Graph topology constraint and multi-modal Bayesian fusion anomaly detection method

Abstract

The invention relates to the technical field of carbon emission anomaly detection, and provides a graph topology constraint and multi-modal Bayesian fusion anomaly detection method, which comprises the steps of collecting electric quantity and carbon factors, and constructing a hierarchical tree graph; based on a time delay propagation matrix, embedding three types of constraints including instantaneous conservation, sliding window accumulation and quota accumulation into layered Lagrangian; calculating a nonlinear weight according to the quota and the time residue and stabilizing the nonlinear weight; a graph neural state space model is trained through graph topology and robust weight, and cross-time coupling loss is constructed; in the prediction stage, multi-target projection of constraint conflict perception is used for digestion according to the priority of carbon emission reduction, and correction is carried out through Pareto and monotonicity projection; and carrying out change point extraction, distribution drift and tail dependency evidence on the residual error, and carrying out Bayesian fusion to obtain an abnormal probability and judgment. According to the invention, the accuracy and stability of carbon emission anomaly detection of the multi-level ammeter network are improved.

Description

Technical Field

[0001] This invention relates to the field of carbon emission anomaly detection technology, and in particular to an anomaly detection method that combines graph topological constraints and multimodal Bayesian fusion. Background Technology

[0002] With the advancement of global carbon neutrality goals and increasingly stringent carbon emission management requirements, carbon emission monitoring of multi-level meter networks has become a key technological means for energy management and environmental protection. A multi-level meter network is a power metering system organized in a hierarchical tree structure, widely used in carbon emission accounting and management in complex electricity consumption scenarios such as smart grids, industrial parks, and commercial complexes. During long-term operation, multi-level meter networks are susceptible to various problems such as abnormal carbon emission data, constraint violations, and energy consumption imbalances due to factors such as load fluctuations, equipment aging, and environmental changes. Graph topology-based carbon emission anomaly detection technology serves as a primary means of monitoring multi-level meter networks, identifying and assessing abnormal states by analyzing the network topology and carbon emission time-series characteristics. However, the unique hierarchical tree structure, multi-timescale constraints, and complex carbon emission propagation mechanisms of multi-level meter networks present numerous challenges to anomaly detection and analysis.

[0003] In existing technologies, carbon emission anomaly detection in multi-level electricity meter networks mainly employs traditional time-series analysis and single-modal detection methods to achieve basic anomaly identification. However, existing methods do not adequately consider the inherent physical correlation mechanism between the complex topology of multi-level electricity meter networks and carbon emission constraint characteristics. This makes it difficult to organically integrate objective carbon emission conservation constraints with actual network topology characteristics, resulting in the inability to achieve high-precision automatic identification of carbon emission anomalies. Summary of the Invention

[0004] In view of this, the present invention proposes an anomaly detection method that combines graph topology constraints and multimodal Bayesian fusion. This method solves the problem that existing methods do not adequately consider the inherent physical correlation mechanism between the complex topology of multi-level electricity meter networks and carbon emission constraint characteristics, making it difficult to organically integrate objective carbon emission conservation constraints with actual network topology characteristics, thus failing to achieve high-precision automatic identification of carbon emission anomalies.

[0005] The technical solution of this invention is implemented as follows: This invention provides an anomaly detection method that fuses graph topological constraints and multimodal Bayesian methods, comprising the following steps: Obtain electricity data and carbon emission factor data from a multi-level electricity meter network, construct a hierarchical tree topology, and output the graph topology. Based on the graph topology, carbon emission time delay accumulation constraints are established. The carbon emission time delay propagation matrix is ​​constructed to describe the cross-time coupling relationship. Instantaneous conservation constraints, sliding window accumulation constraints, and quota accumulation constraints are embedded into the hierarchical Lagrange constraint framework to output carbon emission constraint conditions. Based on the carbon emission constraints, a nonlinear weight stabilization mechanism sensitive to quota surplus is adopted. The original constraint weight coefficients are calculated based on the quota surplus ratio and the time surplus ratio. The weight oscillations are eliminated by a stabilization filter, and the stable constraint weight coefficients are output. The graph topology, electricity data, and carbon emission factor data are input into the graph neural network state space model. The stability constraint weight coefficients are combined to construct a cross-time coupling loss function for model training, and the graph neural network model is output. Carbon emission prediction is performed using the graph neural network model. Constraint conflicts are detected by a multi-objective projection operator that senses constraint conflicts, and conflicts are resolved according to carbon emission reduction priorities. The prediction results are corrected by multi-objective Pareto projection and carbon monotonicity-preserving projection, and the predicted carbon emission values ​​are output. Based on the residuals between the predicted and actual carbon emissions, multimodal anomaly evidence from change point detection, distribution drift detection, and tail dependency detection is integrated, and a Bayesian fusion method is used to calculate the comprehensive anomaly probability, outputting the carbon emission anomaly detection result.

[0006] Based on the above technical solutions, preferably, the step of establishing carbon emission time-delay accumulation constraints based on the graph topology, describing the cross-time coupling relationship by constructing a carbon emission time-delay propagation matrix, and embedding instantaneous conservation constraints, sliding window accumulation constraints, and quota accumulation constraints into a hierarchical Lagrangian constraint framework to output carbon emission constraint conditions, including: Based on the node hierarchy and historical carbon emission data in the graph topology, the carbon emission time delay response characteristics between nodes at different levels are analyzed, and a carbon emission time delay propagation matrix is ​​constructed according to the propagation law of carbon emissions in the graph topology. Based on the carbon emission delay propagation matrix and graph topology, three types of basic constraints are established, including instantaneous conservation constraints, sliding window accumulation constraints, and quota accumulation constraints. The instantaneous conservation constraints are used to describe the carbon emission conservation relationship between parent and child nodes, the sliding window accumulation constraints are used to describe the carbon emission accumulation balance within a time window, and the quota accumulation constraints are used to describe the carbon emission quota consumption progress of nodes. The three types of basic constraints and the carbon emission time delay propagation matrix are embedded into a hierarchical Lagrange constraint framework. The constraint relationship is established across time using the Lagrange multiplier method, forming a unified multi-timescale constraint expression and outputting the carbon emission constraint conditions.

[0007] Based on the above technical solutions, preferably, the construction of the carbon emission time delay propagation matrix includes: Statistical analysis of historical carbon emission data of each node in the graph topology is used to calculate the time delay characteristics and propagation intensity of carbon emission changes between nodes at different levels. Based on the differences in node hierarchy depth and physical connection distance, calculate the time delay parameters and attenuation coefficients for carbon emissions propagating from upper-level nodes to lower-level nodes; A carbon emission delay propagation matrix is ​​constructed based on the delay parameter and the attenuation coefficient to describe the carbon emission delay coupling relationship between nodes. The matrix elements of the carbon emission delay propagation matrix are used to represent the carbon emission influence weights between different nodes at different times.

[0008] Based on the above technical solutions, preferably, the step of employing a quota surplus-sensitive nonlinear weight stabilization mechanism according to the carbon emission constraints, calculating the original constraint weight coefficients based on the quota surplus ratio and the time surplus ratio, and eliminating weight oscillations through a stabilization filter to output stable constraint weight coefficients includes: Based on the quota constraint information in the carbon emission constraints, the remaining quota ratio and time ratio at the current moment are calculated. A nonlinear weight calculation method sensitive to quota remaining is adopted. According to the combination state of the remaining quota ratio and time ratio, the weight adjustment strategy for different constraint types is determined. The instantaneous conservation constraint weight coefficient, sliding window cumulative constraint weight coefficient and quota cumulative constraint weight coefficient are calculated respectively to obtain the original constraint weight coefficient. The temporal fluctuation characteristics and oscillation modes of the original constraint weight coefficients are detected. A stabilization filter designed for carbon emission quota sensitivity is used to filter the original constraint weight coefficients. The filter kernel function eliminates weight oscillations and maintains the quota-sensitive response characteristics of the weights, and outputs stable constraint weight coefficients.

[0009] Based on the above technical solutions, preferably, the step of using a stabilization filter designed for carbon emission quota sensitivity to filter the original constraint weight coefficients includes: Analyze the time series data of the original constraint weight coefficients to identify the weight jump points and oscillation intervals caused by changes in quota surplus; A filter kernel function for carbon emission weight stabilization is constructed, which combines the monotonicity of carbon emissions with the sensitivity of quota constraints. The original constraint weight coefficients are convolved using the filtering kernel function to eliminate high-frequency oscillation noise while maintaining the weights' sensitivity to changes in the remaining quota state, and output stable constraint weight coefficients after stabilization.

[0010] Based on the above technical solutions, preferably, the calculation formula for the original constraint weight coefficient is as follows: ; in, For the first The original constraint weight coefficients of the class constraint; For the first The basic weight adjustment factor for class constraints; This represents the remaining percentage of the quota. The remaining time percentage; and The first Sensitivity index of class constraints to quota surplus and time surplus; For the first Nonlinear combination exponents with class constraints; For the first Quota time coupling adjustment parameters for class constraints.

[0011] Based on the above technical solutions, preferably, the step of acquiring electricity data and carbon emission factor data of a multi-level electricity meter network, constructing a hierarchical tree topology, and outputting the graph topology includes: Acquire the electricity data and carbon emission factor data of each meter node in the multi-level meter network for the corresponding time period. Based on the physical connection relationship of the multi-level meter network, calculate the node hierarchy depth and node importance weight, establish a node attribute table and output it. The node attribute table includes hierarchy depth information and importance weight. Based on the node attribute table and the physical connection relationship of the multi-level meter network, a hierarchical tree graph topology is established using a bottom-up hierarchical construction method. The carbon emission quota of each node is allocated according to the historical carbon emission data and hierarchical depth. The node hierarchical relationship, importance weight and carbon emission quota information are integrated into the graph topology, and the graph topology is output.

[0012] Based on the above technical solutions, preferably, the step of inputting the graph topology, electricity data, and carbon emission factor data into the graph neural network state space model, combining the stability constraint weight coefficients, constructing a cross-time coupling loss function for model training, and outputting a graph neural network model includes: Based on the stability constraint weight coefficients, a multi-timescale loss term is constructed. The multi-timescale loss term is obtained by fusing instantaneous conservation constraint loss, sliding window cumulative constraint loss and quota cumulative constraint loss. The loss correlation between different time steps is established through the time coupling weight matrix to obtain the cross-time coupling loss function. The graph topology, electricity data, and carbon emission factor data are input into the graph neural network state space model for forward propagation calculation. The network parameters of the graph neural network state space model are updated by gradient backpropagation using the cross-temporal coupling loss function. The forward and backpropagation processes are repeated until the loss function converges, and the trained graph neural network model is output.

[0013] Based on the above technical solutions, preferably, the step of using the graph neural network model for carbon emission prediction involves detecting constraint conflicts through a multi-objective projection operator that senses constraint conflicts, resolving conflicts according to carbon emission reduction priorities, and correcting the prediction results using multi-objective Pareto projection and carbon monotonicity-preserving projection to output predicted carbon emission values, including: The graph neural network model is used to perform forward inference calculations on the current electricity data and carbon emission factor data to obtain the initial carbon emission prediction value. The conflict between the initial carbon emission prediction value and the carbon emission constraint conditions is detected by the multi-objective projection operator that senses the constraint conflict. The detected constraint conflicts are prioritized and resolved according to the carbon emission reduction priority strategy, and the constraint conflict resolution result is output. Based on the results of the constraint conflict resolution, the initial carbon emission prediction value is corrected for the first time using the multi-objective Pareto projection algorithm to obtain the Pareto optimal prediction value. The Pareto optimal prediction value after Pareto correction is then corrected for the second time using the carbon monotonicity-preserving projection algorithm to ensure that the prediction value conforms to the monotonicity constraint characteristics of carbon emissions, thus obtaining the carbon emission prediction value.

[0014] Based on the above technical solutions, preferably, the step of fusing multimodal anomaly evidence—based on the residual between the predicted and actual carbon emissions, incorporating change point detection, distribution drift detection, and tail dependency detection—and using a Bayesian fusion method to calculate the comprehensive anomaly probability, outputting the carbon emission anomaly detection result, includes: The predicted residual sequence between the predicted carbon emissions and the actual observed values ​​is calculated. Based on the predicted residual sequence, change point detection analysis, distribution drift detection analysis and tail dependency detection analysis are performed respectively. Anomaly indicators and confidence scores corresponding to each detection method are extracted to generate a multimodal anomaly evidence set, which includes change point anomaly evidence, distribution drift anomaly evidence and tail dependency anomaly evidence. Based on the confidence scores and anomaly indicators of various types of anomaly evidence in the multimodal anomaly evidence set, a joint probability model of multimodal evidence is constructed using a Bayesian fusion algorithm. The comprehensive anomaly probability of the three anomaly detection modes is calculated through Bayesian inference. The carbon emission anomaly status is determined according to the preset anomaly probability threshold, and the carbon emission anomaly detection result is output. The carbon emission anomaly detection result includes anomaly probability value and anomaly determination result.

[0015] The anomaly detection method fused from graph topological constraints and multimodal Bayesian methods of the present invention has the following advantages over existing technologies: (1) By using a graph topology constraint framework and multimodal Bayesian anomaly detection, the carbon emission delay propagation matrix is ​​used to describe the cross-time coupling relationship and embed a hierarchical Lagrangian constraint framework. The nonlinear weight stabilization mechanism sensitive to quota surplus is combined to eliminate weight oscillation. The graph neural state space model is used to construct the cross-time coupling loss function for training, which improves the accuracy and stability of carbon emission anomaly detection in multi-level electricity meter networks. At the same time, through a complete constraint satisfaction and conflict resolution mechanism, the engineering application requirements of carbon emission reduction priority management and real-time monitoring are met. (2) Through time delay propagation analysis and hierarchical Lagrange constraint modeling, the time delay response characteristics analysis and propagation intensity calculation are performed using historical carbon emission data. The time delay parameters and attenuation coefficients are dynamically determined by combining the differences in node hierarchy depth. The three types of basic constraints are embedded in a unified framework based on the cross-time coupling relationship, which improves the accuracy of constraint expression and the scientific nature of time coupling modeling. (3) By integrating quota surplus sensitivity analysis and stabilization filtering technology, the nonlinear weight calculation method is used to analyze the quota surplus ratio and time surplus ratio and calculate the three types of constraint weight coefficients. The filter kernel function parameters are dynamically adjusted in combination with the carbon emission monotonicity characteristics. The stabilization filter is specially designed based on the weight jump point and oscillation interval identification results, thereby improving the quota sensitivity response capability of the constraint weight coefficients. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of an anomaly detection method that combines graph topological constraints and multimodal Bayesian fusion according to the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1This invention provides an anomaly detection method that combines graph topological constraints and multimodal Bayesian methods, comprising the following steps: Obtain electricity data and carbon emission factor data from a multi-level electricity meter network, construct a hierarchical tree topology, and output the graph topology. Based on the graph topology, carbon emission time delay accumulation constraints are established. The carbon emission time delay propagation matrix is ​​constructed to describe the cross-time coupling relationship. Instantaneous conservation constraints, sliding window accumulation constraints, and quota accumulation constraints are embedded into the hierarchical Lagrange constraint framework to output carbon emission constraint conditions. Based on the carbon emission constraints, a nonlinear weight stabilization mechanism sensitive to quota surplus is adopted. The original constraint weight coefficients are calculated based on the quota surplus ratio and the time surplus ratio. The weight oscillations are eliminated by a stabilization filter, and the stable constraint weight coefficients are output. The graph topology, electricity data, and carbon emission factor data are input into the graph neural network state space model. The stability constraint weight coefficients are combined to construct a cross-time coupling loss function for model training, and the graph neural network model is output. Carbon emission prediction is performed using the graph neural network model. Constraint conflicts are detected by a multi-objective projection operator that senses constraint conflicts, and conflicts are resolved according to carbon emission reduction priorities. The prediction results are corrected by multi-objective Pareto projection and carbon monotonicity-preserving projection, and the predicted carbon emission values ​​are output. Based on the residuals between the predicted and actual carbon emissions, multimodal anomaly evidence from change point detection, distribution drift detection, and tail dependency detection is integrated, and a Bayesian fusion method is used to calculate the comprehensive anomaly probability, outputting the carbon emission anomaly detection result.

[0020] Specifically, this embodiment utilizes a graph topology constraint framework and multimodal Bayesian anomaly detection. It describes the cross-temporal coupling relationship using the carbon emission delay propagation matrix and embeds a hierarchical Lagrangian constraint framework. A nonlinear weight stabilization mechanism sensitive to quota surplus is combined to eliminate weight oscillations, and a graph neural state-space model is used to construct a cross-temporal coupling loss function for training. By employing a constraint conflict-aware multi-objective projection operator and a multimodal Bayesian fusion method, it solves the problem of differentiable realization of carbon emission conservation constraints at multiple time scales and the problem of multimodal anomaly evidence fusion, improving the accuracy and stability of carbon emission anomaly detection in multi-level meter networks. Furthermore, through a complete constraint satisfaction and conflict resolution mechanism, it meets the engineering application requirements of carbon emission reduction priority management and real-time monitoring.

[0021] The process of acquiring electricity data and carbon emission factor data from a multi-level electricity meter network, constructing a hierarchical tree topology, and outputting the graph topology includes: The system acquires electricity data and carbon emission factor data for each meter node in a multi-level meter network for the corresponding time period. Based on the physical connection relationship of the multi-level meter network, it calculates the node hierarchy depth and node importance weight, establishes a node attribute table, and outputs it. The node attribute table includes hierarchy depth information and importance weight.

[0022] In one specific embodiment, the node hierarchy depth is calculated based on the physical connection relationship of the multi-level meter network, including: Starting from the root node, traverse the connection paths of the multi-level meter network downwards, and calculate the shortest path length from each meter node to the root node as the node level depth. The importance weight of a node is calculated based on its hierarchy depth and the number of its downstream child nodes. The smaller the hierarchy depth and the more downstream child nodes a node has, the greater its importance weight. The identification information, hierarchical depth, importance weight, and corresponding electricity data and carbon emission factor data of each meter node are integrated to obtain a node attribute table.

[0023] Based on the node attribute table and the physical connection relationship of the multi-level meter network, a hierarchical tree graph topology is established using a bottom-up hierarchical construction method. The carbon emission quota of each node is allocated according to the historical carbon emission data and hierarchical depth. The node hierarchical relationship, importance weight and carbon emission quota information are integrated into the graph topology, and the graph topology is output.

[0024] In one specific embodiment, a bottom-up hierarchical construction method is used to establish a hierarchical tree diagram topology, including: Starting from the bottom leaf node, the parent-child node relationship is constructed layer by layer upwards based on the physical connection relationship; Based on the statistical characteristics of historical carbon emission data of each node and the depth of node hierarchy, carbon emission quotas are allocated to each node according to the principle of proportional allocation. The carbon emission quota of an upper-level node is equal to the sum of the carbon emission quotas of all its lower-level child nodes. Store node identification information, hierarchy depth, importance weight, carbon emission quotas, and parent-child connection relationships between nodes in a graph topology data format.

[0025] Specifically, this embodiment integrates physical connection analysis with hierarchical depth calculation. It utilizes the shortest path algorithm for node hierarchical depth calculation and importance weight analysis based on the number of downstream child nodes. It dynamically allocates node quotas based on historical carbon emission data statistical characteristics and constructs the hierarchical tree graph topology layer by layer using a bottom-up hierarchical construction method. By employing proportional allocation principles and a quota conservation mechanism between upper and lower layers, it solves the problems of multi-level meter network topology modeling and reasonable carbon emission quota allocation, improving the accuracy of the graph topology and the scientific nature of quota allocation. Simultaneously, the integration of complete node attribute information and storage of parent-child connection relationships meet the data foundation requirements for constraint establishment and anomaly detection.

[0026] Based on the graph topology, a carbon emission time-delay accumulation constraint is established. A carbon emission time-delay propagation matrix is ​​constructed to describe the cross-time coupling relationship. Instantaneous conservation constraints, sliding window accumulation constraints, and quota accumulation constraints are embedded into a hierarchical Lagrangian constraint framework, outputting carbon emission constraint conditions, including: Based on the node hierarchy and historical carbon emission data in the graph topology, the carbon emission time delay response characteristics between nodes at different levels are analyzed, and a carbon emission time delay propagation matrix is ​​constructed according to the propagation law of carbon emissions in the graph topology. Based on the carbon emission delay propagation matrix and graph topology, three types of basic constraints are established, including instantaneous conservation constraints, sliding window accumulation constraints, and quota accumulation constraints. The instantaneous conservation constraints are used to describe the carbon emission conservation relationship between parent and child nodes, the sliding window accumulation constraints are used to describe the carbon emission accumulation balance within a time window, and the quota accumulation constraints are used to describe the carbon emission quota consumption progress of nodes. The three types of basic constraints and the carbon emission time delay propagation matrix are embedded into a hierarchical Lagrange constraint framework. The constraint relationship is established across time using the Lagrange multiplier method, forming a unified multi-timescale constraint expression and outputting the carbon emission constraint conditions.

[0027] The construction of the carbon emission time delay propagation matrix includes: Statistical analysis of historical carbon emission data of each node in the graph topology is used to calculate the time delay characteristics and propagation intensity of carbon emission changes between nodes at different levels. Based on the differences in node hierarchy depth and physical connection distance, calculate the time delay parameters and attenuation coefficients for carbon emissions propagating from upper-level nodes to lower-level nodes; A carbon emission delay propagation matrix is ​​constructed based on the delay parameter and the attenuation coefficient to describe the carbon emission delay coupling relationship between nodes. The matrix elements of the carbon emission delay propagation matrix are used to represent the carbon emission influence weights between different nodes at different times.

[0028] In one specific embodiment, the carbon emission delay propagation matrix is ​​calculated as follows: ; in, for Time Node To the node The carbon emission delay propagation matrix elements; For nodes The carbon emission propagation intensity coefficient; For nodes The carbon emission reception sensitivity coefficient; The coefficient representing the influence of differences in hierarchical depth; and They are nodes and The depth of the hierarchy; This is the physical connection distance attenuation parameter; For nodes To the node The delay parameter; This is the time decay coefficient; This is the baseline time point.

[0029] The calculation formula for the multi-timescale constraint expression is: ; in, For a unified expression of multi-timescale constraints; For the first Class-constrained Lagrange multipliers These correspond to instantaneous conservation, sliding window accumulation, and quota accumulation constraints, respectively. For the first Basic constraints of the class; For time step Next node To the node The time-coupled Lagrange multiplier; for Time Node To the node The carbon emission delay propagation matrix elements; For time step Next node To the node Carbon emission coupling constraints; This represents the total number of network nodes. This represents the maximum time delay step.

[0030] Specifically, this embodiment utilizes time-delay propagation analysis and hierarchical Lagrange constraint modeling. It analyzes time-delay response characteristics and calculates propagation intensity using historical carbon emission data, dynamically determines time-delay parameters and attenuation coefficients based on node-level depth differences, and embeds a unified framework for the three types of fundamental constraints according to cross-temporal coupling relationships. By employing the Lagrange multiplier method and the carbon emission time-delay propagation matrix, it solves the problem of differentiable realization of multi-timescale carbon emission conservation constraints and the problem of modeling cross-temporal coupling relationships, improving the accuracy of constraint expression and the scientific rigor of time-coupled modeling.

[0031] The step involves employing a quota surplus-sensitive nonlinear weight stabilization mechanism based on the carbon emission constraints. This mechanism calculates the original constraint weight coefficients according to the quota surplus ratio and the time surplus ratio, and eliminates weight oscillations through a stabilization filter to output stable constraint weight coefficients. The steps include: Based on the quota constraint information in the carbon emission constraints, the remaining quota ratio and time ratio at the current moment are calculated. A nonlinear weight calculation method sensitive to quota remaining is adopted. According to the combination state of the remaining quota ratio and time ratio, the weight adjustment strategy for different constraint types is determined. The instantaneous conservation constraint weight coefficient, sliding window cumulative constraint weight coefficient and quota cumulative constraint weight coefficient are calculated respectively to obtain the original constraint weight coefficient. The temporal fluctuation characteristics and oscillation modes of the original constraint weight coefficients are detected. A stabilization filter designed for carbon emission quota sensitivity is used to filter the original constraint weight coefficients. The filter kernel function eliminates weight oscillations and maintains the quota-sensitive response characteristics of the weights, and outputs stable constraint weight coefficients.

[0032] The stabilization filter, designed specifically for carbon emission quota sensitivity, filters the original constraint weight coefficients, including: Analyze the time series data of the original constraint weight coefficients to identify the weight jump points and oscillation intervals caused by changes in quota surplus; A filter kernel function for carbon emission weight stabilization is constructed, which combines the monotonicity of carbon emissions with the sensitivity of quota constraints. The original constraint weight coefficients are convolved using the filtering kernel function to eliminate high-frequency oscillation noise while maintaining the weights' sensitivity to changes in the remaining quota state, and output stable constraint weight coefficients after stabilization.

[0033] In one specific embodiment, the original constraint weight coefficient is calculated as follows: ; in, For the first The original constraint weight coefficients of the class constraint; For the first The basic weight adjustment factor for class constraints; This represents the remaining percentage of the quota. The remaining time percentage; and The first Sensitivity index of class constraints to quota surplus and time surplus; For the first Nonlinear combination exponents with class constraints; For the first Quota time coupling adjustment parameters for class constraints.

[0034] The formula for calculating the stability constraint weight coefficient is: ; ; ; in, For the first Stability constraint weight coefficients of class constraints; A Gaussian filter kernel function that takes into account quota sensitivity; The carbon emission monotonicity retention factor; The length is half the length of the filtering window; The standard deviation parameter of the Gaussian filter kernel; This is the coefficient for enhancing the quota-sensitive response. for The change in remaining quota at any given time; It is a symbolic function; The intensity coefficient is maintained to ensure monotonicity; This refers to the time-distance attenuation parameter; This is the threshold for sensitivity to quota changes.

[0035] Specifically, this embodiment integrates quota surplus sensitivity analysis with stabilization filtering technology. It utilizes a nonlinear weighting calculation method to analyze the remaining quota ratio and time surplus ratio, and calculates the weight coefficients for three types of constraints. It dynamically adjusts the filter kernel function parameters based on the monotonicity characteristics of carbon emissions, and designs a dedicated stabilization filter based on the identification results of weight jump points and oscillation intervals. Through convolution operations and a sensitivity-preserving mechanism, it solves the weight oscillation problem and high-frequency noise interference problem under quota surplus sensitivity, thereby improving the quota sensitivity response capability of the constraint weight coefficients.

[0036] The process of inputting the graph topology, electricity data, and carbon emission factor data into the graph neural network state space model, combining the stability constraint weight coefficients, constructing a cross-time coupling loss function for model training, and outputting a graph neural network model includes: Based on the stability constraint weight coefficients, a multi-timescale loss term is constructed. The multi-timescale loss term is obtained by fusing instantaneous conservation constraint loss, sliding window cumulative constraint loss and quota cumulative constraint loss. The loss correlation between different time steps is established through the time coupling weight matrix to obtain the cross-time coupling loss function.

[0037] In one specific embodiment, the process of constructing the cross-time coupling loss function includes: Based on the instantaneous conservation constraint weight coefficient, sliding window cumulative constraint weight coefficient, and quota cumulative constraint weight coefficient in the stability constraint weight coefficient, calculate the corresponding instantaneous conservation constraint loss, sliding window cumulative constraint loss, and quota cumulative constraint loss respectively. A time-coupled weight matrix is ​​constructed based on the time-delay propagation characteristics of carbon emissions. The time-coupled weight matrix is ​​used to describe the mutual influence relationship between loss terms at different time steps. The three types of constraint losses are weighted and combined using a time-coupled weight matrix to form a unified loss expression that considers the coupling relationship between time steps, thus obtaining a cross-time-coupled loss function.

[0038] The graph topology, electricity data, and carbon emission factor data are input into the graph neural network state space model for forward propagation calculation. The network parameters of the graph neural network state space model are updated by gradient backpropagation using the cross-temporal coupling loss function. The forward and backpropagation processes are repeated until the loss function converges, and the trained graph neural network model is output.

[0039] In one specific embodiment, the training method for the graph neural state space model includes: The graph topology is used as the graph structure input, and the electricity data and carbon emission factor data are used as node features input to the graph neural state space model for forward computation to obtain the carbon emission prediction results output by the model. The cross-temporal coupling loss function value between the carbon emission prediction results and the real labels is calculated, and the gradient of the loss function with respect to the graph convolutional layer weights, state transition matrix and output layer parameters in the graph neural state space model is calculated through an automatic differentiation mechanism. The gradient descent optimizer is used to update the model parameters based on the calculated gradient. The above process is repeated for iterative training until the cross-time coupling loss function reaches the preset convergence condition.

[0040] Specifically, this embodiment integrates multi-timescale loss construction with graph neural state-space modeling. It utilizes stable constraint weight coefficients to calculate three types of constraint losses and establishes a time-coupled weight matrix. It dynamically adjusts the loss correlation between time steps based on the carbon emission delay propagation characteristics and employs a dedicated training design for the graph neural state-space model according to the cross-time-coupled loss function. The automatic differentiation mechanism and gradient descent optimizer address the issues of unified expression of multi-timescale constraint losses and neural network parameter optimization under graph topology, improving model training convergence and constraint satisfaction consistency. Furthermore, the complete forward and backward propagation iterative mechanism meets the technical requirements of multi-constraint fusion and high-precision modeling in complex carbon emission prediction tasks.

[0041] The process of using the graph neural network model for carbon emission prediction involves detecting constraint conflicts through a multi-objective projection operator that senses these conflicts, resolving conflicts according to carbon reduction priorities, and correcting the prediction results using multi-objective Pareto projection and carbon monotonicity-preserving projection. The resulting carbon emission prediction value is then output. The graph neural network model is used to perform forward inference calculations on the current electricity data and carbon emission factor data to obtain the initial carbon emission prediction value. The conflict between the initial carbon emission prediction value and the carbon emission constraint conditions is detected by the multi-objective projection operator that senses the constraint conflict. According to the carbon emission reduction priority strategy, the detected constraint conflicts are prioritized and conflict resolution is processed, and the constraint conflict resolution result is output.

[0042] In one specific embodiment, the multi-objective projection operator that detects conflicts between the initial carbon emission prediction and carbon emission constraints using a constraint conflict perception method, and performs priority ranking and conflict resolution processing on the detected constraint conflicts according to a carbon emission reduction priority strategy, including: The initial carbon emission forecasts are compared and calculated with instantaneous conservation constraints, sliding window accumulation constraints, and quota accumulation constraints respectively to identify the forecast nodes and time points that violate the constraints. A priority ranking rule for constraint conflicts is established based on the carbon emission reduction priority strategy. The quota accumulation constraint conflict has the highest priority, the instantaneous conservation constraint conflict has the medium priority, and the sliding window accumulation constraint conflict has the lowest priority. The conflicts are resolved step by step according to priority to obtain the constraint conflict resolution results, which include the conflict type, conflict degree and resolution strategy.

[0043] Based on the results of the constraint conflict resolution, the initial carbon emission prediction value is corrected for the first time using the multi-objective Pareto projection algorithm to obtain the Pareto optimal prediction value. The Pareto optimal prediction value after Pareto correction is then corrected for the second time using the carbon monotonicity-preserving projection algorithm to ensure that the prediction value conforms to the monotonicity constraint characteristics of carbon emissions, thus obtaining the carbon emission prediction value.

[0044] In one specific embodiment, the method for dual-layer projection correction includes: Based on the results of constraint conflict resolution, a multi-objective optimization problem is constructed. Different types of constraints are used as multiple objective functions. The multi-objective Pareto projection algorithm is used to find the prediction correction scheme that satisfies the Pareto optimal condition. Based on the physical characteristics of carbon emissions, carbon monotonicity constraints are established. A carbon monotonicity-preserving projection algorithm is used to constrain the Pareto-corrected predicted values ​​to ensure that the corrected predicted values ​​maintain the monotonicity of carbon emission changes over time. The output is a carbon emission prediction value that simultaneously satisfies the multi-objective Pareto optimality and carbon monotonicity constraints.

[0045] Specifically, this embodiment integrates constraint conflict sensing projection with two-layer correction technology. It utilizes multi-objective projection operators for constraint conflict detection and carbon emission reduction priority ranking, dynamically adjusts the prediction correction scheme based on Pareto optimality theory, and designs a dedicated projection for monotonicity constraints according to the physical characteristics of carbon emissions. Through a step-by-step conflict resolution mechanism and a two-layer projection correction algorithm, it solves the problems of conflict detection and resolution under multiple constraints and maintaining the physical consistency of prediction results, improving the constraint satisfaction and physical rationality of carbon emission predictions. Simultaneously, the complete priority management and monotonicity maintenance mechanism meets the policy orientation of carbon emission reduction.

[0046] The method, based on the residuals between the predicted and actual carbon emissions, integrates multimodal anomaly evidence from change point detection, distribution drift detection, and tail dependency detection, and uses a Bayesian fusion method to calculate the comprehensive anomaly probability, outputting carbon emission anomaly detection results, including: The predicted residual sequence between the predicted carbon emissions and the actual observed values ​​is calculated. Based on the predicted residual sequence, change point detection analysis, distribution drift detection analysis, and tail dependency detection analysis are performed respectively. Anomaly indicators and confidence scores corresponding to each detection method are extracted to generate a multimodal anomaly evidence set, which includes change point anomaly evidence, distribution drift anomaly evidence, and tail dependency anomaly evidence.

[0047] In one specific embodiment, the method for generating multimodal anomaly evidence includes: A sliding window change point detection algorithm is used for the predicted residual sequence to identify abrupt changes in the statistical characteristics of the residuals, calculate the anomaly index and corresponding confidence score for change point detection, and generate anomaly evidence for change points. The probability distribution change of the predicted residual sequence is analyzed by using the distribution parameter estimation method, the degree of drift of the distribution mean, variance and skewness is detected, the anomaly index and confidence score of distribution drift detection are calculated, and the evidence of distribution drift anomaly is obtained. The tail dependency structure between the predicted residual and the historical residual is analyzed based on the copula function modeling method. The dependency anomalies of tail extreme events are detected, and the anomaly index and confidence score of the tail dependency detection are calculated to obtain evidence of tail dependency anomalies.

[0048] Based on the confidence scores and anomaly indicators of various types of anomaly evidence in the multimodal anomaly evidence set, a joint probability model of multimodal evidence is constructed using a Bayesian fusion algorithm. The comprehensive anomaly probability of the three anomaly detection modes is calculated through Bayesian inference. The carbon emission anomaly status is determined according to the preset anomaly probability threshold, and the carbon emission anomaly detection result is output. The carbon emission anomaly detection result includes anomaly probability value and anomaly determination result.

[0049] In one specific embodiment, the step of constructing a joint probability model of multimodal evidence using a Bayesian fusion algorithm, and calculating the comprehensive anomaly probability of fusing three anomaly detection modalities through Bayesian inference, includes: Based on the confidence scores of point change anomaly evidence, distribution drift anomaly evidence, and tail dependency anomaly evidence in the multimodal anomaly evidence set, a prior probability distribution for anomaly detection of each modality is established. A Bayesian network structure describing the relevance of multimodal evidence is constructed, and a joint probability model for multimodal evidence fusion is established through the assumption of conditional independence between evidence and joint probability decomposition. The Bayesian inference algorithm is used to perform probability fusion calculation on the abnormal evidence of each modality to obtain the comprehensive abnormal probability. The comprehensive abnormal probability is compared with the preset abnormal threshold to generate a complete carbon emission abnormality detection result that includes numerical abnormal probability and binary abnormal state.

[0050] Specifically, this embodiment combines multimodal anomaly detection with Bayesian probabilistic inference. It utilizes a sliding window change-point detection algorithm for residual statistical characteristic analysis and distribution parameter estimation methods for probability distribution change calculation. Combined with the copula function modeling method, it dynamically adjusts the tail-dependent structure detection parameters and designs a joint probability model for multimodal evidence fusion based on the conditional independence assumption. By employing Bayesian inference algorithms and a prior probability distribution establishment mechanism, it addresses the multimodal anomaly evidence fusion problem and the limitations of single detection methods, improving the accuracy of carbon emission anomaly detection and the reliability of comprehensive judgment. Furthermore, through a complete confidence scoring system and numerical anomaly probability output, it satisfies the need for multi-dimensional evidence integration in real-time carbon emission monitoring.

[0051] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for anomaly detection that combines graph topological constraints and multimodal Bayesian methods, characterized in that, Includes the following steps: Obtain electricity data and carbon emission factor data from a multi-level electricity meter network, construct a hierarchical tree topology, and output the graph topology. Based on the graph topology, carbon emission time delay accumulation constraints are established. The carbon emission time delay propagation matrix is ​​constructed to describe the cross-time coupling relationship. Instantaneous conservation constraints, sliding window accumulation constraints, and quota accumulation constraints are embedded into the hierarchical Lagrange constraint framework to output carbon emission constraint conditions. Based on the carbon emission constraints, a nonlinear weight stabilization mechanism sensitive to quota surplus is adopted. The original constraint weight coefficients are calculated based on the quota surplus ratio and the time surplus ratio. The weight oscillations are eliminated by a stabilization filter, and the stable constraint weight coefficients are output. The graph topology, electricity data, and carbon emission factor data are input into the graph neural network state space model. The stability constraint weight coefficients are combined to construct a cross-time coupling loss function for model training, and the graph neural network model is output. Carbon emission prediction is performed using the graph neural network model. Constraint conflicts are detected by a multi-objective projection operator that senses constraint conflicts, and conflicts are resolved according to carbon emission reduction priorities. The prediction results are corrected by multi-objective Pareto projection and carbon monotonicity-preserving projection, and the predicted carbon emission values ​​are output. Based on the residuals between the predicted and actual carbon emissions, multimodal anomaly evidence from change point detection, distribution drift detection, and tail dependency detection is integrated, and a Bayesian fusion method is used to calculate the comprehensive anomaly probability, outputting the carbon emission anomaly detection result.

2. The anomaly detection method fusion of graph topological constraints and multimodal Bayesian methods as described in claim 1, characterized in that, Based on the graph topology, a carbon emission time-delay accumulation constraint is established. A carbon emission time-delay propagation matrix is ​​constructed to describe the cross-time coupling relationship. Instantaneous conservation constraints, sliding window accumulation constraints, and quota accumulation constraints are embedded into a hierarchical Lagrangian constraint framework, outputting carbon emission constraint conditions, including: Based on the node hierarchy and historical carbon emission data in the graph topology, the carbon emission time delay response characteristics between nodes at different levels are analyzed, and a carbon emission time delay propagation matrix is ​​constructed according to the propagation law of carbon emissions in the graph topology. Based on the carbon emission delay propagation matrix and graph topology, three types of basic constraints are established, including instantaneous conservation constraints, sliding window accumulation constraints, and quota accumulation constraints. The instantaneous conservation constraints are used to describe the carbon emission conservation relationship between parent and child nodes, the sliding window accumulation constraints are used to describe the carbon emission accumulation balance within a time window, and the quota accumulation constraints are used to describe the carbon emission quota consumption progress of nodes. The three types of basic constraints and the carbon emission time delay propagation matrix are embedded into a hierarchical Lagrange constraint framework. The constraint relationship is established across time using the Lagrange multiplier method, forming a unified multi-timescale constraint expression and outputting the carbon emission constraint conditions.

3. The anomaly detection method fusion of graph topological constraints and multimodal Bayesian methods as described in claim 2, characterized in that, The construction of the carbon emission time delay propagation matrix includes: Statistical analysis of historical carbon emission data of each node in the graph topology is used to calculate the time delay characteristics and propagation intensity of carbon emission changes between nodes at different levels. Based on the differences in node hierarchy depth and physical connection distance, calculate the time delay parameters and attenuation coefficients for carbon emissions propagating from upper-level nodes to lower-level nodes; A carbon emission delay propagation matrix is ​​constructed based on the delay parameter and the attenuation coefficient to describe the carbon emission delay coupling relationship between nodes. The matrix elements of the carbon emission delay propagation matrix are used to represent the carbon emission influence weights between different nodes at different times.

4. The anomaly detection method fusion of graph topological constraints and multimodal Bayesian methods as described in claim 1, characterized in that, The step involves employing a quota surplus-sensitive nonlinear weight stabilization mechanism based on the carbon emission constraints. This mechanism calculates the original constraint weight coefficients according to the quota surplus ratio and the time surplus ratio, and eliminates weight oscillations through a stabilization filter to output stable constraint weight coefficients. The steps include: Based on the quota constraint information in the carbon emission constraints, the remaining quota ratio and time ratio at the current moment are calculated. A nonlinear weight calculation method sensitive to quota remaining is adopted. According to the combination state of the remaining quota ratio and time ratio, the weight adjustment strategy for different constraint types is determined. The instantaneous conservation constraint weight coefficient, sliding window cumulative constraint weight coefficient and quota cumulative constraint weight coefficient are calculated respectively to obtain the original constraint weight coefficient. The temporal fluctuation characteristics and oscillation modes of the original constraint weight coefficients are detected. A stabilization filter designed for carbon emission quota sensitivity is used to filter the original constraint weight coefficients. The filter kernel function eliminates weight oscillations and maintains the quota-sensitive response characteristics of the weights, and outputs stable constraint weight coefficients.

5. The anomaly detection method fusion of graph topological constraints and multimodal Bayesian methods as described in claim 4, characterized in that, The stabilization filter, designed specifically for carbon emission quota sensitivity, filters the original constraint weight coefficients, including: Analyze the time series data of the original constraint weight coefficients to identify the weight jump points and oscillation intervals caused by changes in quota surplus; A filter kernel function for carbon emission weight stabilization is constructed, which combines the monotonicity of carbon emissions with the sensitivity of quota constraints. The original constraint weight coefficients are convolved using the filtering kernel function to eliminate high-frequency oscillation noise while maintaining the weights' sensitivity to changes in the remaining quota state, and output stable constraint weight coefficients after stabilization.

6. The anomaly detection method fusion of graph topological constraints and multimodal Bayesian methods as described in claim 5, characterized in that, The formula for calculating the original constraint weight coefficients is: ； in, For the first The original constraint weight coefficients of the class constraint; For the first The basic weight adjustment factor for class constraints; This represents the remaining percentage of the quota. The remaining time percentage; and The first Sensitivity index of class constraints to quota surplus and time surplus; For the first Nonlinear combination exponents with class constraints; For the first Quota time coupling adjustment parameters for class constraints.

7. The anomaly detection method fusion of graph topological constraints and multimodal Bayesian methods as described in claim 1, characterized in that, The process of acquiring electricity data and carbon emission factor data from a multi-level electricity meter network, constructing a hierarchical tree topology, and outputting the graph topology includes: Acquire the electricity data and carbon emission factor data of each meter node in the multi-level meter network for the corresponding time period. Based on the physical connection relationship of the multi-level meter network, calculate the node hierarchy depth and node importance weight, establish a node attribute table and output it. The node attribute table includes hierarchy depth information and importance weight. Based on the node attribute table and the physical connection relationship of the multi-level meter network, a hierarchical tree graph topology is established using a bottom-up hierarchical construction method. The carbon emission quota of each node is allocated according to the historical carbon emission data and hierarchical depth. The node hierarchical relationship, importance weight and carbon emission quota information are integrated into the graph topology, and the graph topology is output.

8. The anomaly detection method fusion of graph topological constraints and multimodal Bayesian methods as described in claim 1, characterized in that, The process of inputting the graph topology, electricity data, and carbon emission factor data into the graph neural network state space model, combining the stability constraint weight coefficients, constructing a cross-time coupling loss function for model training, and outputting a graph neural network model includes: Based on the stability constraint weight coefficients, a multi-timescale loss term is constructed. The multi-timescale loss term is obtained by fusing instantaneous conservation constraint loss, sliding window cumulative constraint loss and quota cumulative constraint loss. The loss correlation between different time steps is established through the time coupling weight matrix to obtain the cross-time coupling loss function. The graph topology, electricity data, and carbon emission factor data are input into the graph neural network state space model for forward propagation calculation. The network parameters of the graph neural network state space model are updated by gradient backpropagation using the cross-temporal coupling loss function. The forward and backpropagation processes are repeated until the loss function converges, and the trained graph neural network model is output.

9. The anomaly detection method fusion of graph topological constraints and multimodal Bayesian methods as described in claim 1, characterized in that, The process of using the graph neural network model for carbon emission prediction involves detecting constraint conflicts through a multi-objective projection operator that senses these conflicts, resolving conflicts according to carbon reduction priorities, and correcting the prediction results using multi-objective Pareto projection and carbon monotonicity-preserving projection. The resulting carbon emission prediction value is then output. The graph neural network model is used to perform forward inference calculations on the current electricity data and carbon emission factor data to obtain the initial carbon emission prediction value. The conflict between the initial carbon emission prediction value and the carbon emission constraint conditions is detected by the multi-objective projection operator that senses the constraint conflict. The detected constraint conflicts are prioritized and resolved according to the carbon emission reduction priority strategy, and the constraint conflict resolution result is output. Based on the results of the constraint conflict resolution, the initial carbon emission prediction value is corrected for the first time using the multi-objective Pareto projection algorithm to obtain the Pareto optimal prediction value. The Pareto optimal prediction value after Pareto correction is then corrected for the second time using the carbon monotonicity-preserving projection algorithm to ensure that the prediction value conforms to the monotonicity constraint characteristics of carbon emissions, thus obtaining the carbon emission prediction value.

10. The anomaly detection method fusion of graph topological constraints and multimodal Bayesian methods as described in claim 1, characterized in that, The method, based on the residuals between the predicted and actual carbon emissions, integrates multimodal anomaly evidence from change point detection, distribution drift detection, and tail dependency detection, and uses a Bayesian fusion method to calculate the comprehensive anomaly probability, outputting carbon emission anomaly detection results, including: The predicted residual sequence between the predicted carbon emissions and the actual observed values ​​is calculated. Based on the predicted residual sequence, change point detection analysis, distribution drift detection analysis and tail dependency detection analysis are performed respectively. Anomaly indicators and confidence scores corresponding to each detection method are extracted to generate a multimodal anomaly evidence set, which includes change point anomaly evidence, distribution drift anomaly evidence and tail dependency anomaly evidence. Based on the confidence scores and anomaly indicators of various types of anomaly evidence in the multimodal anomaly evidence set, a joint probability model of multimodal evidence is constructed using a Bayesian fusion algorithm. The comprehensive anomaly probability of the three anomaly detection modes is calculated through Bayesian inference. The carbon emission anomaly status is determined according to the preset anomaly probability threshold, and the carbon emission anomaly detection result is output. The carbon emission anomaly detection result includes anomaly probability value and anomaly determination result.

Images