjacent to the heavy load electrical node in the ring network through the tie switch are searched. When the reconfigured load current value of the adjacent electrical node is less than the corresponding light load threshold, the adjacent electrical node is determined to be a light load electrical node and recorded in the light load electrical node set.

6. The method according to claim 5, characterized in that, The difference between the reconfigured load current value of the heavily loaded electrical node and the heavily loaded threshold is taken as the load overload, and the difference between the heavily loaded threshold and the reconfigured load current value of the lightly loaded electrical node is taken as the load margin.

7. The method according to claim 1, characterized in that, The historical feature vector includes the time-series difference vector, historical load current change rate, historical dominance participation, number of node topologies, historical prediction error, and historical load current standard deviation. The historical dominance participation is the product of the row vector of the corresponding electrical node in the topology matrix and the filter coefficient vector of the previous sampling period. The number of node topologies is the total number of other electrical nodes connected to the electrical node in the ring network.

8. A ring main unit load dynamic balancing optimization system, used to implement the method as described in any one of claims 1-7, characterized in that, The system includes: The topology construction module constructs an admittance matrix based on the electrical connection relationship and resistance between electrical nodes in the ring network. It then performs singular value decomposition on the admittance matrix and extracts the left singular vectors corresponding to the top K largest singular values ​​to form the topology matrix. The timing difference module obtains the load current of each electrical node in the ring network and constructs the original load current vector for the current sampling period. It obtains the receiving time of the most recent measurement data of each electrical node, calculates the time difference between the current sampling period and the receiving time, and constructs the current timing difference vector based on the time difference. The delay estimation module obtains the historical feature vector of the previous sampling period, inputs the historical feature vector into the improved graph neural network, and calculates the delay factor vector of the current sampling period. The load reconfiguration module calculates a reliability matrix based on the delay factor vector and the current timing difference vector. Based on the reliability matrix and the topology matrix, it filters the original load current vector and calculates the reconfigured load current vector. The reconfiguration distribution module calculates the reconfiguration confidence level based on the reconfiguration load current vector. When the reconfiguration confidence level is greater than a preset confidence threshold, the reconfiguration load current vector is sent to the cloud optimizer so that the cloud optimizer can calculate and distribute the state change strategy of the interconnection switch based on the reconfiguration load current vector.

Images