Visual display method and system for power system based on knowledge graph

Abstract

The invention discloses a power system visualization display method and system based on a knowledge graph, and particularly relates to the technical field of power system intelligent visualization, and the method comprises the steps: constructing a power system knowledge graph, and carrying out the initialization configuration; user interaction behaviors are recorded and quantified, and the dynamic interestingness and the common-view relation are calculated; mining a user operation time sequence mode by using an LSTM network, and constructing and updating a user interest model; according to the interest model, dynamically adjusting a force-oriented layout parameter and a visual channel priority, and realizing adaptive visual rendering; and a reinforcement learning closed loop is established, and a visualization strategy is continuously optimized through a DQN algorithm. According to the method, user behaviors are perceived and the visualization effect is adaptively adjusted, so that the expression definition and the interaction efficiency of complex power grid data are effectively improved, and the cognitive load of the user is reduced.

Description

Technical Field

[0001] This invention relates to the field of intelligent visualization technology for power systems, and more specifically, to a method and system for visualizing power systems based on knowledge graphs. Background Technology

[0002] As a critical national infrastructure, the power system is complex in structure, contains numerous devices, and generates massive amounts of operational data. With the advancement of smart grid construction, the multi-source heterogeneous data generated in the power system, including device status, topology connections, and operating parameters, is growing exponentially. Traditional display methods such as two-dimensional tables and static wiring diagrams are insufficient to effectively express the complex relationships between devices and the system's operating status. This leads to information overload for dispatching and operation personnel when faced with massive amounts of data, affecting their accurate judgment of the system status and rapid decision-making.

[0003] Current knowledge graph-based power system visualization methods mainly employ static layout and fixed visual encoding strategies. While these methods can display the power grid topology to some extent, they have significant shortcomings: First, the use of a uniform force-directed layout algorithm ignores the inherent characteristics of power equipment in terms of voltage level, electrical distance, etc., resulting in unclear locations of key equipment and an unclear hierarchical structure. Second, the visualization strategies lack adaptability and cannot be dynamically adjusted according to differences in user roles and changes in analysis needs; users with different professional backgrounds need to perform a lot of manual operations to obtain effective information. Third, they lack the ability to perceive and learn from user behavior, failing to create a personalized interactive experience.

[0004] Most existing visualization systems remain at the data display level, lacking intelligent analysis capabilities. These systems cannot understand user intent or predict user needs based on historical interactions, resulting in repetitive zooming, filtering, and searching operations with each use. Furthermore, they lack self-optimization mechanisms, failing to continuously improve visualization effects based on feedback during use. These problems severely restrict the application of power system visualization in actual production operations, urgently requiring an intelligent visualization solution capable of adaptive optimization and learning. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method and system for visualizing power systems based on knowledge graphs.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A knowledge graph-based method for visualizing power systems includes the following steps:

[0008] S1: Knowledge Graph Construction and Initialization. Construct a knowledge graph of the power system and establish an entity relationship network of devices, connections, and states. Calculate the overall importance of devices and introduce centripetal force and hierarchical force optimization layout algorithms to generate an initial visualization view that conforms to the power grid hierarchy and the distribution of key devices.

[0009] S2: Behavior recording and quantification, recording user view operations and node interaction events; calculating dynamic interest degree through time decay model and operation weight, and mining co-view relationships between nodes based on co-occurrence frequency;

[0010] S3: Interest Modeling and Update. The user session is divided into analysis units, and the temporal patterns of the operation sequence are mined using an LSTM network. User interest vectors are generated through self-supervised learning to quantify the strength of user preferences for various entities, attributes, and relationships.

[0011] S4: Adaptive visualization rendering, dynamically adjusts force-oriented layout parameters according to the interest model, applies centripetal force to core nodes, and aggregates edge nodes; allocates visual channels based on interest priority;

[0012] S5: Evaluate and optimize the closed loop, establish a reinforcement learning framework to evaluate the visualization effect, and quantify the interaction efficiency through the reward function; the DQN algorithm dynamically adjusts the layout and visual strategy to continuously optimize the user experience.

[0013] Specifically, in S1:

[0014] The overall importance of the equipment is calculated by combining the topological importance and operational importance of the equipment. The topological importance is obtained by calculating the electrical betweenness centrality of the node in the power transmission path of the power grid, and the operational importance is calculated based on the product of the logarithm of the equipment's rated capacity and the current load rate.

[0015] The intensity of centripetal force and hierarchical force can be adjusted by setting the centripetal force coefficient and hierarchical force coefficient, respectively.

[0016] Specifically, in S1:

[0017] The layout algorithm introduces centripetal force and hierarchical force on the basis of the classic force-directed algorithm; the centripetal force acts on the core devices whose overall importance is higher than a set threshold and gathers towards the center of the canvas.

[0018] The hierarchical force sets the target hierarchy according to the voltage level of the equipment, so that the equipment is arranged in layers perpendicular to the hierarchical direction to form a clear display of the power grid transmission and distribution structure.

[0019] Specifically, in S2:

[0020] Recorded user interaction events include view operations, node or edge operations, query and filter operations, and time navigation operations;

[0021] Operation weight allocation includes setting different weight values ​​for click operations, long and short hover operations, box selection operations, and attribute filtering operations;

[0022] The co-occurrence relationship is established by statistically analyzing the co-occurrence frequency of nodes within the same time window and using an association rule algorithm to mine frequent itemsets.

[0023] Specifically, in S2:

[0024] The dynamic interest calculation uses an exponential decay function to weight historical interaction events over time, and combines operation type weights and duration factors for comprehensive calculation.

[0025] For hovering operations, the duration factor is the actual hovering duration; for other instantaneous operations, the duration factor is a fixed value.

[0026] Specifically, in S3:

[0027] LSTM networks abstract user sessions into feature sequences of interaction event types, target entity types, and target attribute types;

[0028] The network is trained using self-supervised learning to predict the target type of the next event from the previous interaction event.

[0029] During training, the network weights are updated by minimizing the prediction error, so that the hidden layer states of the network can encode user interest features.

[0030] Specifically, in S4:

[0031] The dynamic adjustment of the layout parameters includes:

[0032] Apply a centripetal force pointing towards the layout center to nodes whose user interest exceeds the first threshold;

[0033] Clusters of nodes whose user interest is below the second threshold are aggregated to form super nodes;

[0034] To enhance the spring attraction coefficient for node pairs with strong co-visual relationships, thereby shortening the visual distance.

[0035] Specifically, in S4:

[0036] The visual channel allocation assigns the attributes that users care about most to the visual channels that are prominent in terms of node size or color saturation, and downgrades the attributes that users do not care about to the standard to the secondary visual channels of shape or label.

[0037] At the same time, the visual channel priority of each attribute is dynamically adjusted according to changes in the user's interest vector.

[0038] Specifically, in S5:

[0039] The state space of the reinforcement learning framework includes the cluster identifier of the current user interest vector and the currently activated visualization policy ID;

[0040] The action space includes adjusting the centripetal force coefficient, adjusting the aggregation threshold, and switching the primary visual encoding channel;

[0041] The reward function is calculated based on a combination of task completion efficiency, content recommendation adoption rate, and negative user feedback, and a deep Q-network algorithm is used to learn the optimal strategy.

[0042] A knowledge graph-based power system visualization system includes the following modules:

[0043] The data acquisition and quantification module is used to capture user interaction behavior and calculate interaction popularity scores and co-visibility relationships;

[0044] The user interest modeling module is used to run LSTM machine learning models and complete the calculation and updating of user interest vectors.

[0045] An adaptive visualization engine is used to dynamically adjust force-guided layout parameters and visual encoding schemes based on interest vectors;

[0046] The reinforcement learning optimization module is used to run the DQN algorithm and manage the experience replay pool and policy network that interact with the environment.

[0047] The personalized configuration management module is used to store and load user visual profiles;

[0048] The modules interact and collaborate with each other through API interfaces or message queues.

[0049] The technical effects and advantages of this invention are as follows:

[0050] By optimizing the hierarchical layout based on power system characteristics and employing adaptive visualization rendering, the system effectively enhances the clarity and readability of complex power grid data. It automatically highlights key equipment and optimizes visual hierarchy, significantly reducing the cognitive load on users and enabling dispatchers to quickly grasp the system's operational status and key issues, thereby improving analytical and decision-making efficiency.

[0051] By employing a continuous optimization mechanism that combines user behavior awareness with reinforcement learning, the system possesses self-evolution capabilities. Through learning and adapting to users' operating habits and preferences, the system can dynamically adjust its visualization strategies, creating a highly personalized analytical environment. This not only enhances the interactive experience but also achieves continuous performance improvement over long-term use. Attached Figure Description

[0052] Figure 1 This is a flowchart of the method of the present invention;

[0053] Figure 2 This is a system block diagram of the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] like Figure 1 As shown, the steps of the knowledge graph-based power system visualization method are as follows:

[0056] Step 1: Knowledge Graph Construction and Initialization. A knowledge graph of the power system is constructed, establishing a network of entity relationships among devices, connections, and states. By calculating the overall importance of devices and introducing centripetal and hierarchical force optimization algorithms, an initial visualization view conforming to the power grid hierarchy and the distribution of key equipment is generated. This view serves as the baseline starting point for learning.

[0057] First, a standard power system knowledge graph is constructed, which includes entities, relationships, and attributes such as devices, connections, and states.

[0058] Subsequently, an initial, general visual configuration is provided, including:

[0059] Global layout based on force-oriented algorithm; in this process, a hierarchical layout optimization strategy based on power network topology characteristics and equipment importance is implemented, the specific process is as follows:

[0060] Device Comprehensive Importance Measurement: First, calculate a comprehensive importance score for each device node in the knowledge graph. It is derived by weighting topological importance and operational importance, and the calculation formula is as follows:

[0061]

[0062] in: Topological importance is obtained by calculating the electrical betweenness centrality of a node, reflecting the criticality of the device in the power transmission path of the power grid; For operational importance, based on the equipment's rated capacity With current load rate calculate: To demonstrate the impact of the equipment in its current operating state; and These are the weighting coefficients, and their sum is 1;

[0063] Layout force model optimization: Based on the repulsive and attractive forces in the classic force-directed algorithm, two custom forces are introduced:

[0064] Centripetal force Overall Importance Higher than the set threshold Core equipment (such as hub substations and main power plants) applies a centripetal force pointing towards the center of the canvas. :

[0065]

[0066] in: The centripetal force coefficient, The coordinates of the center of the canvas are: For nodes The current location; this force ensures that key equipment is prominently displayed in the visual center area.

[0067] Hierarchical force Predefine the target level based on the device's voltage level. (For example, 500kV corresponds to level 0, 220kV to level 1, and 110kV to level 2). Apply a force perpendicular to the voltage level direction to each node. :

[0068]

[0069] in: For hierarchical force coefficients, For node i, The current level of node i. This is the unit normal vector pointing in the direction of the target level. This force causes devices of different voltage levels to be naturally layered in the vertical or horizontal direction, clearly showing the power grid's transmission and distribution structure.

[0070] Initial layout generation: Using merged and The optimization force model is iteratively calculated until the layout stabilizes. The final generated initial layout not only avoids node overlap, but also intuitively presents the backbone-branch hierarchical structure of the power grid and the core position of key equipment.

[0071] Basic visual coding schemes (e.g., generators are green circles, loads are blue rectangles).

[0072] The default information panel displays the core attributes of the nodes.

[0073] This initial view serves as the starting point and benchmark for system learning.

[0074] Step Two: Behavior Recording and Quantification. This involves recording user view operations, node interactions, and other behavioral events throughout the process. Dynamic interest levels are calculated using a time decay model and operation weights, and co-occurrence frequencies are used to mine co-view relationships between nodes. The raw behavioral data is then transformed into quantified user interest features; the specific process is as follows:

[0075] Behavioral event capture: Continuously and silently recording all user interaction events in the background. These events include, but are not limited to:

[0076] View manipulation: zoom, pan, view switching trajectory and frequency; Node / edge manipulation: click, hover, selected targets and their attributes; Query and filter: executed search keywords, applied attribute filters; Time navigation: timeline operation range and playback speed.

[0077] Behavior Quantification and Association: The captured raw events are transformed into quantified feature vectors and associated with elements in the knowledge graph. An interaction heat score is calculated for each graph element (node / edge) based on the frequency and duration of its clicks and hovers.

[0078] Simultaneously, a dynamic interest calculation model based on time decay and operational weights is introduced, the specific process of which is as follows:

[0079] Assign different interest weights to different types of interactive operations. ,include:

[0080] Click operation 1.0; hovering time greater than three seconds 0.8; hovering time less than or equal to three seconds 0.4; Box selection operation 0.6; Attribute filtering It is 0.9;

[0081] An exponential decay function is used to simulate the natural decay process of user interest. For an interaction event occurring at time t, its decay weight at the current time T is: ,in The attenuation coefficient controls the rate at which interest decays. The time interval from the time the event occurred to the present.

[0082] For each graph element i, its dynamic interest degree Calculated using the following formula:

[0083]

[0084] Where: summation iterates through all interaction events involving element i; Interest weights for interaction types; This is the decay weight for the interaction event; This is a duration factor, equal to the hover duration (seconds) for hover operations, and 1 for other instantaneous operations;

[0085] The dynamic interest scores of all elements are max-min normalized to fall within the [0,1] interval. This interest score model is updated incrementally on a daily basis to ensure that it reflects the user's latest interest preferences while preserving long-term stable interest patterns.

[0086] Co-occurrence relationships are calculated, which nodes are frequently observed simultaneously within the same screen view or the same interactive session. Specifically, a configurable time window (default 30 seconds) is used as a statistical unit. If two or more nodes appear simultaneously within the visible area within this time window and are not moved out, it is counted as a co-occurrence event. The co-occurrence frequency of all nodes throughout the session is recorded to form a co-occurrence matrix. The Apriori algorithm is then used to mine frequent itemsets from this matrix that satisfy the minimum support (default 5 times). Nodes in this set are defined as having strong co-occurrence relationships.

[0087] Record the observation scale of user preferences (e.g., always zooming in to the substation level for analysis) and navigation paths.

[0088] Step 3: Interest Modeling and Update. User sessions are divided into analysis units, and an LSTM network is used to mine temporal patterns in operation sequences. User interest vectors are generated through self-supervised learning, quantifying the strength of user preferences for various entities, attributes, and relationships. The model is continuously and incrementally updated to track interest evolution; the specific process is as follows:

[0089] Session segmentation: Divide a user's continuous sequence of operations into different analysis sessions. For example, the period from logging into the system to starting fault analysis until logging out or having no operation for a long time is considered as one session.

[0090] Pattern Mining: Within each session and across sessions, machine learning models (such as Hidden Markov Models (HMMs) or Long Short-Term Memory Networks (LSTMs) are used to analyze temporal patterns of user interaction behavior. Specifically, each user session is abstracted as a feature sequence consisting of interaction event type, target entity type, and target attribute type. This sequence is fed into a two-layer LSTM network, and the output of the last layer is mapped to a fixed-length user interest vector through a fully connected layer. The dimensions of this vector are predefined, corresponding to different entity types, attribute types, and relation types in the knowledge graph. The value of each dimension of the vector represents the intensity of the user's predicted interest in that type of element. The model is trained using a self-supervised learning approach, taking the first k interaction events as input to predict the target type of the (k+1)th event. The network weights are updated by minimizing the prediction error, enabling the hidden layer states to encode the user's interest preferences.

[0091] Identifying and analyzing patterns: For example, a typical pattern for User A might be clicking on a substation in the global map -> zooming in -> filtering to display all circuit breakers -> viewing their status history. This is identified as a pattern for equipment status inspection.

[0092] Identify the focus of interest: The model outputs a user interest vector that quantifies the strength of the user's preference for various entities (such as "500kV line" and "transformer"), attributes (such as "load rate" and "oil temperature"), and relationships (such as "connection relationship" and "protection relationship").

[0093] Incremental Model Updates: The user interest model is not a static file, but is incrementally updated with each new interaction session, thus tracking the evolution of user interests.

[0094] Step 4: Adaptive visualization rendering. Dynamically adjust force-guided layout parameters based on the interest model, apply centripetal force to core nodes, and aggregate edge nodes. Allocate visual channels based on interest priority to highlight key attributes. Intelligently predict user intent and proactively push relevant device information; the specific process is as follows:

[0095] Layout optimization: Based on the learned user interest model, the parameters in the force-directed algorithm are dynamically adjusted. Specifically, based on the classic force-directed algorithm, an additional force based on user interest is introduced. Let the user interest score of node i be . Its calculation formula is ,in and These are the weighting coefficients for interaction popularity and entity type interest.

[0096] for Above the threshold The node is additionally subjected to a centripetal force pointing towards the center C of the canvas. ,in The centripetal force coefficient, The current coordinates of the node, where C is the coordinate of the canvas center.

[0097] for Below the threshold If the node community is fragmented, it is aggregated into a supernode, the position of which is determined by the centroid of the original community nodes. For node pairs (i,j) with strong co-visibility, the elastic coefficient of the spring force between them is determined. Increase by a factor In order to shorten their visual distance.

[0098] Apply an attraction force to the center of the screen to nodes with high user interaction activity.

[0099] Apply stronger attraction or shorten the default distance between frequently co-existing nodes.

[0100] For clusters of nodes that receive very little attention from users, apply a repulsive force to the edge region or directly aggregate them into a super node.

[0101] Visual enhancement: Dynamically adjust the priority of visual encoding based on the user's interest vector.

[0102] The attributes that users care about most (such as "load rate") will be given the most prominent visual channels (such as node size or color saturation).

[0103] Attributes that users pay less attention to are relegated to secondary visual channels (such as shape or label).

[0104] Content preloading and push notifications: When a user is detected to have started a familiar analysis pattern, intelligent predictions are made. For example, when a user clicks on a transformer, based on historical patterns, information closely related to that transformer, such as "cooler status" and "oil chromatography data," which the user frequently views, is preloaded and automatically displayed in the sidebar, instead of a generic attribute list.

[0105] Step 5: Evaluate and optimize the closed loop, establish a reinforcement learning framework to evaluate the visualization effect, and quantify the interaction efficiency through a reward function. The DQN algorithm dynamically adjusts the layout and visual strategy to continuously optimize the user experience. A personalized visualization configuration profile is created and maintained long-term.

[0106] Performance Evaluation: The system monitors user interaction efficiency after adaptive optimization. Positive feedback includes: shorter user dwell time on the optimized view, fewer clicks to complete specific tasks (such as locating faulty equipment), and a higher adoption rate of system-recommended content.

[0107] Policy adjustment: Feed these positive / negative feedback signals into a reinforcement learning model.

[0108] Specifically, a reinforcement learning (RL) framework was constructed; its state space S is defined as the cluster identifier of the current user interest vector and the ID of the currently activated visualization policy. Its action space A includes: adjusting the centripetal force coefficient μ and adjusting the aggregation threshold. This includes switching the primary visual encoding channel, etc. The reward function R is defined as: R = g1 * (1 - average number of task clicks in this session / historical baseline) + g2 * (content recommendation adoption rate) - g3 * (number of times the user manually resets the layout), where g1, g2, and g3 are the weight coefficients for task efficiency, recommendation adoption rate, and negative feedback, respectively. The DeepQ-Network (DQN) algorithm is used to learn a policy π: S->A, which can select the action that maximizes the long-term cumulative reward based on the current state, i.e., continuously optimize the visual adaptive strategy. This model evaluates the effectiveness of the current visual adaptive strategies (such as layout parameter adjustment rules and visual encoding mapping rules) and fine-tunes these strategies to pursue the highest user interaction efficiency and user satisfaction in the long run.

[0109] Personalized Configuration: After multiple iterations, the system will create distinctly different, highly customized visual configuration profiles for different users or user roles. Upon login, the user's personalized visual environment will be loaded immediately.

[0110] A knowledge graph-based power system visualization system, comprising:

[0111] The data acquisition and quantification module is responsible for capturing all user interactions on the visualization interface in real time, including view operations, node clicks, attribute filtering, and other events. This module performs feature extraction and quantification on the raw behavioral data, calculating key indicators such as node interaction popularity and co-view relationships. As the system's perception layer, it provides a structured user behavior data foundation for subsequent analysis.

[0112] The user interest modeling module employs an LSTM neural network to perform deep temporal pattern mining on user behavior sequences, constructing a dynamically updated user interest model. It generates multi-dimensional interest vectors through self-supervised learning, accurately quantifying users' attention levels to various devices, attributes, and relationships. This module is the intelligent core for achieving personalized visualization, continuously tracking the evolution of user interests.

[0113] The adaptive visualization engine dynamically adjusts the visualization rendering strategy based on interest models, and integrates an improved force-guided layout algorithm to achieve spatial distribution optimization based on interest. It intelligently manages visual channel resources, assigning appropriate encoding methods to key attributes to ensure the clarity and effectiveness of information presentation. It also possesses content prediction capabilities, proactively pushing related information to improve interaction efficiency.

[0114] The reinforcement learning optimization module establishes a complete evaluation-optimization closed loop by constructing a reinforcement learning framework through defining a state space, action set, and reward function. It employs the DQN algorithm to analyze user feedback data and automatically adjusts layout parameters and visual strategies to enhance the interactive experience. This module enables the system to continuously improve itself and gradually optimize visualization effects.

[0115] The personalized configuration management module is responsible for storing and maintaining each user's unique visual configuration profile, including layout preferences, visual coding schemes, content push strategies, and other settings. It supports configuration isolation and rapid switching between multi-user environments, ensuring that each user can immediately access their personalized analysis environment upon login. This module is a key support for achieving long-term personalized services within the system.

[0116] The above formulas are all dimensionless calculations. Dimensionless calculations can be performed using various methods such as standardization, which will not be elaborated here. The formulas are derived from software simulations based on a large amount of collected data, and the preset parameters in the formulas can be set by those skilled in the art according to the actual situation.

[0117] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, ATA hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state ATA hard disk.

[0118] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0119] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0120] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0121] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0122] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0123] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable ATA hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0124] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A knowledge graph-based method for visualizing power systems, characterized in that, Includes the following steps: S1: Knowledge Graph Construction and Initialization. Construct a knowledge graph of the power system and establish an entity relationship network of devices, connections, and states. Calculate the overall importance of devices and introduce centripetal force and hierarchical force optimization layout algorithms to generate an initial visualization view that conforms to the power grid hierarchy and the distribution of key devices. S2: Behavior recording and quantification, recording user view operations and node interaction events; calculating dynamic interest degree through time decay model and operation weight, and mining co-view relationships between nodes based on co-occurrence frequency; S3: Interest Modeling and Update. The user session is divided into analysis units, and the temporal patterns of the operation sequence are mined using an LSTM network. User interest vectors are generated through self-supervised learning to quantify the strength of user preferences for various entities, attributes, and relationships. S4: Adaptive visualization rendering, dynamically adjusts force-oriented layout parameters according to the interest model, applies centripetal force to core nodes, and aggregates edge nodes; allocates visual channels based on interest priority; S5: Evaluate and optimize the closed loop, establish a reinforcement learning framework to evaluate the visualization effect, and quantify the interaction efficiency through the reward function; the DQN algorithm dynamically adjusts the layout and visual strategy to continuously optimize the user experience.

2. The knowledge graph-based power system visualization method according to claim 1, characterized in that, In S1: The overall importance of the equipment is calculated by combining the topological importance and operational importance of the equipment. The topological importance is obtained by calculating the electrical betweenness centrality of the node in the power transmission path of the power grid, and the operational importance is calculated based on the product of the logarithm of the equipment's rated capacity and the current load rate. The intensity of centripetal force and hierarchical force can be adjusted by setting the centripetal force coefficient and hierarchical force coefficient, respectively.

3. The knowledge graph-based power system visualization method according to claim 1, characterized in that, In S1: The layout algorithm introduces centripetal force and hierarchical force on the basis of the classic force-directed algorithm; the centripetal force acts on the core devices whose overall importance is higher than a set threshold and gathers towards the center of the canvas. The hierarchical force sets the target hierarchy according to the voltage level of the equipment, so that the equipment is arranged in layers perpendicular to the hierarchical direction to form a clear display of the power grid transmission and distribution structure.

4. The knowledge graph-based power system visualization method according to claim 1, characterized in that, In S2: Recorded user interaction events include view operations, node or edge operations, query and filter operations, and time navigation operations; Operation weight allocation includes setting different weight values ​​for click operations, long and short hover operations, box selection operations, and attribute filtering operations; The co-occurrence relationship is established by statistically analyzing the co-occurrence frequency of nodes within the same time window and using an association rule algorithm to mine frequent itemsets.

5. The knowledge graph-based power system visualization method according to claim 1, characterized in that, In S2: The dynamic interest calculation uses an exponential decay function to weight historical interaction events over time, and combines operation type weights and duration factors for comprehensive calculation. For hovering operations, the duration factor is the actual hovering duration; for other instantaneous operations, the duration factor is a fixed value.

6. The knowledge graph-based power system visualization method according to claim 1, characterized in that, In S3: LSTM networks abstract user sessions into feature sequences of interaction event types, target entity types, and target attribute types; The network is trained using self-supervised learning to predict the target type of the next event from the previous interaction event. During training, the network weights are updated by minimizing the prediction error, so that the hidden layer states of the network can encode user interest features.

7. The knowledge graph-based power system visualization method according to claim 1, characterized in that, In S4: The dynamic adjustment of the layout parameters includes: Apply a centripetal force pointing towards the layout center to nodes whose user interest exceeds the first threshold; Clusters of nodes whose user interest is below the second threshold are aggregated to form super nodes; To enhance the spring attraction coefficient for node pairs with strong co-visual relationships, thereby shortening the visual distance.

8. The knowledge graph-based power system visualization method and system according to claim 1, characterized in that, In S4: The visual channel allocation assigns the attributes that users care about most to the visual channels that are prominent in terms of node size or color saturation, and downgrades the attributes that users do not care about to the standard to the secondary visual channels of shape or label. At the same time, the visual channel priority of each attribute is dynamically adjusted according to changes in the user's interest vector.

9. The knowledge graph-based power system visualization method according to claim 1, characterized in that, In S5: The state space of the reinforcement learning framework includes the cluster identifier of the current user interest vector and the currently activated visualization policy ID; The action space includes adjusting the centripetal force coefficient, adjusting the aggregation threshold, and switching the primary visual encoding channel; The reward function is calculated based on a combination of task completion efficiency, content recommendation adoption rate, and negative user feedback, and the optimal strategy is learned using a deep Q-network algorithm.

10. A system applied to the knowledge graph-based power system visualization method according to any one of claims 1-9, characterized in that, Includes the following modules: The data acquisition and quantification module is used to capture user interaction behavior and calculate interaction popularity scores and co-visibility relationships; The user interest modeling module is used to run LSTM machine learning models and complete the calculation and updating of user interest vectors. An adaptive visualization engine is used to dynamically adjust force-guided layout parameters and visual encoding schemes based on interest vectors; The reinforcement learning optimization module is used to run the DQN algorithm and manage the experience replay pool and policy network that interact with the environment. The personalized configuration management module is used to store and load user visual profiles; The modules interact and collaborate with each other through API interfaces or message queues.

Images