Power distribution system operation optimization control method and system, device, storage medium

Abstract

The application provides a power distribution system operation optimization control method and system, equipment and a storage medium, and relates to the technical field of power distribution network operation control. The method comprises the following steps: performing space-time alignment and feature fusion on the multi-source space-time data of each node for the current control period and the physical power flow boundary parameters of the last control period to obtain a space-time feature matrix; constructing a dynamic adjacency matrix based on the multi-source space-time data and the nominal node admittance of the power distribution system; inputting the space-time feature matrix and the dynamic adjacency matrix into a space-time prediction model to output the active power prediction interval boundary of each node in the future period, and converting the probability prediction parameters; inputting the prediction interval and the probability prediction parameters as constraint parameters into a differentiable convex optimization layer, constructing a convex optimization problem with the minimum operation cost of the power distribution system as the target, and solving the convex optimization problem to obtain the optimization control instructions of each node and send the optimization control instructions to the end-side equipment for execution. The application solves the problems of low accuracy and safety of the operation control of the power distribution system.

Description

Technical Field

[0001] This invention relates to the field of power distribution network operation control technology, specifically to a power distribution system operation optimization control method, system, equipment, and storage medium. Background Technology

[0002] Distribution systems are evolving towards new types of systems with a high proportion of renewable energy and power electronic equipment. The dual random fluctuations of source and load make the system's operating state constantly changing. Operation control technology is the core of ensuring the safe and economical operation of the distribution system. Traditional distribution network control mainly relies on centralized optimization methods based on physical models, such as Optimal Power Flow (OPF) calculation, which requires collecting data from the entire network for unified solution. As the scale of distribution network nodes expands to tens of thousands and source and load fluctuations intensify, centralized methods face bottlenecks such as significantly increased computational complexity with the number of nodes, high communication latency, and high risk of single-point failures, making it difficult to meet the real-time control needs of massive distributed photovoltaic, energy storage, and flexible loads.

[0003] The existing common approach to address these uncertainties is a two-stage, sequential model of prediction-then-optimization (PTO). However, this approach treats prediction and optimization as two isolated steps. The prediction model is trained to minimize statistical indicators such as mean square error, while the control decision model aims to minimize system operating costs or voltage over-limit risks. This results in high-precision predictions not necessarily leading to optimal decisions, exhibiting a structural mismatch between prediction and decision objectives. Furthermore, the predicted output data serves as direct input to control optimization, amplifying prediction errors through constraints such as nonlinear power flow relationships in the distribution network and limitations on equipment capacity and voltage limits, ultimately leading to decision-making errors. Summary of the Invention

[0004] To overcome the problems of low accuracy and safety in the operation control of the aforementioned power distribution system, this invention provides a method, system, equipment, and storage medium for optimizing the operation control of a power distribution system.

[0005] In a first aspect, the present invention provides an operation optimization control method for a power distribution system, applied to an edge computing device, comprising: The multi-source spatiotemporal data of each node in the power distribution system for the current control cycle and the physical power flow boundary parameters of the previous control cycle are spatiotemporally aligned and feature fused to obtain the spatiotemporal feature matrix of the current control cycle; the physical power flow boundary parameters are determined based on the voltage over-limit and equipment overload conditions of each node. Based on the multi-source spatiotemporal data and the nominal node admittance of the power distribution system, a dynamic adjacency matrix for the current control cycle is constructed; the nominal node admittance is a node admittance matrix constructed based on a preset benchmark topology. The spatiotemporal feature matrix and the dynamic adjacency matrix are input into the spatiotemporal prediction model, and the active power prediction interval boundary of each node in the future time period is output. The probability prediction parameters of active power are calculated based on the prediction interval boundary. The prediction interval and the probability prediction parameters are used as constraint parameters to input a differentiable convex optimization layer. A convex optimization problem with the goal of minimizing the operating cost of the power distribution system is constructed and solved. The optimized control commands for each node in the current control cycle are obtained and sent to the end-side equipment for execution.

[0006] In a second aspect, the present invention also provides an edge computing device for a power distribution system, comprising: The feature space construction module is used to perform spatiotemporal alignment and feature fusion on the multi-source spatiotemporal data of each node of the power distribution system for the current control cycle and the physical power flow boundary parameters of the previous control cycle to obtain the spatiotemporal feature matrix of the current control cycle; the physical power flow boundary parameters are determined based on the voltage over-limit and equipment overload conditions of each node. The topology modeling module is used to construct a dynamic adjacency matrix for the current control cycle based on the multi-source spatiotemporal data and the nominal node admittance of the power distribution system; the nominal node admittance is a node admittance matrix constructed based on a preset benchmark topology. The prediction module is used to input the spatiotemporal feature matrix and the dynamic adjacency matrix into the spatiotemporal prediction model, output the active power prediction interval boundary of each node in the future time period, and calculate the probability prediction parameters of active power based on the prediction interval boundary. The optimization module is used to input the prediction interval and the probability prediction parameters as constraint parameters into the differentiable convex optimization layer, construct and solve a convex optimization problem with the goal of minimizing the operating cost of the power distribution system, obtain the optimized control commands for each node in the current control cycle, and send them to the end-side equipment for execution.

[0007] On the other hand, the present invention also provides an operation optimization control system for a power distribution system, comprising: an edge computing device as described in any of the preceding claims.

[0008] On the other hand, the present invention also provides an electronic device, comprising: at least one processor and a memory; the memory and the processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, the method described in any of the foregoing is implemented.

[0009] On the other hand, the present invention also provides a readable storage medium having an executable program stored thereon, wherein when the executable program is executed, it implements the method described in any one of the above.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention provides a method, device, and system for optimizing the operation of a power distribution system. By using the physical power flow boundary parameters determined based on voltage exceedances and equipment overload conditions from the previous control cycle as model input features, the spatiotemporal prediction network can explicitly perceive nodes in a constrained critical state during the previous control cycle when extracting spatiotemporal features. This guides the network to automatically increase its focus on critical sections during inference, improving the consistency between prediction and control optimization objectives. Furthermore, by using a dynamic adjacency matrix constructed based on multi-source spatiotemporal data and nominal node admittance as model input features, the spatiotemporal prediction model can adaptively track topology changes caused by switch reconfiguration, fault isolation, etc., in new power distribution systems, improving the dynamic applicability of the prediction results.

[0011] By directly inputting the probability prediction parameters (the boundary of the prediction interval and its converted mean and covariance) as the constraint parameters of the differentiable convex optimization layer, the range information and uncertainty information of the prediction output can participate in the optimization decision, fundamentally solving the problem of prediction deviation being amplified by nonlinear power flow, and realizing the synergistic optimization of prediction and decision. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating an operation optimization control method for a power distribution system according to the present invention. Figure 2 This is an example of the overall architecture diagram of a cloud-edge-device collaborative system according to the present invention; Figure 3 This is a schematic diagram of the data flow of the operation optimization control process in a spatiotemporal prediction model, as an example of the present invention. Figure 4 This is a schematic diagram of the feature fusion and spatiotemporal encoder internal structure and data flow, as exemplified by the present invention. Figure 5 This is a schematic diagram of the internal structure and data flow of a dual-branch probability prediction head according to an example of the present invention; Figure 6 This is a structural block diagram of an electronic device according to the present invention. Detailed Implementation

[0013] Modern power distribution network control typically relies on a cloud-edge-device collaborative architecture, performing rapid power prediction and optimization decisions based on real-time and historical measurement data at both the edge (substation / distribution area) and the endpoint (equipment). Currently, the engineering and academic communities widely employ a two-stage approach (Predict-then-Optimize) that combines data-driven approaches with mechanistic models to address control problems under uncertain conditions. The first stage (prediction) uses machine learning models such as Long Short-Term Memory Network (LSTM) or Graph Neural Network to predict future photovoltaic power output and load demand. The training objective is to minimize the mean square error (MSE). The second stage (optimization) involves substituting the prediction results as deterministic constants into the optimal power flow (OPF) model and using solvers such as the traditional interior point method to calculate the control commands.

[0014] This approach has the advantages of clear logic and easy modular deployment. However, this practice of separating complex business logic into two independent modules exposes key technical bottlenecks in practical applications: (1) Structural mismatch between prediction and decision-making objectives: Existing methods treat prediction and optimization as two isolated processes. The prediction model is trained with the goal of minimizing statistical indicators such as mean squared error (MSE), while the control decision model aims to minimize system operating costs or voltage over-limit risks. The inconsistency of their objective functions means that high-precision prediction may not lead to optimal decision-making. For example, when the node voltage is at the center of the safety margin, a large prediction error has no impact on the decision; however, when the node voltage is at the critical point of over-limit, even a tiny prediction deviation can cause the control command to violate physical constraints. Existing technologies cannot make the prediction model aware of this asymmetric decision sensitivity.

[0015] (2) Nonlinear amplification of error and gradient interface blockage: The predicted output is the node output that directly participates in the control optimization process. The power flow relationship of the distribution network and constraints such as equipment capacity and voltage upper and lower limits may turn the prediction deviation into a significant decision deviation.

[0016] (3) Waste of computing resources at the edge: In order to pursue the ultimate prediction accuracy at all times in the two-stage method, complex deep learning models with a large number of parameters are often deployed at the edge. However, since the model fails to focus on the key features that are truly useful for decision-making, a lot of computing power is wasted on redundant accuracy improvement that does not help control, which is difficult to adapt to the objective conditions of limited computing power and memory of edge devices (such as smart terminals in the distribution area).

[0017] Existing technologies suffer from problems such as mismatch between prediction and decision objectives, amplification of nonlinear errors, and low utilization of edge computing power. A single two-stage technology cannot simultaneously achieve the accuracy of distribution network control, physical safety constraints, and computational efficiency.

[0018] Furthermore, existing end-to-end differentiable control schemes often rely on fixed grid topology admittance matrices for forward feature extraction, which cannot adapt to time-varying topology switching caused by dynamic network self-healing and active switch reconfiguration in new distribution systems. This results in the gradients in the backpropagation process failing to detect abrupt changes in physical boundaries. To address these issues, this invention provides an end-to-end differentiable prediction and decision-making joint optimization control method.

[0019] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0020] Example 1 The present invention provides a method for optimizing the operation and control of a power distribution system, the schematic diagram of which is shown below. Figure 1 As shown, the method includes: Step S110: The multi-source spatiotemporal data of each node in the power distribution system for the current control cycle and the physical power flow boundary parameters of the previous control cycle are spatiotemporally aligned and feature fused to obtain the spatiotemporal feature matrix of the current control cycle; the physical power flow boundary parameters are determined based on the voltage over-limit and equipment overload conditions of each node. Step S120: Based on the multi-source spatiotemporal data and the nominal node admittance of the power distribution system, construct the dynamic adjacency matrix for the current control cycle; the nominal node admittance is a node admittance matrix constructed based on a preset reference topology. Step S130: Input the spatiotemporal feature matrix and the dynamic adjacency matrix into the spatiotemporal prediction model, output the active power prediction interval boundary of each node in the future time period, and calculate the probability prediction parameters of active power based on the prediction interval boundary. Step S140: Input the prediction interval and the probability prediction parameters as constraint parameters into the differentiable convex optimization layer, construct and solve a convex optimization problem with the goal of minimizing the operating cost of the power distribution system, obtain the optimized control instructions for each node in the current control cycle, and send them to the end-side equipment for execution.

[0021] In this example implementation, the method is executed by an edge computing device (such as a smart transformer terminal). The system architecture of this invention relies on a cloud-edge-device collaborative system to achieve deep integration of physical and computational technologies, such as... Figure 2 As shown. The cloud can be deployed at the city or provincial level panoramic digital twin platform and cloud computing center, responsible for aggregating and storing massive historical data across regions, performing global status monitoring, cross-regional resource coordination, and model training and distribution. Cloud computing devices train feature fusion networks and spatiotemporal prediction models. After training convergence, the trained model and optimizer configuration of the differentiable convex optimization layer are distributed to various edge computing devices on the edge side. The edge side can rely on edge computing gateways such as substation / transformer area intelligent fusion terminals (TTUs). This layer is responsible for real-time optimization control within the area (obtaining optimization control commands through online inference of multi-source inputs and distributing them). The terminal side includes smart meters, photovoltaic inverter controllers, energy storage PCS, and flexible load terminals in the distribution transformer area. They collect multi-source meteorological and electrical measurements (such as local power and status measurements) according to the sampling period configured in the metering and control system, upload them to the edge layer, and execute the control action commands issued by the edge layer. The control object of this invention is a power distribution system with a high proportion of new energy sources and flexible loads (such as electric vehicles) connected, which is a power distribution network with bidirectional power flow and strong random fluctuation characteristics.

[0022] Each edge computing device first acquires multi-source spatiotemporal data from each node on the edge for the current control cycle. This multi-source spatiotemporal data includes historical measurement data for the current control cycle, such as historical node voltage, historical power injection, and other electrical measurement information, as well as meteorological and environmental information for future periods. Simultaneously, it reads the physical power flow boundary parameters cached from the previous control cycle. These parameters quantitatively characterize the true physical state of the system after the end of the previous control cycle based on the degree to which the actual voltage amplitude of each node exceeds the safe upper and lower limits and whether the equipment has experienced overload operation. The edge computing device performs spatiotemporal alignment and feature fusion processing on these two types of data with different sources and physical meanings. At the same time, the edge computing device also constructs a dynamic adjacency matrix based on the multi-source spatiotemporal data of the current control cycle and the nominal node admittance of the power distribution system. This nominal node admittance is a node admittance matrix pre-constructed based on a preset benchmark topology (such as a fully connected topology when all operational branches are closed), and it participates in the construction of the dynamic adjacency matrix as a physical topology prior. The introduction of the dynamic adjacency matrix enables the spatiotemporal prediction model to perceive topological changes in the distribution network caused by switching operations, fault isolation, or network reconfiguration, avoiding model errors introduced by static topology assumptions. Subsequently, the fused spatiotemporal feature matrix and the dynamic adjacency matrix are input into the spatiotemporal prediction model, which employs a spatiotemporal graph convolutional network architecture. This model extracts spatiotemporal features on the graph structure defined by the dynamic adjacency matrix and outputs the active power prediction interval boundary for each uncertain node in the future time period. This interval boundary is not a single point prediction value but an interval estimate containing both the lower and upper bounds of the prediction interval, quantifying the range of prediction uncertainty. Based on this, probabilistic prediction parameters for active power are calculated from the prediction interval boundary, including the mean vector and covariance matrix. The mean represents the central tendency estimate of power, and the covariance represents the magnitude of prediction uncertainty. Finally, the prediction interval boundary (serving as a boundary constraint on the functional capacity of the equipment) and the probabilistic prediction parameters (mean written into the objective function and the right-hand side of the power flow equation, covariance written into the voltage risk margin) are jointly input as constraint parameters into the differentiable convex optimization layer to construct and solve a convex optimization problem with the goal of minimizing the operating cost of the power distribution system. This convex optimization problem, under constraints such as equipment capacity and voltage safety, yields optimized control commands, which are then sent to the end-side equipment for execution. Through the above end-to-end process, the output of the prediction model is no longer merely a fixed value used by the optimization module, but participates in optimization decision-making in a probabilistic form with uncertainty, fundamentally overcoming the mismatch between prediction and decision objectives in traditional two-stage methods. By fusing the physical power flow boundary parameters of the previous control cycle, when a node experiences voltage exceedance or equipment overload in the previous cycle, the corresponding boundary parameter values ​​are larger. After feature fusion, these parameters will generate significant activation responses in the corresponding channels of the spatiotemporal feature matrix, enabling the subsequent prediction model to perceive the physical weakness of that node; thus achieving collaborative optimization of prediction and decision-making.

[0023] For example, such as Figure 3 The diagram shown illustrates the data processing flow of the end-to-end differentiable predictive and decision-based joint optimization control method of this invention. It consists of multi-source inputs, a spatiotemporal encoder, a dual-branch probability predictor head, a differentiable convex optimization layer, and end-side execution devices, distinguishing between the forward channel for online control and the backward gradient channel for the offline training phase.

[0024] Input and Encoding: The basic spatiotemporal tensor, real-time switching state, voltage over-limit residual from the previous control cycle, and physical constraint shadow price obtained by mapping the physical constraint dual are spatiotemporally aligned and fused to form the input of the spatiotemporal encoder. Dynamic adjacency matrix The independent inputs of the dynamic graph space convolutional sublayer in the first ST-GCN (Spatiotemporal Graph Convolutional Network) block are generated from the basic spatiotemporal tensor, real-time switching state, and nominal admittance matrix; the second cascaded ST-GCN block receives the output of the first ST-GCN block.

[0025] Prediction and Decision-Making: The node feature representation output by the spatiotemporal encoder enters the dual-branch probabilistic prediction head. After passing through a shared MLP (Multilayer Perceptron), fully connected circuitry, and ReLU linear correction, it generates physically valid upper and lower bounds, mean, and covariance of the prediction interval, which form the unified probabilistic prediction parameters passed to the optimization layer. The differentiable convex optimization layer writes the probabilistic prediction parameters into the target, power flow parameters, equipment capacity boundaries, and risk margins, and assembles them into a finite-dimensional SOCP (Second-Order Cone Programming) to obtain optimized control commands for the photovoltaic inverter, energy storage PCS (converter), and SVC (Static Var Compensator) / SVG (Static Var Generator), and sends them to the end-side equipment for execution.

[0026] During offline training or validation in the cloud, the gradient of the task loss is updated at the SOCP convergence point via KKT (Kuhn-Tucker conditions) implicit sensitivity back to the probability prediction head, spatiotemporal encoder, and feature space construction module. Simultaneously, the dual quantities of the physical inequality constraints are mapped to form the physical feedback vector for the next control cycle. During the online execution phase, only the forward link from input, prediction, optimization to control commands is executed; backward updates are not performed within a single real-time control cycle.

[0027] In some implementations, the multi-source spatiotemporal data includes: a basic spatiotemporal tensor and real-time switching status for the current control cycle; the basic spatiotemporal tensor includes historical electrical data of each node, meteorological data for future periods, and grid spatial topology; the historical electrical data includes at least one of node voltage amplitude, active power injection, and reactive power injection for historical periods.

[0028] In this example implementation, the multi-source spatiotemporal data specifically includes a basic spatiotemporal tensor and real-time switching status collected for the current control cycle. The basic spatiotemporal tensor is the core input of the spatiotemporal prediction model, and it is constructed as follows: at each node of the distribution network, edge computing devices continuously collect node electrical data from multiple historical time steps according to a preset sampling period (e.g., once every 15 minutes), while simultaneously acquiring meteorological forecast data for future periods (such as irradiance, temperature, wind speed, etc.), and incorporating the grid spatial topology information (i.e., the network connection relationship represented by the node admittance matrix) to jointly constitute the basic spatiotemporal tensor. Among them, the historical electrical data includes at least one of the voltage amplitude, active power injection, and reactive power injection of each node in historical periods. These electrical quantities reflect the actual state change trajectory of the distribution system during historical operation.

[0029] In some implementations, the physical power flow boundary parameters include: voltage over-limit residuals and physical constraint shadow prices for each node; the voltage over-limit residuals are determined based on the deviation between the actual voltage amplitude of each node and preset voltage safety upper and lower limits; the physical constraint shadow prices are determined based on voltage over-limits and equipment overload conditions for each node.

[0030] In this example implementation, the physical power flow boundary parameters include the voltage over-limit residuals of each node and the physical constraint shadow price. The voltage over-limit residuals are determined based on the deviation between the actual voltage amplitude of each node in the previous control cycle and the preset voltage safety upper and lower limits. For example, in the previous control cycle... System true node voltage over-limit residual vector , For the number of nodes, For the previous control cycle The voltage over-limit residual of node N is used to quantitatively characterize the physical boundary default history memory of each node. For any node i, its voltage over-limit residual... The structure is as follows:

[0031] In the formula, The per-unit value of the node i voltage amplitude obtained from online measurement or AC power flow assessment in the previous control cycle t-1; and These are the preset upper and lower safety limits for node voltages. This residual vector is used to provide the encoder with information on physical boundary violations from the previous cycle.

[0032] The physical constraint shadow price is determined based on the voltage exceedance and equipment overload conditions of each node in the previous control cycle. Essentially, it is a feedback signal formed by a specific mapping of the dual variables corresponding to each physical constraint (including node voltage upper and lower limits, equipment capacity constraints, etc.) after the convergence of the convex optimization problem in the previous cycle. For voltage constraints, the dual component reflects the marginal value of a unit voltage margin to the objective function; a positive value indicates that the system tends to raise the voltage (i.e., voltage is too low), and a negative value indicates that the system tends to lower the voltage (i.e., voltage is too high). For equipment capacity constraints, the corresponding dual component reflects the scarcity of equipment capacity margin. Therefore, the shadow price provides more refined physical sensitivity information than the residual, enabling the prediction model to detect potential risks before the voltage exceeds the limit.

[0033] In some implementations, spatiotemporal alignment and feature fusion are performed on the multi-source spatiotemporal data of the current control cycle and the physical power flow boundary parameters of the previous control cycle, including: The projected spacetime tensor is obtained by channel projection onto the fundamental spacetime tensor of the current control cycle. The real-time switching state of the current control cycle, the voltage over-limit residual and the physical constraint shadow price of the previous control cycle are projected onto a unified feature space through independent linear mapping to obtain the corresponding spatial alignment features. Temporal alignment is performed on each spatial alignment feature to obtain each spatiotemporal alignment feature; The spatiotemporal alignment features are concatenated with the projected spatiotemporal tensor along the feature channel dimension and then fused by convolution to obtain the spatiotemporal feature matrix.

[0034] In this example implementation, to support the feature extractor's adaptive perception of time-varying topology and temporal physical loop correction, this example employs a structured fusion approach combining physical position alignment, independent projection, time broadcasting, channel stitching, and convolution to transform real-time switching states, previous cycle voltage over-limit residuals, and physical constraint shadow prices into enhanced inputs compatible with the temporal feature dimensions of the nodes. First, the basic spatiotemporal tensor of the current control cycle is... Channel projection can be performed, for example, using a 1×1 convolution operation, to project the original feature channels of the underlying spatiotemporal tensor. Projected onto a preset number of uniform channels The projected spacetime tensor is obtained. :

[0035] in, This represents a 1×1 convolution operation. The purpose of this channel projection is to map the heterogeneous feature space of the original measurement data (voltage, active power, reactive power, meteorological data, etc.) to a unified feature dimension, laying the foundation for subsequent multi-source feature fusion. Simultaneously, three types of auxiliary information—the real-time switching state of the current control cycle, the voltage limit remnant of the previous control cycle, and the physical constraint shadow price of the previous control cycle—are projected onto the unified feature space through independent linear mappings. Independent linear mappings mean that each of the three types of auxiliary information has an independent, trainable fully connected projection matrix. Using independent projection matrices allows each type of information to retain its unique semantic features during the mapping process, avoiding semantic confusion caused by forced shared projection. After independent linear mapping, all three types of auxiliary information are projected to the same embedding dimension. This yields the spatial alignment features corresponding to the real-time switching state, the voltage over-limit residual of the previous control cycle, and the physical constraint shadow price of the previous control cycle. , , After spatial alignment is completed, temporal alignment is required: the spatial alignment features of each auxiliary information element are then aligned. , , Copy H times along the time dimension, so that it is consistent with... Alignment in the time dimension ensures that auxiliary information participates in the calculation with the same feature vector at each historical time step, resulting in spatiotemporal alignment features corresponding to the real-time switching state, the voltage over-limit residual of the previous control cycle, and the physical constraint shadow price of the previous control cycle. , , After completing time alignment, project the spacetime tensor. The three spatiotemporally aligned features are concatenated along the feature channel dimension to obtain a channel number of ( The concatenated tensor is then used. Finally, a 1×1 convolutional blending operation is performed on this concatenated tensor to compress the channels to the preset number of model channels. The final spatiotemporal feature matrix is ​​obtained after nonlinear activation. :

[0036] in, This represents the activation function. Indicates channel splicing. This represents a 1×1 convolutional mixing operation.

[0037] In this way, input information from different sources, with different physical meanings, and with different time granularities retains its own independent processing path. It is only fused after being mapped to a unified feature space, thus avoiding the problems of dimension mismatch and semantic confusion caused by directly splicing heterogeneous data or adding them element by element. At the same time, by explicitly injecting the physical boundary default information (residual) and constraint sensitivity information (shadow price) of the previous cycle as external condition channels into each time step of the encoder, the predictive model can perceive the physical feedback of the control results of the previous cycle at every historical moment.

[0038] In some implementations, before projecting the real-time switching state of the current control cycle and the voltage over-limit residual and physical constraint shadow price of the previous control cycle onto a unified feature space via independent linear mappings, the method further includes: The real-time switch status is located to the switch type channel of each node using a first positioning operator; the first positioning operator is constructed based on the association between the node and the switch branch; The physical constraint shadow price of the previous control cycle is located to the constraint type channel of the corresponding node using the second positioning operator; the second positioning operator is constructed based on the constraint conditions of the differentiable convex optimization layer and the location of the node, branch or device associated with the constraint conditions. The voltage over-limit residual from the previous control cycle is located to the corresponding node channel according to the node identifier of each node.

[0039] In this example implementation, before performing independent linear mapping on the three types of auxiliary information (real-time switch status, voltage over-limit residual from the previous control cycle, and physical constraint shadow price from the previous control cycle), the three types of auxiliary information undergo physical location alignment to ensure that each type of information is correctly associated with its corresponding spatial location. Specifically, for the real-time switch status of the current control cycle... (Values ​​can be 0 or 1), using the first positioning operator Position it to the switch type channel of each node. First positioning operator It is a fixed sparse matrix pre-constructed based on the association between nodes and switch branches. For each controllable switch branch in the distribution network, The state value of the switch is assigned to the corresponding switch type channel of its two endpoints. For example, if a switch branch connects node i and node j, then... The switch type channel corresponding to node i and the switch type channel corresponding to node j will both receive the state value of the switch. For the physical constraint shadow price of the previous control cycle... Using the second positioning operator Position it to the constraint type channel of the corresponding node. Second positioning operator. It is a fixed sparse matrix pre-constructed based on the constraints of the differentiable convex optimization layer and the locations of the nodes, branches, or devices associated with those constraints. For node voltage constraints, The voltage shadow price is directly allocated to the voltage constraint type channel of the corresponding node; for equipment capacity constraints (such as photovoltaic inverter capacity, energy storage converter capacity, SVC / SVG capacity, etc.). The capacity shadow price is allocated to the capacity constraint type channel of the node to which the device is connected. This applies to the voltage over-limit residual from the previous control cycle. Since the residuals themselves are vectors defined by nodes (each element corresponds to a node), they can be directly located to the corresponding node channel based on the node identifier. After the above location processing, all three types of auxiliary information are mapped onto a spatial structure with nodes as the first dimension, aligning them spatially with the node dimension of the underlying spatiotemporal tensor. This physical alignment ensures that subsequent independent linear mappings and feature fusion are performed on the correct spatial correspondence, avoiding misassociations of information due to spatial misalignment.

[0040] In some implementations, a dynamic adjacency matrix for the current control cycle is constructed based on the multi-source spatiotemporal data and the nominal node admittance of the power distribution system, including: Using the off-diagonal branch admittance amplitude of the nominal node admittance as the physical topology prior, the physical topology prior is adjusted using the physical connectivity relationship determined based on the real-time switch state to obtain the physical adjacency matrix; Based on the correlation of historical electrical data of each node in the basic spatiotemporal tensor and the real-time switching status, a data-driven adjacency matrix is ​​constructed. The physical adjacency matrix and the data-driven adjacency matrix are weighted and fused to obtain the dynamic adjacency matrix.

[0041] In this example implementation, the construction of the dynamic adjacency matrix integrates both physical topology priors and data-driven information. First, the nominal node admittance matrix is ​​used... The off-diagonal branch admittance magnitude is used as a physical topology prior. Nominal nodal admittance matrix. It is a node admittance matrix pre-constructed based on a preset benchmark topology of the distribution network (i.e., a fully connected topology when all operational branches are closed). Its off-diagonal elements reflect the electrical admittance characteristics of each branch. Because Given a complex matrix, this example uses only the magnitudes of its off-diagonal elements (i.e., the magnitudes of branch admittances) as the quantification index for the physical topology prior. Then, this physical topology prior is adjusted using the physical connectivity determined based on real-time switching states. For example, a symmetric connectivity mask is generated based on the real-time switching states. For branches that do not exist (i.e., branches that are not in the topology) and switching branches that are in the open state, The corresponding off-diagonal position is set to 0; for switch branches and fixed branches in the closed state, The corresponding off-diagonal positions are set to 1. The physical topology prior (branch admittance magnitude matrix) and the connectivity mask are then compared. The physical adjacency matrix can be obtained by multiplying each element. In this matrix, only currently connected branches retain their admittance magnitude as edge weights, while the edge weights corresponding to disconnected branches and non-existent branches are 0. Based on this, for... After symmetry normalization, the physical adjacency matrix is ​​obtained. .

[0042] Meanwhile, a data-driven adjacency matrix is ​​constructed based on the correlation of historical electrical data of each node in the fundamental spatiotemporal tensor and the real-time switching status. The data-driven adjacency matrix is ​​determined by both measurement correlation and the current switching state. This is achieved after obtaining the physical adjacency matrix. and data-driven adjacency matrix Then, the two are weighted and fused to obtain the dynamic adjacency matrix. The fusion method is as follows: For each operating branch, a learnable gating value is calculated based on the historical electrical data differences between nodes and the physical adjacency matrix. This gating value, between 0 and 1, is used to dynamically adjust the fusion ratio of physical prior and data-driven information. For example, when the physical topology prior has strong guiding significance for this branch, the gating value is close to 1, relying more on the physical prior; when the correlation of measurement data shows a pattern inconsistent with the physical prior, the gating value is close to 0, relying more on data-driven information. The final dynamic adjacency matrix is ​​obtained after longitude normalization of the fused edge weight matrix. The construction method of this dynamic adjacency matrix ensures that the off-diagonal connections corresponding to disconnected switch branches are strictly set to zero (through connectivity masks). (This ensures that) the dynamic adjacency matrix always reflects the current physical connectivity; at the same time, the weights on the running branches can be dynamically adjusted according to the changes in real-time measurement data, enabling the model to capture the changes in the system operating state reflected by the power flow distribution and node measurement changes.

[0043] For example, the data-driven adjacency matrix construction process includes: Node-by-node temporal attention convergence is performed on the historical electrical data of each node in the basic spatiotemporal tensor to obtain converged features; Based on the convergence features and the real-time switching status of the switch type channels located to each node, node context features are generated. The data-driven adjacency matrix is ​​generated by scaling dot product attention based on the node context features.

[0044] In this example implementation, the specific construction method of the data-driven adjacency matrix is ​​as follows: First, the basic spatiotemporal tensor... The historical electrical data of each node is converged node-by-node time attention. Specifically, for each node i, its measurement feature sequence over H historical time steps is as follows: ,in, The measurement feature vector (containing multi-dimensional information such as voltage, active power, reactive power, and meteorological data) of node i in the current control cycle t at the h-th historical time step. This represents the measurement feature vector of node i at the H-th historical time step in the current control period t. The temporal attention convergence mechanism assigns an attention weight to each historical time step. The weights are calculated using a learnable temporal attention network. Specifically, the measurement feature vector at each time step is linearly transformed and activated, then multiplied by a learnable attention vector, and finally normalized using Softmax to obtain the weights for each time step. Then, the measurement features at each time step are weighted and summed according to their corresponding weights to obtain the convergent features of that node. The details are as follows:

[0045]

[0046] in, , These are the learnable parameters corresponding to the linear transformation. For attention weights, For activation function, Let h be the normalized exponential function for the h-th historical time step. Let i be the convergence feature of node i.

[0047] Arrange the convergence features of all N nodes in rows to obtain the convergence feature matrix. This converged feature matrix compresses information from H time steps into a single time step, while retaining all N nodes and all feature channels, achieving information convergence in the time dimension while keeping the spatial and feature dimensions unchanged. Then, the converged feature matrix... Real-time switching status of each node's switch type channel. Concatenate along the feature channel dimension to generate node context features , , This involves an operator to restore a vector to a rearrangement with N nodes as the first dimension. Node context features. Each row corresponds to a node, containing the node's historical measurement aggregation information and current switch status information. This is based on node context features. A data-driven adjacency matrix is ​​generated through a scaled dot product attention mechanism. Specifically, the data-driven adjacency matrix is ​​generated by scaling the dot product attention mechanism. Trainable projection matrix , Project the query matrix and key matrix, and calculate the relevance score matrix. . Each element in the matrix reflects the similarity between two nodes in the measurement feature space; a larger value indicates a more similar measurement pattern between the two nodes, potentially suggesting a strong electrical coupling relationship. Finally, in the connectivity mask... MaskSoftmax normalization is performed at the allowed locations (i.e., currently physically connected node pairs and self-loops) to obtain the data-driven adjacency matrix. Specifically:

[0048]

[0049] in, The data-driven adjacency matrix for the current control period t. It is a symmetric function. , for The independent variable of the function, the symmetry function, is used to preserve the symmetry of the undirected topology representation of the distribution network. Indicates only Normalized attention is calculated based on allowed physical adjacency nodes and self-loop locations. It is the identity matrix. , , These are the query matrix, key matrix, and relevance score matrix for the current control period t, respectively. For the attention dimension.

[0050] The weights in this data-driven adjacency matrix not only reflect the statistical correlation between historical measurement data, but also ensure, through a connectivity mask, that the weights are distributed only on currently physically connected node pairs. Therefore, power flow distribution and changes in node measurements dynamically alter the weights on existing physical connections, and disconnections are not reconnected by data correlations.

[0051] In some implementations, the physical adjacency matrix and the data-driven adjacency matrix are weighted and fused to obtain the dynamic adjacency matrix, including: On each operational branch, a learnable gating value is calculated based on the differences in historical electrical data between nodes and the physical adjacency matrix; The physical adjacency matrix and the data-driven adjacency matrix are adaptively weighted and fused using the learnable gating value, and the fusion result is normalized to obtain the dynamic adjacency matrix.

[0052] In this example implementation, the weighted fusion of the physical adjacency matrix and the data-driven adjacency matrix is ​​achieved through an edge-by-edge learnable gating mechanism. Specifically, for each operating branch (i.e., connectivity mask) in the distribution network... The corresponding node pair with position 1 ), calculate a learnable gating value The gating value is calculated as follows:

[0053] in, For node pairs Learnable gating values, , Features of node i and node j in the node context The corresponding row vector in, To take the absolute value, For element-wise multiplication, , These are the learnable parameters for the fully connected layer. For the physical adjacency matrix and The normalized admittance magnitude at the corresponding position. (The three features are then considered.) After being concatenated, the result is mapped to a scalar value through a trainable gated network (a single-layer fully connected network with Sigmoid activation).

[0054] After obtaining the gating value, the physical adjacency matrix and the data-driven adjacency matrix are weighted and fused, and the fusion result is normalized for symmetry, as follows:

[0055]

[0056] in, The fusion result for the current control period t, i.e., the edge weight matrix, Let be the learnable gating matrix for the current control period t, given by composition; For the current control period t The degree matrix, Let be the dynamic adjacency matrix for the current control period t.

[0057] This invention achieves an adaptive combination of physical priors and data-driven approaches. It does not rely entirely on a fixed physical topology (which cannot adapt to changes in the system reflected by measurements) nor entirely on pure data-driven approaches (which may learn spurious correlations). Instead, it dynamically determines the fusion ratio of the two based on real-time data on each branch.

[0058] In some implementations, the spatiotemporal feature matrix and the dynamic adjacency matrix are input into the spatiotemporal prediction model to output the active power prediction interval boundary for each node in the future time period, including: Spatiotemporal encoding is performed on the spatiotemporal feature matrix and the dynamic adjacency matrix to extract node feature representations containing spatiotemporal correlation information; Multi-branch prediction is performed on the node feature representation, and the active power prediction interval boundary of each node in the future time period is output.

[0059] In this example implementation, the spatiotemporal prediction model includes a spatiotemporal encoder and a multi-branch prediction head. After the spatiotemporal feature matrix and dynamic adjacency matrix are input into the spatiotemporal prediction model, the spatiotemporal encoder first extracts spatiotemporal features. The spatiotemporal encoder adopts a spatiotemporal graph convolutional network architecture, using the spatiotemporal feature matrix... As node feature input, the dynamic adjacency matrix As input to the graph structure, spatial information propagation is performed on the graph topology defined by the dynamic adjacency matrix, while sequence modeling is performed in the temporal dimension. The spatiotemporal encoder extracts higher-level spatiotemporal features with a larger receptive field by stacking multiple layers of spatiotemporal convolutional blocks: in each layer, node features are aggregated along the edges defined by the dynamic adjacency matrix in the spatial dimension, and temporal evolution patterns are captured through gated convolution in the temporal dimension. After multiple stacking layers, the feature representation of each node has integrated its own historical information, the spatial information of neighboring nodes, and the evolution patterns of these information in the temporal dimension, forming a node feature representation containing rich spatiotemporal correlation information. The node feature representation is used as input to the prediction head, and the extracted node feature representation is input to the multi-branch prediction head, which maps it to the active power prediction interval boundary of each node in the future time period. The prediction head adopts a multi-branch structure design, where the first branch is responsible for generating the lower bound of the prediction interval, and the second branch is responsible for generating the width of the prediction interval (indirectly controlled by the proportion of the remaining available interval), and the two together determine the complete boundary of the prediction interval.

[0060] In some implementations, spatiotemporal encoding is performed on the spatiotemporal feature matrix and the dynamic adjacency matrix to extract node feature representations containing spatiotemporal correlation information, including: The spatiotemporal feature matrix and the dynamic adjacency matrix are sequentially spatiotemporally convolved through multiple cascaded spatiotemporal convolutional blocks to obtain the spatiotemporal feature representation of each node; Extract the features of the last time step in the spatiotemporal feature representation of each node as the node feature representation; Within each spatiotemporal convolutional block, a first gated temporal convolution, a spatial graph convolution, and a second gated temporal convolution are performed sequentially. In the first spatiotemporal convolutional block, the spatiotemporal feature matrix is ​​subjected to the first gated temporal convolution and then subjected to a spatial graph convolution with the dynamic adjacency matrix.

[0061] In this example embodiment, the spatiotemporal encoder includes multiple cascaded spatiotemporal convolutional blocks; within each spatiotemporal convolutional block, a first gated temporal convolution, a spatial graph convolution, and a second gated temporal convolution are performed sequentially. After inputting the spatiotemporal feature matrix and the dynamic adjacency matrix into the spatiotemporal encoder, the spatiotemporal feature matrix... First, we enter the first spatiotemporal convolution block. In this block, the spatiotemporal feature matrix first undergoes the first gated temporal convolution: the gated temporal convolution uses one-dimensional convolution along the time axis and employs a gated linear unit (GLU) as the activation function. Specifically:

[0062]

[0063] in, For the first The output features of the first gated temporal convolution of each spatiotemporal convolution block , These are the linear branch outputs of the first gated temporal convolution; For the first The first gated temporal convolution operation of each spatiotemporal convolution block, i.e., convolution along the time axis and using... Zero padding on the left, For the first The temporal kernel length of each spatiotemporal convolution block For the first The input to a spatiotemporal convolutional block.

[0064] After the first gated temporal convolution, the features undergo preliminary sequence modeling in the time dimension. Subsequently, the output features of the first gated temporal convolution are fed into the spatial graph convolutional sub-layer. In this layer, the output of the first gated temporal convolution is compared with the dynamic adjacency matrix. Perform graph convolution operations as follows:

[0065] in, For the first The output of the spatial graph convolutional layer of each spatiotemporal convolutional block The Middle Each historical time step corresponds to a feature. for The Each historical time step corresponds to a feature. , The first Trainable spatial transformation matrix and bias of each spatiotemporal convolutional block It is a non-linear activation function. H is the index of the historical time step corresponding to the current control cycle t, and H is the length of the historical time window.

[0066] The features after spatial graph convolution enter the second gated temporal convolution, which has the same structure as the first gated temporal convolution, to perform further sequence modeling in the time dimension, as follows:

[0067] in, These are the linear branch outputs of the second gated temporal convolution, respectively. For the first The second gated temporal convolution operation of each spatiotemporal convolution block For the first The output of the _th spatiotemporal convolutional block, i.e. the _th The input to a spatiotemporal convolutional block.

[0068] Each spatiotemporal convolutional block is executed according to the above process, and the output of the previous spatiotemporal convolutional block enters the next spatiotemporal convolutional block to repeat the above process. It is worth noting that the dynamic adjacency matrix... It is used only in the spatial graph convolutional sublayer of the first spatiotemporal convolutional block. After stacking multiple spatiotemporal convolutional blocks, the features of the last time step (i.e., the most recent time step within the history window) are extracted from the output of the last spatiotemporal convolutional block as the final output of the spatiotemporal encoder.

[0069] For example, such as Figure 4 As shown, Figure 4 The data processing flow for feature fusion and spatiotemporal encoder includes a topology dynamic evolution layer, two cascaded spatiotemporal feature extraction blocks, and a feature extraction block.

[0070] Structured Input Fusion: Fundamental Spatiotemporal Tensors Obtained by 1×1 convolution Real-time on / off status The shadow price of physical constraints in the previous control cycle After positioning and mapping, the voltage over-limit residual from the previous control cycle After projection, After broadcasting the mapping or projection results along each time step of H, the channels are concatenated, and then subjected to mixed 1×1 convolution processing to obtain the spatiotemporal feature matrix. .

[0071] Dynamic topological evolution: fundamental spacetime tensor Real-time on / off status and nominal node admittance As input, these two matrices are used together to construct data-related adjacency relationships and physical topology priors. This input is then processed by temporal attention and physical masking to generate a physical adjacency matrix and a data-driven adjacency matrix. These two matrices are fused and normalized by edge-by-edge gating to obtain the dynamic adjacency matrix for the current control cycle. .

[0072] Spatiotemporal feature extraction: Spatiotemporal feature matrix The first ST-GCN block is input and sequentially subjected to a first-gated temporal convolution, a spatial graph convolution, and a second-gated temporal convolution. The dynamic adjacency matrix and the output of the first-gated temporal convolution are then subjected to a spatial graph convolution. The second ST-GCN block continues this process, sequentially performing the first-gated temporal convolution, spatial graph convolution, and second-gated temporal convolution on the output of the first ST-GCN block. The dynamic adjacency matrix is ​​only input from the first ST-GCN block. Finally, the last valid time step within the historical window is selected from the output of the second ST-GCN block to form the node feature representation used by the dual-branch probability prediction head. .

[0073] In some implementations, the active power prediction interval boundary includes an upper bound and a lower bound; multi-branch prediction is performed on the node feature representation to output the active power prediction interval boundary for each node in the future time period, including: The node feature representation is subjected to a multi-layer nonlinear transformation to obtain shared features; The shared features are subjected to a first linear mapping and a first nonlinear activation to obtain the lower bound of the prediction interval. The shared features are subjected to a second linear mapping and a second nonlinear activation to obtain the proportion of the remaining available intervals; The remaining capacity between the lower bound and the upper bound of the prediction interval is scaled using the remaining available interval ratio to obtain the prediction interval width, and then the upper bound of the prediction interval is obtained.

[0074] In this example implementation, the multi-branch prediction head includes a shared transform layer, a first prediction branch, and a second prediction branch. The multi-branch prediction head is used to output the active power prediction interval boundaries (including the upper and lower bounds of the prediction interval) for each node in the future time period. Exemplarily, the multi-branch prediction head adopts a parallel dual-branch multilayer perceptron (Dual-BranchMLP) structure. The shared front-end consists of a fully connected layer and a ReLU activation layer. It represents the node features output by the spatiotemporal encoder. After the multi-branch prediction head is input, it first undergoes a multi-layer nonlinear transformation through a shared transform layer. This shared transform layer consists of multiple stacked fully connected layers, with ReLU activation functions used between layers. Its function is to further abstract the high-dimensional spatiotemporal features output by the encoder into shared features suitable for the prediction task. These shared features are then input into the first and second prediction branches, respectively. The first prediction branch is responsible for generating the lower bound of the prediction interval: for node i, the shared features are mapped using a first linear mapping to obtain the first original scalar. Then, it is mapped to the (0,1) interval by the Sigmoid function (first nonlinear activation), and then transformed to the physically feasible power range of the device by linear scaling. , The specific values ​​within the range are used to obtain the lower bound of the prediction interval. , , These represent the lower and upper limits of physical feasibility for the device corresponding to node i, respectively. The second prediction branch is responsible for generating the prediction interval width: for node i, the shared features are mapped using the second linear mapping to obtain the second original scalar. Then, the Sigmoid function (the second non-linear activation) maps it to the (0,1) interval, yielding the proportion of the remaining usable interval. This proportion represents what percentage of the remaining capacity between the prediction lower bound and the physical upper bound is used as the prediction interval width. Specifically, the prediction interval width is calculated... for:

[0075] Then predict the upper bound of the interval. for:

[0076] In summary, the dual-branch structure design, which uses the first branch to determine the lower bound and the second branch to determine the relative width, ensures that the prediction head can structurally guarantee... This means that the prediction interval is always within the physical feasible range of the device, and the lower bound is always smaller than the upper bound, avoiding the problems of predicted values ​​going out of bounds or the upper and lower bounds being reversed that may occur in traditional methods.

[0077] In some implementations, the probabilistic prediction parameters include the mean vector and covariance matrix of the predicted active power values ​​for each node; the probabilistic prediction parameters for active power are calculated based on the prediction interval boundaries, including: Using a preset two-tailed probability, the lower bound and upper bound of the prediction interval are converted into the mean vector and standard deviation matrix of the active power prediction values. The covariance matrix of the active power prediction value is constructed based on the standard deviation matrix.

[0078] In this example implementation, the probabilistic prediction parameters include the mean vector and covariance matrix of the predicted active power values ​​for each node. These two parameters are calculated based on the prediction interval boundaries. Specifically, a preset two-tailed probability is uniformly used in the training, validation, and online simulation phases. (For example (corresponding to a 95% confidence level), the lower bound of the prediction interval. and the upper bound of the prediction interval The mean vector and standard deviation matrix are converted into predicted active power values. Details are as follows:

[0079]

[0080]

[0081]

[0082] in, , Let be the mean vector and standard deviation of the predicted active power at node i in the current control period t, respectively. For the reason The standard deviation matrix formed Let be the covariance matrix of the predicted active power values ​​for the current control period t. For standard normal quantiles, It is the quantile function of the standard normal distribution.

[0083] It should be noted that the above conversion is only a symmetric normal position-scale parameterization method under a fixed confidence level. The actual prediction error may not strictly follow a normal distribution. In this case, the protection against distribution deviation and tail uncertainty can be ensured through other methods (such as Wasserstein split-bar optimization layers). Through this conversion, the interval boundary of the prediction head output is transformed into two probability parameters with clear statistical significance: the mean, as the central estimate of power prediction, drives the economic objective of the optimization problem; the covariance, as a measure of the uncertainty magnitude, drives the risk margin of the optimization problem.

[0084] For example, such as Figure 5 The diagram shown illustrates the data processing procedure of the dual-branch probability prediction head. (ST-GCN output) First, shared features are extracted using two fully connected networks (one 32-dimensional and one 16-dimensional) in the shared transform layer and ReLU activation. These shared features are then processed by the lower bound of the prediction interval branch (branch 1) and the remaining available interval proportion branch (branch 2). The lower bound of the prediction interval branch outputs a first raw scalar, which is then mapped using a sigmoid function to ensure the lower bound of the prediction interval is strictly between the lower and upper limits of the corresponding device's physical power. The proportion branch outputs a second raw scalar, which is then mapped using a sigmoid function to represent the remaining available interval proportion. The outputs of these two branches are used to construct the physically valid interval. The remaining capacity between the lower bound of the prediction interval and the upper limit of the physical power is scaled using the remaining available interval proportion to obtain the interval width and the upper bound of the prediction interval. Structurally, this ensures that the lower bound of the prediction interval is smaller than the upper bound, and that neither exceeds the device's physical boundary.

[0085] During the training, validation, and online simulation phases, the same fixed two-tailed probability is used to uniquely convert the upper and lower bounds of the prediction interval into a mean vector and a standard deviation vector, and then construct the covariance matrix from the standard deviation vector. Thus, the prediction head consistently outputs probability prediction parameters consisting of the lower and upper bounds of the predicted active power values ​​at each node, the mean, and the covariance matrix.

[0086] In some implementations, a convex optimization problem is constructed with the objective of minimizing the operating cost of the power distribution system, including: Taking photovoltaic inverters, energy storage converters, and dynamic reactive power compensation devices as control objects, a convex optimization problem is constructed with the active and reactive power control quantities of each device as optimization variables. Based on the expected voltage deviation, expected network loss, and control action cost of the power distribution system, the operating cost is determined, and an objective function is constructed with the goal of minimizing the operating cost. The constraints of the convex optimization problem are constructed based on the power flow constraints, operational safety constraints, and equipment capacity constraints of the power distribution system. Based on the objective function and the constraints, the convex optimization problem is constructed.

[0087] In this example implementation, firstly, the photovoltaic inverter, energy storage converter, and dynamic reactive power compensation device (such as SVC / SVG) are taken as the controlled objects, and the active and reactive power control quantities of each device are taken as optimization variables. Specifically, the control vector... ,in, This refers to the reactive power control quantity of the photovoltaic inverter. This is the active power control quantity of the energy storage converter (positive value indicates discharging, negative value indicates charging). This refers to the reactive power control quantity of the energy storage converter. This refers to the reactive power control quantity of the dynamic reactive power compensation device. Then, the operating cost objective function is determined based on the expected voltage deviation, expected network loss, and control action cost of the distribution system. The expected voltage deviation term measures the degree to which the voltage at each node deviates from the reference value, expressed as the mean vector of the predicted active power values. (by each node) Composition) and covariance matrix The covariance matrix is ​​calculated after mapping with the voltage sensitivity matrix. The expected value of voltage deviation is influenced by the trace term; the expected network loss term measures the active power loss in the distribution system, by... and The objective function is calculated after mapping with the branch power flow sensitivity matrix; the control action cost term penalizes the magnitude of the control variable to prevent the optimizer from generating excessive control commands. Specifically, the objective function is as follows:

[0088]

[0089]

[0090]

[0091]

[0092]

[0093]

[0094] In the formula, To optimize variables, These represent the photovoltaic reactive power injection control quantity, energy storage active power control quantity, energy storage reactive power control quantity, and reactive power injection control quantity of the reactive power compensation device at time t, respectively. Let be the probability prediction parameters at time t. For probability prediction parameters The expected voltage deviation is as follows. For probability prediction parameters The expected network loss is as follows. To control the cost of actions, , These are the weights for the voltage deviation term and the network loss term, respectively. Indicates the expectation. Let be the linearized predicted voltage at time t. Reference voltage; This is the reference voltage when there is no control action. To predict the sensitivity matrix injected into the node voltage, Let be the random injection amount of the source charge at time t. , , This represents finding the covariance matrix. , These are the mean vector and covariance matrix of the active power prediction values ​​in the probability prediction parameters, respectively. To control the action The sensitivity matrix to the node voltage. For the linearized branch power flow at time t, For the linearized branch power flow at time t without control action, These are the sensitivity matrices for predicting the injection to the branch power flow. To control the sensitivity matrix of actions to branch power flow, This is an approximate matrix for network loss; This is the penalty coefficient matrix for control actions, with the superscript T indicating the transpose operation. The sensitivity matrices listed above are, respectively, the predefined sensitivity matrices for linearized predicted voltage, branch power flow, predicted injection, and control action-to-voltage / power flow under the linearized power flow model without control action.

[0095] Regarding constraints, a complete set of constraints for the convex optimization problem is constructed based on the power flow constraints of the power distribution system (using linearized power flow models such as LinDistFlow to construct nodal power balance equations and branch power flow equations), operational safety constraints (the voltage of each node must be maintained between preset safety upper and lower limits), and equipment capacity constraints (apparent power capacity constraints of photovoltaic inverters, apparent power capacity constraints of energy storage converters, and reactive power upper and lower limit constraints of SVC / SVG).

[0096] Through the complete construction of the above objective function and constraints, a parameterized convex optimization problem is formed, with control variables as optimization variables and probability prediction parameters as input parameters.

[0097] In some implementations, the process of solving the convex optimization problem includes: The convex optimization problem is transformed into a second-order cone programming form. The cone optimization solver is then called to perform a forward solution, which yields the optimized control instructions for each node in the current control cycle.

[0098] In this example implementation, the constructed convex optimization problem is transformed into the standard form of second-order cone programming (SOCP). Specifically, the objective function is divided according to the control variables. Expanding the quadratic form and combining like terms, we obtain the objective function in standard form:

[0099] in, It is the coefficient matrix of the quadratic terms. For probability prediction parameters The vector of coefficients of the first-order terms that affect the effect. This is a constant term that is independent of the control variables and can vary with probability prediction parameters. The standard form objective function mentioned above can also include a partial Bruker loss to cover tail-end overruns or scenario-based penalties. The standard form convex optimization problem also includes equality constraints and inequality constraints. Equality constraints include network equality constraints constructed based on linearized power flow models such as LinDistFlow, mainly including the active / reactive power balance equations at nodes and the linearized voltage drop equations between branch power flow and node voltage; the control quantities of photovoltaic inverters, energy storage converters, and SVC / SVG enter the power injection terms of the corresponding nodes through the device-node association matrix. Inequality constraints include the first type as distribution network operation safety constraints, such as node voltage upper and lower limits. The second type is device operation boundaries, including photovoltaic inverters (active / reactive power boundaries and apparent power capacity); energy storage converters (charge and discharge active power upper and lower limits, reactive power regulation boundaries, apparent power capacity, and dynamic active power boundaries determined by the previous cycle's SOC, efficiency, and control step size); SVC / SVG (reactive power upper and lower limits), etc.

[0100] This invention employs linearized power flow models such as LinDistFlow to linearize the nonlinear power flow equations and represents equipment capacity constraints as second-order cone constraints, thereby transforming the original problem into a second-order cone programming problem with convexity guarantees. After the transformation, a cone optimization solver is invoked for forward solving to obtain the values ​​of the control variables that minimize the objective function, i.e., the optimized control commands for each node in the current control cycle. During the forward solving process, the solver not only returns the optimal solution for the original variables but also the dual variables (i.e., Lagrange multipliers) corresponding to each constraint. For inequality constraints (such as voltage upper and lower limit constraints and equipment capacity constraints), the corresponding dual variables reflect the activity level of the constraint at the optimal solution. If the constraint is inactive (i.e., the constraint has a margin), the corresponding dual variable is zero; if the constraint is active (i.e., the constraint is tightened), the corresponding dual variable is positive, and the larger the value, the stronger the constraint's restriction on the objective function. These dual variables will be mapped to physical constraint shadow prices in the next control cycle and fed into the encoder as physical feedback signals. After the solution is completed, the optimized control command is sent to the end-side devices (such as photovoltaic inverter controllers, energy storage PCS controllers, SVC / SVG controllers, etc.) for execution. The above solution process is applicable in both offline training and online inference stages: during offline training, the solver also provides the implicit derivative capability of the KKT conditions, so that the task loss gradient can be backpropagated to the prediction network through the convergence point of the solver; during online inference, the solver only performs forward solving.

[0101] In some implementations, the operational safety constraints include voltage boundary constraints at each node; the generation process of the physical constraint shadow price includes: Extract the dual variables corresponding to the voltage boundary constraints of each node after the convergence of the convex optimization problem in the previous control cycle; The voltage shadow price is obtained by subtracting the voltage upper limit constraint dual component and the voltage lower limit constraint dual component in the dual variables corresponding to the voltage boundary constraints of each node. Extract the dual variable corresponding to the equipment capacity constraint as the capacity shadow price; The physical constraint shadow price is generated based on the voltage shadow price and the capacity shadow price.

[0102] In this example implementation, the generation process of the physical constraint shadow price is based on the dual variables after the convergence of the convex optimization problem in the previous control cycle. Specifically, after the convex optimization problem in the previous control cycle is solved, the cone optimization solver returns the dual variables corresponding to each constraint condition. For each node voltage boundary constraint (including upper voltage limit constraint and lower voltage limit constraint) in the operational safety constraints, its corresponding dual components are extracted. and The upper voltage dual component. Dual component with lower voltage limit By subtracting the values, we can obtain a composite voltage shadow price with sign information. The sign of the shadow price indicates the overall bias of the system voltage. A value greater than 0 indicates that the system is more inclined to raise the voltage (i.e., the current voltage is too low, and the lower limit constraint is tighter). A value less than 0 indicates that the system is more inclined to reduce the voltage (i.e., the current voltage is too high, and the upper limit constraint is tighter). A value near 0 indicates that the voltage is in the middle of the safe zone, and neither the upper nor lower limit constraints are stringent. Simultaneously, the dual variable corresponding to the equipment capacity constraint is extracted as the capacity shadow price. This value reflects the scarcity of equipment capacity margin; a higher value indicates that the equipment's capacity is closer to its limit, and further increases in the equipment's output will be severely constrained by capacity limitations. Finally, the voltage shadow price and capacity shadow price are combined to form the complete physical constraint shadow price. As one of the physical power flow boundary parameters of the previous control cycle, it is used for feature fusion in the current control cycle. Specifically:

[0103]

[0104]

[0105] In the formula, Select a matrix for equipment capacity constraints and directly extract the corresponding cone dual components; The dual vector returned by the solver in the conical canonical form from the previous control cycle t-1 and inversely mapped to the explicit physical inequality constraints via the modeling layer. These are the dual components of the upper limit and lower limit of equipment capacity, respectively. This forms... This information is then fed back to the encoder. If the underlying solver returns the original dual vector of the normalized cone problem, the inverse mapping saved in the modeling layer should be used to restore the above explicit constraint order before performing component subtraction or selection.

[0106] In some implementations, the update process of the spatiotemporal prediction model includes: Based on the statistical deviation between the mean vector of the active power prediction values ​​in the probability prediction parameters and the measured power after executing the optimized control command, the prediction loss of the current control cycle is determined. Based on the deviation between the evaluated voltage obtained after AC power flow evaluation of the optimized control command and the preset upper and lower voltage safety limits, the voltage limit over-limit hinge penalty is determined; based on the voltage limit over-limit hinge penalty and the network loss penalty generated by executing the optimized control command, the decision loss of the current control cycle is determined; Based on the predicted loss and the decision loss, determine the current task loss; If the current task loss is greater than a preset loss threshold, the gradient of the current task loss is uploaded to the cloud so that the cloud can update and distribute the spatiotemporal prediction model based on the gradient uploaded by each edge computing device.

[0107] In this example implementation, after the edge computing device completes online inference for the current control cycle, backpropagation and parameter updates are not immediately performed on the edge side. The edge side only caches key data for this inference in each control cycle. This key data may include the mean vector in the probabilistic prediction parameters, the actual measured power after executing optimized control commands, and the evaluated voltage obtained after AC power flow assessment. Model parameter updates are performed uniformly in the cloud. The cloud first calculates the prediction loss and decision loss for the current control cycle based on the cached data uploaded by each edge computing device: the prediction loss is based on the mean vector of the active power prediction values ​​in the probabilistic prediction parameters. Measured power after executing optimized control commands The decision loss is determined by the statistical deviation between the two (e.g., mean square error, MSE). The decision loss comprises two parts: voltage over-limit hinge penalty and network loss penalty. The voltage over-limit hinge penalty is based on the evaluated voltage obtained after AC power flow assessment of the optimized control command. The penalty is determined by the deviation from the preset safe upper and lower voltage limits. When the voltage is within the safe range, the penalty is zero; when the limit is exceeded, the penalty increases with the square of the exceedance. The network loss penalty is based on the active power loss generated after executing the optimized control command. The predicted loss and the decision loss are then weighted and summed according to preset weights to obtain the current task loss. The details are as follows:

[0108]

[0109] in, To predict losses, For decision-making losses, , The loss weights for prediction loss and decision loss are set before training or calibrated on an independent validation set, respectively. As a weight for network loss penalties, for and The offline evaluation status obtained through power flow mapping; , For the lower and upper limits of voltage safety, It is a positive part function. This indicates a search for the 2-norm.

[0110] If the current task loss exceeds a preset loss threshold (indicating that the model's performance under the current data distribution has deteriorated to the point where it needs to be updated), the edge computing device will apply the gradient of the current task loss. Uploaded to the cloud. The cloud aggregates gradient information from multiple edge computing devices, uses a federated averaging algorithm to weight and aggregate the gradients, updates the global model parameters, and then distributes the updated model to each edge computing device.

[0111] This invention can improve the consistency between prediction and control objectives, reduce the risk of decision-making bias near the safety boundary, and support the efficient utilization of edge-side resources.

[0112] This invention can be embedded in distribution network edge computing terminals (such as smart converged terminals for distribution areas). Without increasing hardware costs, it can significantly reduce the risk of grid voltage exceeding limits and network losses caused by prediction errors, and has significant engineering value for improving the economy and security of new power systems.

[0113] Simulation experiment verification (1) Simulation objects, data and reproduction settings This experiment uses the 12.66 kV IEEE 33-node radial distribution network parameters listed in the MATPOWER standard case 33bw, setting up a baseline topology and two radial topologies after tie switch reconfiguration. Nodes 6, 14, 18, 25, 30, and 33 are configured with distributed photovoltaic (PV) power of a total capacity of 4.65 MW; nodes 18, 25, and 30 are configured with energy storage converters; and nodes 14 and 33 are configured with SVC / SVG. The 4.65 MW capacity is approximately 1.25 times the baseline peak active load of 3.715 MW in this case, used to simultaneously form a high PV reverse overvoltage and a low-voltage section during the evening peak heavy load.

[0114] 4612 continuous operating points were generated at 15-minute sampling intervals, including intraday load and photovoltaic curves, related random disturbances, cloud cluster mutations, concentrated electric vehicle charging events, and topology switching. A continuous 12-step (3-hour) historical window was used to generate 4600 supervised samples, which were then divided chronologically into 2760 training samples, 828 validation samples, and 1012 fully reserved test samples. End-to-end cone layer fine-tuning extracted 1200 samples from the training segment. The above data are reproducible simulation data.

[0115] According to this invention, the prediction network implements two-level gated temporal convolution and one dynamic topology propagation. The inputs are active power, reactive power, voltage, the previous cycle's out-of-bounds residual, and physical constraint dual feedback. The output is a prediction interval that satisfies the physical upper and lower bounds and is converted into mean and covariance.

[0116] Online decision-making employs a compact, differentiable SOCP specialization under the LinDistFlow sensitivity condition of the radial network: photovoltaic reactive power, energy storage active / reactive power, and SVC / SVG reactive power are all constrained by equipment capacity; voltage risk margin consists of the prediction standard deviation term and the preset sensitivity dual norm; energy storage apparent power capability is represented by a second-order cone constraint. The control action is then verified using an independent nonlinear radial AC power flow forward-backward algorithm.

[0117] (2) Comparison methods and evaluation indicators Two sets of control methods were set up under identical data, equipment boundaries, and AC power flow estimators: an uncontrolled method and a two-stage deterministic method that first uses a predictor to predict the output of each node, followed by deterministic control. All joint models (prediction + optimization) started from the same pre-trained weights and used the same fine-tuning sample order and random seed.

[0118] The main indicators are: scenario voltage over-limit rate, i.e., any non-balance node voltage exceeding [0.95, 1.05] pu (per unit value) at a certain test time point is recorded as an over-limit; node voltage over-limit rate; 95th percentile of the maximum over-limit amplitude at each time point; active power loss of nonlinear AC power flow; control vector L2 norm; and SOCP software solution time. The difference confidence interval was estimated using the bootstrap method with 5000 pairwise resamplings at the same time point. The experimental results are shown in Table 1. Table 1

[0119] As shown in Table 1, the method of this invention reduces the scenario voltage over-limit rate from 22.83% to 12.06% compared to the two-stage deterministic method, a decrease of 10.77 percentage points, or a relative decrease of 47.19%. The average AC network loss decreases by 2.953 kW, or 4.21%. Compared to the uncontrolled group, the over-limit rate of this invention decreases by 65.92%, and the average network loss decreases by 10.41%. The node overvoltage rate results under different operating scenarios are shown in Table 2: Table 2

[0120] As shown in Table 2, overvoltage exists in the high photovoltaic back-transmission scenario when uncontrolled. The method of this invention reduces the over-limit rate to 0 in this scenario. In the heavy-load low-voltage scenario, the overvoltage rate of this invention is reduced from 100% in the uncontrolled and two-stage deterministic methods to 21.05%. In the topology switching scenario, the overvoltage rate of this invention is reduced from 20.41% in the two-stage deterministic method to 14.29%. The above results cover the upper and lower voltage boundaries and topology changes, which can support the effectiveness of the closed-loop control of this invention in this simulation example.

[0121] The average software solution time for SOCP in this invention group was 2.43 ms. The test platform was an Apple M1 Pro octa-core processor with 16 GB of memory, and the software used was PyTorch 2.13.0, CVXPY 1.9.2, and CVXPYLayers 1.2.0. All samples returned control solutions during the test.

[0122] Example 2 Based on the same inventive concept, the present invention also provides an edge computing device for a power distribution system, comprising: The feature space construction module is used to perform spatiotemporal alignment and feature fusion on the multi-source spatiotemporal data of each node of the power distribution system for the current control cycle and the physical power flow boundary parameters of the previous control cycle to obtain the spatiotemporal feature matrix of the current control cycle; the physical power flow boundary parameters are determined based on the voltage over-limit and equipment overload conditions of each node. The topology modeling module is used to construct a dynamic adjacency matrix for the current control cycle based on the multi-source spatiotemporal data and the nominal node admittance of the power distribution system; the nominal node admittance is a node admittance matrix constructed based on a preset benchmark topology. The prediction module is used to input the spatiotemporal feature matrix and the dynamic adjacency matrix into the spatiotemporal prediction model, output the active power prediction interval boundary of each node in the future time period, and calculate the probability prediction parameters of active power based on the prediction interval boundary. The optimization module is used to input the prediction interval and the probability prediction parameters as constraint parameters into the differentiable convex optimization layer, construct and solve a convex optimization problem with the goal of minimizing the operating cost of the power distribution system, obtain the optimized control commands for each node in the current control cycle, and send them to the end-side equipment for execution.

[0123] In one possible implementation, the multi-source spatiotemporal data includes: a basic spatiotemporal tensor and real-time switching status for the current control cycle; the basic spatiotemporal tensor includes historical electrical data of each node, meteorological data for future periods, and grid spatial topology; the historical electrical data includes at least one of node voltage amplitude, active power injection, and reactive power injection for historical periods.

[0124] In one possible implementation, the physical power flow boundary parameters include: voltage over-limit residuals and physical constraint shadow prices for each node; the voltage over-limit residuals are determined based on the deviation between the actual voltage amplitude of each node and the preset voltage safety upper and lower limits; the physical constraint shadow prices are determined based on voltage over-limits and equipment overload conditions for each node.

[0125] In one possible implementation, the feature space construction module includes: The channel projection layer is used to project the basic spatiotemporal tensor of the current control cycle into the projected spatiotemporal tensor. The spatial projection layer is used to project the real-time switching state of the current control cycle, the voltage over-limit residual and the physical constraint shadow price of the previous control cycle onto a unified feature space through independent linear mappings to obtain the corresponding spatial alignment features. The temporal alignment layer is used to perform temporal alignment on each spatial alignment feature to obtain each spatiotemporal alignment feature. The feature fusion layer is used to convolve and fuse the spatiotemporal aligned features with the projected spatiotemporal tensor along the feature channel dimension to obtain the spatiotemporal feature matrix.

[0126] In one possible implementation, the feature space construction module further includes: The first positioning layer is used to locate the real-time switch state to the switch type channel of each node using a first positioning operator; the first positioning operator is constructed based on the association relationship between the node and the switch branch; The second positioning layer is used to locate the physical constraint shadow price of the previous control cycle to the constraint type channel of the corresponding node using the second positioning operator; the second positioning operator is constructed based on the constraint conditions of the differentiable convex optimization layer and the location of the node, branch or device associated with the constraint conditions. The third positioning layer is used to locate the voltage over-limit residual of the previous control cycle to the corresponding node channel according to the node identifier of each node.

[0127] In one possible implementation, the topology modeling module includes: The physical modeling layer is used to adjust the physical topology prior by using the off-diagonal branch admittance amplitude of the nominal node admittance as the physical topology prior and the physical connectivity relationship determined based on the real-time switch state to obtain the physical adjacency matrix. The data modeling layer is used to construct a data-driven adjacency matrix based on the correlation of historical electrical data of each node in the basic spatiotemporal tensor and the real-time switching status. The matrix fusion layer is used to perform weighted fusion of the physical adjacency matrix and the data-driven adjacency matrix to obtain the dynamic adjacency matrix.

[0128] In one possible implementation, the data modeling layer is specifically used for: Node-by-node temporal attention convergence is performed on the historical electrical data of each node in the basic spatiotemporal tensor to obtain converged features; Based on the convergence features and the real-time switching status of the switch type channels located to each node, node context features are generated. The data-driven adjacency matrix is ​​generated by scaling dot product attention based on the node context features.

[0129] In one possible implementation, the matrix fusion layer is specifically used for: On each operational branch, a learnable gating value is calculated based on the differences in historical electrical data between nodes and the physical adjacency matrix; The physical adjacency matrix and the data-driven adjacency matrix are adaptively weighted and fused using the learnable gating value, and the fusion result is normalized to obtain the dynamic adjacency matrix.

[0130] In one possible implementation, the prediction module includes a spatiotemporal prediction model, which includes: A spatiotemporal encoder is used to perform spatiotemporal encoding on the spatiotemporal feature matrix and the dynamic adjacency matrix to extract node feature representations containing spatiotemporal correlation information; The multi-branch prediction head is used to perform multi-branch prediction on the node feature representation and output the active power prediction interval boundary of each node in the future time period.

[0131] In one possible implementation, the spatiotemporal encoder includes a plurality of cascaded spatiotemporal convolutional blocks; the spatiotemporal encoder is specifically used for: The spatiotemporal feature matrix and the dynamic adjacency matrix are sequentially spatiotemporally convolved through multiple cascaded spatiotemporal convolutional blocks to obtain the spatiotemporal feature representation of each node; Extract the features of the last time step in the spatiotemporal feature representation of each node as the node feature representation; Within each spatiotemporal convolutional block, a first gated temporal convolution, a spatial graph convolution, and a second gated temporal convolution are performed sequentially. In the first spatiotemporal convolutional block, the spatiotemporal feature matrix is ​​subjected to the first gated temporal convolution and then subjected to a spatial graph convolution with the dynamic adjacency matrix.

[0132] In one possible implementation, the multi-branch prediction head includes: A shared transformation layer is used to perform multi-layer nonlinear transformations on the node feature representation to obtain shared features; The first prediction branch is used to perform a first linear mapping and a first nonlinear activation on the shared features to obtain the lower bound of the prediction interval. The second prediction branch is used to perform a second linear mapping and a second nonlinear activation on the shared features to obtain the proportion of the remaining available interval; The physical interval construction layer is used to scale the remaining capacity between the lower boundary and the upper boundary of the prediction interval using the proportion of the remaining available interval, so as to obtain the width of the prediction interval and thus the upper boundary of the prediction interval.

[0133] In one possible implementation, the probability prediction parameters include the mean vector and covariance matrix of the predicted active power values ​​for each node; the prediction module further includes a conversion layer, which is used for: Using a preset two-tailed probability, the lower bound and upper bound of the prediction interval are converted into the mean vector and standard deviation matrix of the active power prediction values. The covariance matrix of the active power prediction value is constructed based on the standard deviation matrix.

[0134] In one possible implementation, the optimization module includes an optimization problem construction submodule, which is used to: Taking photovoltaic inverters, energy storage converters, and dynamic reactive power compensation devices as control objects, a convex optimization problem is constructed with the active and reactive power control quantities of each device as optimization variables. Based on the expected voltage deviation, expected network loss, and control action cost of the power distribution system, the operating cost is determined, and an objective function is constructed with the goal of minimizing the operating cost. The constraints of the convex optimization problem are constructed based on the power flow constraints, operational safety constraints, and equipment capacity constraints of the power distribution system. Based on the objective function and the constraints, the convex optimization problem is constructed.

[0135] In one possible implementation, the optimization module includes a solution submodule, which is used for: The convex optimization problem is transformed into a second-order cone programming form. The cone optimization solver is then called to perform a forward solution, which yields the optimized control instructions for each node in the current control cycle.

[0136] In one possible implementation, the operational safety constraints include voltage boundary constraints at each node; the optimization module further includes a feedback generation submodule, which is used for: Extract the dual variables corresponding to the voltage boundary constraints of each node after the convergence of the convex optimization problem in the previous control cycle; The voltage shadow price is obtained by subtracting the voltage upper limit constraint dual component and the voltage lower limit constraint dual component in the dual variables corresponding to the voltage boundary constraints of each node. Extract the dual variable corresponding to the equipment capacity constraint as the capacity shadow price; The physical constraint shadow price is generated based on the voltage shadow price and the capacity shadow price.

[0137] In one possible implementation, an update module is also included, for: Based on the statistical deviation between the mean vector of the active power prediction values ​​in the probability prediction parameters and the measured power after executing the optimized control command, the prediction loss of the current control cycle is determined. Based on the deviation between the evaluated voltage obtained after AC power flow evaluation of the optimized control command and the preset upper and lower voltage safety limits, the voltage limit over-limit hinge penalty is determined; based on the voltage limit over-limit hinge penalty and the network loss penalty generated by executing the optimized control command, the decision loss of the current control cycle is determined; Based on the predicted loss and the decision loss, determine the current task loss; If the current task loss is greater than a preset loss threshold, the gradient of the current task loss is uploaded to the cloud so that the cloud can update and distribute the spatiotemporal prediction model based on the gradient uploaded by each edge computing device.

[0138] Example 3 Based on the same inventive concept, the present invention also provides an operation optimization control system for a power distribution system, comprising: an edge computing device according to any one of Embodiment 2. The edge computing device may be a smart transformer terminal (TTU) or an edge computing gateway deployed in a substation or key distribution area, used to run ST-GCN, probabilistic prediction head and structured microscalable optimization layer online, and generate area control commands.

[0139] like Figure 2 As shown, it can also include cloud-based computing devices and edge-side devices. The cloud-based computing devices can be cloud master stations deployed in city-level or provincial-level data centers, responsible for historical data storage, offline joint training, verification, and model version management. Edge-side devices can include distributed photovoltaic inverters, energy storage PCS, and flexible load controllers, responsible for uploading local measurements and executing active / reactive power control commands issued by the edge layer.

[0140] Example 4 like Figure 6 As shown, the present invention also provides an electronic device, which may be a computer device, a microcontroller device, a smart mobile device, etc. The electronic device in this embodiment may include a processor, a memory, a transceiver component, etc. The memory, processor, and transceiver component are connected via a bus; the memory can be used to store executable programs, and an exemplary executable program may include instructions; the processor is used to execute the instructions stored in the memory. The memory can also be used to store data, which can be accessed and / or modified when instructions are executed.

[0141] The processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, and it is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in the storage medium to realize the corresponding method flow or corresponding function, so as to realize the steps of the operation optimization control method of a power distribution system in the above embodiments.

[0142] Example 5 Based on the same inventive concept, this invention also provides a readable storage medium, specifically an electronic device readable storage medium (Memory). An electronic device readable storage medium is a memory device within an electronic device used to store programs and data. It is understood that the storage medium here can include both built-in storage media within the electronic device and extended storage media supported by the electronic device. The storage medium provides storage space, which stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more executable programs (including program code). It should be noted that the storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. Loading and executing one or more instructions stored in the storage medium by the processor can implement the steps of the power distribution system operation optimization control method described in the above embodiments.

[0143] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0144] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0145] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0146] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0147] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit its scope of protection. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading the present invention, they can still make various changes, modifications or equivalent substitutions to the specific implementation methods of the application, but these changes, modifications or equivalent substitutions are all within the scope of protection of the claims pending approval.

Claims

1. A method for optimizing the operation control of a power distribution system, characterized in that, Applications in edge computing devices, including: The multi-source spatiotemporal data of each node in the power distribution system for the current control cycle and the physical power flow boundary parameters of the previous control cycle are spatiotemporally aligned and feature fused to obtain the spatiotemporal feature matrix of the current control cycle; the physical power flow boundary parameters are determined based on the voltage over-limit and equipment overload conditions of each node. Based on the multi-source spatiotemporal data and the nominal node admittance of the power distribution system, a dynamic adjacency matrix for the current control cycle is constructed; the nominal node admittance is a node admittance matrix constructed based on a preset benchmark topology. The spatiotemporal feature matrix and the dynamic adjacency matrix are input into the spatiotemporal prediction model, and the active power prediction interval boundary of each node in the future time period is output. The probability prediction parameters of active power are calculated based on the prediction interval boundary. The prediction interval and the probability prediction parameters are used as constraint parameters to input a differentiable convex optimization layer. A convex optimization problem with the goal of minimizing the operating cost of the power distribution system is constructed and solved. The optimized control commands for each node in the current control cycle are obtained and sent to the end-side equipment for execution.

2. The method according to claim 1, characterized in that, The multi-source spatiotemporal data includes: a basic spatiotemporal tensor and real-time switching status for the current control cycle; the basic spatiotemporal tensor includes historical electrical data of each node, meteorological data for future periods, and power grid spatial topology; the historical electrical data includes at least one of node voltage amplitude, active power injection, and reactive power injection for historical periods.

3. The method according to claim 1, characterized in that, The physical power flow boundary parameters include: voltage over-limit residuals and physical constraint shadow prices for each node; the voltage over-limit residuals are determined based on the deviation between the actual voltage amplitude of each node and the preset voltage safety upper and lower limits; the physical constraint shadow prices are determined based on the voltage over-limits and equipment overload conditions of each node.

4. The method according to claim 2, characterized in that, Spatiotemporal alignment and feature fusion are performed on the multi-source spatiotemporal data of the current control cycle and the physical power flow boundary parameters of the previous control cycle, including: The projected spacetime tensor is obtained by channel projection onto the fundamental spacetime tensor of the current control cycle. The real-time switching state of the current control cycle, the voltage over-limit residual and the physical constraint shadow price of the previous control cycle are projected onto a unified feature space through independent linear mapping to obtain the corresponding spatial alignment features. Temporal alignment is performed on each spatial alignment feature to obtain each spatiotemporal alignment feature; The spatiotemporal alignment features are concatenated with the projected spatiotemporal tensor along the feature channel dimension and then fused by convolution to obtain the spatiotemporal feature matrix.

5. The method according to claim 4, characterized in that, Before projecting the real-time switching state of the current control cycle, the voltage over-limit residual of the previous control cycle, and the shadow price of physical constraints onto a unified feature space through independent linear mappings, the following steps are also included: The real-time switch status is located to the switch type channel of each node using a first positioning operator; the first positioning operator is constructed based on the association between the node and the switch branch; The physical constraint shadow price of the previous control cycle is located to the constraint type channel of the corresponding node using the second positioning operator; the second positioning operator is constructed based on the constraint conditions of the differentiable convex optimization layer and the location of the node, branch or device associated with the constraint conditions. The voltage over-limit residual from the previous control cycle is located to the corresponding node channel according to the node identifier of each node.

6. The method according to claim 5, characterized in that, Based on the multi-source spatiotemporal data and the nominal node admittance of the power distribution system, a dynamic adjacency matrix for the current control cycle is constructed, including: Using the off-diagonal branch admittance amplitude of the nominal node admittance as the physical topology prior, the physical topology prior is adjusted using the physical connectivity relationship determined based on the real-time switch state to obtain the physical adjacency matrix; Based on the correlation of historical electrical data of each node in the basic spatiotemporal tensor and the real-time switching status, a data-driven adjacency matrix is ​​constructed. The physical adjacency matrix and the data-driven adjacency matrix are weighted and fused to obtain the dynamic adjacency matrix.

7. The method according to claim 6, characterized in that, Based on the correlation of historical electrical data of each node in the basic spatiotemporal tensor and the real-time switching status, a data-driven adjacency matrix is ​​constructed, including: Node-by-node temporal attention convergence is performed on the historical electrical data of each node in the basic spatiotemporal tensor to obtain converged features; Based on the convergence features and the real-time switching status of the switch type channels located to each node, node context features are generated. The data-driven adjacency matrix is ​​generated by scaling dot product attention based on the node context features.

8. The method according to claim 6, characterized in that, The physical adjacency matrix and the data-driven adjacency matrix are weighted and fused to obtain the dynamic adjacency matrix, which includes: On each operational branch, a learnable gating value is calculated based on the differences in historical electrical data between nodes and the physical adjacency matrix; The physical adjacency matrix and the data-driven adjacency matrix are adaptively weighted and fused using the learnable gating value, and the fusion result is normalized to obtain the dynamic adjacency matrix.

9. The method according to claim 1, characterized in that, The spatiotemporal feature matrix and the dynamic adjacency matrix are input into the spatiotemporal prediction model, which outputs the active power prediction interval boundary for each node in the future time period, including: Spatiotemporal encoding is performed on the spatiotemporal feature matrix and the dynamic adjacency matrix to extract node feature representations containing spatiotemporal correlation information; Multi-branch prediction is performed on the node feature representation, and the active power prediction interval boundary of each node in the future time period is output.

10. The method according to claim 9, characterized in that, Spatiotemporal encoding is performed on the spatiotemporal feature matrix and the dynamic adjacency matrix to extract node feature representations containing spatiotemporal correlation information, including: The spatiotemporal feature matrix and the dynamic adjacency matrix are sequentially spatiotemporally convolved through multiple cascaded spatiotemporal convolutional blocks to obtain the spatiotemporal feature representation of each node; Extract the features of the last time step in the spatiotemporal feature representation of each node as the node feature representation; Within each spatiotemporal convolutional block, a first gated temporal convolution, a spatial graph convolution, and a second gated temporal convolution are performed sequentially. In the first spatiotemporal convolutional block, the spatiotemporal feature matrix is ​​subjected to the first gated temporal convolution and then subjected to a spatial graph convolution with the dynamic adjacency matrix.

11. The method according to claim 9, characterized in that, The active power prediction interval boundary includes an upper bound and a lower bound; multi-branch prediction is performed on the node feature representation to output the active power prediction interval boundary for each node in the future time period, including: The node feature representation is subjected to a multi-layer nonlinear transformation to obtain shared features; The shared features are subjected to a first linear mapping and a first nonlinear activation to obtain the lower bound of the prediction interval; The shared features are subjected to a second linear mapping and a second nonlinear activation to obtain the proportion of the remaining available intervals; The remaining capacity between the lower bound and the upper bound of the prediction interval is scaled using the remaining available interval ratio to obtain the prediction interval width, and then the upper bound of the prediction interval is obtained.

12. The method according to claim 11, characterized in that, The probabilistic prediction parameters include the mean vector and covariance matrix of the predicted active power values ​​at each node; the probabilistic prediction parameters of active power are calculated based on the boundaries of the prediction interval, including: Using a preset two-tailed probability, the lower bound and upper bound of the prediction interval are converted into the mean vector and standard deviation matrix of the active power prediction values. The covariance matrix of the predicted active power is constructed based on the standard deviation matrix.

13. The method according to claim 3, characterized in that, Constructing a convex optimization problem with the objective of minimizing the operating cost of a power distribution system includes: Taking photovoltaic inverters, energy storage converters, and dynamic reactive power compensation devices as control objects, a convex optimization problem is constructed with the active and reactive power control quantities of each device as optimization variables. Based on the expected voltage deviation, expected network loss, and control action cost of the power distribution system, the operating cost is determined, and an objective function is constructed with the goal of minimizing the operating cost. The constraints of the convex optimization problem are constructed based on the power flow constraints, operational safety constraints, and equipment capacity constraints of the power distribution system. Based on the objective function and the constraints, the convex optimization problem is constructed.

14. The method according to claim 13, characterized in that, The objective function is: In the formula, To optimize variables, These represent the photovoltaic reactive power injection control quantity, energy storage active power control quantity, energy storage reactive power control quantity, and reactive power injection control quantity of the reactive power compensation device at time t, respectively. Let be the probability prediction parameters at time t. For probability prediction parameters The expected voltage deviation is as follows. For probability prediction parameters The expected network loss is as follows. To control the cost of actions, , These are the weights for the voltage deviation term and the network loss term, respectively. Indicates the expectation. Let be the linearized predicted voltage at time t. Reference voltage; This is the reference voltage when there is no control action. To predict the sensitivity matrix injected into the node voltage, Let be the random injection amount of the source charge at time t. , , This represents finding the covariance matrix. , These are the mean vector and covariance matrix of the active power prediction values ​​in the probability prediction parameters, respectively. To control the action The sensitivity matrix to the node voltage. For the linearized branch power flow at time t, For the linearized branch power flow at time t without control action, These are the sensitivity matrices for predicting the injection to the branch power flow. To control the sensitivity matrix of actions to branch power flow, This is an approximate matrix for network loss; This is the penalty coefficient matrix for controlling actions, with the superscript T indicating the transpose operation.

15. The method according to claim 14, characterized in that, The solution process for the convex optimization problem includes: The convex optimization problem is transformed into a second-order cone programming form. The cone optimization solver is then called to perform a forward solution, which yields the optimized control instructions for each node in the current control cycle.

16. The method according to claim 15, characterized in that, The operational safety constraints include voltage boundary constraints at each node; the generation process of the physical constraint shadow price includes: Extract the dual variables corresponding to the voltage boundary constraints of each node after the convergence of the convex optimization problem in the previous control cycle; The voltage shadow price is obtained by subtracting the voltage upper limit constraint dual component and the voltage lower limit constraint dual component in the dual variables corresponding to the voltage boundary constraints of each node. Extract the dual variable corresponding to the equipment capacity constraint as the capacity shadow price; The physical constraint shadow price is generated based on the voltage shadow price and the capacity shadow price.

17. The method according to any one of claims 1-16, characterized in that, The update process of the spatiotemporal prediction model includes: Based on the statistical deviation between the mean vector of the active power prediction values ​​in the probability prediction parameters and the measured power after executing the optimized control command, the prediction loss of the current control cycle is determined. Based on the deviation between the evaluated voltage obtained after AC power flow evaluation of the optimized control command and the preset upper and lower voltage safety limits, the voltage over-limit hinge penalty is determined; based on the voltage over-limit hinge penalty and the network loss penalty generated by executing the optimized control command, the decision loss of the current control cycle is determined; Based on the predicted loss and the decision loss, determine the current task loss; If the current task loss is greater than a preset loss threshold, the gradient of the current task loss is uploaded to the cloud so that the cloud can update and distribute the spatiotemporal prediction model based on the gradient uploaded by each edge computing device.

18. An edge computing device for a power distribution system, characterized in that, include: The feature space construction module is used to perform spatiotemporal alignment and feature fusion on the multi-source spatiotemporal data of each node of the power distribution system for the current control cycle and the physical power flow boundary parameters of the previous control cycle to obtain the spatiotemporal feature matrix of the current control cycle. The physical power flow boundary parameters are determined based on the voltage over-limit and equipment overload conditions of each node; The topology modeling module is used to construct the dynamic adjacency matrix for the current control cycle based on the multi-source spatiotemporal data and the nominal node admittance of the power distribution system. The nominal node admittance is a node admittance matrix constructed based on a preset benchmark topology; The prediction module is used to input the spatiotemporal feature matrix and the dynamic adjacency matrix into the spatiotemporal prediction model, output the active power prediction interval boundary of each node in the future time period, and calculate the probability prediction parameters of active power based on the prediction interval boundary. The optimization module is used to input the prediction interval and the probability prediction parameters as constraint parameters into the differentiable convex optimization layer, construct and solve a convex optimization problem with the goal of minimizing the operating cost of the power distribution system, obtain the optimized control commands for each node in the current control cycle, and send them to the end-side equipment for execution.

19. The edge computing device according to claim 18, characterized in that, The multi-source spatiotemporal data includes: a basic spatiotemporal tensor and real-time switching status for the current control cycle; the basic spatiotemporal tensor includes historical electrical data of each node, meteorological data for future periods, and power grid spatial topology; the historical electrical data includes at least one of node voltage amplitude, active power injection, and reactive power injection for historical periods.

20. The edge computing device according to claim 18, characterized in that, The physical power flow boundary parameters include: voltage over-limit residuals and physical constraint shadow prices for each node; the voltage over-limit residuals are determined based on the deviation between the actual voltage amplitude of each node and the preset voltage safety upper and lower limits; the physical constraint shadow prices are determined based on the voltage over-limits and equipment overload conditions of each node.

21. The edge computing device according to claim 19, characterized in that, The feature space construction module includes: The channel projection layer is used to project the basic spatiotemporal tensor of the current control cycle into the projected spatiotemporal tensor. The spatial projection layer is used to project the real-time switching state of the current control cycle, the voltage over-limit residual and the physical constraint shadow price of the previous control cycle onto a unified feature space through independent linear mappings to obtain the corresponding spatial alignment features. The temporal alignment layer is used to perform temporal alignment on each spatial alignment feature to obtain each spatiotemporal alignment feature. The feature fusion layer is used to convolve and fuse the spatiotemporal aligned features with the projected spatiotemporal tensor along the feature channel dimension to obtain the spatiotemporal feature matrix.

22. The edge computing device according to claim 21, characterized in that, The feature space construction module further includes: The first positioning layer is used to locate the real-time switch state to the switch type channel of each node using a first positioning operator; the first positioning operator is constructed based on the association relationship between the node and the switch branch; The second positioning layer is used to locate the physical constraint shadow price of the previous control cycle to the constraint type channel of the corresponding node using the second positioning operator; the second positioning operator is constructed based on the constraint conditions of the differentiable convex optimization layer and the location of the node, branch or device associated with the constraint conditions. The third positioning layer is used to locate the voltage over-limit residual of the previous control cycle to the corresponding node channel according to the node identifier of each node.

23. The edge computing device according to claim 22, characterized in that, The topology modeling module includes: The physical modeling layer is used to adjust the physical topology prior by using the off-diagonal branch admittance amplitude of the nominal node admittance as the physical topology prior and the physical connectivity relationship determined based on the real-time switch state to obtain the physical adjacency matrix. The data modeling layer is used to construct a data-driven adjacency matrix based on the correlation of historical electrical data of each node in the basic spatiotemporal tensor and the real-time switching status. The matrix fusion layer is used to perform weighted fusion of the physical adjacency matrix and the data-driven adjacency matrix to obtain the dynamic adjacency matrix.

24. The edge computing device according to claim 23, characterized in that, The data modeling layer is specifically used for: Node-by-node temporal attention convergence is performed on the historical electrical data of each node in the basic spatiotemporal tensor to obtain converged features; Based on the convergence features and the real-time switching status of the switch type channels located to each node, node context features are generated. The data-driven adjacency matrix is ​​generated by scaling dot product attention based on the node context features.

25. The edge computing device according to claim 23, characterized in that, The matrix fusion layer is specifically used for: On each operational branch, a learnable gating value is calculated based on the differences in historical electrical data between nodes and the physical adjacency matrix; The physical adjacency matrix and the data-driven adjacency matrix are adaptively weighted and fused using the learnable gating value, and the fusion result is normalized to obtain the dynamic adjacency matrix.

26. The edge computing device according to claim 18, characterized in that, The prediction module includes a spatiotemporal prediction model, which includes: A spatiotemporal encoder is used to perform spatiotemporal encoding on the spatiotemporal feature matrix and the dynamic adjacency matrix to extract node feature representations containing spatiotemporal correlation information; The multi-branch prediction head is used to perform multi-branch prediction on the node feature representation and output the active power prediction interval boundary of each node in the future time period.

27. The edge computing device according to claim 26, characterized in that, The spatiotemporal encoder comprises multiple cascaded spatiotemporal convolutional blocks; the spatiotemporal encoder is specifically used for: The spatiotemporal feature matrix and the dynamic adjacency matrix are sequentially spatiotemporally convolved through multiple cascaded spatiotemporal convolutional blocks to obtain the spatiotemporal feature representation of each node; Extract the features of the last time step in the spatiotemporal feature representation of each node as the node feature representation; Within each spatiotemporal convolutional block, a first gated temporal convolution, a spatial graph convolution, and a second gated temporal convolution are performed sequentially. In the first spatiotemporal convolutional block, the spatiotemporal feature matrix is ​​subjected to the first gated temporal convolution and then subjected to a spatial graph convolution with the dynamic adjacency matrix.

28. The edge computing device according to claim 26, characterized in that, The multi-branch prediction head includes: A shared transformation layer is used to perform multi-layer nonlinear transformations on the node feature representation to obtain shared features; The first prediction branch is used to perform a first linear mapping and a first nonlinear activation on the shared features to obtain the lower bound of the prediction interval. The second prediction branch is used to perform a second linear mapping and a second nonlinear activation on the shared features to obtain the proportion of the remaining available interval; The physical interval construction layer is used to scale the remaining capacity between the lower boundary and the upper boundary of the prediction interval using the proportion of the remaining available interval, so as to obtain the width of the prediction interval and thus the upper boundary of the prediction interval.

29. The edge computing device according to claim 28, characterized in that, The probability prediction parameters include the mean vector and covariance matrix of the predicted active power values ​​for each node; the prediction module also includes a conversion layer, which is used for: Using a preset two-tailed probability, the lower bound and upper bound of the prediction interval are converted into the mean vector and standard deviation matrix of the active power prediction values. The covariance matrix of the predicted active power is constructed based on the standard deviation matrix.

30. The edge computing device according to claim 20, characterized in that, The optimization module includes an optimization problem construction submodule, which is used for: Taking photovoltaic inverters, energy storage converters, and dynamic reactive power compensation devices as control objects, a convex optimization problem is constructed with the active and reactive power control quantities of each device as optimization variables. Based on the expected voltage deviation, expected network loss, and control action cost of the power distribution system, the operating cost is determined, and an objective function is constructed with the goal of minimizing the operating cost. The constraints of the convex optimization problem are constructed based on the power flow constraints, operational safety constraints, and equipment capacity constraints of the power distribution system. Based on the objective function and the constraints, the convex optimization problem is constructed.

31. The edge computing device according to claim 30, characterized in that, The optimization module includes a solution submodule, which is used for: The convex optimization problem is transformed into a second-order cone programming form. The cone optimization solver is then called to perform a forward solution, which yields the optimized control instructions for each node in the current control cycle.

32. The edge computing device according to claim 31, characterized in that, The operational safety constraints include voltage boundary constraints for each node; the optimization module further includes a feedback generation submodule, which is used for: Extract the dual variables corresponding to the voltage boundary constraints of each node after the convergence of the convex optimization problem in the previous control cycle; The voltage shadow price is obtained by subtracting the voltage upper limit constraint dual component and the voltage lower limit constraint dual component in the dual variables corresponding to the voltage boundary constraints of each node. Extract the dual variable corresponding to the equipment capacity constraint as the capacity shadow price; The physical constraint shadow price is generated based on the voltage shadow price and the capacity shadow price.

33. The edge computing device according to any one of claims 18-32, characterized in that, It also includes an update module for: Based on the statistical deviation between the mean vector of the active power prediction values ​​in the probability prediction parameters and the measured power after executing the optimized control command, the prediction loss of the current control cycle is determined. Based on the deviation between the evaluated voltage obtained after AC power flow evaluation of the optimized control command and the preset upper and lower voltage safety limits, the voltage over-limit hinge penalty is determined; based on the voltage over-limit hinge penalty and the network loss penalty generated by executing the optimized control command, the decision loss of the current control cycle is determined; Based on the predicted loss and the decision loss, determine the current task loss; If the current task loss is greater than a preset loss threshold, the gradient of the current task loss is uploaded to the cloud so that the cloud can update and distribute the spatiotemporal prediction model based on the gradient uploaded by each edge computing device.

34. A power distribution system operation optimization control system, characterized in that, include: The edge computing device as described in any one of claims 18-33.

35. An electronic device, characterized in that, include: At least one processor and memory; The memory and processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, the method as described in any one of claims 1 to 17 is implemented.

36. A readable storage medium, characterized in that, It contains an executable program, which, when executed, implements the method as described in any one of claims 1 to 17.

Images