A power distribution cabinet intelligent power distribution adjusting method and system based on a PLC

Abstract

The application provides a power distribution cabinet intelligent power distribution regulation method and system based on PLC, relates to the power system power distribution automation technical field, in particular to the intelligent power distribution regulation and multi-physical field coupling control technical field of the power distribution cabinet, and solves the technical problems that the environment data and electrical data are mutually separated, the multi-physical field coupling constraint is missing, and the regulation strategy cannot be dynamically and adaptively adjusted according to the environment state in the prior art. The method comprises the following steps: collecting environment sensing data and electrical operation data, constructing an environment semantic graph and a power distribution knowledge graph, and performing space coordinate alignment and graph neural network coding; a fusion feature matrix is generated through a cross-attention mechanism fusion; an environment risk coefficient and a thermal-electric coupling coefficient are extracted, and the dynamic constraint boundary of model predictive control is adjusted accordingly; an optimal control sequence is solved and converted into a PLC instruction to execute a regulation action. The application is used in the real-time intelligent power distribution regulation and safe operation control process of the power distribution cabinet under complex environmental conditions.

Description

Technical Field

[0001] This invention belongs to the field of power system distribution automation technology, specifically a PLC-based intelligent power distribution regulation method and system for distribution cabinets. Background Technology

[0002] With the continuous advancement of intelligent power system construction, distribution cabinets, as key power distribution equipment at the end of the power grid, directly affect the power supply quality of the entire power system through their operational reliability and regulation accuracy. Existing technologies offer various solutions for distribution cabinet regulation, primarily including: regulation methods based on electrical variable feedback, which adjust output power by monitoring changes in electrical parameters such as circuit voltage and current; regulation methods based on environmental parameter compensation, which use temperature and humidity sensors to collect environmental data and compensate for and correct the monitored electrical variable values; and data-driven intelligent regulation methods, which employ machine learning algorithms to learn the influence of operating parameters and environmental factors on the distribution cabinet's state and generate adjustment schemes. Furthermore, in recent years, exploratory technologies such as knowledge graph-based power distribution system optimization control methods and model predictive control-based electrical cabinet temperature management schemes have emerged.

[0003] However, the existing technologies still have the following shortcomings: First, the processing of environmental data and electrical data is fragmented. Existing methods typically treat environmental parameters as auxiliary correction terms for electrical measurements or as independent adjustment bases, failing to incorporate the spatial distribution characteristics of the environmental field and the structural characteristics of the electrical topology into a unified graph structure framework for joint representation and collaborative reasoning. This results in environmental information not being able to deeply participate in the structured decision-making process of power distribution regulation. Second, there is a lack of systematic modeling of multi-physics coupling constraints. There are complex nonlinear coupling relationships between the temperature field, humidity field, and electric field inside the distribution cabinet. However, existing regulation methods often only control a single physical quantity independently, failing to comprehensively consider the impact of thermal and electrical coupling on the equipment operating boundary within the same optimization framework. This makes it difficult to accurately describe the true safe operating domain of the equipment under complex environmental conditions. Third, the constraint boundaries of the regulation strategy are fixed and lack environmental adaptability. Most existing methods use preset fixed thresholds as control constraints, failing to dynamically adjust the safe operating boundary according to the real-time distribution of the environmental field and the thermal and electrical coupling strength between devices. This causes the regulation strategy to deviate from the optimal or even fail when environmental conditions change.

[0004] The aforementioned shortcomings directly lead to the following technical problems: insufficient adjustment accuracy of the distribution cabinet under complex environmental conditions; inability to effectively perceive and avoid the risks of multi-physics coupling; and lag in the response of the adjustment strategy to environmental changes, seriously affecting the operational safety and adjustment reliability of the power distribution system under harsh environmental conditions. Therefore, there is an urgent need for an intelligent power distribution adjustment method for distribution cabinets that can deeply integrate the spatial distribution of the environmental field with the electrical topology, dynamically perceive multi-physics coupling constraints, and adaptively adjust the adjustment strategy. Summary of the Invention

[0005] This application provides a PLC-based intelligent power distribution regulation method and system for power distribution cabinets, which solves the technical problems in the prior art such as the separation of environmental data and electrical data, the lack of multi-physics coupling constraints, and the inability of regulation strategies to dynamically and adaptively adjust with environmental conditions.

[0006] To achieve the above objectives, this application adopts the following technical solution: Firstly, a PLC-based intelligent power distribution regulation method for distribution cabinets is provided, including: Collect environmental sensor data and electrical operation data inside the distribution cabinet, construct an environmental semantic graph containing spatial grid nodes and environmental features, and a power distribution knowledge graph containing equipment entities and electrical relationships; The environmental semantic graph and the power distribution knowledge graph are spatially aligned, and graph neural networks are used to encode the spatially aligned environmental semantic graph and the power distribution knowledge graph respectively. Feature fusion is performed through a cross-attention mechanism to generate a fused feature matrix. The dynamic constraint boundary of the model predictive control is adjusted based on the environmental risk coefficient of the device node and the thermal-electric coupling coefficient between the device pairs extracted from the fusion feature matrix. Based on the current state of the system and the dynamic constraint boundary, the optimal control sequence of model predictive control is solved, and the optimal control sequence is converted into PLC instructions, which are then used by the PLC controller to execute the adjustment action.

[0007] Based on the above technical solutions, in the PLC-based intelligent power distribution regulation method for distribution cabinets provided in this application, the environmental semantic graph and the power distribution knowledge graph are spatially aligned and encoded separately using a graph neural network. Then, they are fused using a cross-attention mechanism to generate a fusion feature matrix. This enables deep integration of the spatial distribution characteristics of the environmental field and the structural characteristics of the electrical topology within the same framework, giving the power distribution system spatial semantic perception capabilities. By dynamically adjusting the constraint boundaries of model predictive control through the environmental risk coefficient of equipment nodes and the thermal-electric coupling coefficient between equipment pairs, the control decision is not only based on electrical measurements but also integrates the real-time environmental risk of the equipment and the multi-physics coupling relationship between equipment. This achieves a leap from fixed threshold constraints to environmental adaptive constraints, solving the technical problem of the disconnect between regulation strategies and environmental states in traditional methods. By converting the optimal control sequence of model predictive control into PLC instructions and having the PLC controller execute the regulation actions, the optimality of the control strategy and the high real-time performance and high reliability of engineering implementation are taken into account. Therefore, this application realizes intelligent power distribution regulation of distribution cabinets based on field and graph joint embedded driving, which effectively improves the regulation accuracy, adaptability and operation safety of power distribution system under complex environmental conditions, and provides a new technical path for intelligent operation and maintenance of distribution cabinets.

[0008] Furthermore, the process of constructing the environmental semantic graph includes: The internal space of the power distribution cabinet is discretized into three-dimensional grid nodes according to the preset grid spacing. The feature vector of each node is composed of temperature value, humidity value, airflow velocity value, temperature gradient value and humidity gradient value in three axes, semantic label and status flag. Spatial proximity is calculated using a Gaussian kernel function based on Euclidean distance of spatial coordinates, and environmental feature similarity is calculated using a Gaussian kernel function based on Euclidean distance of environmental feature vectors. The edge weights between nodes are obtained by weighting and summing the values ​​of spatial proximity and environmental feature similarity. The feature vectors of each node are used as node attributes, the edge weight matrix is ​​used as the adjacency matrix, and the grid nodes are used as vertices and the lines connecting adjacent nodes in space are used as edges to construct a weighted undirected graph, thus obtaining the environmental semantic graph. The semantic labels are obtained by clustering algorithms based on the environmental sensing data of each grid node, and are used to characterize the area type to which the grid node belongs. The area types include hot spots, cold spots, risk areas, normal areas, ventilation bottleneck areas, and transition areas. The status flag consists of binary 0 or 1 and is used to characterize whether the node is occupied by a device.

[0009] Furthermore, the construction process of the power distribution knowledge graph includes: Entities and entity attributes are extracted from the electrical topology of the distribution cabinet, equipment nameplate parameters, and historical operation records. The entity types include power supply nodes, bus nodes, circuit breaker nodes, switch nodes, load nodes, compensation nodes, and transformer nodes. The entity attributes include rated voltage, rated current, rated capacity, impedance value, health status, spatial coordinates, and current operating status. Extract the relationships between entities, including electrical connection relationships, control relationships, membership relationships, thermal impact relationships, dependency relationships, and spatial proximity relationships; A graph embedding model based on complex vector space is used to map the entities and relations to complex vector space, thereby obtaining the embedding vectors of each entity and the embedding vectors of each relation. Based on the spatial proximity and thermal influence relationships between entities, construct the adjacency matrix of the power distribution knowledge graph, calculate the graph Laplacian matrix of the power distribution knowledge graph, and complete the construction of the power distribution knowledge graph.

[0010] Furthermore, the construction process of the graph embedding model based on complex vector space also includes: Entities and relations are extracted from the power distribution knowledge graph to obtain triples containing head entities, relations, and tail entities; Map the head entity vector, relation vector, and tail entity vector to the complex vector space; The scoring function of the triple is calculated using the negative norm between the Hadamard product of the head entity vector and the relation vector and the tail entity vector; the formula for the scoring function is as follows: ;in, Let be the embedding vector of the head entity in the complex vector space. Let be the embedding vector of the relation in the complex vector space. Let be the embedding vector of the tail entity in the complex vector space. For Hadamard product, Let the norm of the complex vector be the scoring function. The closer the value is to 0, the higher the reliability of the triple.

[0011] Furthermore, the spatial coordinate alignment of the environmental semantic graph with the power distribution knowledge graph specifically includes: Obtain the spatial coordinates of each device entity in the power distribution knowledge graph; Using the spatial coordinates of each device entity as the query vector, search for the nearest Euclidean distance grid node among all grid nodes in the environmental semantic graph, and establish the nearest neighbor mapping relationship from the device entity to the nearest grid node; Calculate the Euclidean distance between the spatial coordinates of each device entity and the spatial coordinates of its nearest neighbor grid node as the mapping error; If the mapping error is greater than a preset threshold, a manual verification warning is triggered; otherwise, the device entity is bound to the nearest neighbor grid node to obtain the binding mapping relationship between the device entity and the environmental semantic graph grid node. The binding mapping relationship is used for cross-graph information transmission in the subsequent feature aggregation process. Spatial coordinate alignment is completed based on the binding mapping relationship.

[0012] Furthermore, the spatial coordinate alignment also includes a topological difference calculation process, specifically: Based on the binding mapping relationship, the graph Laplacian matrix of the power distribution knowledge graph is spatially mapped and resampled to the spatial coordinate system of the environmental semantic graph to obtain the resampled power distribution knowledge graph Laplacian matrix. The topological difference matrix is ​​obtained by calculating the difference between the graph Laplacian matrix of the environmental semantic graph and the Laplacian matrix of the resampled power distribution knowledge graph. Calculate the F-norm of the topology difference matrix. If the F-norm is greater than a preset topology difference threshold, generate a trigger signal to trigger dynamic adjustment of model predictive control constraints. Positive values ​​of the topology difference matrix indicate that the coupling strength between nodes in the environmental semantic graph is greater than the coupling strength between corresponding nodes in the electrical graph, while negative values ​​indicate that the coupling strength between nodes in the electrical graph is greater than the coupling strength between corresponding nodes in the environmental graph.

[0013] Furthermore, the process of generating the fused feature matrix specifically includes: A graph convolutional network is used to encode the environment semantic graph. The node feature matrix and adjacency matrix of the environment semantic graph are used as inputs, and the environment embedding vector of each grid node in the environment semantic graph is output. The node feature matrix of the environment semantic graph is obtained by stacking the feature vectors of each grid node in the order of node index. A graph attention network is used to encode the power distribution knowledge graph. The complex vector space embedding and adjacency matrix of each entity in the power distribution knowledge graph are used as input. The attention weights of the neighbor nodes of each entity are calculated by a multi-head graph attention layer and the neighbor features are aggregated. The output is the electrical embedding vector of each device entity in the power distribution knowledge graph. The environment embedding vectors of the nearest neighbor mesh nodes corresponding to each device entity are aggregated according to the spatial coordinate mapping relationship to obtain the environment aggregation features corresponding to each device entity. Using the electrical embedding vectors of each device entity as the query matrix and the environmental aggregation features as the key and value matrices, the first cross-attention feature from the power distribution knowledge graph to the environmental semantic graph is calculated. Using the environmental aggregation features corresponding to each device entity as the query matrix and the electrical embedding vector of each device entity as the key matrix and value matrix, the second cross-attention matrix from the environmental semantic graph to the power distribution knowledge graph is calculated. The electrical embedding vector of each device entity is added to the first cross-attention feature, and the environmental aggregation feature of each device entity is added to the second cross-attention feature. The two parts of the added feature are then concatenated to obtain the fused feature vector of each device entity. The fused feature vectors of all device entities constitute the fused feature matrix.

[0014] Furthermore, the method for obtaining the environmental risk coefficient includes: The fused feature vector of each device entity in the fused feature matrix is ​​input into a pre-trained multilayer perceptron network, and the environmental risk coefficient of the device entity is output; the environmental risk coefficient is used to characterize the relative magnitude of the device operation risk under the current environmental conditions.

[0015] Furthermore, the method for obtaining the thermal-electric coupling coefficient includes: For each pair of device entities in the power distribution knowledge graph, calculate the cosine similarity between the fused feature vectors; The cosine similarity is divided by a preset temperature coefficient and then input into the Softmax function for normalization to obtain the thermal-electric coupling coefficient between the device entities; the thermal-electric coupling coefficient is used to characterize the degree of mutual influence between the two devices under the coupling of thermal and electric fields.

[0016] Furthermore, the adjustment of the dynamic constraint boundary of the model predictive control specifically includes: Obtain the rated operating boundaries of each device entity, including the lower limit of rated voltage, the upper limit of rated voltage, the upper limit of rated current, and the upper limit of rated temperature; Multiply the environmental risk coefficient by the preset environmental deduction coefficient to obtain the environmental deduction amount for each device. The environmental derating is superimposed on the rated operating boundary of each device to generate the dynamic constraint boundary for the current control cycle. The dynamic constraint boundary is used as the inequality constraint condition for the optimization problem of model predictive control. Determine whether the F-norm of the topological difference matrix is ​​greater than a preset threshold; if yes, increase the environmental depreciation coefficient to a preset emergency depreciation value, recalculate the dynamic constraint boundary, and execute.

[0017] Furthermore, the step of superimposing the environmental derating amount onto the rated operating boundaries of each device to generate the dynamic constraint boundaries for the current control cycle specifically includes: The voltage constraint boundary for the i-th device node is: , ;in, This represents the lower boundary of the voltage constraint. This represents the upper boundary of the voltage constraint. The preset environmental deduction factor. Let i be the environmental risk coefficient of the i-th device node. Let be the lower limit of the rated voltage of the i-th device node. This represents the upper limit of the rated voltage of the i-th device node; The current constraint boundary for the i-th device node is: ;in, This represents the upper boundary of the current constraint. This represents the upper limit of the rated current of the i-th device node; The temperature constraint boundary for the i-th device node is: ;in, This represents the upper boundary of the temperature constraint. The rated temperature limit is the upper limit for the i-th device node. This is the preset temperature margin coefficient.

[0018] Furthermore, the step of using the dynamic constraint boundary as an inequality constraint condition for the optimization problem of model predictive control includes: In the prediction time domain For each prediction step k+t|k within the range, the inequality constraint is expressed as: This indicates that the voltage amplitude of the i-th device node always operates between the dynamic lower voltage limit and the dynamic upper voltage limit in each prediction step. This indicates that the voltage magnitude of the i-th device node is always non-negative in each prediction step and does not exceed the dynamic current limit. This indicates that the current amplitude of the i-th device node never exceeds the dynamic temperature limit in each prediction step. And define the magnitude constraints and increment constraints of the control variables in model predictive control: This indicates that the values ​​of all control variables in each prediction step always lie between the lower limit and the upper limit of the control variables, where the lower limit of the control variables is determined by the minimum physical limit of each actuator, and the upper limit of the control variables is determined by the maximum physical limit of each actuator. This indicates that the change in all control variables between two adjacent control cycles does not exceed the preset maximum increment limit, which is used to limit the action rate of the actuator to avoid shocks and oscillations. in, , , Let these represent the voltage, current, and temperature of the i-th device node, respectively. To control variables, To control the increment, This indicates the lower limit of the control variable. This indicates the upper limit of the control variable. This indicates the maximum increment limit.

[0019] Furthermore, the step of solving for the optimal control sequence of model predictive control based on the current system state and the dynamic constraint boundary specifically includes: A state-space prediction model for a power distribution system is established. The state variables of the state-space prediction model include the voltage amplitude, current amplitude, active power, reactive power, and equipment temperature of each equipment node. The control variables of the state-space prediction model include opening and closing commands, capacitor switching commands, and tap position commands of on-load tap-changing transformers. The state-space prediction model adopts a physical information neural network. The system state value of the current control cycle is obtained, and the future state is recursively predicted in multiple steps based on the state space prediction model in the prediction time domain; the system state value represents the state variables composed of the electrical operating parameters and equipment temperature parameters of each equipment node in the power distribution system obtained at the beginning of the current control cycle. During the recursive process, the dynamic constraint boundary is used as an inequality constraint of the multi-objective optimization function, and the sequential quadratic programming algorithm is used to solve the multi-objective optimization objective function. In each iteration, the environmental risk coefficient, thermal-electric coupling coefficient and dynamic constraint boundary are updated according to the fusion feature matrix at the current time. The first control step in the optimal control sequence obtained in the control time domain is used as the execution instruction of the current control cycle to obtain the optimal control sequence. The optimal control sequence includes: digital instructions, which are mapped to the digital output of the PLC to control the opening and closing of the circuit breaker; analog instructions, which are mapped to the analog output of the PLC to control the switching degree of the capacitor; and step instructions, which are mapped to the pulse output of the PLC to control the tap position of the on-load tap-changing transformer.

[0020] Furthermore, the state-space prediction model employs a physical information neural network, specifically expressed as follows: ;in, Let k be the state variable at time k. Let k be the control variable at time k. This is the uncontrollable disturbance input at time k. For data-driven neural network terms, These are analytical terms based on the physical equations; The analytical terms include at least the power flow equation and the thermal balance equation to ensure that the prediction results satisfy physical laws; the data-driven neural network terms use long short-term memory networks or gated recurrent networks to learn the dynamic characteristics of environmental disturbances. The physical information neural network is trained by minimizing the weighted sum of the prediction error and the physical equation residuals.

[0021] Furthermore, the multi-objective optimization function is used to minimize the tracking error of the output variable relative to the reference value, the control increment, and the physical penalty term, while satisfying the dynamic constraint boundaries. The specific expression is as follows: ;in, To predict the time domain, Let be the predicted output variable at time k+t. This is the reference value vector for the output variables. This is the control increment at time k+t. To output the weight matrix of the tracking error, The weight matrix for controlling the increment, The weighting coefficient for the physical penalty term. To fuse the feature matrix, This is a physical penalty item; The expression for the physical penalty term is: ;in, Let i be the environmental risk coefficient of the i-th device node. Let be the thermal-electric coupling coefficient between device i and device j. This is the control variable that acts on the i-th device node.

[0022] Secondly, this application provides a PLC-based intelligent power distribution and regulation system for distribution cabinets, comprising: a graph construction module, a feature fusion module, a model prediction and control module, and a PLC execution module; wherein, The graph construction module is used to collect environmental sensing data and electrical operation data in the power distribution cabinet, construct an environmental semantic graph containing spatial grid nodes and environmental features, and a power distribution knowledge graph containing equipment entities and electrical relationships. The feature fusion module is used to align the environmental semantic graph with the power distribution knowledge graph in spatial coordinates, and to encode the spatially aligned environmental semantic graph and power distribution knowledge graph respectively using a graph neural network, and to perform feature fusion through a cross-attention mechanism to generate a fused feature matrix. The model predictive control module is used to adjust the dynamic constraint boundary of model predictive control based on the environmental risk coefficient of the device node and the thermal-electric coupling coefficient between the device pairs extracted from the fusion feature matrix, and to solve the optimal control sequence of model predictive control based on the current state of the system and the dynamic constraint boundary. The PLC execution module is used to convert the optimal control sequence into PLC instructions and use the PLC controller to execute adjustment actions.

[0023] Compared with the prior art, the beneficial effects of this application are: This application constructs an environmental semantic graph containing spatial grid nodes and environmental features, and uses a clustering algorithm to perform category semantic annotation on the grid nodes, thus elevating environmental data from discrete sensor values ​​to graph-structured data with spatial topology and regional semantics. Based on this, a power distribution knowledge graph containing multiple types of electrical equipment entities and multidimensional relationships is further constructed. A graph embedding model based on complex vector space is used to model entities and relationships, and a scoring function is used to quantitatively evaluate the credibility of triples, achieving a unified representation and high-precision embedding of structured and unstructured information in the power distribution system. By aligning the environmental semantic graph with the power distribution knowledge graph using spatial coordinates and calculating topological differences, this application can quantitatively evaluate the spatial matching degree between the environmental field distribution and the electrical topology. When the topological difference between the two exceeds a threshold, dynamic tightening of control constraints is automatically triggered, providing prior warning information at the spatial semantic level for subsequent intelligent adjustment decisions. Based on this, this application uses graph convolutional networks to encode the environmental semantic graph and graph attention networks to encode the power distribution knowledge graph. It also uses a bidirectional cross-attention mechanism to achieve deep feature interaction and fusion between the two types of heterogeneous graph structures, so that the fused feature matrix simultaneously contains the electrical attribute information of the device itself, the topological connection information between devices, and the physical field distribution information of the local environment where the device is located.

[0024] This application further extracts the environmental risk coefficient of the device nodes and the thermal-electric coupling coefficient between device pairs from the fusion feature matrix, enabling the system to quantitatively assess the operational risk level of each device under the current environmental conditions, as well as the degree of mutual influence between any two devices under the coupling of thermal and electric fields. Based on the above coefficients, the voltage, current, and temperature constraint boundaries of the model predictive control are dynamically adjusted, so that the control constraints are adaptively tightened or relaxed as the environmental state changes, and the system automatically switches to emergency derating mode when the topology difference exceeds the threshold, realizing the dynamic online update of the safety domain under multi-physics coupling conditions. Building upon this foundation, this application employs a physical information neural network to construct a state-space prediction model. It integrates the data-driven neural network with analytical terms related to physical laws such as power flow equations and thermal balance equations, enabling the prediction results to possess both data-driven adaptive capabilities and satisfy fundamental physical conservation constraints. By introducing a physical penalty term composed of an environmental risk coefficient and a thermo-electric coupling coefficient into the multi-objective optimization function, the optimization decision proactively avoids significant equipment adjustments in high-environment-risk areas while satisfying electrical constraints and suppressing differentiated actions between strongly coupled equipment pairs. This, in turn, enhances the operational safety, regulation stability, and equipment lifespan of the power distribution system under complex environmental conditions, while ensuring voltage quality and power factor. Finally, this application transforms the optimal control sequence of model predictive control into PLC-executable digital, analog, and step instructions, balancing the global optimality of forward-looking optimization decisions with the high real-time performance and high reliability of industrial-grade execution. This provides a complete technical path for intelligent power distribution regulation of distribution cabinets. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 A system architecture diagram of a PLC-based intelligent power distribution and regulation system for a power distribution cabinet is provided for embodiments of this application; Figure 2 A flowchart illustrating a PLC-based intelligent power distribution regulation method for a power distribution cabinet, as provided in an embodiment of this application; Figure 3 A flowchart illustrating another PLC-based intelligent power distribution regulation method for distribution cabinets provided in this application embodiment; Figure 4 This is a flowchart illustrating another PLC-based intelligent power distribution regulation method for distribution cabinets, provided as an embodiment of this application. Detailed Implementation

[0027] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0028] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0029] The PLC-based intelligent power distribution regulation method for distribution cabinets provided in this application embodiment can be applied to, for example... Figure 1 The illustrated example is a PLC-based intelligent power distribution and regulation system for a distribution cabinet, such as... Figure 1 As shown, the system includes: a map construction module, a feature fusion module, a model prediction and control module, and a PLC execution module; among which, The graph construction module is used to collect environmental sensing data and electrical operation data in the power distribution cabinet, and to construct an environmental semantic graph containing spatial grid nodes and environmental features, as well as a power distribution knowledge graph containing equipment entities and electrical relationships. The feature fusion module is used to align the spatial coordinates of the environmental semantic graph and the power distribution knowledge graph, and to encode the spatially aligned environmental semantic graph and power distribution knowledge graph respectively using a graph neural network. Then, the feature is fused through a cross-attention mechanism to generate a fused feature matrix. The model predictive control module is used to adjust the dynamic constraint boundary of model predictive control based on the environmental risk coefficient of the device node and the thermal-electric coupling coefficient between the device pairs extracted from the fusion feature matrix, and to solve the optimal control sequence of model predictive control based on the current system state and the dynamic constraint boundary. The PLC execution module is used to convert the optimal control sequence into PLC instructions, and then use the PLC controller to execute the adjustment actions.

[0030] To address the technical problem in existing technologies where environmental sensing and electrical control in power distribution cabinets are disconnected, making it difficult to achieve coordinated environmental-electrical optimization, this application provides a PLC-based intelligent power distribution regulation method for power distribution cabinets. This method includes: Collect environmental sensor data and electrical operation data inside the distribution cabinet, construct an environmental semantic graph containing spatial grid nodes and environmental features, and a power distribution knowledge graph containing equipment entities and electrical relationships; The environmental semantic graph and the power distribution knowledge graph are spatially aligned, and graph neural networks are used to encode the spatially aligned environmental semantic graph and power distribution knowledge graph respectively. The features are then fused through a cross-attention mechanism to generate a fused feature matrix. The dynamic constraint boundary of the model predictive control is adjusted based on the environmental risk coefficient of the device node and the thermal-electric coupling coefficient between the device pairs extracted from the fusion feature matrix. Based on the current state of the system and the dynamic constraint boundary, the optimal control sequence of model predictive control is solved, and the optimal control sequence is converted into programmable logic controller (PLC) instructions, which are then used to execute the adjustment action.

[0031] Based on the above technical solutions, the PLC-based intelligent power distribution regulation method for distribution cabinets provided in this application unifies environmental perception and electrical control within a graph structure and model predictive control framework, thereby achieving joint perception and collaborative decision-making of environmental risks and electrical coupling, significantly improving the control accuracy, response speed, and operational safety of the distribution cabinet under complex operating conditions.

[0032] like Figure 2 As shown in the embodiment of this application, a PLC-based intelligent power distribution regulation method for power distribution cabinets includes: S1. Collect environmental sensor data and electrical operation data in the distribution cabinet, construct an environmental semantic graph containing spatial grid nodes and environmental features, and a power distribution knowledge graph containing equipment entities and electrical relationships.

[0033] The environmental semantic graph uses the internal space of the distribution cabinet as a baseline, dividing the cabinet space into several spatial grid nodes, and assigning each grid node a graph structure data composed of environmental characteristics such as temperature, humidity, and dust concentration. The environmental semantic graph organizes discrete environmental data collected by distributed environmental sensors into a structured representation with spatial continuity and correlation, thus providing a data foundation for subsequent environmental risk perception and spatial reasoning. The power distribution knowledge graph uses circuit breakers, contactors, busbars, cables, and other equipment entities within the distribution cabinet as nodes, and electrical connections, control relationships, and topological dependencies between these devices as edges. The power distribution knowledge graph organizes the dispersed equipment and their electrical relationships within the distribution cabinet into a machine-understandable structured knowledge system, providing semantic support for subsequent electrical control reasoning and constraint calculation.

[0034] In some implementations, environmental sensor data within the distribution cabinet can be collected in the following ways: Various types of environmental sensors, such as temperature sensors, humidity sensors, and particulate matter concentration sensors, are deployed at predetermined spatial intervals inside the distribution cabinet. Each sensor uploads its collected data to a data acquisition terminal via a fieldbus or wireless communication network. The data acquisition terminal performs analog-to-digital conversion, filtering, noise reduction, and timestamp alignment on the raw sensor data to form a structured environmental sensor dataset. Electrical operation data can be collected in the following ways: Electrical measuring devices such as voltage transformers, current transformers, and power transmitters installed at various equipment nodes collect electrical operation parameters such as voltage, current, active power, reactive power, and power factor of each device. This electrical operation data is then transmitted to a data processing unit via a data acquisition and monitoring control system or an industrial Ethernet network.

[0035] Constructing an environmental semantic graph can be achieved as follows: Using the internal space of the distribution cabinet as a reference, the cabinet space is divided into uniform or non-uniform spatial grids according to a predetermined step size, with each spatial grid serving as a graph node. For each grid node, based on its spatial location and data collected from environmental sensors adjacent to it, a spatial interpolation algorithm is used to calculate the node's temperature, humidity, and dust concentration values, forming the node's environmental feature vector. Finally, the spatial adjacency relationships between grid nodes are used as the edges of the graph to construct the environmental semantic graph. In other implementations, the spatial grid can be adaptively partitioned using an octree structure. This involves using a finer grid resolution in areas with drastic changes in environmental features and a coarser grid resolution in areas with gradual changes, thereby reducing the graph's size while maintaining accuracy.

[0036] The construction of a power distribution knowledge graph can be achieved through the following methods: extracting equipment entity information and electrical connection relationships between equipment from electrical design drawings of distribution cabinets, equipment ledger databases, and field configuration information; using each equipment entity as a graph node, with equipment type, rated parameters, and current operating status as node attributes; using electrical connection relationships, control subordination relationships, and protection coordination relationships between equipment as edges of the graph, and assigning type labels and weight attributes to the edges; and using a resource description framework or attribute graph model to structurally store the above nodes and edges, forming the power distribution knowledge graph. Equipment entities may include, but are not limited to, incoming circuit breakers, feeder circuit breakers, contactors, thermal relays, busbars, power cables, transformers, reactive power compensation devices, and measuring instruments. Electrical connection relationships may include series relationships, parallel relationships, power supply and distribution relationships, control relationships, and protection relationships.

[0037] S2. Align the spatial coordinates of the environmental semantic graph and the power distribution knowledge graph, and use a graph neural network to encode the spatially aligned environmental semantic graph and the power distribution knowledge graph respectively. Then, use a cross-attention mechanism to fuse features and generate a fused feature matrix.

[0038] The environmental semantic graph describes the environmental state at various locations within the distribution cabinet space, while the power distribution knowledge graph describes the various equipment entities and their electrical relationships. The data for both reside in different semantic and coordinate spaces. To establish cross-domain associations between environmental states, equipment entities, and electrical relationships, both need to be unified to the same spatial coordinate reference system, enabling each equipment node to be associated with the environmental characteristics of its location and neighboring areas. The generated fused feature matrix simultaneously includes environmental spatial distribution information, equipment entity information, electrical topology relationship information, and cross-domain association information between the environment and equipment, providing a rich and structured feature representation for subsequent environmental risk assessment, thermoelectric coupling analysis, and control decisions.

[0039] In some implementations, spatial coordinate alignment can be achieved by: pre-labeling the three-dimensional spatial coordinates of each device node in the power distribution knowledge graph within the power distribution cabinet; using these three-dimensional spatial coordinates as anchor points, locating the corresponding grid node or set of grid nodes in the environmental semantic graph; and spatially associating the environmental feature vector of the grid node or set of grid nodes with the attribute features of the corresponding device node to form a unified graph structure after spatial alignment.

[0040] In other implementations, if the spatial coordinates of a device node fall inside a certain grid node, the environmental features of that grid node are directly associated with the device node; if the spatial coordinates of a device node are located at the intersection of multiple grid nodes, the environmental feature value at the device node is calculated by weighted averaging or bilinear interpolation of the environmental features of adjacent grid nodes.

[0041] It should be noted that the combination of the graph neural network encoder and the cross-attention mechanism constitutes an end-to-end trainable feature fusion network, which can be trained through supervised or self-supervised learning. The choice of encoder and parameter settings will affect the quality of the fused features. In practical applications, a suitable graph neural network architecture can be selected based on the size of the distribution cabinet and the data types.

[0042] For example, in one specific implementation, a two-layer graph convolutional network is used as the encoder for the environmental semantic graph, with each graph convolutional layer having a hidden layer dimension of 128. A two-layer graph attention network is used as the encoder for the power distribution knowledge graph, with four attention heads, each with a dimension of 32. The cross-attention mechanism uses eight cross-attention heads to map the environmental semantic encoding vector and the power distribution knowledge encoding vector to a 256-dimensional fusion feature space, ultimately generating a fusion feature matrix containing 57 node feature vectors, each with a dimension of 256.

[0043] S3. Adjust the dynamic constraint boundary of the model predictive control based on the environmental risk coefficient of the device node and the thermal-electric coupling coefficient between the device pairs extracted from the fusion feature matrix.

[0044] This step dynamically narrows or widens these constraint boundaries based on the current environmental risk status and the degree of thermoelectric coupling between devices, allowing model predictive control to obtain a larger feasible solution space or a more stringent safety boundary while ensuring safe equipment operation. When the environmental risk coefficient of a device node is high, it means that the device faces adverse environmental factors such as overheating, moisture, or dust accumulation. The constraint boundaries of its power, current, and other control variables should be appropriately narrowed to reduce operational risks. When the thermoelectric coupling coefficient between two devices is high, it means that their heating significantly affects each other. When controlling one device, the thermal impact on the other device should be considered, thus introducing coupling constraint conditions into the dynamic constraint boundaries. Through the above mechanism, the dynamic constraint boundaries can adaptively adjust according to the real-time environmental state and the device coupling state, achieving a deep integration of environmental perception and electrical control.

[0045] In some implementations, extracting the environmental risk coefficient of a device node from the fused feature matrix can be achieved as follows: For each device node, extract the corresponding fused feature vector from the fused feature matrix; input the environmentally relevant dimensional components of this fused feature vector into an environmental risk assessment model, which uses a pre-trained classifier or regressor to output the environmental risk coefficient of the device node. The environmental risk coefficient can be a continuous value between 0 and 1, with a higher value indicating a higher environmental risk. The environmental risk assessment model can be trained using historical operating data, and training labels can be generated based on equipment failure records or equipment operating temperature exceeding limits records.

[0046] In other implementations, the environmental risk coefficient can be directly calculated by weighting the ambient temperature, humidity, and dust concentration at the location of the device node. The calculation formula can be: Environmental risk coefficient = α × normalized temperature value + β × normalized humidity value + γ × normalized dust concentration value; where α, β, and γ are preset weighting coefficients, and each normalized value is obtained by mapping the measured value to a predetermined normal range.

[0047] Extracting the thermal-electric coupling coefficient between device pairs from the fusion feature matrix can be achieved in the following ways: For any two device nodes, extract their respective fusion feature vectors from the fusion feature matrix; calculate the similarity or correlation between the two fusion feature vectors as the initial value of the thermal-electric coupling coefficient of the device pair; or extract the edge features between the two device nodes from the fusion feature matrix and output the thermal-electric coupling coefficient of the device pair through a pre-trained thermal-electric coupling prediction model.

[0048] In other implementations, the thermal-electric coupling coefficient can be calculated based on the spatial distance between the two devices, their electrical connection relationship, and their respective current load currents, using the following formula: Thermo-electric coupling coefficient = spatial proximity factor × electrical connection factor × load current coupling factor; where the spatial proximity factor is inversely proportional to the distance between the devices, the electrical connection factor depends on whether there is a direct electrical connection between the two devices, and the load current coupling factor is directly proportional to the product of the load currents of the two devices.

[0049] S4. Based on the current state of the system and the dynamic constraint boundary, solve for the optimal control sequence of the model predictive control, and convert the optimal control sequence into programmable logic controller (PLC) instructions, and use the PLC to execute the adjustment action.

[0050] Model predictive control is a model-based control strategy. Its core idea is to solve a constrained optimization problem in a finite time domain based on the current state of the system and the system prediction model at each sampling time, obtain the optimal control sequence for several future times, and then apply only the first control variable in the control sequence to the system. The above process is repeated at the next sampling time to achieve rolling optimization.

[0051] This step, based on the current system state and the dynamic constraint boundary determined in step S3, solves for the optimal control sequence that optimizes the predetermined objective function while satisfying the dynamic constraint boundary, thus achieving closed-loop coordination between environmental perception and electrical control. Converting the optimal control sequence into programmable logic controller (PLC) instructions means converting the numerical control quantities obtained from the optimization solution into control instructions in the form of ladder logic, structured text, or instruction lists that can be executed by the PLC. These instructions then drive the actuators to complete adjustment actions through the PLC's digital or analog output modules, such as circuit breaker opening and closing operations, contactor on / off control, voltage regulator tap adjustment, and reactive power compensation device switching.

[0052] The conversion of the optimal control sequence into programmable logic controller (PLC) instructions can be achieved as follows: Extract the first control variable from the optimal control sequence as the execution control variable for the current cycle; map the execution control variable to the analog or digital output value of the PLC based on its data type and value range; generate the corresponding control instruction frame according to the PLC's communication protocol and instruction format; and send the control instructions to the PLC via industrial Ethernet, fieldbus, or serial communication interface. In other implementations, the PLC has a pre-built control instruction mapping table that maps control variables with different value ranges to different execution actions. After obtaining the optimal control variable, the corresponding PLC instruction is quickly determined by looking up the table.

[0053] The programmable logic controller (PLC) can be used to perform adjustment actions in the following ways: After receiving a control command, the PLC parses the command content through its central processing unit to determine the target actuator and target action parameters; the PLC's digital output module drives relays or contactors to complete switching adjustment actions, such as opening or closing circuit breakers; the PLC's analog output module drives speed control devices or voltage regulation devices to complete analog adjustment actions, such as adjusting the output frequency of a frequency converter or the output voltage of a voltage regulator; after executing the adjustment action, the PLC collects feedback signals from the actuator through its input module to verify whether the adjustment action was successfully executed.

[0054] Based on the above technical solutions, this application provides a PLC-based intelligent power distribution regulation method for distribution cabinets. This method achieves spatial continuity representation of the environmental state within the distribution cabinet through an environmental semantic graph, structured expression of electrical relationships between equipment through a power distribution knowledge graph, establishes cross-domain semantic associations between the environment and electrical systems through spatial coordinate alignment and graph neural network encoding, achieves deep fusion of environmental and electrical features through a cross-attention mechanism, dynamically and adaptively adjusts the model predictive control constraint boundaries through environmental risk coefficients and thermo-electric coupling coefficients, and finally realizes intelligent power distribution regulation driven by environmental perception through a programmable logic controller. The entire solution effectively solves the technical problem of the separation between environmental perception and electrical control in traditional distribution cabinet regulation methods, significantly improving the operational safety, regulation accuracy, and adaptive capability of the distribution cabinet under complex environmental conditions.

[0055] In one possible implementation of this application embodiment, the above-mentioned S1 can be specifically implemented by the following S101 and S102, which are described in detail below: S101. Collect environmental sensor data inside the power distribution cabinet, discretize the internal space of the power distribution cabinet into three-dimensional grid nodes according to the preset grid spacing, and construct an environmental semantic map.

[0056] Among them, the environmental semantic graph is a weighted undirected graph structure constructed by discretizing the internal space of the power distribution cabinet into three-dimensional grid nodes according to a preset grid spacing, using the environmental characteristics of each grid node as node attributes and the lines connecting adjacent nodes in space as edges. Its function is to quantitatively characterize the spatial distribution characteristics and regional semantic information of the temperature field, humidity field and airflow field inside the power distribution cabinet.

[0057] In some implementations, environmental sensing data is collected through a distributed sensor network deployed inside the power distribution cabinet. This sensor network includes at least temperature sensors, humidity sensors, and airflow velocity sensors, with each sensor arranged at key locations inside the power distribution cabinet according to a preset spatial sampling interval.

[0058] In some implementations, the process of constructing an environment semantic graph includes: First, the internal space of the distribution cabinet is discretized into three-dimensional grid nodes according to a preset grid spacing. The feature vector of each grid node is composed of temperature, humidity, airflow velocity, temperature gradient, humidity gradient, semantic labels, and status flags. For grid nodes without directly deployed sensors, the temperature, humidity, and airflow velocity values ​​can be calculated from the measurements of neighboring sensors using inverse distance weighted interpolation. The temperature and humidity gradient values ​​are calculated using the finite difference method. Specifically, for grid node i, its temperature gradient along the x-axis... The calculation method is as follows ,in and These represent the temperature values ​​of the nodes adjacent to this node along the positive and negative x-axis directions, respectively. This represents the grid spacing along the x-axis; similarly, it is calculated... and .

[0059] In the feature vectors of grid nodes, semantic labels are obtained by clustering algorithms based on the temperature, humidity, and temperature gradient values ​​of each grid node. These labels characterize the region type to which the grid node belongs. Region types include hotspot areas, coldspot areas, condensation risk areas, normal areas, ventilation bottleneck areas, and transition areas. Specifically, density-based clustering algorithms are typically used to cluster grid nodes. Nodes with temperature gradients exceeding a preset threshold and temperatures more than two standard deviations above the average temperature are clustered as hotspot areas. Nodes with humidity exceeding a preset threshold and temperatures below a preset threshold are clustered as condensation risk areas. Nodes with airflow velocities below a preset threshold are clustered as ventilation bottleneck areas. The remaining nodes are clustered into coldspot areas, normal areas, and transition areas based on the combined characteristics of temperature, humidity, and airflow velocity.

[0060] The hot spot zone refers to a localized area within the distribution cabinet where the internal temperature is significantly higher than the surrounding area, exhibiting a large temperature gradient. This is typically caused by high current flowing through switch contacts or busbar connections where contact resistance is high. Equipment in this zone faces a significantly increased risk of overheating and insulation aging, requiring careful avoidance of large-scale adjustments during regulation. The cold spot zone refers to a localized area within the distribution cabinet where the internal temperature is significantly lower than the average surrounding temperature. This is usually located near air inlets or away from heat sources. Equipment in this zone experiences lower thermal stress and can handle more regulation tasks. The condensation risk zone refers to a localized area within the distribution cabinet where humidity is high and temperature is low, making it prone to condensation. This is typically located at the bottom of the cabinet, near air inlets / outlets, or in areas of sudden temperature drops. Equipment in this zone faces a high risk of insulation degradation and electrical breakdown. Regulation strategies should prioritize controlling the operating power of equipment in this zone to prevent insufficient temperature rise and exacerbated condensation. The normal zone refers to an ordinary area within the distribution cabinet where temperature, humidity, and airflow velocity are all within the allowable range for normal equipment operation. Equipment in this zone can participate in power distribution regulation under rated operating conditions. A ventilation bottleneck zone refers to a localized area within a distribution cabinet where the airflow velocity is significantly lower than the ventilation design requirements. This area is typically located in a dead zone in the middle of the cabinet or where heat dissipation channels are obstructed. Heat is difficult to dissipate in this area, leading to localized heat accumulation. In regulation decisions, the load rate of equipment in this area should be limited to prevent continuous heat buildup. A transition zone refers to the boundary area where the environmental parameters inside the distribution cabinet fall between the threshold values ​​of the aforementioned zones. Its environmental state is unstable and has the potential to evolve into a hotspot or condensation risk zone. In regulation decisions, the changing status trends of equipment in this area should be appropriately monitored, and a certain adjustment margin should be reserved.

[0061] The status flag consists of binary numbers 0 or 1, used to indicate whether a node is occupied by a device. Specifically, if the spatial location corresponding to the grid node is occupied by an electrical device entity in the distribution cabinet, the status flag is set to 1; otherwise, it is set to 0.

[0062] Next, a weighted undirected graph is constructed with each grid node as a vertex and the lines connecting adjacent nodes in space as edges, and the edge weights between nodes are calculated.

[0063] Specifically, the set of all N discretized grid nodes is taken as the vertex set. The set of edges is the line connecting adjacent nodes in space. Spatially adjacent is defined as two mesh nodes whose coordinate difference in three-dimensional space does not exceed one mesh spacing, i.e., satisfying... ,in For Chebyshev distance, This represents the grid spacing.

[0064] node With nodes Edge weights between This is obtained by weighted summation of spatial proximity and environmental feature similarity: ;in, Spatial proximity Calculated using a Gaussian kernel function based on the Euclidean distance between the spatial coordinates of two nodes: In the formula, This is the spatial bandwidth parameter, with a value of 1.2 times the grid spacing, used to control the decay rate of spatial proximity.

[0065] Environmental feature similarity Calculated using a Gaussian kernel function based on the Euclidean distance between the Euclidean distances of the environmental feature vectors of two nodes: In the formula, The feature bandwidth parameter has a value of 0.3 and is used to control the decay rate of feature similarity.

[0066] and These are the weighting coefficients, with default values ​​of 0.6 and 0.4, respectively, satisfying... This edge weight calculation method gives larger edge weights to nodes that are spatially adjacent and have similar environmental characteristics, thus enabling the effective propagation and aggregation of local environmental information during subsequent graph neural network encoding.

[0067] Next, the feature vectors of each node are used as node attributes, and the edge weight matrix is ​​used as the adjacency matrix to construct an environmental semantic graph. Specifically, the feature vectors of each constructed grid node are... As a node attribute, the edge weight Construct an adjacency matrix If the node With nodes Adjacent, then ,otherwise =0. The feature vectors of all nodes are stacked in node index order to form the node feature matrix. .

[0068] At this point, the environmental semantic graph is complete. Construction complete.

[0069] Finally, the graph Laplacian matrix of the environmental semantic graph is calculated for subsequent topological difference calculation.

[0070] Thulaplatz matrix The calculation method is as follows: ;in, For a degree matrix, its diagonal elements Off-diagonal elements are zero. The graph Laplacian matrix characterizes the spatial coupling relationships and feature propagation paths between nodes in the environmental semantic graph, and will be used in subsequent steps to calculate the topological difference matrix between the environmental semantic graph and the power distribution knowledge graph.

[0071] S102. Collect electrical operation data in the distribution cabinet, extract entities and relationships from the electrical topology of the distribution cabinet, equipment nameplate parameters and historical operation records, and construct a power distribution knowledge graph.

[0072] The power distribution knowledge graph is a heterogeneous graph structure built using electrical equipment within the distribution cabinet as entities and electrical connections and control relationships between devices as edges. Its function is to structurally represent the electrical topology and equipment parameter information of the power distribution system. The environmental semantic graph and the power distribution knowledge graph model the operating status of the distribution cabinet from the environmental spatial dimension and the electrical topology dimension, respectively, providing a structured data foundation for subsequent cross-modal feature fusion.

[0073] In some implementations, electrical operation data is acquired through voltage transformers, current transformers, and power transmitters connected to each electrical circuit of the distribution cabinet. The acquired parameters include voltage amplitude, current amplitude, active power, reactive power, and power factor of each equipment node.

[0074] Then, entities and entity attributes are extracted from the electrical topology design documents, equipment nameplate parameters, and historical operation records of the distribution cabinet. Entity types include power supply nodes, bus nodes, circuit breaker nodes, switch nodes, load nodes, compensation nodes, and transformer nodes. Entity attributes include rated voltage, rated current, rated capacity, impedance value, health status, spatial coordinates, and current operating status. Among them, the health status is obtained by comprehensively evaluating the equipment's historical operating time, number of failures, and maintenance records, with a value ranging from 0 to 1. The closer the value is to 1, the better the equipment's health status. Spatial coordinates are the three-dimensional spatial coordinates of the equipment within the distribution cabinet.

[0075] Relationships between entities are extracted, including electrical connection relationships, control relationships, membership relationships, thermal impact relationships, dependency relationships, and spatial proximity relationships. Electrical connection relationships represent direct electrical connection between two equipment entities via cables or busbars; control relationships represent one device having a control function over another, such as a circuit breaker controlling the switching of its downstream loads; membership relationships represent one device being subordinate to another, such as a load node being connected to a busbar node; thermal impact relationships represent the heat generated by one device during operation affecting the operating state of another device; dependency relationships represent the normal operation of one device depending on the normal operation of another device; and spatial proximity relationships represent two devices being spatially adjacent within the distribution cabinet.

[0076] Next, a graph embedding model based on complex vector space is used to map entities and relations to complex vector space, obtaining the embedding vectors of each entity and each relation.

[0077] Specifically, entities and relations are extracted from the power distribution knowledge graph to obtain a set of triples containing head entities, relations, and tail entities. . Head entity vector relation vectors Tail entity vector All are mapped to complex vector space , where d is the embedding dimension.

[0078] The scoring function of the triple is calculated using the negative norm between the Hadamard product of the head entity vector and the relation vector and the tail entity vector. The formula is as follows: , This is the Hadamard product operator, which performs element-wise multiplication. Let be the norm of the modulus of a complex vector. Scoring function. The closer the value is to 0, the higher the reliability of the triple.

[0079] The graph embedding model is trained by maximizing the scoring function of all positive triples and minimizing the scoring function of negative triples, with the training objective being to minimize the loss function. ,in For the set of positive triples, For the set of negative triples, The Sigmoid activation function is used. This training process enables the graph embedding model to learn the optimal distribution of entities and relations in the complex vector space—the score function value of positive triples is pushed closer to 0, and the score function value of negative triples is pushed further into the negative region, thus ensuring that the embedding vectors can accurately reflect the true semantic relationships between entities.

[0080] Finally, the adjacency matrix of the power distribution knowledge graph can be constructed based on the spatial proximity and thermal influence relationships between entities, and the graph Laplacian matrix of the power distribution knowledge graph can be calculated. ,in For degree matrix, Using an adjacency matrix, complete the construction of the power distribution knowledge graph.

[0081] It should be noted that the graph embedding model transforms the symbolic entities and relations in the power distribution knowledge graph into numerical representations in a continuous vector space, enabling the structured information in the knowledge graph to be processed by subsequent graph neural networks. Specifically, this embedding process serves three purposes: First, it converts discrete symbolic data into continuous vector data, allowing it to serve as input features for graph attention networks in the computation and optimization of the neural network; second, it models the semantic relationships between entities through rotation operations on relation vectors in complex vector space, so that the head entity approximates the vector position of the tail entity after relation rotation, thereby preserving the topological connection semantics between entities in the vector space; third, it provides a unified numerical feature foundation for subsequent cross-modal cross-attention computation between the environmental semantic graph and the power distribution knowledge graph, ensuring that information from the environmental space dimension and the electrical topology dimension can interact and merge in the same vector space.

[0082] Based on the above technical solutions, this application constructs an environmental semantic graph with spatial topology and regional semantic information in S101, enabling environmental data to be upgraded from discrete sensor values ​​to graph-structured data. In S102, a power distribution knowledge graph containing multiple types of electrical equipment entities and multidimensional relationships is constructed, and a graph embedding model based on complex vector space is used to achieve unified representation and high-precision embedding of structured and unstructured information of the power distribution system. The environmental semantic graph and the power distribution knowledge graph respectively provide a complete structured model of the operating status of the distribution cabinet from the environmental spatial dimension and the electrical topology dimension, providing a data foundation for spatial coordinate alignment, graph neural network encoding, and cross-attention feature fusion in subsequent S2 steps.

[0083] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 3 As shown, the above S2 can be implemented through the following S201, S202 and S203, which are explained in detail below: S201. Align the environmental semantic graph with the power distribution knowledge graph in terms of spatial coordinates to establish a spatial mapping relationship between the two heterogeneous graphs.

[0084] In some implementations, the specific process of spatial coordinate alignment is as follows: Obtain the spatial coordinates of each device entity in the power distribution knowledge graph. These spatial coordinates are pre-stored in the attributes of the device entity and are the three-dimensional spatial coordinate values ​​of the device within the distribution cabinet, denoted as... .

[0085] Using the spatial coordinates of each device entity as the query vector, search for the nearest Euclidean distance grid node among all grid nodes in the environment semantic graph, and establish the nearest neighbor mapping relationship from the device entity to the nearest grid node.

[0086] After establishing the nearest neighbor mapping relationship from the device entity to the nearest grid node, the Euclidean distance between the spatial coordinates of each device entity and the spatial coordinates of the nearest grid node is calculated as the mapping error: ; If mapping error If the value exceeds the preset threshold (e.g., twice the grid spacing), it indicates that the device entity lacks sufficient grid node support near its corresponding spatial location in the environmental semantic map. This may be due to incorrect device coordinate labeling or insufficient grid resolution in the environmental semantic map. In this case, a manual verification warning is triggered to remind maintenance personnel to check the device location labeling or adjust the grid resolution.

[0087] If the mapping error is less than or equal to a preset threshold, the device entity is bound to the nearest neighbor grid node, and the binding mapping relationship from the device entity to the environment semantic graph grid node is obtained. This binding mapping relationship records the environment semantic graph grid node index corresponding to each device entity, which is used for cross-graph information transmission in the subsequent feature aggregation process.

[0088] It should be noted that when establishing mapping relationships, there may be multiple device entities bound to the same grid node. This is because when devices are densely arranged within a distribution cabinet, the spatial coordinates of multiple devices may fall near the same grid node. In this case, when calculating the environmental aggregation features of that grid node, the electrical embedding vectors of all device entities bound to that node are averaged and aggregated.

[0089] S202. Calculate the topological difference matrix between the environmental semantic graph and the power distribution knowledge graph to quantitatively evaluate the spatial matching degree between the environmental field distribution and the electrical topology.

[0090] The topology difference calculation is used to quantitatively assess the spatial matching degree between the environmental field distribution and the electrical topology. When there is a significant difference between the spatial topology of the environmental field and the electrical topology, it indicates a structural mismatch between the physical field distribution inside the distribution cabinet and the electrical connection layout. For example, the heat dissipation bottleneck area may happen to be a cluster of heavy-duty equipment, or the heat dissipation capacity may be insufficient in densely populated electrical connection areas corresponding to areas with large temperature gradients. This topology difference information serves as a trigger signal for the dynamic adjustment of subsequent model predictive control constraints.

[0091] In some implementations, based on the binding mapping relationship established in S201, the graph Laplacian matrix of the power distribution knowledge graph is spatially mapped and resampled to the spatial coordinate system of the environmental semantic graph to obtain the resampled power distribution knowledge graph Laplacian matrix.

[0092] Specifically, the graph Laplace matrix of the power distribution knowledge graph Where M is the total number of device entities, and its elements Characterizes the electrical topology coupling strength between device i and device j. Based on the binding mapping relationship... (This mapping relationship is established by the nearest neighbor search in step S201, that is, mapping each device entity to the grid node with the nearest Euclidean distance in the environment semantic graph.) Mapping each device entity to its bound grid node yields the resampled Laplace matrix of the power distribution knowledge graph. , where N is the total number of grid nodes in the environment semantic graph. For any two grid nodes and The corresponding resampled Laplace matrix elements are: That is, the Laplace matrix element values ​​between device entities are aggregated to the corresponding mesh nodes.

[0093] Then, the graph Laplacian matrix of the environment semantic graph is calculated. Laplace matrix of the resampled power distribution knowledge graph The difference is used to obtain the topological difference matrix: Among them, the topological difference matrix Positive values ​​indicate that the coupling strength between corresponding grid node pairs in the environmental semantic graph is greater than the coupling strength between corresponding device entity pairs in the power distribution knowledge graph, indicating that the interaction of the environmental physical field in this region is stronger than the electrical connection relationship, and there is an environmental structural association that is not reflected in the electrical topology; negative values ​​indicate that the coupling strength between corresponding device entity pairs in the power distribution knowledge graph is greater than the coupling strength between corresponding grid node pairs in the environmental semantic graph, indicating that the electrical connection relationship in this region is stronger than the interaction of the environmental physical field.

[0094] Next, in order to quantitatively assess the degree of mismatch between the environmental field and the electrical topology as a whole, the F-norm of the topology difference matrix can also be calculated: If the F-norm is greater than the preset topology difference threshold, a trigger signal is generated to trigger the dynamic adjustment of the model predictive control constraints. This trigger signal will be transmitted to the adjustment step of the model predictive control dynamic constraint boundary in S3, which is used to achieve emergency tightening of the constraint boundary when the environmental field and electrical topology are severely mismatched.

[0095] It should be noted that the topology difference matrix reflects the degree of matching between the environmental field and the electrical topology at the spatial structure level, rather than a simple numerical difference. When the topology difference consistently exceeds the threshold, it is recommended that maintenance personnel check whether the heat dissipation design of the distribution cabinet is reasonable or whether the equipment layout needs to be optimized.

[0096] S203. A graph neural network is used to encode the spatially aligned environmental semantic graph and the power distribution knowledge graph respectively, and the features are fused through a cross-attention mechanism to generate a fused feature matrix.

[0097] The cross-attention mechanism is used to achieve deep feature interaction and fusion between two types of heterogeneous graph structures, enabling environmental information to flow to electrical nodes and electrical information to flow to environmental nodes, thus achieving bidirectional information fusion. The generated fusion feature matrix simultaneously contains the electrical attribute information of the device itself, the topological connection information between devices, and the physical field distribution information of the local environment in which the device is located, providing a feature basis for subsequent extraction of environmental risk coefficients and thermal-electric coupling coefficients.

[0098] In some implementations, the specific process of graph neural network encoding is as follows: A graph convolutional network is used to encode the context semantic graph, using the node feature matrix of the context semantic graph. and adjacency matrix As input, environmental spatial features are extracted through multi-layer graph convolution operations, and the output is the environmental embedding vector of each grid node in the environmental semantic map. After multi-layer graph convolution operations, the output is the environmental embedding vector of each grid node in the environmental semantic map. ,in =256.

[0099] A graph attention network is employed to encode the power distribution knowledge graph. The complex vector space embedding and adjacency matrix of each entity in the power distribution knowledge graph are used as input. A multi-head graph attention layer calculates the attention weights of each entity's neighbor nodes and aggregates neighbor features, outputting the electrical embedding vectors of each equipment entity in the power distribution knowledge graph. After multi-layer graph attention layer encoding, the electrical embedding vectors of each equipment entity in the power distribution knowledge graph are output. ,in =256.

[0100] Then, the environment embedding vectors of the nearest neighbor mesh nodes corresponding to each device entity are aggregated according to the spatial coordinate mapping relationship to obtain the environment aggregation features corresponding to each device entity. Specifically, for device entities... Its corresponding nearest neighbor grid node is Then its environmental aggregation characteristics are: ;in For grid nodes The environment embedding vector. If multiple device entities are bound to the same mesh node, they share the same environment aggregation feature.

[0101] Next, cross-attention fusion is performed. First, the electrical embedding vectors of each device entity are used... As a query matrix Based on the environmental aggregation characteristics corresponding to each device entity as a bond matrix Sum matrix Calculate the first cross-attention feature from the power distribution knowledge graph to the environmental semantic graph: ;in =256 is the dimension of the key vector. This is a row-wise normalized exponential function. This step enables the electrical embedding vector to extract relevant environmental field information from the aggregated environmental features, thus achieving the focus of electrical features on environmental information.

[0102] Then, aggregate the environmental features corresponding to each device entity. As a query matrix Electrical embedding vectors of each device entity as a bond matrix Sum matrix Calculate the second cross-attention feature from the environmental semantic graph to the power distribution knowledge graph: .

[0103] The cross-attention mechanism enables environmental aggregate features to obtain information related to the electrical operating state from the electrical embedding vector, thereby enabling environmental features to focus on electrical information.

[0104] Finally, the electrical embedding vector of each device entity is added to the first cross-attention feature, and the environmental aggregation feature of each device entity is added to the second cross-attention feature. The two added feature parts are then concatenated to obtain the fused feature vector of each device entity. ;in, Let be the electrical embedding vector of the k-th device entity. This is the feature vector corresponding to the k-th device entity in the first cross-attention feature set. This is the feature vector corresponding to the k-th device entity in the second cross-attention feature. This is a vector concatenation operation. The fused feature vectors of all device entities constitute the fused feature matrix. .

[0105] It is important to note that conventional knowledge graphs are typically used as static knowledge bases or semantic retrieval tools to query equipment parameters, topological connections, or perform simple fault rule reasoning. Their function is often limited to knowledge query and relationship display, lacking a mechanism for deep integration with real-time data and direct participation in control decisions. In contrast, the power distribution knowledge graph in this application not only stores equipment entities and their electrical topological relationships but also achieves multiple innovative applications. First, this knowledge graph is not statically fixed but continuously and dynamically updated with the real-time acquisition of electrical operation data. The current operating status in its node attributes is synchronously refreshed in each control cycle, ensuring that the graph is always precisely synchronized with the actual state of the physical system. Second, this knowledge graph performs cross-modal deep feature fusion with the environmental semantic graph through spatial coordinate alignment and cross-attention mechanisms. Each equipment entity obtains the physical field context information of its local environment through spatial coordinate mapping, and then achieves deep interaction between electrical embedding and environmental aggregation features through bidirectional cross-attention. This makes the knowledge graph not only a carrier of electrical relationships but also a comprehensive decision-making foundation that integrates environmental physical field semantics. Furthermore, the complex vector space embedding of this knowledge graph provides a unified numerical feature foundation for graph attention network encoding and cross-modal cross-attention computation, enabling symbolic knowledge to be directly processed by graph neural networks and interact with environmental features in the same vector space, rather than being limited to querying and display as in conventional applications. Through this approach, the power distribution knowledge graph in this application deeply embeds a complete control loop of perception-fusion-decision-execution, fundamentally differentiating it from the conventional application mode of knowledge graphs, which are merely used as auxiliary query tools.

[0106] Based on the above technical solution, this application establishes a spatial mapping relationship between the two heterogeneous graphs by aligning the environmental semantic graph and the power distribution knowledge graph in S201. In S202, it calculates the topology difference matrix to quantitatively evaluate the spatial matching degree between the environmental field distribution and the electrical topology, providing prior warning information for the dynamic adjustment of subsequent control constraints. In S203, it encodes the two types of graphs using a graph convolutional network and an environmental graph attention network respectively, and achieves deep feature interaction and fusion through a bidirectional cross-attention mechanism, generating a fused feature matrix that simultaneously contains electrical attribute information, topological connection information, and local environmental physical field distribution information. These steps together complete the transformation from raw heterogeneous data to unified fused features, providing high-quality feature input for the extraction of environmental risk coefficients and thermal-electric coupling coefficients in S3, as well as for adjusting the dynamic constraint boundaries of model predictive control.

[0107] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 4 As shown, the above S3 can be implemented through the following S301, S302 and S303, which are explained in detail below: S301. Extract the environmental risk coefficient of each device node from the fusion feature matrix.

[0108] The environmental risk coefficient is used to quantitatively assess the relative magnitude of equipment operation risk under current environmental conditions. A value closer to 1 indicates more severe environmental conditions and higher operational risk, while a value closer to 0 indicates better environmental conditions and lower operational risk. The purpose of extracting the environmental risk coefficient is to compress the environmental physical field information contained in the fused feature matrix into a scalar risk value for each equipment node, providing a quantitative basis for subsequent adjustments to dynamic constraint boundaries.

[0109] In some implementations, the fusion feature matrix generated in S2 is used... The fused feature vector of each device entity is input into a pre-trained multilayer perceptron network, which outputs the environmental risk coefficient of the device entity. The output layer has only one neuron, and the sigmoid activation function maps the output value to the range of 0 to 1. .

[0110] The calculation process for the environmental risk coefficient of the i-th device node can be expressed as follows: in, Let be the fused feature vector of the i-th device node. , , These are the weight matrices for each layer. , , These are the bias vectors for each layer.

[0111] In some implementations, the pre-training process of a multilayer perceptron network is as follows: Historical operating data of the distribution cabinet is collected, including the fused feature vectors of each device node and their corresponding actual operating results. The actual operating results are labeled based on events such as whether the equipment malfunctioned, whether an over-temperature alarm was triggered, or whether insulation degradation occurred. Samples that have experienced the above events are labeled as high-risk. =1), samples that have not experienced the above events are labeled as low risk ( =0). The labeled samples are divided into training and validation sets. The network is trained using the backpropagation algorithm and the Adam optimizer to minimize the cross-entropy loss function. After training, the network parameters are fixed for use in extracting environmental risk coefficients during the online inference stage.

[0112] S302. Extract the thermal-electric coupling coefficient between device pairs from the fused feature matrix.

[0113] The thermo-electric coupling coefficient characterizes the degree of mutual influence between two devices under the coupling of thermal and electric fields. A larger value indicates a tighter coupling between their operating states and a higher degree of coordination required during regulation. The purpose of extracting the thermo-electric coupling coefficient is to quantify the multi-physics coupling relationship between devices, so that the regulation actions of strongly coupled device pairs can be coordinated in subsequent model predictive control optimization, avoiding the creation of new temperature gradients or thermal stresses in the same thermal zone due to differential regulation.

[0114] In some implementations, for each pair of device entities (i,j) in the power distribution knowledge graph, the cosine similarity between their fused feature vectors is first calculated. Then, the cosine similarity is divided by a preset temperature coefficient and subjected to Softmax normalization to obtain the thermal-electric coupling coefficient between the device entities. ;in, This is a preset temperature coefficient used to control the smoothness of the output distribution of the Softmax function. Temperature coefficient The larger the value, the smoother the Softmax output distribution, meaning the smaller the difference in thermal-electric coupling coefficients between device pairs; temperature coefficient The smaller the value, the steeper the output distribution, meaning that the coupling coefficient of strongly coupled device pairs is further amplified, while the coupling coefficient of weakly coupled device pairs is further suppressed. The recommended value is between 0.5 and 1.0.

[0115] It should be noted that the thermal-electric coupling coefficient is only calculated for device pairs that have a physical proximity or thermal influence relationship. For device pairs that are far apart in space and have no thermal influence relationship, the coupling coefficient can be set to zero to reduce the computational complexity.

[0116] S303. Adjust the dynamic constraint boundary of the model predictive control based on the environmental risk coefficient and the thermo-electric coupling coefficient.

[0117] In this context, adjusting the dynamic constraint boundary of model predictive control refers to changing the safe operating limits of voltage, current, and temperature at equipment nodes from fixed rated values ​​to adaptive boundary values ​​that change in real time with environmental conditions. Fixed threshold control methods cannot dynamically adjust when environmental conditions change, causing the control strategy to operate at rated values ​​even under harsh conditions such as high temperature and humidity, which can easily lead to electrical faults or equipment damage. Dynamic constraint boundaries adaptively tighten or loosen based on the real-time environmental risks of the equipment and the coupling strength with other equipment, enabling the control strategy to achieve optimal adjustment under various environmental conditions while ensuring safety.

[0118] In some implementations, the specific process of dynamically adjusting constraint boundaries is as follows: Obtain the rated operating boundaries of each device entity, including the lower limit of rated voltage. Upper limit of rated voltage Upper limit of rated current and the upper limit of rated temperature These rated limits are specified by the equipment nameplate parameters or relevant national standards and represent the safe operating limits of the equipment under standard environmental conditions.

[0119] Environmental risk coefficients extracted from S301 Multiply by the preset environmental depreciation factor This yields the environmental degradation amount for the i-th device. Environmental degradation coefficient. These are pre-defined fixed parameters, ranging from 0.1 to 0.3, with a default value of 0.2. They are used to control the overall impact of environmental risks on equipment operating boundaries. The environmental derating is then added to the rated operating boundaries of each device to generate the dynamic constraint boundaries for the current control cycle. The dynamic voltage constraint boundary for the i-th device node is: , The voltage lower limit minus the environmental derating, and the voltage upper limit plus the environmental derating, results in a narrower permissible voltage range for equipment with higher environmental risks, meaning the permissible voltage fluctuation space is compressed. It should be noted that for certain equipment types that are more sensitive to the voltage lower limit, different derating factors can be applied to the voltage lower and upper limits.

[0120] The dynamic current constraint boundary for the i-th device node is: ; Current limit multiplied by This means that equipment with higher environmental risk has a lower permissible current-carrying capacity. When the environmental risk factor approaches 1, the upper limit of current tends to be 80% of the rated current (taking...). =0.2), to achieve proactive reduction of the quota operation.

[0121] The dynamic temperature constraint boundary for the i-th device node is: upper temperature limit minus ,in This is a preset temperature margin factor, expressed in degrees Celsius, with a value between 10°C and 15°C. Equipment with higher environmental risks has a lower upper limit for permissible operating temperature, allowing for a larger safety margin.

[0122] Then, the aforementioned dynamic constraint boundary is used as the inequality constraint condition for the model predictive control optimization problem. Simultaneously, it is determined whether the F-norm of the topology difference matrix calculated in S202 is greater than a preset topology difference threshold. If the F-norm is greater than the threshold, the environmental depreciation coefficient is adjusted. Increase to the preset emergency deduction value Recalculate the dynamic constraint boundaries and execute. Emergency depreciation value. The value is between 0.4 and 0.5, which further narrows the safe operating boundary of the equipment when there is a severe mismatch between the environmental field and the electrical topology, in order to cope with the potential risk of heat accumulation.

[0123] It should be noted that the adjustment of dynamic constraint boundaries is a continuous and reversible process: when environmental conditions improve and the environmental risk coefficient decreases, the constraint boundaries will automatically recover to a more relaxed range; when the topology difference trigger signal is released, the derating coefficient will also automatically recover to its normal value. This method avoids the problem of large jumps once protection is triggered in traditional fixed threshold control, achieving a smooth balance between safety and economy.

[0124] Based on the above technical solution, this application extracts the environmental risk coefficient of the device node from the fusion feature matrix in S301, realizing the quantitative perception of the device operation risk; extracts the thermal-electric coupling coefficient between the device pairs in S302, realizing the structured representation of the multi-physics coupling relationship between the devices; and dynamically adjusts the voltage, current, and temperature constraint boundaries of the model prediction control according to the above coefficients in S303, and determines whether to trigger emergency derating based on the topology difference matrix, realizing the adaptive adjustment of control constraints with environmental conditions and topology matching degree, solving the problem of insufficient adaptability of traditional fixed threshold control methods under complex working conditions.

[0125] In one possible implementation of this application embodiment, the above-mentioned S4 can be specifically implemented by the following S401, S402 and S403, which are described in detail below: S401. Establish a state-space prediction model for the power distribution system.

[0126] The state-space prediction model employs a physical information neural network (PIN) to perform multi-step recursive predictions of the system state at future time points. The role of the state-space prediction model is to predict the trend of system state changes within a finite time domain in each control cycle, based on the current system state and control input, providing forward-looking state information for subsequent optimization decisions. The purpose of using a PSN is to integrate data-driven adaptive learning capabilities with prior constraints from physical laws, enabling the prediction results to both learn complex dynamic characteristics from operational data and satisfy fundamental physical conservation laws, thereby improving prediction accuracy and generalization ability.

[0127] In some implementations, the state variables of the state-space prediction model include the voltage magnitude, current magnitude, active power, reactive power, and equipment temperature of each device node. Then, all the state variables of the device nodes are uniformly represented as a state vector: Where M is the total number of device nodes. , , , and These represent the voltage amplitude, current amplitude, active power, reactive power, and equipment temperature of the i-th device node, respectively. The dimension of the state variables is 5M, and for a power distribution system containing 30 device nodes, the dimension of the state vector is 150.

[0128] The control variables include the opening and closing commands of each circuit breaker, the switching commands of each compensation capacitor, and the tap position commands of the on-load tap-changing transformer: ;in, This is the circuit breaker opening and closing command (0 indicates opening, 1 indicates closing). This is a capacitor switching command (negative values ​​indicate switching off, positive values ​​indicate switching on, and absolute values ​​indicate the switching ratio). This is a command indicating the tap position of an on-load tap-changing transformer.

[0129] The specific expression form of a physical information neural network is as follows: ;in, Let k be the state variable at time k. Let k be the control variable at time k. The input is the uncontrollable disturbance at time k (including random fluctuations in load and changes in external ambient temperature, etc.). For data-driven neural network terms, These are analytical terms based on the physical equations.

[0130] Among them, the data-driven neural network item Gated recurrent unit networks can be used to learn complex dynamic characteristics that are not fully modeled by physical equations, including nonlinear load characteristics, parameter drift caused by equipment aging, unmodeled dynamics in multiphysics coupling, and the nonlinear effects of environmental disturbances.

[0131] Based on the analytical terms of the physical equations It must include at least power flow equations and heat balance equations. Power flow equations describe the distribution of voltage, current, active power, and reactive power among various equipment nodes, and their node power balance equations... ; ;in, and Let be the conductance and susceptance between node i and node j, respectively. Let be the voltage phase angle difference between node i and node j. The heat balance equation describes the dynamic relationship between equipment temperature and heat generation and dissipation; its simplified form is: ;in, Let i be the heat capacity of the i-th device. Let be the power loss of the i-th device. Let be the heat dissipation coefficient of the i-th device. The ambient temperature.

[0132] The training of the physical information neural network is achieved by minimizing the weighted sum of the prediction error and the residuals of the physical equations, and its loss function is: ;in, The mean squared error between the predicted and actual values ​​is used to measure the learning accuracy of the data-driven part, and is calculated as follows: ;in, These are the model's predicted values. These are actual measured values. This represents the number of training samples. The mean square error of the residuals in the physical equations is used to measure the degree to which the prediction results satisfy the physical laws. ;in, and These represent the power imbalance quantities in the power flow equations. This represents the temperature deviation in the heat balance equation. These are the physical constraint weighting coefficients, used to balance the priority between data fitting and physical consistency.

[0133] It's important to note that the analytical terms in the physical information neural network require no training; their parameters are directly determined by the device nameplate parameters and physical constants. This ensures that even with insufficient or out-of-distribution training data, the model output still satisfies fundamental physical laws. The data-driven neural network terms are combined additively with the analytical terms, rather than cascaded or nested. This allows physical laws to be integrated into the prediction process in an interpretable way: the analytical terms provide a baseline prediction, while the data-driven terms learn the residuals not covered by the analytical terms. This architecture avoids the non-physical predictions that pure data-driven models may produce when training data is insufficient, and also avoids the prediction accuracy loss caused by the simplification assumptions of pure physical models.

[0134] It should be noted that the Physical Information Neural Network (PIN), as a state-space prediction model, integrates data-driven neural network terms with analytical terms based on physical equations during application. It uses the current system state variables, control variables, and uncontrollable disturbance inputs as common inputs to perform multi-step recursive predictions of the system state at future moments. Specifically, within each prediction step, the PIN integrates the data-driven features extracted by the gated recurrent unit network with the physical analytical features calculated from the power flow equations and thermal balance equations, directly outputting the predicted state value for the next moment. This predicted value is constrained by both the training data fitting loss and the physical equation residual loss, thus ensuring prediction accuracy while satisfying fundamental physical conservation laws. In this way, the PIN iteratively recursively predicts the system state at multiple future moments, using the prediction result of the previous moment as the input for the next moment, continuously generating a sequence of predicted system states for multiple future moments, providing forward-looking state information for optimizing the model's predictive control.

[0135] S402. Obtain the system state value of the current control cycle, perform multi-step recursion on the future state based on the state space prediction model in the prediction time domain, and use the dynamic constraint boundary as the inequality constraint of the multi-objective optimization function to solve the optimal control sequence using the sequential quadratic programming algorithm.

[0136] The multi-objective optimization function minimizes the tracking error of the output variable relative to the reference value, the control increment, and the physical penalty term, while satisfying the dynamic constraint boundaries. The output tracking error ensures adjustment accuracy, the control increment suppresses frequent actuator movements to extend equipment life, and the physical penalty term suppresses large adjustments by high-environment-risk equipment and differential adjustments between strongly coupled equipment pairs.

[0137] In some implementations, the system state value of the current control cycle k is obtained. In the prediction time domain The internal structure uses a physical information neural network to perform multi-step recursive prediction of future states, specifically including: , ;in, Let k be the predicted state value at time k+t. The control input is the time k+t predicted at time k. The perturbation input is predicted at time k+t at time k. For the prediction time domain, the value is 12 steps (corresponding to 60 seconds).

[0138] After obtaining the state sequence in the prediction time domain, a multi-objective optimization function is used to optimize the control variables in this time domain. The specific expression of the multi-objective optimization function is as follows: ;in, To predict output variables, including voltage amplitude, power factor, and device temperature at each device node, For the corresponding reference value vector (rated voltage, target power factor, safe temperature). To control the increment, To output the weight matrix of the tracking error, The weight matrix for controlling the increment, The weighting coefficient for the physical penalty term. The fusion feature matrix generated in S2, This is a physical penalty item. This represents the weighted square norm.

[0139] In this embodiment, the physical penalty item The specific expression is: ;in, Let i be the environmental risk coefficient of the i-th device node. Let be the thermal-electric coupling coefficient between device i and device j. This is the control variable applied to the i-th device node. The first term of the physical penalty term. This is used to suppress large adjustments to equipment with high environmental risks—the higher the environmental risk of the equipment, the greater the penalty imposed on the amplitude of its control variables in the optimization objective, thereby proactively avoiding large adjustments to the equipment under harsh environmental conditions. The second term of the physical penalty term. This is used to suppress differential regulation between strongly thermo-electrically coupled equipment pairs. The larger the thermo-electrical coupling coefficient of the equipment pair, the greater the penalty is imposed on the difference in the control variables between the two in the optimization objective. This forces the regulation actions of strongly coupled equipment pairs to tend to be consistent, avoiding the creation of new temperature gradients or thermal stresses in the same thermal zone.

[0140] Then, the dynamic constraint boundary generated in S303 is used as the inequality constraint condition for the multi-objective optimization function, and the magnitude constraint and increment constraint of the control variable are defined: ; ; ; ; ; in, , , Let be the voltage, current, and temperature of the i-th device node, respectively. , These are the lower and upper limits of the control variable, respectively, determined by the physical limits of each actuator. To control the upper limit of the increment, it is used to limit the action rate of the actuator to avoid shocks and oscillations.

[0141] Next, a sequential quadratic programming algorithm is used to solve the above multi-objective optimization function online. The core idea of ​​the sequential quadratic programming algorithm is to approximate the original nonlinear optimization problem as a quadratic programming subproblem at the current iteration point, solve this subproblem to obtain the search direction, and then perform a one-dimensional search along the search direction to determine the step size. After updating the iteration point, the above process is repeated until convergence. In each iteration, the environmental risk coefficient, thermal-electric coupling coefficient, and dynamic constraint boundary are updated according to the fusion feature matrix at the current time, so that the optimization problem can reflect the changes in the environmental state in real time. After the solution is completed, the control time domain is obtained. The optimal control sequence within: ;in To control the time domain, the value is set to 4 steps (corresponding to 20 seconds), and .

[0142] It should be noted that the convergence speed and solution accuracy of the sequential quadratic programming algorithm depend on the choice of the initial point and the parameter settings. To accelerate the online solution speed, the optimal solution of the previous control cycle can be used as the initial guess for the current cycle. In addition, considering the real-time requirements of power distribution system control, a maximum number of iterations (e.g., 20 times) and a convergence tolerance can be set. The iteration is terminated when the number of iterations exceeds the maximum value or the change in the objective function is less than the tolerance.

[0143] S403. The first control step in the optimal control sequence is converted into a programmable logic controller instruction, which is then executed by the programmable logic controller to perform the adjustment action. Data is then collected again for rolling optimization when the next control cycle arrives.

[0144] In some implementations, the first control step in the optimal control sequence obtained in S402 is... Extract the execution instructions for the current control cycle. These execution instructions include three types: Digital commands correspond to the opening and closing operations of the circuit breaker, and their values ​​are either 0 or 1. These digital commands are mapped to digital output signals of the programmable logic controller (PLC), and the high and low levels output by the DO module control the opening and closing coils of the circuit breaker, thus achieving the closing or opening of the circuit breaker. For example, when... When =1, the DO module outputs a high level to drive the closing coil to close the circuit breaker; when When the value is 0, the low-level output drives the trip coil to open the circuit breaker.

[0145] Analog commands correspond to the capacitor switching degree, taking continuous values ​​between [-1, 1]. These analog commands are mapped to analog output signals from a programmable logic controller (PLC). The AO module outputs 4-20 mA or 0-10 V analog signals to control the conduction angle of the thyristor switch or the PWM duty cycle of the solid-state relay, achieving continuous stepless switching of the capacitor capacity. For example, when… When the value is 0.6, the AO module outputs 12 mA (corresponding to 60% capacity input).

[0146] The stepping command corresponds to the tap position adjustment of the on-load tap-changing transformer, and its value is an integer (e.g., -5 to +5). The stepping command is mapped to the pulse output of a programmable logic controller (PLC). A certain number of pulse signals are sent through the pulse output module to drive the stepper motor to rotate, thereby switching the tap changer to the target position. For example, when... When the value is +2, the pulse output module sends a positive pulse to adjust the tap changer up two levels.

[0147] After the programmable logic controller completes the adjustment action, the system enters the next control cycle. When the next control cycle arrives, environmental sensor data and electrical operation data in the distribution cabinet are collected again, the environmental semantic graph and power distribution knowledge graph are updated, and the complete process from S2 to S4 is re-executed to form a closed-loop rolling optimization.

[0148] To ensure the safe operation of the system under abnormal conditions, a three-level degradation strategy can be set: Level 1 is the normal operation mode, where model predictive control operates with full functionality, and the environmental semantic graph and power distribution knowledge graph are updated normally; Level 2 is the model predictive control failure mode, where when the model predictive control solution times out (more than 5 seconds) or the solution diverges, it automatically switches to proportional-integral-derivative backup control, where the preset proportional-integral-derivative parameters maintain basic regulation functions, while the environmental semantic graph and power distribution knowledge graph continue to be updated but do not participate in control decisions; Level 3 is the communication interruption mode, where when communication between the programmable logic controller and the edge computing device is completely lost, the programmable logic controller executes the last valid instruction received and triggers local safety logic (such as overcurrent protection, overheat protection, etc.), while simultaneously issuing audible and visual alarm signals to notify maintenance personnel.

[0149] Based on the above technical solution, this application establishes a state-space prediction model based on a physical information neural network in S401, integrating data-driven adaptive learning capabilities with prior constraints of physical laws, so that the prediction results have the ability to learn complex dynamic characteristics and satisfy basic physical conservation laws; in S402, using dynamic constraint boundaries as constraints and a multi-objective optimization function containing physical penalty terms as objectives, a sequential quadratic programming algorithm is used to solve the optimal control sequence, so that the optimization decision actively avoids large-scale equipment adjustments in high-environmental-risk areas and suppresses differentiated actions between strongly coupled equipment pairs while satisfying electrical constraints; in S403, the first control step of the optimal control sequence is converted into a programmable logic controller instruction and executed, while a three-level degradation strategy is set to ensure system safety under abnormal operating conditions, realizing the organic combination of forward-looking optimization decision-making and industrial-grade reliable execution, as well as closed-loop rolling optimization of perception, fusion, decision-making, and execution.

[0150] In one specific application scenario of this application, the aforementioned PLC-based intelligent power distribution and regulation method for power distribution cabinets can be applied to the intelligent power distribution and micro-environment collaborative control of drone-managed hangars. A drone-managed hangar is an automated facility integrating drone storage, automatic charging, environmental control, and remote communication. It typically houses multiple electronic systems, including power supply and charging mechanisms, HVAC temperature control and dehumidification systems, wind speed and rainfall sensors, and edge computing and communication equipment. Hangars are usually deployed in remote or harsh environments such as fields, substations, mines, and ports, where external temperatures can drop to -35 degrees Celsius or reach as high as 50 degrees Celsius. The hangar interior must maintain a temperature range of 0 to 45 degrees Celsius and suitable humidity conditions required for the normal charging and discharging of drone batteries. Traditional hangar power distribution control methods typically manage environmental control, charging management, and equipment power supply as independent subsystems, resulting in fragmented environmental and electrical operation data. This makes it impossible to collaboratively optimize and adjust the power distribution strategy based on the spatial distribution of the internal temperature field, the thermal coupling relationship between equipment, and the dynamic changes in environmental risks.

[0151] In this embodiment, the specific process of the above-mentioned PLC-based intelligent power distribution regulation method for distribution cabinets is as follows: In step S1, environmental sensor data and electrical operation data are collected from the UAV hangar power distribution cabinet to construct an environmental semantic graph and a power distribution knowledge graph. A distributed sensor network, including temperature sensors, humidity sensors, and airflow velocity sensors, is deployed within the hangar power distribution cabinet and hangar cabin. The hangar interior space (including UAV parking areas, battery compartments, charging equipment areas, power distribution cabinet areas, etc.) is discretized into three-dimensional grid nodes according to a preset grid spacing. The feature vector of each node is composed of temperature value, humidity value, airflow velocity value, temperature gradient value in three axes, humidity gradient value, semantic label, and status flag, thus constructing the environmental semantic graph. The semantic labels in the environmental semantic graph are obtained by a clustering algorithm based on the temperature value, humidity value, and temperature gradient value of each grid node. Region types include hotspot areas (such as areas with dense charging equipment), coldspot areas (such as areas near hangar doors or vents), condensation risk areas (such as areas prone to condensation at the bottom of the hangar), normal areas, ventilation bottleneck areas, and transition areas. Simultaneously, entities and entity attributes are extracted from the electrical topology, equipment nameplate parameters, and historical operation records of the hangar power distribution system. Entity types include power nodes, charging module nodes, air conditioning unit nodes, fan nodes, lighting nodes, and drone battery charging nodes, while entity attributes include rated voltage, rated current, rated capacity, spatial coordinates, and current operating status. Relationships between entities are extracted, including electrical connection relationships, control relationships, thermal impact relationships, dependency relationships, and spatial proximity relationships, to construct a power distribution knowledge graph.

[0152] In step S2, the environmental semantic graph and the power distribution knowledge graph are spatially aligned. Each electrical equipment entity within the hangar is mapped to its corresponding grid node in the environmental semantic graph via spatial coordinates, establishing a binding mapping relationship between equipment entities and grid nodes for cross-graph information transfer. Based on this, the topological difference matrix between the environmental semantic graph and the power distribution knowledge graph is calculated, along with its F-norm. When the F-norm exceeds a preset topological difference threshold, a trigger signal is generated to dynamically adjust the model's predictive control constraints. Then, a graph convolutional network is used to encode the environmental semantic graph, and a graph attention network is used to encode the power distribution knowledge graph. A bidirectional cross-attention mechanism enables deep interaction and fusion of environmental and electrical features, generating a fused feature matrix. This fused feature matrix simultaneously contains the electrical attribute information of each electrical equipment, the topological connection information between equipment, and the physical field information of the local environment in which the equipment is located.

[0153] In step S3, the environmental risk coefficient of each device node and the thermal-electric coupling coefficient between device pairs are extracted from the fused feature matrix. For example, when the environmental semantic map of the concentrated area of ​​charging modules in the hangar is marked as a hotspot and the temperature gradient continues to increase, the environmental risk coefficient of the charging modules in that area will increase accordingly; when the fused feature vector similarity between the charging module and the adjacent battery compartment is high, their thermal-electric coupling coefficient is large. The dynamic constraint boundary of the model predictive control is adjusted based on the above coefficients: the voltage constraint boundary of the i-th device node is... , The current constraint boundary is Temperature constraint boundary is If the F-norm of the topological difference matrix is ​​greater than a preset threshold, then the environmental depreciation coefficient will be adjusted. Increase to the preset emergency deduction value The dynamic constraint boundaries are recalculated. For example, when a severe temperature gradient is formed between a cold spot and a hot spot in the area near the hangar door due to the intrusion of cold air from outside, and the topological difference matrix F norm exceeds the threshold, an emergency derating is triggered, which actively reduces the upper limit of the current carrying capacity and the upper limit of the temperature of the charging module and battery compartment in that area.

[0154] In step S4, using the current state values ​​of the hangar power distribution system (including voltage amplitude, current amplitude, active power, reactive power, and equipment temperature at each equipment node) as initial conditions, a multi-step recursive prediction of the future state is performed based on a physical information neural network within the prediction time domain. The physical information neural network superimposes the data-driven features extracted by the gated cyclic unit network with the physical analytical features calculated by the power flow equations and thermal balance equations, and outputs the predicted state value for the next time step.

[0155] During the recursive process, the dynamic constraint boundary generated in S3 is used as the inequality constraint of the multi-objective optimization function, and a sequential quadratic programming algorithm is used for online solution. After obtaining the optimal control sequence in the control time domain, only the first control step is used as the execution instruction for the current control cycle, which is converted into PLC instructions and executed by the PLC controller in the hangar. These execution instructions include: digital instructions, mapped to the PLC's digital output to control the switching of charging modules, and the start / stop of air conditioning units and fans; analog instructions, mapped to the PLC's analog output to control the continuous adjustment of charging power and the cooling / heating level of temperature control equipment; and step instructions, mapped to the PLC's pulse output to control the precise positioning and movement of the battery replacement robotic arm. After execution, the system enters the next control cycle, re-collects data, and repeats the above steps, forming a closed-loop rolling optimization of perception-fusion-decision-execution.

[0156] Based on the above technical solutions, this application achieves the following beneficial effects in the UAV hangar scenario: Through the joint construction of environmental semantic graph and power distribution knowledge graph and cross-modal feature fusion, the hangar power distribution system can simultaneously perceive the spatial distribution characteristics of the temperature field inside the cabin, the electrical operating status of each electrical device, and the thermal-electric coupling relationship between devices; through the adaptive adjustment of the dynamic constraint boundary of model predictive control by environmental risk coefficient and thermal-electric coupling coefficient, the hangar power distribution system can proactively reduce the upper limit of the current carrying capacity of the device in the area when the charging module operates under high load, causing local temperature rise, thus avoiding the risk of overheating; through the state space prediction model constructed by physical information neural network and the multi-objective optimization function containing physical penalty term, the hangar power distribution system can proactively avoid large adjustments of devices in high environmental risk areas while satisfying electrical constraints, and suppress differentiated actions between strongly coupled device pairs; through the reliable execution of the optimal control sequence by PLC, the hangar power distribution system realizes unmanned intelligent power distribution and micro-environment collaborative control in extreme weather and remote environments, effectively improving the operational safety, energy utilization efficiency, and equipment lifespan of the hangar under complex environmental conditions.

[0157] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple instances. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0158] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely illustrative descriptions of the application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and variations.

Claims

1. A PLC-based intelligent power distribution regulation method for distribution cabinets, characterized in that, include: Collect environmental sensor data and electrical operation data inside the distribution cabinet, construct an environmental semantic graph containing spatial grid nodes and environmental features, and a power distribution knowledge graph containing equipment entities and electrical relationships; The environmental semantic graph and the power distribution knowledge graph are spatially aligned, and graph neural networks are used to encode the spatially aligned environmental semantic graph and the power distribution knowledge graph respectively. Feature fusion is performed through a cross-attention mechanism to generate a fused feature matrix. The dynamic constraint boundary of the model predictive control is adjusted based on the environmental risk coefficient of the device node and the thermal-electric coupling coefficient between the device pairs extracted from the fusion feature matrix. Based on the current state of the system and the dynamic constraint boundary, the optimal control sequence of model predictive control is solved, and the optimal control sequence is converted into PLC instructions, which are then used by the PLC controller to execute the adjustment action.

2. The intelligent power distribution regulation method for a PLC-based distribution cabinet according to claim 1, characterized in that, The process of constructing the environment semantic graph includes: The internal space of the power distribution cabinet is discretized into three-dimensional grid nodes according to the preset grid spacing. The feature vector of each node is composed of temperature value, humidity value, airflow velocity value, temperature gradient value and humidity gradient value in three axes, semantic label and status flag. Spatial proximity is calculated using a Gaussian kernel function based on Euclidean distance of spatial coordinates, and environmental feature similarity is calculated using a Gaussian kernel function based on Euclidean distance of environmental feature vectors. The edge weights between nodes are obtained by weighting and summing the values ​​of spatial proximity and environmental feature similarity. The feature vectors of each node are used as node attributes, the edge weight matrix is ​​used as the adjacency matrix, and the grid nodes are used as vertices and the lines connecting adjacent nodes in space are used as edges to construct a weighted undirected graph, thus obtaining the environmental semantic graph. The semantic labels are obtained by clustering algorithms based on the environmental sensing data of each grid node, and are used to characterize the area type to which the grid node belongs. The area types include hot spots, cold spots, risk areas, normal areas, ventilation bottleneck areas, and transition areas. The status flag consists of binary 0 or 1 and is used to characterize whether the node is occupied by a device.

3. The intelligent power distribution regulation method for a PLC-based distribution cabinet according to claim 1, characterized in that, The process of constructing the power distribution knowledge graph includes: Entities and entity attributes are extracted from the electrical topology of the distribution cabinet, equipment nameplate parameters, and historical operation records. The types of entities include power supply nodes, bus nodes, circuit breaker nodes, switch nodes, load nodes, compensation nodes, and transformer nodes. The entity attributes include rated voltage, rated current, rated capacity, impedance value, health status, spatial coordinates, and current operating status. Extract the relationships between entities, including electrical connection relationships, control relationships, membership relationships, thermal impact relationships, dependency relationships, and spatial proximity relationships; A graph embedding model based on complex vector space is used to map the entities and relations to complex vector space, thereby obtaining the embedding vectors of each entity and the embedding vectors of each relation. Based on the spatial proximity and thermal influence relationships between entities, construct the adjacency matrix of the power distribution knowledge graph, calculate the graph Laplacian matrix of the power distribution knowledge graph, and complete the construction of the power distribution knowledge graph.

4. The intelligent power distribution regulation method for a PLC-based distribution cabinet according to claim 1, characterized in that, The step of aligning the environmental semantic graph with the power distribution knowledge graph in terms of spatial coordinates specifically includes: Obtain the spatial coordinates of each device entity in the power distribution knowledge graph; Using the spatial coordinates of each device entity as the query vector, search for the nearest Euclidean distance grid node among all grid nodes in the environmental semantic graph, and establish the nearest neighbor mapping relationship from the device entity to the nearest grid node; Calculate the Euclidean distance between the spatial coordinates of each device entity and the spatial coordinates of its nearest neighbor grid node as the mapping error; If the mapping error is greater than a preset threshold, a manual verification warning is triggered; otherwise, the device entity is bound to the nearest neighbor grid node to obtain the binding mapping relationship between the device entity and the environmental semantic graph grid node. Spatial coordinate alignment is completed based on the binding mapping relationship.

5. The intelligent power distribution regulation method for a PLC-based distribution cabinet according to claim 1, characterized in that, The process of generating the fusion feature matrix specifically includes: A graph convolutional network is used to encode the environment semantic graph. The node feature matrix and adjacency matrix of the environment semantic graph are used as inputs, and the environment embedding vector of each grid node in the environment semantic graph is output. The node feature matrix of the environment semantic graph is obtained by stacking the feature vectors of each grid node in the order of node index. A graph attention network is used to encode the power distribution knowledge graph. The complex vector space embedding and adjacency matrix of each entity in the power distribution knowledge graph are used as input. The attention weights of the neighbor nodes of each entity are calculated by a multi-head graph attention layer and the neighbor features are aggregated. The output is the electrical embedding vector of each device entity in the power distribution knowledge graph. The environment embedding vectors of the nearest neighbor mesh nodes corresponding to each device entity are aggregated according to the spatial coordinate mapping relationship to obtain the environment aggregation features corresponding to each device entity. Using the electrical embedding vectors of each device entity as the query matrix and the environmental aggregation features as the key and value matrices, the first cross-attention feature from the power distribution knowledge graph to the environmental semantic graph is calculated. Using the environmental aggregation features corresponding to each device entity as the query matrix and the electrical embedding vector of each device entity as the key matrix and value matrix, the second cross-attention matrix from the environmental semantic graph to the power distribution knowledge graph is calculated. The electrical embedding vector of each device entity is added to the first cross-attention feature, and the environmental aggregation feature of each device entity is added to the second cross-attention feature. The two parts of the added feature are then concatenated to obtain the fused feature vector of each device entity. The fused feature vectors of all device entities constitute the fused feature matrix.

6. The intelligent power distribution regulation method for a PLC-based distribution cabinet according to claim 5, characterized in that, The method for obtaining the environmental risk coefficient includes: The fused feature vector of each device entity in the fused feature matrix is ​​input into a pre-trained multilayer perceptron network, and the environmental risk coefficient of the device entity is output; the environmental risk coefficient is used to characterize the relative magnitude of the device operation risk under the current environmental conditions.

7. The intelligent power distribution regulation method for a PLC-based distribution cabinet according to claim 5, characterized in that, The method for obtaining the thermal-electric coupling coefficient includes: For each pair of device entities in the power distribution knowledge graph, calculate the cosine similarity between the fused feature vectors; The cosine similarity is divided by a preset temperature coefficient and then input into the Softmax function for normalization to obtain the thermal-electric coupling coefficient between the device entities; the thermal-electric coupling coefficient is used to characterize the degree of mutual influence between the two devices under the coupling of thermal and electric fields.

8. The intelligent power distribution regulation method for a PLC-based distribution cabinet according to claim 1, characterized in that, The adjustment of the dynamic constraint boundaries of the model predictive control specifically includes: Obtain the rated operating boundaries of each device entity, including the lower limit of rated voltage, the upper limit of rated voltage, the upper limit of rated current, and the upper limit of rated temperature; Multiply the environmental risk coefficient by the preset environmental deduction coefficient to obtain the environmental deduction amount for each device. The environmental derating is superimposed on the rated operating boundary of each device to generate the dynamic constraint boundary for the current control cycle. The dynamic constraint boundary is used as the inequality constraint condition for the optimization problem of model predictive control. Determine whether the F-norm of the topology difference matrix is ​​greater than a preset threshold; if yes, increase the environmental derating coefficient to a preset emergency derating value, recalculate the dynamic constraint boundary and execute; the topology difference matrix is ​​obtained by resampling the graph Laplacian matrix of the power distribution knowledge graph to the spatial coordinate system of the environmental semantic graph according to the binding mapping relationship, and then subtracting the graph Laplacian matrix of the environmental semantic graph element by element.

9. The intelligent power distribution regulation method for a PLC-based distribution cabinet according to claim 1, characterized in that, The process of solving for the optimal control sequence of model predictive control based on the current system state and the dynamic constraint boundary specifically includes: A state-space prediction model for a power distribution system is established. The state variables of the state-space prediction model include the voltage amplitude, current amplitude, active power, reactive power, and equipment temperature of each equipment node. The control variables of the state-space prediction model include opening and closing commands, capacitor switching commands, and tap position commands of on-load tap-changing transformers. The state-space prediction model adopts a physical information neural network. Obtain the system state value of the current control cycle, and perform multi-step recursion on the future state based on the state space prediction model in the prediction time domain; During the recursive process, the dynamic constraint boundary is used as an inequality constraint of the multi-objective optimization function, and the sequential quadratic programming algorithm is used to solve the multi-objective optimization objective function. In each iteration, the environmental risk coefficient, thermal-electric coupling coefficient and dynamic constraint boundary are updated according to the fusion feature matrix at the current time. The first control step in the optimal control sequence obtained in the control time domain is used as the execution instruction of the current control cycle to obtain the optimal control sequence. The optimal control sequence includes: digital instructions, which are mapped to the digital output of the PLC to control the opening and closing of the circuit breaker; analog instructions, which are mapped to the analog output of the PLC to control the switching degree of the capacitor; and step instructions, which are mapped to the pulse output of the PLC to control the tap position of the on-load tap-changing transformer.

10. A PLC-based intelligent power distribution and regulation system for distribution cabinets, characterized in that, include: The system comprises a graph construction module, a feature fusion module, a model prediction and control module, and a PLC execution module; among which... The graph construction module is used to collect environmental sensing data and electrical operation data in the power distribution cabinet, construct an environmental semantic graph containing spatial grid nodes and environmental features, and a power distribution knowledge graph containing equipment entities and electrical relationships. The feature fusion module is used to align the environmental semantic graph with the power distribution knowledge graph in spatial coordinates, and to encode the spatially aligned environmental semantic graph and power distribution knowledge graph respectively using a graph neural network, and to perform feature fusion through a cross-attention mechanism to generate a fused feature matrix. The model predictive control module is used to adjust the dynamic constraint boundary of model predictive control based on the environmental risk coefficient of the device node and the thermal-electric coupling coefficient between the device pairs extracted from the fusion feature matrix, and to solve the optimal control sequence of model predictive control based on the current state of the system and the dynamic constraint boundary. The PLC execution module is used to convert the optimal control sequence into PLC instructions and use the PLC controller to execute adjustment actions.

Images