A distributed power distribution management method and platform for power distribution cabinet

Abstract

The present invention discloses a distributed power distribution management method and platform for power distribution cabinets, and relates to the technical field of power distribution management. The method includes: collecting power operation data sets; constructing a first graph neural network and combining it with data set training to obtain a second graph neural network; performing federated learning on multiple power distribution cabinets based on the second graph neural network, generating power learning logs and formulating power dispatching strategies; executing strategies for feedback response, obtaining dispatching feedback results; updating the graph neural network according to the feedback results, dynamically adjusting the dispatching strategy, generating a second dispatching strategy, and realizing distributed power distribution management. The present invention solves the technical problem that the existing power distribution cabinet management cannot realize unified coordinated dispatching and real-time optimization due to data dispersion and lack of effective linkage between the power distribution cabinets, and achieves the technical effect of realizing intelligent and dynamic management of multiple power distribution cabinets through graph neural networks and federated learning, and improving the accuracy and response speed of power dispatching.

Description

Technical Field

[0001] The present invention relates to the technical field of power distribution management, and in particular to a distributed power distribution management method and platform for a power distribution cabinet. Background Art

[0002] With the rapid development of the power system, distribution cabinets play a vital role in the power transmission and distribution process. Traditional distribution cabinet management mainly relies on manual monitoring and centralized control systems. Although this model meets the needs of the power system to a certain extent, it shows some technical limitations when facing the distributed management of multiple distribution cabinets. For example, when multiple distribution cabinets are distributed in different areas and the power load status is complex, the existing centralized control system is difficult to perceive the operating status of each distribution cabinet in time, resulting in the inability to optimize the power dispatching strategy in real time, which in turn affects the overall efficiency of the power system. In addition, the traditional distribution management method is also unable to effectively handle the linkage and coordination between distribution cabinets, and cannot quickly respond and adjust to the load fluctuations of the power system, thereby increasing the risk of power loss and system instability. Summary of the invention

[0003] The present application provides a distributed power distribution management method and platform for power distribution cabinets, which is used to solve the technical problem that the existing power distribution cabinet management cannot achieve unified coordinated scheduling and real-time optimization due to data dispersion and lack of effective linkage between the distribution cabinets.

[0004] In view of the above problems, the present application provides a distributed power distribution management method and platform for a power distribution cabinet.

[0005] In a first aspect of the present application, a method for distributed power distribution management of a distribution cabinet is provided, the method comprising: performing sensor deployment on a plurality of distribution cabinets based on a sensor network, and collecting power operation data sets of the plurality of distribution cabinets; constructing a first graph neural network, and training the first graph neural network in combination with the power operation data set to obtain a second graph neural network; performing federated learning on a plurality of distribution cabinets based on the second graph neural network, generating a power learning log, and formulating a first power dispatching strategy based on the power learning log; executing the first power dispatching strategy to provide real-time feedback response to a plurality of distribution cabinets, and obtaining a dispatching feedback result; updating the second graph neural network based on the dispatching feedback result, dynamically adjusting the first power dispatching strategy according to the updated result, and generating a second power dispatching strategy to perform distributed power distribution management on a plurality of distribution cabinets.

[0006] According to a second aspect of the present application, a distributed power distribution management platform for power distribution cabinets is provided, the platform comprising: a sensor deployment module, the sensor deployment module is used to perform sensor deployment on multiple power distribution cabinets based on a sensor network, and collect power operation data sets of multiple power distribution cabinets; a graph neural network training module, the graph neural network training module is used to construct a first graph neural network, and use the first graph neural network in combination with the power operation data set for training to obtain a second graph neural network; a federated learning module, the federated learning module is used to perform federated learning on multiple power distribution cabinets based on the second graph neural network, generate a power learning log, and formulate a first power dispatching strategy based on the power learning log; a real-time feedback response module, the real-time feedback response module is used to execute the first power dispatching strategy to perform real-time feedback response on multiple power distribution cabinets and obtain a dispatching feedback result; a distributed power distribution management module, the distributed power distribution management module is used to update the second graph neural network according to the dispatching feedback result, dynamically adjust the first power dispatching strategy according to the updated result, and generate a second power dispatching strategy to perform distributed power distribution management on multiple power distribution cabinets.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0008] Based on the sensor network, multiple distribution cabinets are deployed for sensing, and the power operation data sets of multiple distribution cabinets are collected and obtained; a first graph neural network is constructed, and the first graph neural network is used to train the power operation data set to obtain a second graph neural network; based on the second graph neural network, multiple distribution cabinets are federated for learning, and a power learning log is generated, and a first power dispatching strategy is formulated according to the power learning log; the first power dispatching strategy is executed to provide real-time feedback response to multiple distribution cabinets to obtain dispatching feedback results; the second graph neural network is updated according to the dispatching feedback results, and the first power dispatching strategy is dynamically adjusted according to the updated results, and a second power dispatching strategy is generated to perform distributed power distribution management on multiple distribution cabinets. The technical effect of realizing intelligent and dynamic management of multiple distribution cabinets through graph neural networks and federated learning, and improving the accuracy and response speed of power dispatching is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0010] Figure 1A schematic diagram of a distributed power distribution management method for a power distribution cabinet provided in an embodiment of the present application;

[0011] Figure 2 A schematic diagram of the structure of a distributed power distribution management platform for a power distribution cabinet provided in an embodiment of the present application.

[0012] Explanation of the accompanying drawings: sensor deployment module 1, graph neural network training module 2, federated learning module 3, real-time feedback response module 4, distributed power distribution management module 5. DETAILED DESCRIPTION

[0013] The present application provides a distributed power distribution management method and platform for a distribution cabinet, which is used to solve the technical problem that the existing distribution cabinet management cannot achieve unified coordinated scheduling and real-time optimization due to data dispersion and lack of effective linkage between distribution cabinets.

[0014] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0015] Embodiment 1, as Figure 1 As shown, the present application provides a distributed power distribution management method for a power distribution cabinet, the method comprising:

[0016] Based on the sensor network, multiple distribution cabinets are deployed for sensing, and the power operation data sets of multiple distribution cabinets are collected and obtained.

[0017] In one embodiment, in the process of sensor deployment of multiple distribution cabinets based on a sensor network, the system terminal first installs a variety of sensors, such as current sensors, voltage sensors, and temperature sensors, at different parts of each distribution cabinet. These sensors can monitor the operating status of the distribution cabinet in real time and obtain key power parameters such as current, voltage, power, and temperature. Each sensor will collect these parameters regularly and transmit the data to the system terminal through the deployed sensor network. The data collected by these sensors are aggregated through the sensor network to form an electric power operation data set containing the operating conditions of all distribution cabinets. Through this data set, the real-time status of each distribution cabinet can be obtained, providing comprehensive data support for subsequent intelligent scheduling and management.

[0018] A first graph neural network is constructed, and the first graph neural network is used in combination with the power operation data set for training to obtain a second graph neural network.

[0019] In one embodiment, the process of constructing the first graph neural network is implemented based on the physical connection structure of multiple distribution cabinets. In this process, the system terminal regards each distribution cabinet as a node, and the power transmission line between the distribution cabinets is used as the edge between these nodes, and then the first graph neural network is constructed by combining these points and edges. The first graph neural network intuitively reflects the physical layout and mutual relationship between multiple distribution cabinets through the setting of such nodes and edges. In the first graph neural network, the node not only represents the physical location of the distribution cabinet, but also contains various structural data related to the distribution cabinet. The setting of the edge reflects the power transmission path between adjacent distribution cabinets, which can help the first graph neural network understand how power flows between different distribution cabinets. Subsequently, the system terminal uses the power operation data set collected by the sensor to train the first graph neural network. The training process mainly includes two steps: the first is to update the state of each node in the network. This means that each distribution cabinet node in the first graph neural network will perform forward propagation according to the power operation data set to obtain the data propagation trend. The second is to use the data propagation trend to optimize the parameters of the first graph neural network, for example, the weight distribution between nodes. Through this optimization process, the first graph neural network can more accurately simulate the power transmission relationship between distribution cabinets and improve the accuracy of overall power dispatching. After the above training, the first graph neural network will gradually evolve into the second graph neural network. The second graph neural network propagates and aggregates the information of adjacent nodes (i.e., adjacent distribution cabinets) so that each node not only relies on its own sensor data, but also can obtain information from other distribution cabinets connected to it. This way of information propagation and aggregation can better reflect the global state of the power system, thus laying the foundation for federated learning and the formulation of power dispatching strategies.

[0020] In a possible implementation, the first graph neural network is trained in combination with the power operation data set to obtain a second graph neural network, including:

[0021] The first graph neural network is traversed to record multiple nodes and multiple edges, and a topological structure of multiple distribution cabinets is constructed, wherein the topological structure includes multiple weight coefficients; based on the power operation data set, feature extraction is performed on the first graph neural network to determine a feature data set, wherein the feature data set includes multiple node feature data and multiple edge feature data; based on the multiple node feature data and the multiple edge feature data, forward propagation is performed according to the topological structure of multiple distribution cabinets to generate a data propagation trend.

[0022] Optionally, the system terminal traverses the nodes and edges in the first graph neural network, checks each node and each edge in the network one by one, and records their status and attribute information, thereby constructing a topological structure of multiple distribution cabinets. The topological structure of the distribution cabinet is composed of distribution cabinet nodes and power transmission lines (edges) between them. Each node represents a distribution cabinet, and the edge represents the power transmission path between adjacent distribution cabinets. During the traversal process, weights are assigned to each node and edge. These weights reflect the intensity or efficiency of power transmission between different distribution cabinets and become key parameters for subsequent analysis. Subsequently, based on the power operation data set collected by the sensor, feature extraction is performed on the first graph neural network. The purpose of feature extraction is to match the corresponding information of each node and edge from the power operation data set, such as the voltage, current, power, etc. of the node, as well as the load and transmission efficiency of power transmission on the edge. These extracted data will constitute a feature data set. The feature data set is divided into node feature data and edge feature data. The node feature data describes the operating status of each distribution cabinet, and the edge feature data reflects the power transmission between adjacent distribution cabinets. After determining the feature data set, the system terminal performs the forward propagation process based on these node feature data and edge feature data according to the previously constructed distribution cabinet topology. Forward propagation is a computational method that aims to propagate information from one node to its connected nodes through the hierarchical structure of the graph neural network. Specifically, each distribution cabinet node obtains feature data from its directly connected adjacent nodes (distribution cabinets connected by edges). The data of the adjacent nodes are propagated to the target node through the edges. At this time, the feature data of the edges will affect the information propagation process. For example, the weight of the edge can represent the path information with higher transmission efficiency. After obtaining the feature data of the adjacent nodes, the target node aggregates the data. Aggregation is a process of merging feature values ​​from different adjacent nodes. Commonly used methods include weighted average, summation, or maximum aggregation. In this way, each node integrates the feature information of itself and its adjacent nodes to form a feature representation containing more global context information. This process not only considers the data of the node itself, but also integrates the information of the adjacent nodes to reflect the overall status of multiple distribution cabinets in the power system. After aggregating the features of adjacent nodes, the node will combine these aggregation results with its own feature data to further update the node's feature representation. This step allows each node to not only reflect its own status, but also capture the status changes of adjacent distribution cabinets. With multiple rounds of aggregation and propagation, each node not only obtains the data of the distribution cabinet directly connected to itself, but also absorbs more status information of remote distribution cabinets. This enables the system terminal to capture the global power flow trend rather than the local perspective of a single distribution cabinet. After multiple rounds of aggregation, the load feature data of each node can reflect how the load is transferred.For example, when the load of a distribution cabinet in a certain area increases, the neighboring distribution cabinets may be affected by the change in power demand. This local change gradually expands to the entire system through multiple rounds of propagation, generating a global load change trend. In addition, power flow is not only reflected in the change of power load, but also involves how power is transmitted between different distribution cabinets. With multiple rounds of forward propagation and aggregation, each distribution cabinet node can capture the changes in power flowing into or out of its neighboring distribution cabinets. The integration of this information enables the system terminal to obtain the global power flow trend. Finally, the global load change trend and power flow trend are summarized to obtain the data propagation trend, which reflects the overall state of power system operation and can provide an effective reference for distribution impact analysis.

[0023] Perform power distribution impact analysis on the multiple nodes and the multiple edges of the first graph neural network according to the data propagation trend, update the multiple weight coefficients according to the analysis results, and determine multiple optimized weight coefficients; perform data transmission training in combination with the multiple nodes and the multiple edges of the first graph neural network according to the multiple optimized weight coefficients to construct the second graph neural network.

[0024] Optionally, after obtaining the data propagation trend, the system terminal performs a distribution impact analysis on each node (i.e., each distribution cabinet) and edge (power transmission line between adjacent distribution cabinets) in the first graph neural network. First, the system terminal extracts the power load and transmission efficiency of each distribution cabinet node and edge from the data propagation trend. For the node, the power load of each distribution cabinet is checked and its current load is compared with the average load of the entire system. If the load of a distribution cabinet is significantly higher than the average, it indicates that it plays a more important role in power transmission. The system terminal will increase the weight of the node to reflect its influence on the adjacent distribution cabinet. On the contrary, the weight of the node with lower load will be reduced to ensure that its role in information transmission is moderately reduced. In the process of adjusting the weight coefficient, the difference between the current load and the average load is used to calculate the ratio with the average load, and the node adjustment factor is multiplied by the calculated ratio to obtain the node influence, and then the node influence is added to the corresponding current weight coefficient to determine the optimized weight coefficient of the node. For the edge, the system terminal will evaluate the transmission efficiency of the power transmission line between adjacent distribution cabinets. If the transmission efficiency of a line is lower than the target transmission efficiency, the weight of the edge will be reduced to reduce its impact in power transmission and avoid the negative impact of inefficient transmission on the entire network. On the contrary, if the line transmission efficiency is higher than the target transmission efficiency, the weight of the edge will be appropriately increased so that the transmission efficiency of the line in the network can be fully utilized. The specific adjustment method is similar to the above, except that the adjustment factor of the edge and the target transmission efficiency are used. After power distribution analysis and weight adjustment, the system terminal obtains multiple optimized weight coefficients. These optimized weights can not only more accurately reflect the actual contribution of each distribution cabinet and its adjacent lines in power transmission, but also provide more accurate data input for subsequent graph neural network training. Finally, the optimized weight coefficients will be applied to the first graph neural network for subsequent training and data transmission. In this process, the data transmission method and node status update will be based on the new weight distribution, which can better reflect the actual operation of the power system. Through multiple rounds of data transmission training, the feature updates of nodes and edges gradually converge, that is, the adjustment of weight coefficients tends to be stable. At this time, the system terminal obtains the final second graph neural network. The nodes and edges in this network not only reflect the local characteristics of the power system, but also capture the power flow pattern and load trend of the entire system through optimized weights combined with feature updates. Compared with the first graph neural network, the second graph neural network has stronger global perception capabilities and dynamic adaptability to the distribution network, and can better provide support for intelligent distribution scheduling and management.

[0025] Based on the second graph neural network, federated learning is performed on multiple distribution cabinets to generate a power learning log, and a first power dispatching strategy is formulated according to the power learning log.

[0026] In one embodiment, federated learning is a distributed learning method that is different from traditional centralized data collection and processing. In the scenario of distribution cabinets, federated learning means that each distribution cabinet node can process power data and model training locally without aggregating all data to a central server. Through the nodes and edges in the second graph neural network, the system terminal can connect multiple distribution cabinets to achieve federated learning. First, the system terminal will traverse multiple distribution cabinets, aggregate the information of adjacent nodes through the second graph neural network, and each node will generate aggregate features based on the information of its neighboring distribution cabinets. These aggregate features correspond to each distribution cabinet one by one, reflecting the position of each distribution cabinet in the network and the power transmission status around it. Subsequently, based on the aggregate features of these nodes, local power operation data is selected from multiple distribution cabinets to construct local data training sets, and then these local data training sets are sent to the federated server, which centrally calculates the collected local training data to construct a global power learning model. The model reflects the global state of the entire power system and is optimized in combination with the local data characteristics of each distribution cabinet. Afterwards, the system terminal shares the global power learning model based on multiple nodes, records the learning process of the model, and generates a power learning log. This log records in detail the power transmission mode, load distribution and other key information in the power system, providing data support for subsequent power dispatch strategies. After obtaining the power learning log, the system terminal will analyze and summarize the key information, and formulate the first power dispatch strategy through steps such as power transfer analysis and power supply distribution. This strategy will combine the learned global load distribution and power demand forecast to balance the load and power transmission efficiency of each distribution cabinet, ensuring that power resources can be reasonably distributed among different distribution cabinets.

[0027] In a possible implementation, a plurality of power distribution cabinets are federatedly learned based on the second graph neural network to generate a power learning log, including:

[0028] The second graph neural network is used to traverse multiple distribution cabinets to aggregate information of adjacent nodes and generate aggregate features of multiple nodes, where the aggregate features of the multiple nodes correspond to the multiple distribution cabinets; local data is selected based on the multiple distribution cabinets according to the aggregate features of the multiple nodes to construct a local data training set; the local data training set is received through a federal server for calculation to construct a global power learning model; the global power learning model is shared based on the multiple nodes, and learning records of the global power learning model are recorded to generate the power learning log.

[0029] Optionally, the system terminal traverses multiple distribution cabinets through the second graph neural network to obtain the adjacent node information of each distribution cabinet. The information of the adjacent nodes is aggregated by the same method as above to generate the aggregated features of each distribution cabinet. These aggregated features reflect the position of each distribution cabinet in the network and its relationship with the adjacent distribution cabinets, and are used to capture the local characteristics of power transmission and load. Subsequently, according to the aggregated features of each distribution cabinet, the local power operation data related to its features are selected to construct a local data training set. The local data training set only includes the model weights and update information used for training, rather than the original power data. This ensures that the privacy data is not uploaded to the federal server, thereby protecting the operation privacy of the distribution cabinet. After that, these local training sets are sent to the federal server, which receives these locally processed model weights and training results, and calculates based on these model weights and training results, merges the models of each distribution cabinet, and constructs a global power learning model. This global model integrates the power operation characteristics of multiple distribution cabinets and reflects the power flow and load conditions of the entire power system. After the construction is completed, the global power learning model will be shared back to each distribution cabinet, and each distribution cabinet will adjust and optimize the local operating status according to the global model. At the same time, the system terminal will record the data update of each distribution cabinet during the learning process and generate a power learning log. This log records in detail the power operation status of each distribution cabinet, the model changes during the federated learning process, and the optimization details of the global model, providing data support for subsequent power dispatch strategies.

[0030] In a possible implementation, the federal server includes:

[0031] Establish communication links for the multiple nodes to obtain a node communication network; perform collaborative verification on the multiple distribution cabinets based on the node communication network, perform aggregation analysis on the multiple nodes according to the verification results, and generate power dynamic operation status information of the multiple nodes, wherein the power dynamic operation status information includes power load data and power loss data; calculate the fluctuation extreme value of the power loss data according to the power load data, identify it, and determine the aggregation period; perform calculations on the local data training set according to the aggregation period to generate multiple training calculation results; and construct the federal server based on the update frequency and the multiple training calculation results.

[0032] Optionally, the system terminal establishes communication links based on multiple nodes of the second graph neural network, and constructs a node communication network based on these communication links to ensure that each distribution cabinet node can communicate with each other and share information. Through this node communication network, multiple distribution cabinets are collaboratively verified, and the verification process includes comparative analysis of the power operation status of each node to ensure the consistency of power data and transmission stability of each node. Specifically, the system terminal collects current power operation data from each distribution cabinet node. These data include key power parameters such as voltage, current, and power load, and check the synchronization of these parameters between nodes to ensure that the status reports of all distribution cabinets are consistent, without lag or abnormal data delay. In the synchronous state, the local data of each distribution cabinet node is verified for consistency, that is, the power transmission data between each distribution cabinet is checked to see if it conforms to the physical connection relationship. For example, the power output of a distribution cabinet node should match the power received by the distribution cabinet to which it is connected to avoid data inconsistency. In this way, the accuracy of data during power transmission can be ensured. In addition, the power load of each node will be balanced and verified, and the load of each distribution cabinet will be analyzed to check whether there is uneven load or abnormal fluctuation. For example, if the power load of some nodes is abnormally high or low, these abnormalities will be marked for further analysis and processing. Through load balancing verification, potential power bottlenecks or power waste can be prevented. The power transmission efficiency between nodes is also analyzed through the communication network, whether there is loss in the power transmission process, and the power loss from one distribution cabinet node to another is calculated and marked. Subsequently, based on the verification results, multiple nodes are aggregated and analyzed, that is, the power load data and power loss data of adjacent nodes are aggregated in the same way as above to generate the power dynamic operation status information of multiple nodes. After that, based on the power load data, the fluctuation extreme values ​​of the power loss data of each node are calculated. The fluctuation extreme values ​​identify the maximum and minimum points of load changes and energy losses in the power system. These extreme values ​​help the system terminal better understand the key nodes and problem areas in power transmission. Based on these fluctuation data, the system terminal determines the aggregation period, that is, the time interval between the maximum point and the adjacent minimum point of the power load data, to ensure that the operation strategy can be adjusted in time. Then, according to the determined aggregation period, the local data training set of each node is calculated. Each node aggregates local training data during its aggregation cycle, shares this data with the federated server through the node communication network, and generates multiple training calculation results. These training results integrate the power load and loss conditions of each node and reflect the overall operating status of the power system. Then, the update frequency of federated learning is dynamically determined based on the real-time load conditions of the power system. The update frequency determines the time interval at which the federated server receives training calculation results from each distribution cabinet node.If the load of the power system fluctuates greatly, the update frequency will be increased to ensure that the model can respond quickly to load changes; if the load is relatively stable, the update frequency will be reduced to save computing resources. Finally, by combining the training calculation results with the dynamically adjusted update frequency, a federated server is built to ensure that the entire power system can adapt to changes in power demand and transmission status in real time during the federated learning process, thereby achieving more efficient distributed power management.

[0033] In a possible implementation, formulating a first power dispatching strategy according to the power learning log includes:

[0034] Based on the power learning log and the power operation data set, a multi-level analysis is performed to generate a multi-level power state; according to the multi-level power state, multiple distribution cabinets are traversed to perform load balancing judgment and determine the judgment result; according to the judgment result, power transfer analysis is performed on multiple distribution cabinets to obtain multiple load distribution results; based on the power learning log, power demand forecast information is extracted, the multiple load distribution results are combined with the power demand forecast information to perform power supply allocation, and the first power dispatching strategy is formulated.

[0035] Optionally, based on the power learning log and the power operation data set, the system terminal first performs a multi-level analysis. The multi-level analysis includes detailed recording and analysis of the load information of each distribution cabinet. Specifically, the system terminal obtains the load fluctuation data of each distribution cabinet in a specific time period from the power operation data set, and compares it with the corresponding threshold to identify high-load areas and low-load areas. This load information helps to understand the current status of power distribution and find out which distribution cabinets are overloaded and which are relatively low. Then, the prediction results combined with the global model are obtained from the power learning log. These results are power demand trend predictions based on graph neural networks. The system terminal will analyze the possible changes in the load based on the trend prediction of future power demand and identify the distribution cabinets that may bear a large load in the future. At the same time, through the historical fault and anomaly detection in the power operation data set, the nodes or faulty distribution cabinets that have shown abnormalities are extracted from the power learning log to identify whether these nodes are likely to have problems again. Subsequently, the results of the multi-level analysis are summarized to form a multi-level power state. After completing the multi-level analysis, the system terminal traverses multiple distribution cabinets based on these data to make load balancing decisions. The purpose of load balancing judgment is to determine whether the load of each distribution cabinet meets expectations and whether its power distribution needs to be adjusted, thereby generating a judgment result. In this process, if the load of a distribution cabinet exceeds the set safety threshold, it will be marked as a high-load area. A high-load distribution cabinet may cause overheating, equipment aging or damage when carrying too much power for a long time. If the load of a distribution cabinet is lower than the set minimum load threshold, it will be marked as a low-load area. A low-load distribution cabinet may waste power resources or reduce efficiency when the power demand is insufficient. If the load fluctuation frequency of the distribution cabinet is higher than the fluctuation frequency threshold or the fluctuation amplitude is higher than the fluctuation amplitude threshold, the node will be marked as an unstable area, indicating that there may be problems with the operation status of the distribution cabinet. The system terminal can understand the area of ​​unbalanced load through the obtained judgment results, and determine which distribution cabinets need to reduce the load and which distribution cabinets are capable of taking on more loads. Based on this judgment result, the system terminal performs power transfer analysis on multiple distribution cabinets, analyzes the physical connection between the high-load distribution cabinet and the low-load distribution cabinet through the topological structure of the second graph neural network, and determines the feasible path for power transfer. In this process, the system terminal selects the low-load distribution cabinets directly connected to the high-load distribution cabinets, and selects the one with the largest weight coefficient as the first candidate for power transfer. If there is no directly connected low-load distribution cabinet, the low-load distribution cabinet with the largest cumulative weight coefficient among the low-load distribution cabinets indirectly connected to the high-load distribution cabinet is selected. After determining the transfer path, the system terminal calculates the load that needs to be reduced for each high-load distribution cabinet to reach the safe load threshold. Usually, the load reduction is directly related to the excess load in the high-load area. Then calculate the power load that the low-load distribution cabinet can undertake.This load increase depends on the current load of the low-load distribution cabinet and its maximum load capacity. After that, the determined transfer paths and power transfer amounts are integrated to generate multiple load distribution results, and the load adjustment plan for each distribution cabinet is clarified. Then, based on the power learning log, the power demand forecast information is extracted, which reflects the trend of power demand changes in the entire power system in the future. These forecast information are combined with the previously obtained load distribution results to identify whether there is a possibility of overload in the future. If the forecast shows that the power demand will continue to rise in the future, the power distribution in these areas will be further optimized to ensure that they can handle the future load increase. For the distribution cabinets with lower loads, it will be predicted whether these areas will experience load increase or decrease at some point in the future based on the power demand forecast results. If future demand growth is predicted, additional power will be allocated to these distribution cabinets in advance to cope with future load fluctuations. Finally, the system terminal formulates the first power dispatching strategy based on all the analysis results and forecast information. This strategy clarifies how power is reasonably distributed among different distribution cabinets to ensure load balance and stable power supply of the entire power system.

[0036] The first power dispatching strategy is executed to provide real-time feedback response to multiple distribution cabinets to obtain dispatching feedback results.

[0037] In one embodiment, after formulating the first power dispatching strategy, the system terminal applies the first power dispatching strategy to the entire power distribution network. Each distribution cabinet adjusts its power receiving and supplying status according to the strategy. Specific operations include adjusting the power load, optimizing the power transmission path, and changing the power supply in real time. The system terminal will instruct the high-load distribution cabinet to reduce the load and transfer part of the power to the low-load area to ensure that the load of the entire network can be balanced. During the execution of the strategy, the system terminal monitors the operating status of each distribution cabinet in real time and continuously collects power operation data through the sensor network. The key parameters such as current, voltage, load fluctuation, etc. of each distribution cabinet will be monitored in real time. The system terminal can evaluate the effect of the dispatching strategy in real time based on these data, including load balancing, power transmission efficiency, response speed and stability, etc., and summarize these evaluation data to obtain the dispatching feedback results, and provide data basis for subsequent dynamic adjustments to ensure the balanced distribution of power load and maximize supply efficiency.

[0038] The second graph neural network is updated according to the scheduling feedback result, the first power scheduling strategy is dynamically adjusted according to the update result, and a second power scheduling strategy is generated to perform distributed power distribution management on multiple distribution cabinets.

[0039] In one embodiment, after executing the first power dispatching strategy and obtaining feedback, the system terminal analyzes the feedback results. The feedback results include information such as the power operation status, load distribution effect, transmission efficiency and response speed of each distribution cabinet. By analyzing these results, problems or deficiencies in the current dispatching strategy can be identified. Subsequently, the dispatching feedback results are synchronized to the second graph neural network for interaction. The interaction process is completed by generating adversarial networks (GANs). The second graph neural network is then used to extract multiple nodes and edges with optimized weight coefficients, and data distribution analysis is performed on them. According to the analysis results, the second graph neural network and the adversarial network are combined to generate and judge data to generate multiple discriminant data. Afterwards, the second graph neural network is updated based on these discriminant data, the weights and features of the nodes and edges are adjusted, and the final update results are generated. The updated graph neural network is then used for power forecasting to obtain power demand forecast distribution information, load change trend forecast information, etc. Then, the system terminal dynamically adjusts the first power dispatching strategy based on these forecast information, thereby generating an optimized second power dispatching strategy to achieve more efficient distributed power distribution management.

[0040] In a possible implementation, updating the second graph neural network according to the scheduling feedback result includes:

[0041] The scheduling feedback result is synchronized to the second graph neural network for interaction, and an adversarial network is generated according to the interactive data set, wherein the adversarial network includes a generator and a discriminator; based on the second graph neural network, the multiple nodes and the multiple edges with the multiple optimization weight coefficients are extracted for analysis to determine multiple data distribution graphs; based on the second graph neural network in combination with the generator, data is generated according to the multiple data distribution graphs to obtain multiple generated data; the multiple generated data are synchronized with the second graph neural network to the discriminator for data screening to obtain multiple discriminant data; the second graph neural network is updated according to the multiple discriminant data to generate the updated result.

[0042] Optionally, the system terminal synchronizes the obtained dispatch feedback results to the second graph neural network to update the node and edge information of the graph neural network. The dispatch feedback results are combined with the existing feature information of each node and edge in the second graph neural network, including the power load status of each node, the weight and efficiency of the transmission path, etc. This process enables the graph neural network to reflect the current state of the power system more accurately based on the latest dispatch feedback. After synchronizing the feedback results, an interactive data set is generated by combining the node features and edge features in the second graph neural network. This interactive data set contains the synchronized power load information, transmission path data, and the latest state of the power system operation. The interactive data set is a combination of the second graph neural network and the real-time power system feedback information, which can provide basic data for the next generation adversarial network (GAN). After obtaining the interactive data set, the generation adversarial network is created. GAN consists of two core components: a generator and a discriminator, which interact and optimize the prediction ability of the power system state through competition. For the generator, the system terminal is designed through the architecture of the graph neural network, including a node aggregation layer, a data processing layer, and a generation layer. Among them, the node aggregation layer extracts input features from the nodes of the graph neural network, and uses the aggregation mechanism of the graph neural network to combine the features of each node with the features of its neighboring nodes to extract global context information. The data processed by the data processing layer will pass through a series of fully connected layers or convolutional layers for nonlinear transformation. These layers generate new node and edge features by adjusting weights. The generation layer generates new node and edge features, that is, simulates the power load and transmission conditions. Through this layer, the output result of the generator is the possible state of the future power system. For the discriminator, it is also designed through the architecture of the graph neural network, including node and edge feature processing layer, feature fusion layer, classification layer, etc. Among them, the feature fusion layer fuses node features and edge features, and further processes them through the fully connected layer. This layer will integrate the global information in the generated data and the real data to judge its authenticity. The classification layer classifies the input data as real or generated. If the discriminator believes that the input data comes from the real power system, the output is real data, otherwise the output is generated data. During the training process, for the generator, the weights in the generator are adjusted through steps such as forward propagation, back propagation, and minimizing cross entropy loss, and learn how to generate more and more realistic simulated power system data. For the discriminator, by constantly learning how to better distinguish between real data and generated data, its classification ability is optimized. Similarly, the internal parameters are adjusted through steps such as forward propagation, back propagation, and minimizing cross entropy loss to improve the discrimination accuracy. The generator and the discriminator are in a state of mutual confrontation during the training process. The generator tries to deceive the discriminator to generate seemingly real data, while the discriminator tries to identify the generated data.This confrontation process promotes continuous optimization of both parties. The simulated data generated by the generator is becoming more and more realistic, and the discriminator's discrimination ability is becoming stronger and stronger. After the construction of the generator and the discriminator is completed, the system terminal obtains the adversarial network.

[0043] After obtaining the adversarial network, the system terminal extracts multiple nodes and edges with optimized weight coefficients through the second graph neural network. These nodes and edges represent the key elements of the power load in the power system. By analyzing these nodes and edges, the respective power transmission conditions can be determined and multiple data distribution graphs can be generated. The data distribution graph reflects key information such as the power flow and load change trend of the nodes and edges. Subsequently, multiple simulation data are generated according to the data distribution graph using the second graph neural network and the generator. These generated data reflect the possible power load and transmission status in the power system. The generator generates these data through optimized node and edge information so that future power demand and load distribution can be predicted. The generated data is then synchronized to the discriminator for screening. The discriminator identifies which data conforms to the actual situation of the current power system and which does not by evaluating the validity of the generated data. The task of the discriminator is to ensure that the generated data has a high degree of authenticity and can truly reflect the operating status of the power system. According to the discriminant data screened by the discriminator, the system terminal updates the second graph neural network to obtain an update result, which includes the updated second graph neural network. The update process includes adjusting the weights of nodes and edges, optimizing transmission paths, etc. This process is constantly iterating, and by generating and filtering data multiple times, the second graph neural network gradually adapts to the latest status of the power system. Through this iterative process, the second graph neural network will continue to update back and forth. With each iteration, new dispatch feedback results will be synchronized to the network, generating new interactive data sets, and then continuously optimized through the generative adversarial network. This iterative process ensures that the second graph neural network can be continuously optimized as the operating status of the power system changes, thereby dynamically adjusting the power dispatch strategy and achieving more intelligent distributed distribution management.

[0044] In a possible implementation, dynamically adjusting the first power dispatching strategy according to the update result to generate a second power dispatching strategy includes:

[0045] Perform power forecasting based on the power dynamic operation status information of multiple nodes in combination with the update results to obtain a power forecast result, wherein the power forecast result includes power demand forecast distribution information and load change trend forecast information; traverse the first power dispatching strategy based on the power demand forecast distribution information to generate first information to be adjusted; traverse the first power dispatching strategy based on the load change trend forecast information to generate second information to be adjusted; dynamically allocate according to the first information to be adjusted in combination with the second information to be adjusted to generate a power allocation result to be adjusted; dynamically adjust the first power dispatching strategy according to the power allocation result to be adjusted to generate the second power dispatching strategy.

[0046] Optionally, the system terminal performs power forecasting for the entire power system based on the power dynamic operation status information of multiple nodes and the latest update result of the second graph neural network. The forecasting process analyzes the load change trend and power demand situation between nodes through the second graph neural network in the update result to generate a power forecast result, including power demand forecast distribution information and load change trend forecast information. Among them, the power demand forecast distribution information is generated by analyzing the power demand historical data, current load situation and future power demand trend of each distribution cabinet to generate the power demand forecast of different distribution cabinets in the future period of time. The load change trend forecast information is based on the load change history of each distribution cabinet and the influence of the adjacent distribution cabinets, and predicts the load change trend of each distribution cabinet in the future, especially the increase or decrease of the load. Subsequently, the first power dispatching strategy is traversed through the power demand forecast distribution information. During the traversal process, the system terminal analyzes the difference between the current power distribution and the future demand, checks which distribution cabinets are expected to increase in power demand, and evaluates whether the current dispatching strategy has reserved enough power for these distribution cabinets, identifies the areas where power demand is expected to decrease, and adjusts the excess power distribution to avoid power waste. Through the above analysis, the system terminal determines the nodes that need to be adjusted and generates the first information to be adjusted. This first information to be adjusted indicates the distribution cabinets that need to adjust the power supply and the corresponding power distribution changes. Similarly, the system terminal also traverses the existing first power dispatching strategy according to the load change trend prediction information. The traversal process analyzes the load change trend between the distribution cabinets, identifies the distribution cabinets that are about to face high loads, evaluates whether the current power distribution is sufficient to cope with future load changes, and identifies the areas where the load may drop, reduces power distribution, and optimizes the dispatching strategy. Through the above analysis, the system terminal also determines the nodes that need to be adjusted and generates the second information to be adjusted to ensure that the power distribution can meet the needs of future load changes. After that, the first information to be adjusted is combined with the second information to be adjusted for dynamic allocation. The dynamic allocation process will re-plan the power distribution of each distribution cabinet based on the first information to be adjusted and the second information to be adjusted, and generate the power distribution result to be adjusted. The result will list in detail the amount of power that needs to be increased or decreased for each distribution cabinet, and optimize the power resource allocation of the entire distribution network to ensure that the power supply is more matched with the demand. Then, according to the result of the power distribution to be adjusted, the first power dispatching strategy is dynamically adjusted to generate the second power dispatching strategy. This new dispatching strategy better reflects the future power demand and load change trend, ensuring that the power system can efficiently cope with future power dispatching needs. The new strategy will optimize power distribution on a global scale to avoid overload and power waste.

[0047] Embodiment 2 is based on the same inventive concept as the distributed power distribution management method of a power distribution cabinet in the above embodiment. Figure 2As shown, the present application provides a distributed power distribution management platform for power distribution cabinets, and the platform and method embodiments in the embodiments of the present application are based on the same inventive concept. Among them, the platform includes: sensor deployment module 1: the sensor deployment module 1 is used to perform sensor deployment on multiple power distribution cabinets based on a sensor network, and collect and obtain power operation data sets of multiple power distribution cabinets; graph neural network training module 2: the graph neural network training module 2 is used to construct a first graph neural network, and use the first graph neural network combined with the power operation data set for training to obtain a second graph neural network; federated learning module 3: the federated learning module 3 is used to perform federated learning on multiple power distribution cabinets based on the second graph neural network, generate a power learning log, and formulate a first power dispatching strategy based on the power learning log; real-time feedback response module 4: the real-time feedback response module 4 is used to execute the first power dispatching strategy to perform real-time feedback response on multiple power distribution cabinets and obtain a dispatching feedback result; distributed power distribution management module 5: the distributed power distribution management module 5 is used to update the second graph neural network according to the dispatching feedback result, dynamically adjust the first power dispatching strategy according to the updated result, and generate a second power dispatching strategy to perform distributed power distribution management on multiple power distribution cabinets.

[0048] Furthermore, the graph neural network training module 2 also includes:

[0049] The first graph neural network is traversed to record multiple nodes and multiple edges, and a topological structure of multiple distribution cabinets is constructed, wherein the topological structure includes multiple weight coefficients; based on the power operation data set, feature extraction is performed on the first graph neural network to determine a feature data set, wherein the feature data set includes multiple node feature data and multiple edge feature data; based on the multiple node feature data and the multiple edge feature data, forward propagation is performed according to the topological structure of multiple distribution cabinets to generate a data propagation trend; according to the data propagation trend, a power distribution impact analysis is performed on the multiple nodes and the multiple edges of the first graph neural network, and according to the analysis results, the multiple weight coefficients are updated to determine a multiple optimized weight coefficients; according to the multiple optimized weight coefficients, data transmission training is performed in combination with the multiple nodes and the multiple edges of the first graph neural network to construct the second graph neural network.

[0050] Furthermore, the federated learning module 3 also includes:

[0051] The second graph neural network is used to traverse multiple distribution cabinets to aggregate information of adjacent nodes and generate aggregate features of multiple nodes, where the aggregate features of the multiple nodes correspond to the multiple distribution cabinets; local data is selected based on the multiple distribution cabinets according to the aggregate features of the multiple nodes to construct a local data training set; the local data training set is received through a federal server for calculation to construct a global power learning model; the global power learning model is shared based on the multiple nodes, and learning records of the global power learning model are recorded to generate the power learning log.

[0052] Furthermore, the federated learning module 3 also includes:

[0053] Establish communication links for the multiple nodes to obtain a node communication network; perform collaborative verification on the multiple distribution cabinets based on the node communication network, perform aggregation analysis on the multiple nodes according to the verification results, and generate power dynamic operation status information of the multiple nodes, wherein the power dynamic operation status information includes power load data and power loss data; calculate the fluctuation extreme value of the power loss data according to the power load data, identify it, and determine the aggregation period; perform calculations on the local data training set according to the aggregation period to generate multiple training calculation results; and construct the federal server based on the update frequency and the multiple training calculation results.

[0054] Furthermore, the federated learning module 3 also includes:

[0055] Based on the power learning log and the power operation data set, a multi-level analysis is performed to generate a multi-level power state; according to the multi-level power state, multiple distribution cabinets are traversed to perform load balancing judgment and determine the judgment result; according to the judgment result, power transfer analysis is performed on multiple distribution cabinets to obtain multiple load distribution results; based on the power learning log, power demand forecast information is extracted, the multiple load distribution results are combined with the power demand forecast information to perform power supply allocation, and the first power dispatching strategy is formulated.

[0056] Furthermore, the distributed power distribution management module 5 also includes:

[0057] The scheduling feedback result is synchronized to the second graph neural network for interaction, and an adversarial network is generated according to the interactive data set, wherein the adversarial network includes a generator and a discriminator; based on the second graph neural network, the multiple nodes and the multiple edges with the multiple optimization weight coefficients are extracted for analysis to determine multiple data distribution graphs; based on the second graph neural network in combination with the generator, data is generated according to the multiple data distribution graphs to obtain multiple generated data; the multiple generated data are synchronized with the second graph neural network to the discriminator for data screening to obtain multiple discriminant data; the second graph neural network is updated according to the multiple discriminant data to generate the updated result.

[0058] Furthermore, the distributed power distribution management module 5 also includes:

[0059] Perform power forecasting based on the power dynamic operation status information of multiple nodes in combination with the update results to obtain a power forecast result, wherein the power forecast result includes power demand forecast distribution information and load change trend forecast information; traverse the first power dispatching strategy based on the power demand forecast distribution information to generate first information to be adjusted; traverse the first power dispatching strategy based on the load change trend forecast information to generate second information to be adjusted; dynamically allocate according to the first information to be adjusted in combination with the second information to be adjusted to generate a power allocation result to be adjusted; dynamically adjust the first power dispatching strategy according to the power allocation result to be adjusted to generate the second power dispatching strategy.

[0060] It should be noted that the above-mentioned sequence of the embodiments of the present application is only for description and does not represent the advantages and disadvantages of the embodiments. And the above-mentioned specific embodiments of this specification are described. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0061] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

[0062] This specification and the drawings are merely exemplary illustrations of the present application and are deemed to cover any and all modifications, variations, combinations or equivalents within the scope of the present application. Obviously, a person skilled in the art may make various modifications and variations to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present application and its equivalents, the present application intends to include these modifications and variations.

Claims

1. A distributed power distribution management method for a power distribution cabinet, characterized in that: The method comprises: Based on the sensor network, multiple distribution cabinets are deployed for sensing, and the power operation data sets of multiple distribution cabinets are collected and obtained; Constructing a first graph neural network, and using the first graph neural network in combination with the power operation data set for training to obtain a second graph neural network; Performing federated learning on multiple power distribution cabinets based on the second graph neural network to generate a power learning log, and formulating a first power dispatching strategy according to the power learning log; Execute the first power dispatching strategy to provide real-time feedback response to multiple distribution cabinets to obtain dispatching feedback results; The second graph neural network is updated according to the dispatch feedback result, the first power dispatch strategy is dynamically adjusted according to the update result, and a second power dispatch strategy is generated to perform distributed power distribution management on multiple distribution cabinets; The method of training the first graph neural network in combination with the power operation data set to obtain a second graph neural network includes: Traversing the first graph neural network to record multiple nodes and multiple edges, and constructing a topological structure of multiple distribution cabinets, wherein the topological structure includes multiple weight coefficients; Performing feature extraction on the first graph neural network based on the power operation data set to determine a feature data set, wherein the feature data set includes a plurality of node feature data and a plurality of edge feature data; Based on the plurality of node feature data and the plurality of edge feature data, forward propagation is performed according to the topological structure of the plurality of power distribution cabinets to generate a data propagation trend; Performing a power distribution impact analysis on the multiple nodes and the multiple edges of the first graph neural network according to the data propagation trend, updating the multiple weight coefficients according to the analysis results, and determining multiple optimized weight coefficients; Perform data transfer training in combination with the multiple nodes and the multiple edges of the first graph neural network according to the multiple optimized weight coefficients to construct the second graph neural network; The method of updating the second graph neural network according to the scheduling feedback result includes: Synchronize the scheduling feedback result to the second graph neural network for interaction, and generate an adversarial network according to the interaction data set, wherein the adversarial network includes a generator and a discriminator; Extracting the plurality of nodes and the plurality of edges having the plurality of optimized weight coefficients based on the second graph neural network for analysis, and determining a plurality of data distribution graphs; Based on the second graph neural network combined with the generator, data is generated according to the multiple data distribution graphs to obtain multiple generated data; Synchronizing the plurality of generated data in combination with the second graph neural network to the discriminator for data screening to obtain a plurality of discriminant data; The second graph neural network is updated according to the multiple discrimination data to generate the update result.

2. A distributed power distribution management method for a power distribution cabinet as claimed in claim 1, characterized in that: Based on the second graph neural network, federated learning is performed on multiple distribution cabinets to generate a power learning log, and the method includes: Traversing multiple power distribution cabinets through the second graph neural network to aggregate information of adjacent nodes, generating aggregate features of multiple nodes, where the aggregate features of the multiple nodes correspond to the multiple power distribution cabinets; Selecting local data based on the aggregation features of the multiple nodes based on the multiple distribution cabinets to construct a local data training set; The local data training set is received by the federal server for calculation to build a global power learning model; The global power learning model is shared based on the multiple nodes, and learning records are recorded for the global power learning model to generate the power learning log.

3. A distributed power distribution management method for a power distribution cabinet as claimed in claim 2, characterized in that: The federal server method includes: Establishing communication links for the multiple nodes to obtain a node communication network; Based on the node communication network, a plurality of distribution cabinets are collaboratively verified, and according to the verification results, the plurality of nodes are aggregated and analyzed to generate power dynamic operation status information of the plurality of nodes, wherein the power dynamic operation status information includes power load data and power loss data; Calculate the fluctuation extreme value of the power loss data according to the power load data, mark it, and determine the aggregation period; Execute calculations on the local data training set according to the aggregation period to generate multiple training calculation results; The federated server is constructed based on an update frequency and in combination with the plurality of training calculation results.

4. A distributed power distribution management method for a power distribution cabinet as claimed in claim 1, characterized in that: Formulating a first power dispatching strategy according to the power learning log, the method includes: Perform multi-level analysis based on the power learning log combined with the power operation data set to generate multi-level power status; According to the multi-level power status, multiple distribution cabinets are traversed to perform load balancing determination and determine the determination result; Performing power transfer analysis on multiple power distribution cabinets according to the determination results to obtain multiple load distribution results; The power demand forecast information is extracted based on the power learning log, the multiple load distribution results are combined with the power demand forecast information to perform power supply distribution, and the first power dispatching strategy is formulated.

5. A distributed power distribution management method for a power distribution cabinet as claimed in claim 3, characterized in that: The first power dispatching strategy is dynamically adjusted according to the update result to generate a second power dispatching strategy, the method comprising: Based on the power dynamic operation status information of multiple nodes combined with the update result, power forecasting is performed to obtain a power forecasting result, wherein the power forecasting result includes power demand forecasting distribution information and load change trend forecasting information; Traversing the first power dispatching strategy based on the power demand forecast distribution information to generate first information to be adjusted; Traversing the first power dispatching strategy based on the load change trend prediction information to generate second information to be adjusted; Dynamically allocate the power according to the first information to be adjusted and the second information to be adjusted, and generate a power allocation result to be adjusted; The first power dispatching strategy is dynamically adjusted according to the to-be-adjusted power allocation result to generate the second power dispatching strategy.

6. A distributed power distribution management platform for a power distribution cabinet, characterized in that: The platform is used to execute a distributed power distribution management method for a power distribution cabinet according to any one of claims 1 to 5, comprising: Sensor deployment module: Deploy sensors on multiple distribution cabinets based on the sensor network to collect and obtain power operation data sets of multiple distribution cabinets; A graph neural network training module: constructing a first graph neural network, and using the first graph neural network in combination with the power operation data set for training to obtain a second graph neural network; Federated learning module: performing federated learning on multiple distribution cabinets based on the second graph neural network, generating a power learning log, and formulating a first power dispatching strategy according to the power learning log; Real-time feedback response module: executes the first power dispatching strategy to provide real-time feedback response to multiple distribution cabinets to obtain dispatching feedback results; Distributed power distribution management module: update the second graph neural network according to the scheduling feedback result, dynamically adjust the first power scheduling strategy according to the update result, and generate a second power scheduling strategy to perform distributed power distribution management on multiple distribution cabinets.

Images