Electric power information operation violation risk supervision system based on knowledge graph

Abstract

The invention discloses an electric power information operation violation risk supervision system based on a knowledge graph, which relates to the field of violation risk supervision and comprises a dynamic graph construction module, a causal analysis module, a strategy analysis module, a strategy modeling module and a supervision decision module. The method comprises the following steps: obtaining multi-type electric power operation field sensing data and information system data, and carrying out data processing and knowledge graph dynamic construction to obtain a dynamic knowledge graph; performing risk situation quantification and risk decision point inference based on the dynamic knowledge graph to obtain a key causal decision point set; based on the key causal decision point set, through strategy logic analysis, obtaining a logic rule set which can be directly deployed and executed; according to the method, a logic rule set which can be directly deployed and executed is subjected to dynamic strategy evolution of a stochastic differential equation to obtain a dynamic strategy model, and the dynamic strategy model is subjected to decision optimal screening of measurement transformation to obtain an optimal supervision decision set, so that a risk value can be accurately calculated, and the supervision response speed can be increased.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of risk supervision, in particular to a power information operation risk supervision system based on a knowledge graph. BACKGROUND

[0002] The prior art has the following deficiencies in the aspect of power operation risk supervision: in data processing, it can only collect on-site sensing data or information system data separately, and often has problems of data time dislocation and space mismatch. In the risk identification link, the prior art relies on artificial experience judgment, such as observing the equipment state by a patrol personnel, which is difficult to comb the deep correlation among personnel, equipment and environment from a large amount of data, and cannot identify potential relationships and accurately calculate risk values, and can only find surface risks. In the intervention strategy formulation, the prior art mostly follows fixed disposal processes, does not systematically learn effective experience from historical risk events, and lacks accurate analysis of the effectiveness of the strategy, so when encountering a new risk scenario, the staff can only choose intervention measures based on experience, which is difficult to match the optimal solution in time, resulting in missing the best opportunity to block the risk. The prior art has low automation, and many links from data collection, risk analysis to strategy generation need human participation, which not only increases the burden of the staff, but also leads to slow response of supervision and inability to keep pace with the dynamic changes of the operation site in real time.

[0003] In order to solve the above-mentioned defects, a technical solution is provided. SUMMARY

[0004] In order to solve the technical problems proposed in the background, the present application is proposed. The embodiments of the present application provide a power information operation risk supervision system based on a knowledge graph.

[0005] The purpose of the present application can be achieved by the following technical solution: a power information operation risk supervision system based on a knowledge graph, comprising a dynamic graph construction module, a causal analysis module, a strategy analysis module, a strategy modeling module and a supervision decision module, the dynamic graph construction module is used for obtaining multiple types of power operation site sensing data and information system data for data processing and dynamic construction of a knowledge graph, to obtain a dynamic knowledge graph;

[0006] The causal analysis module performs risk situation quantification and risk decision point inference based on the dynamic knowledge graph, to obtain a set of key causal decision points;

[0007] The strategy analysis module performs strategy logic analysis based on the set of key causal decision points, to obtain a set of logic rules that can be directly deployed and executed;

[0008] The strategy modeling module performs dynamic strategy evolution of the set of logic rules that can be directly deployed and executed by using a stochastic differential equation, to obtain a dynamic strategy model;

[0009] The regulatory decision module performs metric transformation on the dynamic strategy model to select the optimal regulatory decision set.

[0010] Furthermore, the dynamic knowledge graph analysis steps are as follows:

[0011] Deep representation learning and node similarity analysis are performed on the initial heterogeneous information graph to obtain updated relation triples. The updated relation triples are then integrated and aligned with the attribute triples to obtain the initial knowledge graph pattern layer.

[0012] Based on the initial knowledge graph pattern layer, incremental updates and entity links are performed on the real-time continuously flowing panoramic fusion data pool to obtain a dynamic knowledge graph.

[0013] Furthermore, the initial heterogeneous information graph analysis steps are as follows:

[0014] Acquire sensor data and information system data from various types of power operation sites, perform data fusion and alignment, and obtain a panoramic fusion data pool;

[0015] Based on the panoramic fusion data pool, entities, their corresponding types, relationships between entities, and entity attributes in power operation information are identified, resulting in original relation triples and attribute triples. Based on the original triples, the extracted entities are used as nodes, and the identified explicit relationships are used as edges to construct an initial heterogeneous information graph.

[0016] Furthermore, the steps for setting the key causal decision points are as follows:

[0017] A causal relationship graph structure between variables is constructed based on the risk pattern feature tensor and dynamic structural causal model. Nodes represent variables, and edges represent the causal direction between variables. The NOTEARS causal discovery algorithm is used to solve and optimize the causal relationship graph structure, outputting an optimized causal graph structure. The in-degree of each node is calculated based on the optimized causal graph structure, and a candidate set of key causal nodes is obtained through threshold determination. Causal effects are quantified through do-calculus, and the average causal effect of the nodes in the candidate set of key causal nodes is calculated. A threshold determination is then performed to obtain the set of key causal decision points.

[0018] Furthermore, the steps of the enhanced dynamic risk knowledge graph are as follows:

[0019] A core abstract risk set is created, and each core abstract risk in the core abstract risk set is targetedly mined using natural language processing technology to obtain specific risk factors and their critical values. Measured values ​​of the specific risk factors are obtained through a panoramic fusion data pool. Based on the critical values ​​of the specific risk factors, the measured values ​​are standardized to obtain standardized risk indices. The standardized risk indices are mapped using membership functions to obtain membership vectors for the specific risk factors. Based on historical violation datasets, conditional probability values ​​for each specific risk factor are calculated using kernel density estimation and normalized to obtain weight vectors for each specific factor. Fuzzy operations are performed on the membership vectors and weight vectors of the specific risk factors to obtain a scalar comprehensive risk value. Through entity parsing and attribute mapping mechanisms, the scalar comprehensive risk value is matched with relation triples and attribute triples in a dynamic knowledge graph to obtain an enhanced dynamic risk knowledge graph.

[0020] Furthermore, the risk pattern feature tensor steps are as follows:

[0021] A fourth-order tensor is constructed based on the enhanced dynamic risk knowledge graph, with dimensions corresponding to time series, spatial region, entity type, and risk type, respectively. The scalar comprehensive risk value of each entity corresponding to each risk type in the enhanced dynamic risk knowledge graph is calculated using a preset aggregation function. The representation value is then filled into the four-dimensional coordinate points of the tensor corresponding to time, space, entity type, and risk type to obtain the risk pattern feature tensor.

[0022] Furthermore, the steps for the directly deployable and executable logical rule set are as follows:

[0023] The global initial parameters of the operation intervention meta-policy library are mapped into deterministic rules through a symbolic regression algorithm. Logical conflict detection and simplification optimization are performed on the deterministic rules to obtain a set of logical rules that can be directly deployed and executed.

[0024] Furthermore, the steps of the operation intervention meta-policy library are as follows:

[0025] Risk event records are extracted from historical databases. The key causal decision point set is calculated using the causal analysis module and used as the support set for the task. The effective specific intervention sequence is used as the query set for the task, resulting in the structural machine learning task of risk events. The structural machine learning task of risk events is then used to construct a meta-policy library through an irrelevant meta-learning framework, resulting in the operational intervention meta-policy library.

[0026] Furthermore, the steps of the dynamic strategy model are as follows:

[0027] A basic trigger strength is set for each rule in the set of directly deployable and executable logical rules. The activation state of the set of directly deployable and executable logical rules at time t is defined as a multivariate stochastic process. A jump-diffusion process model is established for the trigger strength of each rule. Based on the basic trigger strength, the jump-diffusion process model is solved using the Euler-Maruyama method to obtain the trigger strength of the stochastic differential equation with jumps. Based on the trigger strength of the stochastic differential equation with jumps and the multivariate stochastic process, the probability of a rule transitioning from an inactive state to an active state in the time interval is calculated. Monte Carlo simulation and rule statistics are performed on the probability of a rule transitioning from an inactive state to an active state in the time interval to obtain a dynamic policy model.

[0028] Furthermore, the steps for obtaining the optimal regulatory decision set are as follows:

[0029] The strength stochastic process of each strategy in the dynamic strategy model is mathematically defined as a specific Iton process. A mathematical model is established for each stochastic process to obtain a formally defined Iton process. The Gilsanov theorem is applied to the formally defined Iton process to construct the Laden-Nicotim derivative. A new probability measure is defined based on the Laden-Nicotim derivative. The new probability measure is substituted into the formally defined Iton process to obtain the strategy process under the new probability measure. An ideal baseline process is defined under the new probability measure. The relative entropy between the probability distribution of the new probability measure and the distribution of the ideal baseline process is calculated based on the ideal baseline process and the strategy process under the new probability measure. A strategy list sorted by relative entropy value is obtained. The optimal strategy is selected from the strategy list by a threshold judgment to obtain the optimal regulatory decision set.

[0030] Compared with the prior art, the beneficial effects of the present invention are:

[0031] 1. This invention acquires various types of field sensor data and information system data for power operations, processes the data, and dynamically constructs a knowledge graph to obtain a dynamic knowledge graph. Based on the dynamic knowledge graph, it quantifies risk situations and infers risk decision points to obtain a set of key causal decision points. Based on this set of key causal decision points, it analyzes strategic logic to obtain a set of directly deployable and executable logical rules. It then performs dynamic strategy evolution using stochastic differential equations on this set of directly deployable and executable logical rules to obtain a dynamic strategy model. This approach avoids solely collecting field sensor data or information system data, reducing issues of data temporal misalignment and spatial mismatch. In the risk identification phase, it can extract deep connections between personnel, equipment, and the environment from a large amount of data, identify potential relationships, and accurately calculate risk values, thereby discovering deep-seated risks. In intervention strategy formulation, it can systematically learn effective experiences from historical risk events and accurately analyze the effectiveness of strategies. When encountering new risk scenarios, it can promptly match the optimal solution, ensuring that the best opportunity to mitigate risks is not missed.

[0032] 2. This invention obtains the optimal regulatory decision set by performing metric transformation on the dynamic strategy model and selecting the optimal decision. From data collection and risk analysis to strategy generation, it can not only reduce the burden on staff, but also improve the speed of regulatory response and synchronize the dynamic changes of the work site in real time. Attached Figure Description

[0033] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. The following drawings are not drawn to scale according to the actual size, but are intended to show the main idea of ​​the present invention.

[0034] Figure 1 This is a system block diagram of the present invention;

[0035] Figure 2 This is a flowchart of the causal analysis module method of the present invention;

[0036] Figure 3 This is a flowchart of the regulatory decision-making module method of the present invention. Detailed Implementation

[0037] The technical solutions in 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 also within the scope of protection of the present invention.

[0038] like Figure 1 As shown, the knowledge graph-based power information operation violation risk monitoring system includes a dynamic graph construction module, a causal analysis module, a strategy analysis module, a strategy modeling module, and a regulatory decision-making module.

[0039] The dynamic knowledge graph construction module is used to acquire various types of power operation site sensor data and information system data, perform data processing and dynamic knowledge graph construction, and obtain a dynamic knowledge graph.

[0040] The causal analysis module uses a dynamic knowledge graph to quantify risk status and infer risk decision points, thereby obtaining a set of key causal decision points.

[0041] The strategy analysis module uses strategy logic analysis based on the set of key causal decision points to obtain a set of logical rules that can be directly deployed and executed.

[0042] The strategy modeling module performs dynamic strategy evolution using stochastic differential equations on a set of logical rules that can be directly deployed and executed, resulting in a dynamic strategy model.

[0043] The regulatory decision module performs metric transformation on the dynamic strategy model to select the optimal regulatory decision set.

[0044] In this embodiment, the dynamic map construction module is analyzed in detail as follows:

[0045] Acquire sensor data and information system data from various types of power operation sites, perform data fusion and alignment, and obtain a panoramic fusion data pool;

[0046] Based on the panoramic fusion data pool, entities, their corresponding types, relationships, and attributes in power operation information are identified, resulting in original relation triples and attribute triples. Based on the original triples, an initial heterogeneous information graph is constructed using the extracted entities as nodes and the identified explicit relationships as edges. Deep representation learning and node similarity analysis are performed on the initial heterogeneous information graph to obtain updated relation triples. The updated relation triples are then integrated and aligned with the attribute triples to obtain the initial knowledge graph pattern layer.

[0047] Based on the initial knowledge graph pattern layer, incremental updates and entity links are performed on the real-time continuously flowing panoramic fusion data pool to obtain a dynamic knowledge graph.

[0048] In this embodiment, multiple types of field sensor data for power operations are acquired. Specifically, spatial coordinate data is obtained through a UWB positioning terminal; behavior recognition and status monitoring data are obtained through a video analysis unit; equipment status sensors collect electrical and mechanical parameters in real time; and environmental monitoring units monitor conditions such as temperature and humidity to obtain and acquire information system data. Specifically, this includes work tickets and operation tickets (two tickets) used to standardize power operation procedures and clarify safety responsibilities, as well as power production management system data covering equipment management, operation management, and other aspects. A spatiotemporal stamp service based on a precision clock protocol or network time protocol is used to mark all multi-source data with unified and traceable microsecond-level timestamps and spatial coordinate labels, achieving synchronization and unification of multi-source data in the spatiotemporal dimension. Based on this, a streaming data fusion algorithm is applied. This algorithm performs semantic parsing, windowing, conflict detection, feature extraction, and confidence-weighted fusion of real-time multi-source data, dynamically eliminating contradictions and redundancies between original data to obtain a panoramic fusion data pool. Based on the panoramic fusion data pool, the semantic understanding capabilities of a pre-trained language model (PLM) are used to perform deep parsing of unstructured text. This process is completed synchronously through a joint extraction framework: the model first extracts entities and then extracts relations or attributes. In a single traversal, it identifies entities in the text in parallel, determines their types, identifies relations between entities, and extracts entity attributes, resulting in original relation triples (entity-relation-entity) and attribute triples (entity-attribute-attribute value). Based on the original triples, using the extracted entities as nodes and the identified explicit relations as edges, an initial heterogeneous information graph is constructed. The message passing and neighborhood aggregation mechanisms of Graph Neural Networks (GNNs) are used to perform deep representation learning on this initially constructed graph. GNNs provide high-quality vectorized representations of known entities and relations, capturing deep structural and semantic information in the graph, thus serving downstream tasks. For example, by calculating the similarity between node representations, GNNs can infer potential relations that are not explicitly mentioned in the data. Entities A and B are related to entity C. By calculating cosine similarity, if it exceeds a set threshold, it is determined that they belong to the same working group, thereby discovering new relation triples to complete and expand the knowledge graph. The updated relation triples and attribute triples are integrated and aligned to obtain the initial knowledge graph pattern layer.Based on the constructed initial knowledge graph pattern layer, an incremental graph computation framework is introduced to process the continuously flowing panoramic fusion data stream. This framework listens for newly arriving data through dynamic time window division and triggering mechanisms, and calls the entity linking service. It uses semantic similarity calculation and context disambiguation to solve the matching and ambiguity problems between entity references extracted from new data and existing entity objects in the knowledge graph, thus enabling entity alignment and fusion. For successfully linked entities, the framework uses a real-time graph update interface to efficiently calculate and insert new nodes, edges, and attributes in an incremental manner, or refresh the attribute states of existing graph elements, thereby avoiding the overhead of recalculating everything. This ensures that the topology and attribute states of the knowledge graph can synchronously reflect the latest dynamic changes in the work site, resulting in a dynamic knowledge graph.

[0049] like Figure 2 As shown, in this embodiment, the specific analysis of the causal analysis module is as follows:

[0050] A core abstract risk set is created, and each core abstract risk in the core abstract risk set is targetedly mined using natural language processing technology to obtain specific risk factors and their critical values. Measured values ​​of the specific risk factors are obtained through a panoramic fusion data pool. Based on the critical values ​​of the specific risk factors, the measured values ​​are standardized to obtain standardized risk indices. The standardized risk indices are mapped using membership functions to obtain membership vectors for the specific risk factors. Based on historical violation datasets, conditional probability values ​​for each specific risk factor are calculated using kernel density estimation and normalized to obtain weight vectors for each specific factor. Fuzzy operations are performed on the membership vectors and weight vectors of the specific risk factors to obtain a scalar comprehensive risk value. Through entity parsing and attribute mapping mechanisms, the scalar comprehensive risk value is matched with relation triples and attribute triples in a dynamic knowledge graph to obtain an enhanced dynamic risk knowledge graph.

[0051] In this embodiment, a core abstract risk set is created, including violation probability, equipment failure rate, environmental hazard level, safety distance risk, authorization risk, and cross-operation risk. To quantify each core risk, the system processes text specifications, such as safety regulations and operating procedures, using Natural Language Processing (NLP) technology. Specifically, for each core risk, the system performs targeted mining in the text to obtain specific risk factors and their critical values. Measured values ​​of these risk factors are obtained through a panoramic fusion data pool. Different standardization functions are used; for positive indicators (higher values ​​indicate greater danger, such as temperature, humidity, and concentration), different standardization functions are applied. The standardized risk index is calculated using a half-rising function: Risk Index = max(0, (Current Value - Warning Threshold) / (Critical Value - Warning Threshold). The warning threshold is a safety boundary below the critical value, specifically taken as 95% of the critical value. For negative indicators (the smaller the value, the more dangerous, such as insulation resistance, safety distance, and voltage stability), the standardized risk index is calculated using a half-falling function: Risk Index = max(0, (Warning Threshold - Current Value) / (Warning Threshold - Critical Value)). The warning threshold is a safety boundary above the critical value, specifically taken as 105% of the critical value, thus obtaining the specific factor. The standardized risk index is calculated. For example, for environmental hazard, all clauses related to environmental conditions are identified, such as the working environment temperature not exceeding 40℃, with a temperature threshold of 40℃. The standardized risk index for each specific factor is mapped to the membership degrees of low, medium, and high risk levels through a semi-trapezoidal membership function, forming a membership vector for the specific factor [low: 0, medium: 0.1, high: 0.9] (0, 0.1, and 0.9 are membership degrees). Based on historical violation datasets, the conditional probability of a violation event occurring under different values ​​is calculated for each specific factor using kernel density estimation, thus obtaining the conditional probability of each specific factor. After obtaining the probability values, the system normalizes them using the softmax function, converting them into weight coefficients to obtain the weight vectors of each factor. Then, the membership vectors of each factor and the weight vectors are subjected to fuzzy operations, and a scalar comprehensive risk value within the range of 0-1 is synthesized using a weighted average algorithm. The system uses entity parsing and attribute mapping mechanisms to map these scalar comprehensive risk values ​​to elements in the dynamic risk knowledge graph. Specifically, the system matches the specific factors used in calculating the risk value with the relation triples and attribute triples in the dynamic knowledge graph to obtain an enhanced dynamic risk knowledge graph of node and edge scalar comprehensive risk values.

[0052] A fourth-order tensor is constructed based on the enhanced dynamic risk knowledge graph, with dimensions corresponding to time series, spatial region, entity type and risk type, respectively. The scalar comprehensive risk value of each entity corresponding to each risk type in the enhanced dynamic risk knowledge graph is calculated using a preset aggregation function. The representation value is then filled into the four-dimensional coordinate points of time, space, entity type and risk type in the tensor to obtain the risk pattern feature tensor.

[0053] A causal relationship graph structure between variables is constructed based on the risk pattern feature tensor and dynamic structural causal model. Nodes represent variables, and edges represent the causal direction between variables. The NOTEARS causal discovery algorithm is used to solve and optimize the causal relationship graph structure, outputting an optimized causal graph structure. The in-degree of each node is calculated based on the optimized causal graph structure, and a candidate set of key causal nodes is obtained through threshold determination. Causal effects are quantified through do-calculus, and the average causal effect of the nodes in the candidate set of key causal nodes is calculated. A threshold determination is then performed to obtain the set of key causal decision points.

[0054] In this embodiment, a fourth-order tensor is constructed using an enhanced dynamic risk knowledge graph. The dimensions correspond to time series, spatial region, entity type, and risk type, respectively. The time series and spatial region are obtained from the enhanced dynamic risk knowledge graph through time window slicing and spatial grid aggregation operations. The time series is generated by segmenting continuously flowing real-time data according to defined time intervals (e.g., every minute). Spatial grid aggregation continuously records the real-time coordinates (x, y, z) of tagged personnel, tools, and equipment, storing them as attributes on the corresponding entity nodes of the enhanced dynamic risk knowledge graph. The physical space of the entire work site is digitized and gridded, for example, divided into 1m×1m units. The risk type represents the core abstract risk set. The scalar comprehensive risk value corresponding to each entity and risk type in the enhanced dynamic risk knowledge graph is calculated using a preset aggregation function (e.g., taking the maximum value). Finally, this value is filled into the four-dimensional coordinate points of the tensor corresponding to time, space, entity type, and risk type. The entire structured tensor is constructed by traversing all possible dimension combinations. Based on the risk pattern feature tensor, a graph structure reflecting the causal relationships between personnel, equipment, and environmental variables at the work site is constructed using a dynamic structural causal model (SCM). Nodes represent variables (specifically, operating steps, equipment status, and environmental parameters), and edges represent the direction of causal interaction between variables. A causal discovery algorithm based on NOTEARS is used to identify the causal graph structure from the data by solving the following optimization problem:

[0055] min W L(W;X)+λ||W||1

[0056]

[0057] Here, W is the weighted adjacency matrix, used to characterize the causal relationship graph structure between personnel, equipment, and environmental variables, including the weights of causal interactions between nodes (variables). It is the core parameter to be obtained through optimization. L(W; X) represents the loss function, where W is the parameter to be optimized, and X is the risk pattern feature tensor. The loss function measures the degree of fit between the causal graph constructed based on the current W and the actual data (carried by X). It is constructed by combining data distribution, prediction error, etc., to make the model fit the real causal relationship as closely as possible. λ is the regularization coefficient, ||·||1 is the L1 regularization term, and h(W) is the acyclic constraint function. It is a matrix exponentiation operation. The matrix is ​​multiplied element-wise, tr represents the trace of the matrix (the sum of the elements on the main diagonal), and d is the number of variables (i.e., the dimension of matrix W). Based on the optimized identification of the causal graph structure, the in-degree of each node is calculated; the in-degree is the number of edges pointing to that node. All nodes with an in-degree less than a set threshold are marked as key causal node candidate sets. Do-calculus is used to quantify causal effects, and the values ​​of nodes X in the key causal node candidate set are calculated. i For the average causal effect of outcome variable Y, the set of key causal decision points is selected when the absolute value of ACE exceeds a preset threshold. This set identifies core operational links or equipment states that can effectively interrupt the risk transmission chain once intervention is implemented. The specific formula used is... ACE is the average causal effect, the final required quantity. It quantitatively represents the magnitude of the causal influence of the cause variable on the outcome variable Y. This indicates a causal intervention operation, which forcibly sets nodes in the candidate set of key causal nodes to specific values. or The simulation of the impact of intervention on risk, with the outcome variable Y being the scalar comprehensive risk value calculated by a fuzzy algorithm at the current moment, is directly read from the attributes of the corresponding entity or relation in the enhanced dynamic risk knowledge graph. To express the mathematical expectation, force X to be... i Set as When, what is the average result of Y?

[0058] In this embodiment, the strategy analysis module performs the following specific analysis:

[0059] Risk event records are extracted from historical databases. The key causal decision point set is calculated using the causal analysis module and used as the support set for the task. The effective specific intervention sequence is used as the query set for the task, resulting in the structural machine learning task of risk events. The structural machine learning task of risk events is then used to construct a meta-policy library through an irrelevant meta-learning framework, resulting in an operational intervention meta-policy library.

[0060] The global initial parameters of the operation intervention meta-policy library are mapped into deterministic rules through a symbolic regression algorithm. Logical conflict detection and simplification optimization are performed on the deterministic rules to obtain a set of logical rules that can be directly deployed and executed.

[0061] In this embodiment, each complete risk event record is extracted from the historical database. This record contains full-cycle data from the occurrence of the risk to the end of its handling. The set of key causal decision points ultimately identified by the causal analysis module in this event is used as the support set for this task, representing the core features of the risk scenario. The sequence of specific intervention measures that have been verified as effective in the event is used as the query set for this task, i.e., the target output to be learned. Each historical risk event is transformed into a structural machine learning task T. j =(S j Q j) Among them, support set S j For input features, query set Q j To provide computationally achievable training samples for the desired output, a meta-policy library is constructed using a non-relevant meta-learning framework. This framework comprises two core processes: an inner loop and an outer loop. The inner loop handles a structured machine learning task T for each risk event. j To support the S set j Using this as input, the parameters of the base policy network fθ are further updated with gradients in one or more steps to obtain the task-specific adaptation parameters θ. j The outer loop calculates the adaptation parameter θ across all tasks. j In their respective query sets Q j The sum of losses is used to optimize the initial parameters θ of the base policy network through backpropagation, enabling it to quickly adapt to new tasks with a small number of gradient steps. After multiple iterations of optimization, the global initial parameters θ* are obtained, forming the operation intervention meta-policy library. The symbolic regression algorithm is as follows: The global initial parameters of the operation intervention meta-policy library are used to generate sample pairs of input decision points and output intervention actions through forward propagation. Using genetic programming as the core framework, a function set containing mathematical operators, comparison operators, logical operators, and decision point variables is initialized. The sample pairs generated by forward propagation are used as training data. With the optimization objectives of maximizing fitting accuracy and minimizing expression complexity, the population is iteratively evolved through genetic operations such as selection, crossover, and mutation to obtain a mathematical analytical expression. This analytical expression is converted into deterministic rules, such as if <conditional expression> then <execution action>. Through a rule engine, a set of logical rules that can be directly deployed and executed is obtained. Specifically, the rule engine is based on the RETE algorithm to build a rule network, detect the inclusion, overlap, and contradiction relationships between different rule preconditions, identify and mark rule pairs with logical conflicts, perform static analysis on the rule set, and merge redundant rules with the same execution action and overlapping preconditions.

[0062] In this embodiment, the strategy modeling module is analyzed in detail as follows:

[0063] A basic trigger strength is set for each rule in the set of logical rules that can be directly deployed and executed. The activation state of the set of logical rules that can be directly deployed and executed at time t is defined as a multivariate stochastic process. A jump-diffusion process model is established for the trigger strength of each rule. Based on the basic trigger strength, the jump-diffusion process model is solved using the Euler-Maruyama method to obtain the trigger strength of the stochastic differential equation with jumps. Based on the trigger strength of the stochastic differential equation with jumps and the multivariate stochastic process, the probability of a rule transitioning from an inactive state to an active state in the time interval is calculated. Monte Carlo simulation and rule statistics are performed on the probability of a rule transitioning from an inactive state to an active state in the time interval to obtain a dynamic policy model.

[0064] In this embodiment, the set of logical rules that can be directly deployed and executed contains N rules, each rule having R... n Set base trigger strength Specifically, the current time rule R n The degree to which the real-time risk conditions are met directly derives from the scalar comprehensive risk value of the corresponding entity in the enhanced dynamic risk knowledge graph. The activation state of the set of directly deployable and executable logical rules at time t is defined as a multivariate stochastic process S(t): S(t) = (S1(t), S2(t), ..., S... N (t)) T S n (t)∈{0,1} is a random variable, representing the rule R. n At time t, whether the rule is activated (1 for activation, 0 for inactivation), T is the vector transpose, and the trigger strength λ of each rule is... n (t) is not constant, but changes dynamically with environmental uncertainties. Its dynamics are described by the following jump-diffusion process model:

[0065] It is the mean recovery term, α n It is the response rate. It is the long-term average intensity, ensuring that intensity fluctuations revolve around a steady-state value, σ. n dW n (t) is the diffusion term. It simulates continuous, small-amplitude fluctuations in intensity. σ n It's volatility, W n (t) is a standard Wiener process. γ n dJ(t) is the jump term, used to simulate the violent impact of a sudden event. J(t) is a Poisson process with parameter β. dJ(t) indicates whether a jump occurs within a small time interval (1 if it occurs, 0 otherwise). n It refers to the jump range. Rule Rn The probability of transitioning from an inactive state to an active state within the time interval [t, t+dt) is determined by its transient triggering strength λ. n (t) determines: P(S) n (t+dt))=1|S n (t)=0)=λ n (t)dt+o(dt), where P(*) represents probability, S n (t+dt))=1|S n (t) = 0 represents the conditional probability. It represents the probability that the rule will be activated within the next small time interval [t, t+dt] given that it is not activated at the current time (t). o(dt) represents a higher-order infinitesimal. It represents a term that approaches 0 faster than dt when dt approaches 0, and can be ignored in modeling. A numerical method based on the Euler-Maruyama method for solving stochastic differential equations with jumps is used to discretize and iteratively solve the above equations:

[0066] Where Z is a random variable that follows a standard normal distribution, i.e., Z ~ N(0,1). ΔJ is a scaling factor for a standard normal random variable, and ΔJ is a random variable following a Poisson distribution, i.e., ΔJ ~ Poisson(βΔt), where β is the jump intensity, representing the average number of jump events per unit time. The increment dJ(t) used to approximate the Poisson process takes non-negative integer values. Through numerous Monte Carlo simulations, the activation probability distribution P(S) of each rule at any future time can be obtained. n (t))=1, the set of activation probability distributions of all rules at all future times, yields the dynamic policy model, which is the policy effectiveness probability distribution obtained after countless simulations and includes time dimension and uncertainty information.

[0067] like Figure 3 As shown, the specific analysis of the regulatory decision-making module in this embodiment is as follows:

[0068] The strength stochastic process of each strategy in the dynamic strategy model is mathematically defined as a specific Iton process. A mathematical model is established for each stochastic process to obtain a formally defined Iton process. The Gilsanov theorem is applied to the formally defined Iton process to construct the Laden-Nicotim derivative. A new probability measure is defined based on the Laden-Nicotim derivative. The new probability measure is substituted into the formally defined Iton process to obtain the strategy process under the new probability measure. An ideal baseline process is defined under the new probability measure. The relative entropy between the probability distribution of the new probability measure and the distribution of the ideal baseline process is calculated based on the ideal baseline process and the strategy process under the new probability measure. A strategy list sorted by relative entropy value is obtained. The optimal strategy is selected from the strategy list by a threshold judgment to obtain the optimal regulatory decision set.

[0069] In this embodiment, the strength stochastic process of each policy n in the dynamic policy model is mathematically defined as a specific Iton process: dλ n (t)=u n (t)+σ n dW n (t), where the drift term and diffusion term σ n All are directly derived from the jump-diffusion process model. An accurate mathematical model is established for each stochastic process, resulting in the formally defined Iton process {λ1(t),λ2(t),...,λ...}. N} is a mathematical description of policy behavior under the original probability measure P. Applying the Girsanov theorem to the formally defined Itō process, we construct a Laden-Nicodim derivative H(t), which is a stochastic process dH(t) = -θ(t)H(t)dW(t), where θ(t) represents the Girsanov kernel. This eliminates the common random noise source faced by all policy processes. u m (t), σ m (t) represents the drift rate and volatility of the average risk faced by all strategies, which are obtained by solving for... s is a temporary placeholder for the integral variable. A new probability measure Q is defined through H(t). For any event A that occurs in the future time T, its probability under the Q measure is defined as: Q(A) = E P [H(T)·Π A ], where E P [*] indicates the expectation under the original measure P, Π A Let Q be the indicator function of event A. According to Girsanov's theorem, under this new measure Q, the new Brownian motion W... Q (t) is defined as Substituting the original specific Itoh process dλ n (t)=u n (t)+σ n dW n (t) can be used to obtain the policy process under measure Q. Under the measure Q, an ideal reference process λ is selected. ndeal (t), a deterministic process with a constant positive drift rate and zero diffusivity, λ ndeal (t)=u ndeal dt, where u ndeal To maintain a constant positive drift rate, a small constant of 0.01 is used, representing a stable, low-risk, and predictable ideal growth pattern for each strategy process. Calculate the relative entropy between its probability distribution and the ideal distribution. KLDn It is the divergence value of the nth strategy, relative entropy. Represents the Radon-Nikodym derivative, a random variable. This represents the policy process of policy n under measure Q. The probability distribution dp of all behaviors within a specific time range. ndeal This represents the probability distribution corresponding to the ideal baseline process under the same measure Q. A list of strategies sorted by relative entropy value is obtained. A screening threshold is set (such as selecting the top 10% of strategies). The optimal strategy is selected from the top of the list to obtain the optimal regulatory decision set. This set contains the strategies that are proven to be the most robust and effective in a probabilistic sense and can be directly executed by the system.

[0070] Specifically, it can integrate on-site sensor data and information system data, eliminate data contradictions through spatiotemporal synchronization, and form a panoramic data pool. It leverages knowledge graphs to mine entity relationships and dynamically update risk information, combines causal analysis to deconstruct abstract risks and locate key issues, and extracts effective experiences from historical events to generate and select optimal intervention strategies. Most processes are completed collaboratively, achieving early risk detection and accurate handling, reducing manual workload, improving regulatory efficiency, and effectively ensuring the safety of power operations.

[0071] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0072] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0073] The foregoing description is illustrative of the invention and should not be construed as limiting it. Although several exemplary embodiments of the invention have been described, those skilled in the art will readily understand that many modifications can be made to the exemplary embodiments without departing from the novel teachings and advantages of the invention. Therefore, all such modifications are intended to be included within the scope of the invention as defined in the claims. It should be understood that the foregoing description is illustrative of the invention and should not be construed as limiting it to the specific embodiments disclosed, and modifications to the disclosed embodiments and other embodiments are intended to be included within the scope of the appended claims. The invention is defined by the claims and their equivalents.

Claims

1. A knowledge graph-based power information operation violation risk monitoring system, comprising a dynamic graph construction module, a causal analysis module, a strategy analysis module, a strategy modeling module, and a regulatory decision-making module, characterized in that: The dynamic knowledge graph construction module is used to acquire various types of power operation site sensor data and information system data, perform data processing and dynamic knowledge graph construction, and obtain a dynamic knowledge graph. The causal analysis module uses a dynamic knowledge graph to quantify risk status and infer risk decision points, thereby obtaining a set of key causal decision points. The strategy analysis module uses strategy logic analysis based on the set of key causal decision points to obtain a set of logical rules that can be directly deployed and executed. The strategy modeling module performs dynamic strategy evolution using stochastic differential equations on a set of logical rules that can be directly deployed and executed, resulting in a dynamic strategy model. The regulatory decision module performs metric transformation on the dynamic strategy model to select the optimal regulatory decision set.

2. The knowledge graph-based power information operation violation risk monitoring system according to claim 1, characterized in that, The dynamic knowledge graph analysis steps are as follows: Deep representation learning and node similarity analysis are performed on the initial heterogeneous information graph to obtain updated relation triples. The updated relation triples are then integrated and aligned with the attribute triples to obtain the initial knowledge graph pattern layer. Based on the initial knowledge graph pattern layer, incremental updates and entity links are performed on the real-time continuously flowing panoramic fusion data pool to obtain a dynamic knowledge graph.

3. The knowledge graph-based power information operation violation risk monitoring system according to claim 2, characterized in that, The initial heterogeneous information graph analysis steps are as follows: Acquire sensor data and information system data from various types of power operation sites, perform data fusion and alignment, and obtain a panoramic fusion data pool; Based on the panoramic fusion data pool, entities, their corresponding types, relationships between entities, and entity attributes in power operation information are identified, resulting in original relation triples and attribute triples. An initial heterogeneous information graph is constructed based on the original triples, with the extracted entities as nodes and the identified explicit relationships as edges.

4. The knowledge graph-based power information operation violation risk monitoring system according to claim 1, characterized in that, The steps for setting up the key causal decision points are as follows: A causal relationship graph structure between variables is constructed based on the risk pattern feature tensor and dynamic structural causal model. Nodes represent variables, and edges represent the causal direction between variables. The NOTEARS causal discovery algorithm is used to solve and optimize the causal relationship graph structure, outputting an optimized causal graph structure. The in-degree of each node is calculated based on the optimized causal graph structure, and a candidate set of key causal nodes is obtained through threshold determination. Causal effects are quantified through do-calculus, and the average causal effect of the nodes in the candidate set of key causal nodes is calculated. A threshold determination is then performed to obtain the set of key causal decision points.

5. The knowledge graph-based power information operation violation risk monitoring system according to claim 4, characterized in that, The process of obtaining the risk pattern feature tensor is as follows: A fourth-order tensor is constructed based on the enhanced dynamic risk knowledge graph, with dimensions corresponding to time series, spatial region, entity type, and risk type, respectively. The scalar comprehensive risk value of each entity corresponding to each risk type in the enhanced dynamic risk knowledge graph is calculated using a preset aggregation function. The representation value is then filled into the four-dimensional coordinate points of the tensor corresponding to time, space, entity type, and risk type to obtain the risk pattern feature tensor.

6. The knowledge graph-based power information operation violation risk monitoring system according to claim 5, characterized in that, The steps of the enhanced dynamic risk knowledge graph are as follows: A core abstract risk set is created. For each core abstract risk in this set, natural language processing (NLP) is used for targeted mining to obtain specific risk factors and their critical values. Measured values ​​of these specific risk factors are obtained from a panoramic fusion data pool. Based on the critical values, the measured values ​​are standardized to obtain standardized risk indices. Membership vectors of these standardized risk indices are mapped using membership functions. Based on historical violation datasets, the conditional probability values ​​of each specific risk factor are calculated using kernel density estimation and normalized to obtain weight vectors. Fuzzy operations are performed on the membership vectors and weight vectors to obtain a scalar comprehensive risk value. Through entity parsing and attribute mapping mechanisms, the scalar comprehensive risk value is matched with relation triples and attribute triples in a dynamic knowledge graph to obtain an enhanced dynamic risk knowledge graph.

7. The knowledge graph-based power information operation violation risk monitoring system according to claim 1, characterized in that, The steps for the directly deployable and executable logical rule set are as follows: The global initial parameters of the operation intervention meta-policy library are mapped into deterministic rules through a symbolic regression algorithm. Logical conflict detection and simplification optimization are performed on the deterministic rules to obtain a set of logical rules that can be directly deployed and executed.

8. The knowledge graph-based power information operation violation risk monitoring system according to claim 7, characterized in that, The steps for obtaining the operational intervention meta-strategy are as follows: Risk event records are extracted from historical databases. The key causal decision point set is calculated using the causal analysis module and used as the support set for the task. The effective specific intervention sequence is used as the query set for the task, resulting in the structural machine learning task of risk events. The structural machine learning task of risk events is then used to construct a meta-policy library through an irrelevant meta-learning framework, resulting in the operational intervention meta-policy library.

9. The knowledge graph-based power information operation violation risk monitoring system according to claim 1, characterized in that, The steps of the dynamic strategy model are as follows: A basic trigger strength is set for each rule in the set of directly deployable and executable logical rules. The activation state of the set of directly deployable and executable logical rules at time t is defined as a multivariate stochastic process. A jump-diffusion process model is established for the trigger strength of each rule. Based on the basic trigger strength, the jump-diffusion process model is solved using the Euler-Maruyama method to obtain the trigger strength of the stochastic differential equation with jumps. Based on the trigger strength of the stochastic differential equation with jumps and the multivariate stochastic process, the probability of a rule transitioning from an inactive state to an active state in the time interval is calculated. Monte Carlo simulation and rule statistics are performed on the probability of a rule transitioning from an inactive state to an active state in the time interval to obtain a dynamic policy model.

10. The knowledge graph-based power information operation violation risk monitoring system according to claim 1, characterized in that, The steps for the optimal regulatory decision set are as follows: The strength stochastic process of each strategy in the dynamic strategy model is mathematically defined as a specific Iton process. A mathematical model is established for each stochastic process to obtain a formally defined Iton process. The Gilsanov theorem is applied to the formally defined Iton process to construct the Laden-Nicotim derivative. A new probability measure is defined based on the Laden-Nicotim derivative. The new probability measure is substituted into the formally defined Iton process to obtain the strategy process under the new probability measure. An ideal baseline process is defined under the new probability measure. The relative entropy between the probability distribution of the new probability measure and the distribution of the ideal baseline process is calculated based on the ideal baseline process and the strategy process under the new probability measure. A strategy list sorted by relative entropy value is obtained. The optimal strategy is selected from the strategy list by a threshold judgment to obtain the optimal regulatory decision set.

Images