Edge ai-based low-voltage distribution network adaptive regulation method and system, and storage medium

Abstract

The application provides a low-voltage power distribution network adaptive regulation method and system based on edge AI, a storage medium, and the like. The method includes collecting voltage, current, and power time series data of key nodes of a power grid, constructing a block Hankel matrix after standardization, and folding the block Hankel matrix into a high-order space-time tensor. Core feature tensors are extracted through Tucker decomposition. The feature tensors are input into a deep separable convolution network. A deep convolution layer uses a filter group in a matrix product operator format to extract spatial features. A tensor core is updated in real time through a variational algorithm based on a prediction error and a Riemann manifold. A point-by-point convolution layer further reorganizes channel features. A voltage-power sensitivity matrix pseudo-inverse is embedded into a weight core in a tensor column decomposition format of the point-by-point convolution layer through tensor contraction. Mapping constraints of feature channels and topological nodes are realized. A regression prediction layer calculates active and reactive power set values of each controllable device according to a fused feature map. The set values are issued to a device controller to complete closed-loop regulation.

Description

Technical Field

[0001] This application belongs to the field of regulation and control, and in particular relates to an adaptive regulation and control method, system and storage medium for low-voltage distribution networks based on edge AI. Background Technology

[0002] With the advancement of the "dual-carbon" goals and the construction of new power systems, low-voltage distribution networks are facing large-scale integration of distributed energy sources with high penetration rates, such as distributed photovoltaics and electric vehicle charging piles, as well as flexible loads. This random fluctuation on both the source and load sides causes the operating conditions of the distribution network to exhibit strong time-varying and non-stationary characteristics, leading to frequent problems such as voltage exceeding limits and backflow of power. Existing distribution network control methods heavily rely on accurate grid topology and line impedance parameters. However, low-voltage distribution networks suffer from problems such as missing topology files and untimely parameter updates, severely limiting the accuracy and convergence speed of model-based methods in practical applications. While methods represented by deep convolutional neural networks can fit nonlinear mapping relationships using historical data, traditional deep learning models typically treat spatiotemporal data as low-dimensional matrices, resulting in models with a huge number of parameters, high training costs, and difficulty in deployment on edge control equipment with limited computing resources.

[0003] Furthermore, when facing the time-varying and non-stationary characteristics of low-voltage distribution networks, most existing deep learning control strategies adopt a static model of offline training and online inference. When the operating conditions of the power grid undergo conceptual drift or minor topological changes, the fixed-weight network is difficult to adjust, and control commands are prone to deviation. Existing models are mostly black-box structures, lacking interpretability and failing to incorporate prior knowledge such as Kirchhoff's laws and voltage-power sensitivity into the network structure. This makes the generated control strategies potentially violate constraints and pose safety hazards. Although deep separable convolution and tensor decomposition techniques have been applied in the field of model compression, how to combine these techniques with topological constraints and plan a mechanism that can perform low-rank parameter iterative updates in non-Euclidean spaces such as Riemannian manifolds, and solve the problems of difficult high-dimensional data feature extraction, weak model adaptability, and poor consistency, remains a critical technical bottleneck that urgently needs to be overcome in the field of intelligent control of low-voltage distribution networks. Summary of the Invention

[0004] To address the problems that existing technologies neglect the high-order spatiotemporal evolution laws contained in the multi-node measurement data of the distribution network, and that the model parameters are huge, the training cost is high, and it is difficult to deploy on control equipment with limited computing resources at the edge.

[0005] In the first aspect, the present invention proposes an adaptive control method for low-voltage distribution networks based on edge AI, comprising the following steps: Voltage, current and power time-series operation data of key nodes in low-voltage distribution network are collected. After standardization of the data, a block Hankel matrix is ​​constructed, and the matrix is ​​folded into a high-order spatiotemporal tensor. Tucker decomposition is performed on the high-order spatiotemporal tensor to extract the core feature tensor. The core feature tensor is input into a depthwise separable convolutional network, and spatial features are extracted through deep convolutional layers. The filter bank of the deep convolutional layer is parameterized into a matrix multiplication operator format. During the control process, a variational algorithm based on prediction error and Riemann manifold is used to iteratively update the local tensor kernel of the matrix multiplication operator in real time to adapt to the time-varying non-stationary characteristics of the distribution network. By performing channel-dimensional feature reorganization on the output of the deep convolutional layer through pointwise convolutional layers, the pseudo-inverse of the voltage-power sensitivity matrix of the distribution network under the current operating conditions is calculated. The pseudo-inverse matrix is ​​used as prior knowledge, and the pseudo-inverse matrix is ​​embedded into the weight kernel of the pointwise convolutional layer based on the tensor column decomposition format using tensor shrinkage operation. The mapping relationship between feature channels and topology nodes is established, and a feature map with fusion constraints is generated. The fused feature map is input into the regression prediction layer to calculate the active and reactive power setpoints of each controllable device, and then sent to the device controller to perform closed-loop control of the low-voltage distribution network.

[0006] Optionally, the step of collecting voltage, current, and power time-series operating data of key nodes in the low-voltage distribution network, and constructing a block-based Hankel matrix after standardizing the data, includes: Calculate the mean and standard deviation of the collected voltage, current and power time-series operating data, and perform Z-score standardization, which is to subtract the mean from each data point and then divide by the standard deviation; Set the sliding window length L and the sliding step size S. Starting from the starting point of the standardized time series data, extract data segments as column vectors according to the set length and step size. The multiple column vectors generated in chronological order are arranged sequentially to construct a block Hankel matrix in which each row corresponds to a time step and each column corresponds to a time window feature.

[0007] Optionally, the step of performing Tucker decomposition on the high-order spatiotemporal tensor to extract the core feature tensor includes: A high-order orthogonal iterative algorithm is used to decompose the high-order spatiotemporal tensor and initialize the core tensor and the factor matrices of each mode. During the iteration process, the factor matrices of other modes except the current mode are fixed in turn, and the matrix of the current mode is updated by calculating the product of the higher-order spatiotemporal tensor and the fixed factor matrix modulo n. Perform singular value decomposition on the updated matrix and take the first k left singular vectors as the new factor matrix of the current mode; The iteration stops when the reconstruction error is less than the preset convergence threshold, and the converged core tensor is output as the extracted feature tensor.

[0008] Optionally, the step of using a variational algorithm based on prediction error and Riemannian manifold to iteratively update the local tensor kernel of the matrix multiplication operator in real time includes: Construct a loss function based on the deviation between the network's predicted output and the actual measured value, and calculate the gradient tensor of each local tensor kernel in the matrix multiplication operator in Euclidean space; By utilizing the left and right regularized forms of the matrix product operator, a tangent space projection operator is constructed at the current parameter point; Projecting the Euclidean gradient tensor onto the tangent space yields the Riemann gradient; The shrinkage operator is used to map the updated tangent vector along the Riemann gradient direction back to the manifold surface where the matrix multiplication operator is located, preserving the low-rank structure constraint of the matrix multiplication operator and completing the update of the local tensor kernel.

[0009] Optionally, the calculation of the pseudo-inverse of the voltage-power sensitivity matrix of the distribution network under the current operating conditions includes: Based on the current node admittance matrix of the distribution network, calculate the partial derivatives of the voltage amplitude of each node with respect to the changes in active power injection and reactive power injection. The partial derivatives are arranged and combined to form the Jacobian matrix, which serves as the voltage-power sensitivity matrix. Singular value decomposition is performed on the Jacobian matrix to obtain a left singular vector matrix, a singular value diagonal matrix, and a right singular vector matrix. Take the reciprocal of the non-zero singular values ​​in the singular value diagonal matrix to construct the inverse singular value matrix, and then multiply it by the transpose of the right singular vector matrix and the left singular vector matrix in turn to obtain the Moore-Penrose pseudo-inverse matrix of the sensitivity matrix.

[0010] Optionally, embedding the pseudo-inverse matrix into the weight kernel of the pointwise convolutional layer based on the tensor column decomposition format using tensor shrinkage operation includes: The weight matrix of the pointwise convolutional layer is decomposed into tensor column format to obtain a tensor column sequence containing two edge tensor kernels and several intermediate tensor kernels, where the dimension index of the edge tensor kernel at the input end corresponds to the channel dimension of the feature map. Select the edge tensor kernel at the input end, and align and map the channel dimension index of the edge tensor kernel with the node dimension index of the pseudo-inverse of the sensitivity matrix; Perform a tensor shrinkage operation, multiplying the pseudo-inverse matrix with the selected edge tensor kernel along the aligned dimensions to generate a new edge tensor kernel that incorporates topological constraints; Keeping the other tensor kernel parameters in the tensor column sequence unchanged, the tensor column format weights of the pointwise convolutional layer are reconstructed using the new edge tensor kernel.

[0011] Optionally, the step of inputting the fused feature map into the regression prediction layer to calculate the active and reactive power setpoints for each controllable device includes: Using a fully connected neural network as the regression prediction layer, the original output vector is obtained by flattening the fused feature map and mapping it through linear weighting and activation functions; The elements in the original output vector are mapped one-to-one to each photovoltaic inverter and reactive power compensation device in the distribution network through index mapping. Read the current rated power upper and lower limits of each device; The mapped power values ​​are corrected for exceeding limits. Values ​​exceeding the upper limit of the rated power are corrected to the upper limit value, and values ​​below the lower limit of the rated power are corrected to the lower limit value. Values ​​within the rated power range are retained. The processed values ​​are used as the issued active and reactive power setting values.

[0012] In a second aspect, the present invention also proposes an adaptive control system for low-voltage distribution networks based on edge AI, comprising the following modules: The execution module is used to collect voltage, current and power time-series operation data of key nodes in the low-voltage distribution network, construct a block Hankel matrix after standardizing the data, fold the matrix into a high-order spatiotemporal tensor, and perform Tucker decomposition on the high-order spatiotemporal tensor to extract the core feature tensor. The update module is used to input the core feature tensor into the depthwise separable convolutional network and extract spatial features through the deep convolutional layer. The filter bank of the deep convolutional layer is parameterized into a matrix multiplication operator format. During the control process, the local tensor kernel of the matrix multiplication operator is updated in real time using a variational algorithm based on prediction error and Riemann manifold to adapt to the time-varying non-stationary characteristics of the power distribution network. The generation module is used to perform channel-dimensional feature reorganization on the output of the deep convolutional layer through pointwise convolutional layers, calculate the pseudo-inverse of the voltage-power sensitivity matrix of the distribution network under the current operating conditions, use the pseudo-inverse matrix as prior knowledge, and embed the pseudo-inverse matrix into the weight kernel of the pointwise convolutional layer based on the tensor column decomposition format using tensor shrinkage operation, establish the mapping relationship between feature channels and topology nodes, and generate a feature map with fusion constraints. The control module is used to input the fused feature map into the regression prediction layer, calculate the active and reactive power setpoints of each controllable device, and send them to the device controller to perform closed-loop control of the low-voltage distribution network.

[0013] Preferably, the step of collecting voltage, current, and power time-series operational data from key nodes of the low-voltage distribution network, and constructing a block-based Hankel matrix after standardizing the data, includes: Calculate the mean and standard deviation of the collected voltage, current and power time-series operating data, and perform Z-score standardization, which is to subtract the mean from each data point and then divide by the standard deviation; Set the sliding window length L and the sliding step size S. Starting from the starting point of the standardized time series data, extract data segments as column vectors according to the set length and step size. The multiple column vectors generated in chronological order are arranged sequentially to construct a block Hankel matrix in which each row corresponds to a time step and each column corresponds to a time window feature.

[0014] Preferably, the step of performing Tucker decomposition on the high-order spatiotemporal tensor to extract the core feature tensor includes: A high-order orthogonal iterative algorithm is used to decompose the high-order spatiotemporal tensor and initialize the core tensor and the factor matrices of each mode. During the iteration process, the factor matrices of other modes except the current mode are fixed in turn, and the matrix of the current mode is updated by calculating the product of the higher-order spatiotemporal tensor and the fixed factor matrix modulo n. Perform singular value decomposition on the updated matrix and take the first k left singular vectors as the new factor matrix of the current mode; The iteration stops when the reconstruction error is less than the preset convergence threshold, and the converged core tensor is output as the extracted feature tensor.

[0015] Preferably, the step of using a variational algorithm based on prediction error and Riemannian manifold to perform real-time iterative updates of the local tensor kernel of the matrix multiplication operator includes: Construct a loss function based on the deviation between the network's predicted output and the actual measured value, and calculate the gradient tensor of each local tensor kernel in the matrix multiplication operator in Euclidean space; By utilizing the left and right regularized forms of the matrix product operator, a tangent space projection operator is constructed at the current parameter point; Projecting the Euclidean gradient tensor onto the tangent space yields the Riemann gradient; The shrinkage operator is used to map the updated tangent vector along the Riemann gradient direction back to the manifold surface where the matrix multiplication operator is located, preserving the low-rank structure constraint of the matrix multiplication operator and completing the update of the local tensor kernel.

[0016] Preferably, the calculation of the pseudo-inverse of the voltage-power sensitivity matrix of the distribution network under the current operating conditions includes: Based on the current node admittance matrix of the distribution network, calculate the partial derivatives of the voltage amplitude of each node with respect to the changes in active power injection and reactive power injection. The partial derivatives are arranged and combined to form the Jacobian matrix, which serves as the voltage-power sensitivity matrix. Singular value decomposition is performed on the Jacobian matrix to obtain a left singular vector matrix, a singular value diagonal matrix, and a right singular vector matrix. Take the reciprocal of the non-zero singular values ​​in the singular value diagonal matrix to construct the inverse singular value matrix, and then multiply it by the transpose of the right singular vector matrix and the left singular vector matrix in turn to obtain the Moore-Penrose pseudo-inverse matrix of the sensitivity matrix.

[0017] Preferably, the step of embedding the pseudo-inverse matrix into the weight kernel of the pointwise convolutional layer based on the tensor column decomposition format using tensor shrinkage operation includes: The weight matrix of the pointwise convolutional layer is decomposed into tensor column format to obtain a tensor column sequence containing two edge tensor kernels and several intermediate tensor kernels, where the dimension index of the edge tensor kernel at the input end corresponds to the channel dimension of the feature map. Select the edge tensor kernel at the input end, and align and map the channel dimension index of the edge tensor kernel with the node dimension index of the pseudo-inverse of the sensitivity matrix; Perform a tensor shrinkage operation, multiplying the pseudo-inverse matrix with the selected edge tensor kernel along the aligned dimensions to generate a new edge tensor kernel that incorporates topological constraints; Keeping the other tensor kernel parameters in the tensor column sequence unchanged, the tensor column format weights of the pointwise convolutional layer are reconstructed using the new edge tensor kernel.

[0018] Preferably, the step of inputting the fused feature map into the regression prediction layer to calculate the active and reactive power setpoints for each controllable device includes: Using a fully connected neural network as the regression prediction layer, the original output vector is obtained by flattening the fused feature map and mapping it through linear weighting and activation functions; The elements in the original output vector are mapped one-to-one to each photovoltaic inverter and reactive power compensation device in the distribution network through index mapping. Read the current rated power upper and lower limits of each device; The mapped power values ​​are corrected for exceeding limits. Values ​​exceeding the upper limit of the rated power are corrected to the upper limit value, and values ​​below the lower limit of the rated power are corrected to the lower limit value. Values ​​within the rated power range are retained. The processed values ​​are used as the issued active and reactive power setting values.

[0019] This invention eliminates data redundancy and extracts the spatiotemporal evolution features of low-voltage distribution networks by constructing a block-based Hankel matrix and a Tucker decomposition of high-order spatiotemporal tensors. Utilizing a matrix multiplication operator format to parameterize deep convolutional layers reduces model complexity. Combined with a Riemannian manifold-based variational algorithm update mechanism, numerical stability is ensured while adapting to the time-varying and non-stationary characteristics of the distribution network. Using the voltage-power sensitivity matrix as a priori, a deep mapping between data features and topology is constructed by embedding it into the pointwise convolutional layer weights through tensor shrinkage. This data fusion architecture uses constraints to guide the model optimization direction, improving the accuracy and convergence speed of active and reactive power setpoint calculations, thereby achieving closed-loop regulation of the low-voltage distribution network's operating state and ensuring the stable operation of the power grid. Attached Figure Description

[0020] Figure 1 A flowchart of the first embodiment; Figure 2 A schematic diagram of a physically constrained pointwise convolutional layer; Detailed Implementation

[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] In the first embodiment, the present invention proposes an adaptive control method for low-voltage distribution networks based on edge AI, such as... Figure 1 This includes the following steps: S1. Collect voltage, current and power time-series operation data of key nodes in low-voltage distribution network, construct block Hankel matrix after standardizing the data, and fold the matrix into a high-order spatiotemporal tensor. Perform Tucker decomposition on the high-order spatiotemporal tensor to extract core feature tensors. Specifically, miniature synchronous phasor measurement units are installed at the common connection point of the low-voltage distribution network, the low-voltage side outlet of the distribution transformer, and each user terminal. The sampling frequency is set to 50 times per second to synchronously collect time series data of three-phase voltage amplitude, three-phase current amplitude, active power, and reactive power. The mean and standard deviation of each data channel are calculated using the Z-score standardization method. The original data is subtracted from the mean and divided by the standard deviation to complete the standardization process. The time delay window length and sliding step size are set, and the standardized multi-channel one-dimensional time series data are arranged according to the time delay embedding method to construct a block Hankel matrix containing historical time series correlation. The block Hankel matrix is ​​reshaped according to the node dimension, time delay dimension, and feature channel dimension to obtain a third-order or higher-order spatiotemporal tensor. The high-order spatiotemporal tensor is decomposed using a high-order orthogonal iterative algorithm. The target rank of the core tensor is set, and the projection matrix of each mode is calculated iteratively by alternating least squares method. The high-order spatiotemporal tensor is projected onto the low-dimensional subspace to output the core feature tensor with compressed dimensions and retention of the main spatiotemporal energy distribution.

[0023] In an optional embodiment, the step of collecting voltage, current, and power time-series operational data from key nodes of the low-voltage distribution network, and constructing a block-based Hankel matrix after standardizing the data, includes: Calculate the mean and standard deviation of the collected voltage, current and power time-series operating data, and perform Z-score standardization, which is to subtract the mean from each data point and then divide by the standard deviation; Set the sliding window length L and the sliding step size S. Starting from the starting point of the standardized time series data, extract data segments as column vectors according to the set length and step size. The multiple column vectors generated in chronological order are arranged sequentially to construct a block Hankel matrix in which each row corresponds to a time step and each column corresponds to a time window feature.

[0024] The raw data collected by synchronous phasor measurement units or smart meters deployed in low-voltage distribution networks is preprocessed. Assuming a sampling frequency of once every 15 minutes, the average values ​​of voltage, current, and power data within the historical time window are calculated. and standard deviation For each newly acquired data point x, perform a standardization operation. This eliminates the order-of-magnitude differences between different dimensions, preventing gradient vanishing or exploding.

[0025] A segmented Hankel matrix is ​​constructed to detect trajectory features in a time series. The sliding window length L is set to 96, and the sliding step size S is 1. Starting from the beginning of the standardized data sequence, the first data segment of length 96 is extracted as the first column vector of the matrix. , , , Slide backward according to step size S to intercept. , , , As the second column, and so on. Arranging these column vectors horizontally in sequence yields an L×K dimensional block Hankel matrix. This matrix effectively reconstructs the system's phase space, with each column representing a complete daily operating cycle characteristic and each row representing the state at the same moment under different historical cycles, thus transforming one-dimensional time-series data into a two-dimensional data structure containing more information.

[0026] In an optional embodiment, the step of performing Tucker decomposition on the higher-order spatiotemporal tensor to extract the core feature tensor includes: A high-order orthogonal iterative algorithm is used to decompose the high-order spatiotemporal tensor and initialize the core tensor and the factor matrices of each mode. During the iteration process, the factor matrices of other modes except the current mode are fixed in turn, and the matrix of the current mode is updated by calculating the product of the higher-order spatiotemporal tensor and the fixed factor matrix modulo n. Perform singular value decomposition on the updated matrix and take the first k left singular vectors as the new factor matrix of the current mode; The iteration stops when the reconstruction error is less than the preset convergence threshold, and the converged core tensor is output as the extracted feature tensor.

[0027] The constructed Hankel matrix is ​​folded into a third-order tensor. ,in Represents the time window dimension. Represents the spatial node dimension. Representing the feature dimension. A higher-order orthogonal iterative algorithm is used for processing, and the factor matrix is ​​initialized using the results of the higher-order singular value decomposition. , , Set the target rank of the core tensor to be... , , This rank is much smaller than the original dimension and is used to remove measurement noise and redundant information.

[0028] The algorithm enters an alternating least squares iterative loop. For example, when updating the factor matrix of the spatial pattern, the factor matrices of the temporal pattern and the feature pattern are fixed, and the tensor is calculated. and , The intermediate tensor is obtained by modular multiplication, and then expanded into a matrix along the spatial node dimensions. Truncated singular value decomposition is performed on the expanded matrix to extract the values ​​corresponding to the previous nodes. The left singular vectors of the maximal singular values ​​form a new factor matrix. The iteration stops when the reconstruction error is less than a preset threshold, and the output core tensor is... It is a highly compressed tensor that removes high-frequency noise while retaining the main voltage fluctuation modes and power flow patterns in the distribution network.

[0029] S2 inputs the core feature tensor into a depthwise separable convolutional network and extracts spatial features through deep convolutional layers. The filter bank of the deep convolutional layers is parameterized into a matrix multiplication operator format. During the control process, a variational algorithm based on prediction error and Riemann manifold is used to iteratively update the local tensor kernel of the matrix multiplication operator in real time to adapt to the time-varying non-stationary characteristics of the power distribution network. Specifically, a depthwise separable convolutional network model is constructed, and the core feature tensor is fed into the first deep convolutional layer as input data. The original four-dimensional convolutional kernel tensor in the deep convolutional layer is reshaped into a matrix form, and then the matrix is ​​further decomposed into a sequence of matrix product operators composed of a series of low-rank third-order tensor kernels connected in sequence using singular value decomposition. This sequence is used as the parameterized representation of the filter. In the online adjustment stage, the prediction error loss function between the network output value and the actual measurement value is calculated in real time. The sequence of matrix product operators is regarded as points located on a fixed-rank Riemannian manifold, and the Euclidean gradient of the loss function with respect to the current matrix product operator parameters is calculated. The Euclidean gradient is projected onto the tangent space where the current parameter point is located to obtain the Riemann gradient. The parameters of the local tensor kernel are updated along the tangent space direction using the Riemann gradient descent method, and the updated parameters are mapped back from the tangent space to the matrix product operator manifold through a shrinking mapping operation to complete the real-time iterative update of the filter parameters.

[0030] The input to the depthwise separable convolutional network is the core feature tensor after Tucker decomposition. The network structure mainly consists of matrix multiplication operator deep convolutional layers, tensor column pointwise convolutional layers, and fully connected regression prediction layers. The matrix multiplication operator deep convolutional layers extract spatial features using convolution kernels in the form of matrix multiplication operators; the tensor column pointwise convolutional layers fuse channel features using a weight matrix in tensor column format; and the fully connected regression prediction layers map the feature maps to control variables. The training set consists of historical voltage, current, and power time-series data of the distribution network, and a sliding window method is used to construct the input samples and labels. The training process aims to minimize the prediction error, calculating the mean squared error loss function between the predicted output and the true value, and updating the network parameters using a gradient descent algorithm on a Riemannian manifold. The network output is the active and reactive power setpoints required by each controllable device in the distribution network.

[0031] In an optional embodiment, the step of using a variational algorithm based on prediction error and Riemannian manifold to iteratively update the local tensor kernel of the matrix multiplication operator in real time includes: Construct a loss function based on the deviation between the network's predicted output and the actual measured value, and calculate the gradient tensor of each local tensor kernel in the matrix multiplication operator in Euclidean space; By utilizing the left and right regularized forms of the matrix product operator, a tangent space projection operator is constructed at the current parameter point; Projecting the Euclidean gradient tensor onto the tangent space yields the Riemann gradient; The shrinkage operator is used to map the updated tangent vector along the Riemann gradient direction back to the manifold surface where the matrix multiplication operator is located, preserving the low-rank structure constraint of the matrix multiplication operator and completing the update of the local tensor kernel.

[0032] At time t, the actual measured state of the power grid is obtained. And read the prediction value made by the previous time t-1 for the current time. The mean squared error between the two is calculated as the loss function. At this point, the local tensor kernels of the matrix multiplication operators in the convolutional layer are calculated through backpropagation. Euclidean gradient .

[0033] To preserve the low-rank structure of the matrix product operator, a Riemannian manifold optimization method is employed. The matrix product operator is rearranged into a mixed normalized form through singular value decomposition, such as a left-right orthogonalized mixed normalized form, ensuring the uniqueness of the tangent space projection operator. This is used to construct the tangent space projection operator. Projecting the Euclidean gradient onto the tangent space of the fixed-rank manifold containing the current parameter point W yields the Riemannian gradient. Set the online learning rate. The updated tangent vector is calculated. A retraction operator, such as the SVD retraction operator or the QR retraction operator, is applied to remap the updated tangent vector back to the manifold surface. After the parameter update is complete, the updated model is used to predict and control the state at the next time step t+1. This process ensures that the model parameters evolve in real time with the power grid's operating state and logically conform to causality.

[0034] S3, through pointwise convolutional layers, the output of the deep convolutional layer is reorganized in the channel dimension, the pseudo-inverse of the voltage-power sensitivity matrix of the distribution network under the current working condition is calculated, the pseudo-inverse matrix is ​​used as prior knowledge, and the pseudo-inverse matrix is ​​embedded into the weight kernel of the pointwise convolutional layer based on the tensor column decomposition format using tensor contraction operation, the mapping relationship between feature channels and topology nodes is established, and a feature map with fusion constraints is generated. Specifically, the spatial feature map output from the deep convolutional layer is received and linearly combined between channels using pointwise convolutional layers with a kernel size of 1×1. Based on the voltage and power data collected at the current moment, the Jacobian matrix is ​​calculated using the Newton-Raphson power flow method or the partial derivatives of the voltage amplitude with respect to active and reactive power are calculated using the incremental perturbation method to construct a voltage-power sensitivity matrix. The Moore-Penrose pseudo-inverse matrix of this sensitivity matrix is ​​calculated using the singular value decomposition method. The weight matrix of the pointwise convolutional layer is decomposed into a tensor column format, that is, expressed as a product of multiple third-order tensor kernels, such as... Figure 2 As shown; select the tensor kernel corresponding to the input channel dimension in the tensor column, and perform tensor shrinkage operation on the calculated sensitivity pseudo-inverse matrix and the tensor kernel, that is, multiply the elements of the pseudo-inverse matrix as constraint coefficients by the previous weight parameters, forcing the network weights to include the inverse dynamics of the topology; perform convolution operation to output a fused feature map that integrates topological constraints and data features.

[0035] In an optional embodiment, calculating the pseudo-inverse of the voltage-power sensitivity matrix of the distribution network under the current operating conditions includes: Based on the current node admittance matrix of the distribution network, calculate the partial derivatives of the voltage amplitude of each node with respect to the changes in active power injection and reactive power injection. The partial derivatives are arranged and combined to form the Jacobian matrix, which serves as the voltage-power sensitivity matrix. Singular value decomposition is performed on the Jacobian matrix to obtain a left singular vector matrix, a singular value diagonal matrix, and a right singular vector matrix. Take the reciprocal of the non-zero singular values ​​in the singular value diagonal matrix to construct the inverse singular value matrix, and then multiply it by the transpose of the right singular vector matrix and the left singular vector matrix in turn to obtain the Moore-Penrose pseudo-inverse matrix of the sensitivity matrix.

[0036] Using the topology information of the distribution network, a node admittance matrix is ​​generated based on the latest topology and line parameters. The sensitivity coefficient is calculated by solving the partial derivatives of the power flow equation. For a network with N nodes, the voltage amplitude at each node is calculated. Active power injection for all nodes and reactive power injection The partial derivatives are then combined into an N×2N Jacobian matrix J.

[0037] Since the Jacobian matrix is ​​usually not a square matrix and may be close to singular, singular value decomposition is performed on matrix J, i.e. To enhance numerical stability, a truncation threshold is set, and singular values ​​smaller than this threshold are treated as noise and set to zero. For singular values ​​larger than the threshold, the reciprocal of the singular value is calculated. Construct a diagonal matrix Through matrix multiplication The Moore-Penrose pseudo-inverse matrix was calculated. It represents the ideal power adjustment required to produce a unit voltage change, mapping the voltage deviation space back to the power control space.

[0038] In an optional embodiment, embedding the pseudo-inverse matrix into the weight kernel of the pointwise convolutional layer based on the tensor column decomposition format using tensor shrinkage operation includes: The weight matrix of the pointwise convolutional layer is decomposed into tensor column format to obtain a tensor column sequence containing two edge tensor kernels and several intermediate tensor kernels, where the dimension index of the edge tensor kernel at the input end corresponds to the channel dimension of the feature map. Select the edge tensor kernel at the input end, and align and map the channel dimension index of the edge tensor kernel with the node dimension index of the pseudo-inverse of the sensitivity matrix; Perform a tensor shrinkage operation, multiplying the pseudo-inverse matrix with the selected edge tensor kernel along the aligned dimensions to generate a new edge tensor kernel that incorporates topological constraints; Keeping the other tensor kernel parameters in the tensor column sequence unchanged, the tensor column format weights of the pointwise convolutional layer are reconstructed using the new edge tensor kernel.

[0039] The original fully connected weight matrix of the pointwise convolutional layer Decomposed into a tensor column format, consisting of a series of low-rank core tensors , , , Composition. Identify the first edge tensor kernel corresponding to the channel dimension of the input feature map. The dimension is usually 1× × Due to the channel dimension of the feature map The previous steps preserved the node-related topological information, and the calculated pseudo-inverse of the sensitivity matrix was then used. As a projection operator.

[0040] calculate ,in, To shrink along the second mode of the tensor, matrix-tensor multiplication is performed along the feature channel dimension. The data-driven weight space of the neural network is projected onto a subspace defined by sensitivity, forcing the network to focus on the node features most sensitive to voltage regulation. The generated... Replace the original Meanwhile, the other core tensors in the sequence remain unchanged. The reconstructed convolutional layer weights incorporate the topological constraints of the distribution network, making the network output regulation strategy naturally conform to the rules.

[0041] S4 inputs the fused feature map into the regression prediction layer, calculates the active and reactive power setpoints of each controllable device, and sends them to the device controller for closed-loop control of the low-voltage distribution network.

[0042] Specifically, a regression prediction layer consisting of a global average pooling layer and a fully connected layer is constructed. The fused feature map is input into the global average pooling layer to compress the spatial dimension into a single numerical vector. This vector is then input into the fully connected layer, where the number of output nodes corresponds to twice the number of photovoltaic inverters, energy storage converters, and static var generators in the distribution network, corresponding to active and reactive power commands, respectively. At the end of the fully connected layer, a hyperbolic tangent activation function is used to scale the output to the adjustable power range of each device, obtaining the active and reactive power setpoints for each controllable device. The calculated power setpoints are packaged into control messages using the Modbus TCP communication protocol or the IEC61850 protocol and sent to the local controllers of each device in the field via Ethernet. After parsing the messages, the local controllers adjust the duty cycle of the power electronic switches to change the output power, thereby achieving closed-loop regulation of voltage deviation and power flow distribution in the low-voltage distribution network.

[0043] In an optional embodiment, the step of inputting the fused feature map into the regression prediction layer and calculating the active and reactive power setpoints for each controllable device includes: Using a fully connected neural network as the regression prediction layer, the original output vector is obtained by flattening the fused feature map and mapping it through linear weighting and activation functions; The elements in the original output vector are mapped one-to-one to each photovoltaic inverter and reactive power compensation device in the distribution network through index mapping. Read the current rated power upper and lower limits of each device; The mapped power values ​​are corrected for exceeding limits. Values ​​exceeding the upper limit of the rated power are corrected to the upper limit value, and values ​​below the lower limit of the rated power are corrected to the lower limit value. Values ​​within the rated power range are retained. The processed values ​​are used as the issued active and reactive power setting values.

[0044] After undergoing global average pooling or flattening, the fused feature maps are input into a fully connected layer, with an output dimension of 2×. The vector is used. An activation function is used to map the output range to the device's capacity range. Real-time readings are performed on the operating constraint parameters of each device at the current moment, as well as the current thermal stability limits of critical branches. .

[0045] Perform hard thresholding correction on the raw output of the neural network, if Then set the value ;like Then let For reactive power, the capacity circle constraint must be satisfied. Furthermore, in order to utilize the current data collected above, a sensitivity matrix can be used to quickly verify whether the current setting will cause branch current overload. If the current of a certain branch is predicted... Based on the current-power sensitivity, the power setting value of the device that contributes the most to the current of that branch is reduced until the safety constraints are met. The generated instruction sequence is sent to the local controller of each device through the communication protocol to complete the closed-loop control.

[0046] In a second embodiment, the present invention also provides an adaptive control system for low-voltage distribution networks based on edge AI, comprising the following modules: The execution module is used to collect voltage, current and power time-series operation data of key nodes in the low-voltage distribution network, construct a block Hankel matrix after standardizing the data, fold the matrix into a high-order spatiotemporal tensor, and perform Tucker decomposition on the high-order spatiotemporal tensor to extract the core feature tensor. The update module is used to input the core feature tensor into the depthwise separable convolutional network and extract spatial features through the deep convolutional layer. The filter bank of the deep convolutional layer is parameterized into a matrix multiplication operator format. During the control process, the local tensor kernel of the matrix multiplication operator is updated in real time using a variational algorithm based on prediction error and Riemann manifold to adapt to the time-varying non-stationary characteristics of the power distribution network. The generation module is used to perform channel-dimensional feature reorganization on the output of the deep convolutional layer through pointwise convolutional layers, calculate the pseudo-inverse of the voltage-power sensitivity matrix of the distribution network under the current operating conditions, use the pseudo-inverse matrix as prior knowledge, and embed the pseudo-inverse matrix into the weight kernel of the pointwise convolutional layer based on the tensor column decomposition format using tensor shrinkage operation, establish the mapping relationship between feature channels and topology nodes, and generate a feature map with fusion constraints. The control module is used to input the fused feature map into the regression prediction layer, calculate the active and reactive power setpoints of each controllable device, and send them to the device controller to perform closed-loop control of the low-voltage distribution network.

[0047] In this specification, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise limited, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. In this document, "a," "an," "the," "the," and "its" may also include plural forms unless the context clearly indicates otherwise. "Multiple" refers to at least two, such as 2, 3, 5, or 8, etc. "And / or" includes any and all combinations of the associated listed items.

[0048] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0049] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An adaptive control method for low-voltage distribution networks based on edge AI, characterized in that, Includes the following steps: Voltage, current and power time-series operation data of key nodes in low-voltage distribution network are collected. After standardization of the data, a block Hankel matrix is ​​constructed, and the matrix is ​​folded into a high-order spatiotemporal tensor. Tucker decomposition is performed on the high-order spatiotemporal tensor to extract the core feature tensor. The core feature tensor is input into a depthwise separable convolutional network, and spatial features are extracted through deep convolutional layers. The filter bank of the deep convolutional layer is parameterized into a matrix multiplication operator format. During the control process, a variational algorithm based on prediction error and Riemann manifold is used to iteratively update the local tensor kernel of the matrix multiplication operator in real time to adapt to the time-varying non-stationary characteristics of the distribution network. By performing channel-dimensional feature reorganization on the output of the deep convolutional layer through pointwise convolutional layers, the pseudo-inverse of the voltage-power sensitivity matrix of the distribution network under the current operating conditions is calculated. The pseudo-inverse matrix is ​​used as prior knowledge, and the pseudo-inverse matrix is ​​embedded into the weight kernel of the pointwise convolutional layer based on the tensor column decomposition format using tensor shrinkage operation. The mapping relationship between feature channels and topology nodes is established, and a feature map with fusion constraints is generated. The fused feature map is input into the regression prediction layer to calculate the active and reactive power setpoints of each controllable device, and then sent to the device controller to perform closed-loop control of the low-voltage distribution network.

2. The method according to claim 1, characterized in that, The process involves collecting voltage, current, and power time-series operational data from key nodes in the low-voltage distribution network, standardizing the data, and constructing a block-based Hankel matrix, including: Calculate the mean and standard deviation of the collected voltage, current and power time-series operating data, and perform Z-score standardization, which is to subtract the mean from each data point and then divide by the standard deviation; Set the sliding window length L and the sliding step size S. Starting from the starting point of the standardized time series data, extract data segments as column vectors according to the set length and step size. The multiple column vectors generated in chronological order are arranged sequentially to construct a block Hankel matrix in which each row corresponds to a time step and each column corresponds to a time window feature.

3. The method according to claim 1, characterized in that, The process of extracting core feature tensors by performing Tucker decomposition on high-order spatiotemporal tensors includes: A high-order orthogonal iterative algorithm is used to decompose the high-order spatiotemporal tensor and initialize the core tensor and the factor matrices of each mode; During the iteration process, the factor matrices of other modes except the current mode are fixed in turn, and the matrix of the current mode is updated by calculating the product of the higher-order spatiotemporal tensor and the fixed factor matrix modulo n. Perform singular value decomposition on the updated matrix and take the first k left singular vectors as the new factor matrix of the current mode; The iteration stops when the reconstruction error is less than the preset convergence threshold, and the converged core tensor is output as the extracted feature tensor.

4. The method according to claim 2, characterized in that, The method of using a variational algorithm based on prediction error and Riemannian manifold to perform real-time iterative updates of the local tensor kernel of the matrix multiplication operator includes: Construct a loss function based on the deviation between the network's predicted output and the actual measured value, and calculate the gradient tensor of each local tensor kernel in the matrix multiplication operator in Euclidean space; By utilizing the left and right regularized forms of the matrix product operator, a tangent space projection operator is constructed at the current parameter point; Projecting the Euclidean gradient tensor onto the tangent space yields the Riemann gradient; The shrinkage operator is used to map the updated tangent vector along the Riemann gradient direction back to the manifold surface where the matrix multiplication operator is located, preserving the low-rank structure constraint of the matrix multiplication operator and completing the update of the local tensor kernel.

5. The method according to claim 1, characterized in that, The calculation of the pseudo-inverse of the voltage-power sensitivity matrix of the distribution network under the current operating conditions includes: Based on the current node admittance matrix of the distribution network, calculate the partial derivatives of the voltage amplitude of each node with respect to the changes in active power injection and reactive power injection. The partial derivatives are arranged and combined to form the Jacobian matrix, which serves as the voltage-power sensitivity matrix. Singular value decomposition is performed on the Jacobian matrix to obtain a left singular vector matrix, a singular value diagonal matrix, and a right singular vector matrix. Take the reciprocal of the non-zero singular values ​​in the singular value diagonal matrix to construct the inverse singular value matrix, and then multiply it by the transpose of the right singular vector matrix and the left singular vector matrix in turn to obtain the Moore-Penrose pseudo-inverse matrix of the sensitivity matrix.

6. The method according to claim 1, characterized in that, The step of embedding the pseudo-inverse matrix into the weight kernel of the pointwise convolutional layer based on the tensor column decomposition format using tensor shrinkage operation includes: The weight matrix of the pointwise convolutional layer is decomposed into tensor column format to obtain a tensor column sequence containing two edge tensor kernels and several intermediate tensor kernels, where the dimension index of the edge tensor kernel at the input end corresponds to the channel dimension of the feature map. Select the edge tensor kernel at the input end, and align and map the channel dimension index of the edge tensor kernel with the node dimension index of the pseudo-inverse of the sensitivity matrix; Perform a tensor shrinkage operation, multiplying the pseudo-inverse matrix with the selected edge tensor kernel along the aligned dimensions to generate a new edge tensor kernel that incorporates topological constraints; Keeping the other tensor kernel parameters in the tensor column sequence unchanged, the tensor column format weights of the pointwise convolutional layer are reconstructed using the new edge tensor kernel.

7. The method according to claim 5, characterized in that, The step of inputting the fused feature map into the regression prediction layer to calculate the active and reactive power setpoints for each controllable device includes: Using a fully connected neural network as the regression prediction layer, the original output vector is obtained by flattening the fused feature map and mapping it through linear weighting and activation functions; The elements in the original output vector are mapped one-to-one to each photovoltaic inverter and reactive power compensation device in the distribution network through index mapping. Read the current rated power upper and lower limits of each device; The mapped power values ​​are corrected for exceeding limits. Values ​​exceeding the upper limit of the rated power are corrected to the upper limit value, and values ​​below the lower limit of the rated power are corrected to the lower limit value. Values ​​within the rated power range are retained. The processed values ​​are used as the issued active and reactive power setting values.

8. A low-voltage distribution network adaptive control system based on edge AI, characterized in that, Includes the following modules: The execution module is used to collect voltage, current and power time-series operation data of key nodes in the low-voltage distribution network, construct a block Hankel matrix after standardizing the data, fold the matrix into a high-order spatiotemporal tensor, and perform Tucker decomposition on the high-order spatiotemporal tensor to extract the core feature tensor. The update module is used to input the core feature tensor into the depthwise separable convolutional network and extract spatial features through the deep convolutional layer. The filter bank of the deep convolutional layer is parameterized into a matrix multiplication operator format. During the control process, the local tensor kernel of the matrix multiplication operator is updated in real time using a variational algorithm based on prediction error and Riemann manifold to adapt to the time-varying non-stationary characteristics of the power distribution network. The generation module is used to perform channel-dimensional feature reorganization on the output of the deep convolutional layer through pointwise convolutional layers, calculate the pseudo-inverse of the voltage-power sensitivity matrix of the distribution network under the current operating conditions, use the pseudo-inverse matrix as prior knowledge, and embed the pseudo-inverse matrix into the weight kernel of the pointwise convolutional layer based on the tensor column decomposition format using tensor shrinkage operation, establish the mapping relationship between feature channels and topology nodes, and generate a feature map with fusion constraints. The control module is used to input the fused feature map into the regression prediction layer, calculate the active and reactive power setpoints of each controllable device, and send them to the device controller to perform closed-loop control of the low-voltage distribution network.

9. The system according to claim 8, characterized in that, The process involves collecting voltage, current, and power time-series operational data from key nodes in the low-voltage distribution network, standardizing the data, and constructing a block-based Hankel matrix, including: Calculate the mean and standard deviation of the collected voltage, current and power time-series operating data, and perform Z-score standardization, which is to subtract the mean from each data point and then divide by the standard deviation; Set the sliding window length L and the sliding step size S. Starting from the starting point of the standardized time series data, extract data segments as column vectors according to the set length and step size. The multiple column vectors generated in chronological order are arranged sequentially to construct a block Hankel matrix in which each row corresponds to a time step and each column corresponds to a time window feature.

10. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program, when executed by a processor, implements the method as described in any one of claims 1-7.

Images