A method and system for multi-source data fusion in power distribution networks enhanced by a large language model.

By establishing spatiotemporal indexes and comprehensive quality scores, unstructured operation and maintenance texts are transformed into structured semantic constraints, solving the problem of dynamic changes in data quality in the fusion of multi-source data in the distribution network, and realizing a state awareness and auxiliary decision-making closed loop with high robustness and security.

CN123115254APending Publication Date: 2026-11-17ELECTRIC POWER RESEARCH INSTITUTE OF STATE GRID NINGXIA ELECTRIC POWER COMPANY +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611278173.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-21
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing multi-source data fusion methods for distribution networks cannot uniformly utilize structured measurement data and unstructured operation and maintenance semantics, cannot adaptively adjust the fusion mechanism according to dynamic changes in data quality, lack state perception results with confidence, and are difficult to form a closed loop for safety auxiliary decision-making.

Method used

Acquire multi-source heterogeneous data, establish a spatiotemporal index, calculate the comprehensive quality score of the data source, transform unstructured operation and maintenance text into structured semantic constraint terms, construct a joint weighted matrix of quality and uncertainty, perform state estimation under physical constraints, and output uncertainty and auxiliary decision-making scheme.

Benefits of technology

It improves the robustness of fusion perception and the security of auxiliary decision-making in the distribution network under conditions of missing measurements, communication delays, and the presence of unstructured operation and maintenance semantics, forming a closed loop that is constrained, verifiable, and interpretable.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115254A_ABST
    Figure CN123115254A_ABST
Patent Text Reader

Abstract

A method and system for multi-source data fusion in distribution networks enhanced by a large language model is proposed. This method acquires heterogeneous data from multiple sources, including SCADA, synchronization phasors, AMI, equipment monitoring, meteorological, GIS / topology data, work orders, alarms, and defect records, and establishes a unified spatiotemporal index. Dynamic comprehensive quality scores are obtained based on each data source. Unstructured operation and maintenance text is converted into structured text results such as equipment objects, event types, topology states, operational intentions, and text confidence levels. These results are further converted into semantic constraint terms, topology correction terms, or pseudo-observation terms in the state estimation model. A joint weighted matrix of quality and uncertainty is constructed based on the comprehensive quality score and prediction uncertainty. State estimation is performed under physical constraints, and the results output uncertainty, anomaly identification, and auxiliary decision-making schemes. This invention improves the robustness of fusion perception and the security of auxiliary decision-making in distribution networks under conditions of measurement loss, communication delay, inter-source conflicts, and the presence of unstructured operation and maintenance semantics.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of power distribution network data processing and intelligent operation technology, and in particular relates to a method and system for power distribution network operation perception, anomaly identification and auxiliary decision-making that integrates large language model, physical constraint state estimation and multi-source heterogeneous data fusion. Background Technology

[0002] With the large-scale integration of distributed power sources, energy storage, electric vehicle charging facilities, and new distribution terminals, distribution network operation data exhibits significant characteristics such as diverse sources, inconsistent time scales, dynamic quality changes, and the coexistence of structured and unstructured data. Existing publicly available research has shown that the real-time sensing task of distribution networks is evolving from traditional state estimation based primarily on pseudo-measurements and low-frequency measurements to a comprehensive estimation framework that combines smart meters, micro-PMUs, measurements of varying precision, and topology discrimination. However, under conditions of missing measurements, communication delays, inter-source conflicts, and weak observability, estimation accuracy and robustness remain significantly constrained.

[0003] Existing multi-source fusion methods for distribution networks mainly include fixed-weighted fusion, state estimation fusion based on Kalman filtering or weighted least squares, and decision-level fusion based on evidence theory. These methods are effective under structured numerical measurement conditions, but they typically struggle to address the following issues simultaneously: First, the data quality from different data sources is not constant under different operating conditions, and fixed-weighted or static noise modeling cannot adaptively reflect changes in reliability. Second, a large amount of unstructured operational semantics exists, such as work order texts, defect records, dispatch passwords, and alarm statements, but traditional numerical fusion devices cannot effectively absorb this information. Third, most methods only provide point estimates, lacking credible uncertainty outputs for operational decision-making. Fourth, anomaly detection and auxiliary decision-making are often decoupled from front-end state perception, making it difficult to form a closed loop.

[0004] Meanwhile, published literature indicates that large language models and large basic models have considerable potential in power system knowledge retrieval, context awareness, simulation tool invocation, and active distribution network voltage regulation. However, published research also points out that without retrieval enhancement, domain constraints, and physical feasibility verification, LLMs are difficult to use directly as final decision-makers in high-reliability distribution tasks.

[0005] Furthermore, even when existing methods incorporate text processing or intelligent question-answering models, they typically remain at the level of alarm explanation, knowledge-based question answering, or supplementary explanations. They struggle to transform the equipment objects, topology changes, operational intentions, fault types, and time constraints contained in unstructured maintenance texts into constraint variables, topology corrections, or pseudo-observations that can be directly invoked by the state estimation model. Consequently, unstructured semantic information and structured measurement information remain isolated, making it difficult for them to jointly influence the distribution network state estimation results within the same physical constraint optimization model.

[0006] Therefore, there is an urgent need for a large language model-enhanced multi-source data fusion method and system that is oriented towards power distribution network scenarios, can uniformly process structured measurement data and unstructured operation and maintenance texts, can output credible uncertainty and form a closed loop for safety auxiliary decision-making. Summary of the Invention

[0007] The technical problem to be solved by this invention is: to address the shortcomings of existing multi-source data fusion methods for distribution networks, such as the inability to uniformly utilize structured measurement data and unstructured operation and maintenance semantics, the inability to dynamically adjust the fusion mechanism according to changes in data quality, the lack of state awareness results with confidence, and the difficulty in forming a closed loop for safety auxiliary decision-making. This invention provides a multi-source data fusion method and system for distribution networks enhanced by a large language model.

[0008] Furthermore, the technical problems to be solved by this invention also include: how to transform the unstructured operation and maintenance semantic results output by the retrieval-enhanced large language model into semantic constraint terms, topology correction terms, or pseudo-observation terms that can enter the physical constraint state estimation model; how to use the integrated quality score of the data source and the prediction uncertainty of the observation to jointly construct the state estimation weighting matrix; and how to avoid the large language model directly outputting operation control suggestions that have not been physically verified, thereby forming a constrained, verifiable, and interpretable closed loop between distribution network operation perception, anomaly identification, and auxiliary decision-making.

[0009] To address the shortcomings of existing technologies, this invention provides a method and system for multi-source data fusion in distribution networks enhanced by a large language model. It acquires heterogeneous data from multiple sources, including SCADA, synchronization phasors, AMI, equipment monitoring, meteorological, GIS / topology data, work orders, alarms, and defect records, and establishes a unified spatiotemporal index. Dynamic comprehensive quality scores are obtained based on each data source. Unstructured operation and maintenance text is converted into structured text results such as equipment objects, event types, topology states, operational intentions, and text confidence levels, and further converted into semantic constraint terms, topology correction terms, or pseudo-observation terms in the state estimation model. A joint weighted matrix of quality and uncertainty is constructed based on the comprehensive quality score and prediction uncertainty. State estimation is performed under physical constraints, and the uncertainty, anomaly identification results, and auxiliary decision-making schemes are output. This invention can improve the robustness of fusion perception and the security of auxiliary decision-making in distribution networks under conditions of missing measurements, communication delays, inter-source conflicts, and the presence of unstructured operation and maintenance semantics.

[0010] The present invention adopts the following technical solution.

[0011] This invention proposes the following steps: Acquire multi-source heterogeneous data of the power distribution network and establish a spatiotemporal index of the multi-source heterogeneous data, which includes structured measurement data and unstructured operation and maintenance text data; Based on the spatiotemporal index, the integrity score, accuracy score, timeliness score, and semantic consistency score of each data source are calculated respectively. Based on the integrity score, accuracy score, timeliness score, and semantic consistency score, the comprehensive quality score of each data source at the current moment is calculated. A structured state representation is obtained based on the structured measurement data; a textual semantic representation is obtained based on the unstructured operation and maintenance text data. The semantic representation of the text is verified based on a large language model with retrieval enhancement, and the verified results are converted into semantic constraints, topological corrections or pseudo-observations for state estimation. Based on the comprehensive quality score of each data source and the prediction uncertainty of the corresponding observation, a joint weighted matrix of quality and uncertainty is constructed. The structured state representation is used as the observation input, and the semantic constraint term, topology correction term, or pseudo-observation term is used as the constraint term or observation term and introduced into the physical constraint state estimation model to obtain the distribution network state estimation result. The comprehensive anomaly score is calculated based on the power distribution network state estimation results, and anomaly discrimination information is output by combining the uncertainty quantification results. Based on the power distribution network state estimation results and anomaly identification information, a large language model with retrieval enhancement is used to generate auxiliary decision candidate schemes; physical security verification is performed on the auxiliary decision candidate schemes, and only auxiliary decision schemes that pass the verification are output.

[0012] More preferably, the structured measurement data includes at least three of the following: SCADA measurement data, synchronous phasor measurement data, AMI data, equipment online monitoring data, meteorological data, and GIS / topology data; The unstructured operation and maintenance text data includes at least one of the following: work order text, alarm text, defect record, scheduling record, and maintenance record.

[0013] More preferably, the integrity score is calculated by dividing the number of data points actually received within a preset time window by the number of data points that should have been received within that time window to obtain the integrity score; The accuracy score is calculated as follows: the absolute value of the difference between each observation and the corresponding reference prediction value is divided by a preset accuracy threshold to obtain a ratio result, and the smaller value between 1 and the ratio result is taken; the mean of the smaller values ​​corresponding to all observations is calculated, and the accuracy score is obtained by subtracting the mean from 1. The timeliness score is calculated as follows: the product of the delay attenuation coefficient and the current time delay is calculated, and the negative of the product is taken as the input of the exponential function to calculate the timeliness score. The semantic consistency score is calculated as follows: calculate the cosine similarity between the current semantic vector and the predicted semantic vector of the normal pattern, add 1 to the cosine similarity and divide by 2 to obtain the semantic consistency score.

[0014] More preferably, the semantic consistency score is calculated as follows: extract the device number, topology node, timestamp, switch status and over-limit alarm corresponding to the measurement data, and cross-match it with the domain knowledge base and other data sources; if the match is consistent, it is assigned a value of 1.0; if there is a clear conflict, it is assigned a value of 0; if there is partial field conflict, it is assigned a value by linear interpolation according to the proportion of the number of conflicting fields to the total number of matched fields.

[0015] More preferably, the method for calculating the comprehensive quality score is as follows: pre-setting a completeness weight, an accuracy weight, a timeliness weight, and a semantic consistency weight; multiplying the completeness score by the completeness weight, the accuracy score by the accuracy weight, the timeliness score by the timeliness weight, and the semantic consistency score by the semantic consistency weight, and then adding them together to obtain the comprehensive quality score; wherein, the sum of the completeness weight, the accuracy weight, the timeliness weight, and the semantic consistency weight is 1.

[0016] More preferably, the calculation method for the structured state representation is as follows: feature extraction is performed on the structured measurement data through a numerical coding network to obtain the structured state representation; The text semantic representation calculation method is as follows: entity extraction and event extraction are performed on the unstructured operation and maintenance text data to obtain intermediate results containing device objects, event types, event times, topology status, operation intentions and text confidence; the intermediate results are semantically encoded through a retrieval-enhanced large language model to obtain the text semantic representation. The structured state representation and the text semantic representation are fused through an attention mechanism to obtain a fused representation vector. The fused representation vector is used for anomaly detection, retrieval query and generation of auxiliary decision context, and is not directly used as the observation input of the physical constraint state estimation model.

[0017] More preferably, the calculation method for knowledge verification of the text semantic representation based on the retrieval-enhanced large language model is as follows: Retrieve topology constraints, equipment ledgers, protection settings, scheduling rules, and historical similar cases that match the current moment from the domain knowledge base. Input the retrieval results and the text semantic representation into the large language model, and output a structured text result that has been verified by knowledge constraints. The structured text result includes equipment object field, event type field, time constraint field, topology status field, operation intention field, and text confidence field.

[0018] More preferably, the method for calculating the semantic constraint terms used for state estimation by converting the verified result into a variable is as follows: When the structured text result represents node injection power, device operating status, or the time of occurrence of abnormal events, it is mapped to a semantic constraint term for solving the constraint state estimation. The method for converting the verified result into a topology correction term for state estimation is as follows: When the structured text result represents the status of switch opening / closing, line outage, equipment maintenance, or equipment commissioning / decommissioning, it is mapped to a topology correction term used to correct the topology connection relationship in the state estimation model. The method for converting the verified results into pseudo-observations for state estimation is as follows: When the structured text results represent missing measurements, low-confidence measurements, or load changes confirmed by work orders, they are mapped to pseudo-observations used to supplement the state estimation input.

[0019] More preferably, the specific method for obtaining the distribution network state estimation results is as follows: The overall quality scores of each data source are truncated to upper and lower limits to obtain the truncated reliability weights; the standard deviation of the prediction error corresponding to each observation is calculated based on the historical effective prediction residuals of each observation before the current time; the square of the standard deviation of the prediction error is summed with the corresponding lower limit of variance, the reliability weight is divided by the summation result, and the quotient is used as the diagonal element to construct a quality-uncertainty joint weighting matrix composed of the diagonal elements; The structured state representation is input into the observation function of the physical constraint state estimation model to obtain the expected observation value; Construct a physical constraint equation, and add the physical constraint equation, the semantic constraint term, the topology correction term, and the pseudo-observation term as penalty terms into the optimization objective function of the physical constraint state estimation model; The state variable vector of the distribution network is obtained by minimizing the optimization objective function, and is used as the state estimation result of the distribution network.

[0020] More preferably, a comprehensive anomaly score is calculated based on the distribution network state estimation results, and anomaly discrimination information is output in combination with the uncertainty quantification results. The specific method is as follows: Calculate the measurement residual vector corresponding to the power distribution network state estimation result, and calculate the product of the measurement residual vector and the inverse of the residual covariance matrix to obtain the weighted residual term; calculate the cosine similarity between the current text embedding vector and the normal pattern predicted text embedding vector, and subtract the cosine similarity from 1 to obtain the text semantic bias term; obtain the anomaly isolation score; assign weight coefficients to the weighted residual term, the text semantic bias term, and the anomaly isolation score respectively, and then perform a weighted summation to obtain the comprehensive anomaly score; The mean and variance of the state estimation results are calculated through multiple state estimation sub-models. The conformal calibration quantiles are calculated based on the historical calibration dataset. The prediction interval at a preset confidence level is generated based on the mean, variance and conformal calibration quantiles as the uncertainty quantification result. The comprehensive anomaly score is compared with a preset anomaly judgment threshold, and anomaly discrimination information including anomaly label, anomaly type, scope of impact, and confidence level is generated based on the comparison result.

[0021] More preferably, the method for obtaining the isolated anomaly score includes: Based on historical operation records, select fusion representation vectors that are in normal operation to form a historical normal training sample set, and use the historical normal training sample set to train an isolated forest model containing multiple isolated trees. Input the fused representation vector at the current moment into each isolated tree to obtain the path length of the fused representation vector from the root node to the terminal node in each isolated tree, and calculate the average path length in all isolated trees. The expected path length adjustment coefficient is calculated based on the number of historical normal training samples randomly selected during the training of each isolated tree. The outlier score is calculated based on the average path length and the expected path length adjustment coefficient. Specifically, the outlier score is calculated by dividing the average path length by the negative value of the expected path length adjustment coefficient by 2 as the base, and the resulting power value is the outlier score.

[0022] More preferably, based on the power distribution network state estimation results and anomaly identification information, a large language model with retrieval enhancement is used to generate auxiliary decision-making candidate schemes; physical security verification is performed on the auxiliary decision-making candidate schemes, and only the auxiliary decision-making schemes that pass the verification are output. The specific method is as follows: The distribution network state estimation results, anomaly identification information, scheduling rules in the domain knowledge base, and historical similar cases are input into the retrieval-enhanced big language model to generate auxiliary decision-making candidate schemes that include at least one of on-load voltage regulation, parallel capacitor switching, distributed power reactive power regulation, flexible interconnection power regulation, and network reconfiguration suggestions. The multiple physical security checks include: Based on power flow calculation, it is determined whether the node voltage after the candidate scheme is executed meets the upper and lower voltage limits, whether the branch power flow meets the line capacity constraints, whether the switch action meets the topology connectivity and radial operation constraints, and whether the candidate scheme meets the constraints of equipment action number, action interval and protection setting. The system filters candidate solutions that simultaneously satisfy all constraints in multiple physical security checks. The filtered candidate solutions are output as the final auxiliary decision solution. If no candidate solution is found, the system returns a message indicating that no feasible solution can be generated.

[0023] This invention also proposes a large language model-enhanced distribution network multi-source data fusion system, including a data access and alignment module, a data quality assessment module, a structured coding module, a text retrieval and large language model module, a semantic constraint conversion module, a physical constraint state estimation module, an anomaly detection module, and an auxiliary decision-making and safety verification module. The data access and alignment module is used to acquire multi-source heterogeneous data of the distribution network and establish a spatiotemporal index of the multi-source heterogeneous data, which includes structured measurement data and unstructured operation and maintenance text data. The data quality assessment module is used to calculate the integrity score, accuracy score, timeliness score and semantic consistency score of each data source based on the spatiotemporal index, and to calculate the comprehensive quality score of each data source at the current time based on the integrity score, accuracy score, timeliness score and semantic consistency score. A structured coding module is used to obtain a structured state representation based on the structured measurement data; The text retrieval and large language model module is used to obtain text semantic representations based on the unstructured operation and maintenance text data, and to perform knowledge verification on the text semantic representations based on the retrieval-enhanced large language model. The semantic constraint transformation module is used to convert the verified results into semantic constraint terms, topology correction terms, or pseudo-observation terms for state estimation. The physical constraint state estimation module is used to construct a joint weighted matrix of quality and uncertainty based on the comprehensive quality score of each data source and the prediction uncertainty of the corresponding observation. The structured state representation is used as the observation input, and the semantic constraint term, topology correction term or pseudo-observation term is used as the constraint term or observation term and introduced into the physical constraint state estimation model to obtain the distribution network state estimation result. An anomaly detection module is used to calculate a comprehensive anomaly score based on the power distribution network state estimation results, and output anomaly discrimination information in combination with the uncertainty quantification results; The auxiliary decision-making and safety verification module is used to generate auxiliary decision-making candidate schemes based on the power distribution network state estimation results and anomaly identification information using a retrieval-enhanced large language model, perform physical safety verification on the auxiliary decision-making candidate schemes, and output only the auxiliary decision-making schemes that pass the verification.

[0024] The present invention also proposes a computer device, including a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein the processor executes the computer program to implement the above-described method.

[0025] The present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention unifies structured measurement and unstructured operation and maintenance text into the same fusion framework, which can transform work orders, alarms, maintenance and scheduling semantics that cannot be directly used in traditional state estimation into structured constraints that can be used for state perception and anomaly identification, thereby improving the information utilization rate of multi-source fusion.

[0027] 2. This invention corrects the weight allocation of each data source in real time through a dynamic data quality assessment mechanism, no longer assuming that the reliability of different data sources is constant, thus maintaining high fusion accuracy even under conditions of missing data, drift, sudden noise increase, and time delay changes.

[0028] 3. This invention limits the output of the retrieval-enhanced large language model to structured intermediate results, and further converts these structured intermediate results into semantic constraint terms, topology correction terms, or pseudo-observation terms, so that unstructured operation and maintenance semantics can enter the physical constraint state estimation model together with structured measurements, thereby improving the state perception capability under conditions of weak observability, missing measurements, and inter-source conflicts.

[0029] 4. This invention can output status results with confidence intervals and support operation and maintenance decisions with anomaly scores, coverage and explanatory text, thereby forming a closed loop of "perception-diagnosis-recommendation".

[0030] 5. In the experimental scenarios described in this specification, the present invention, through joint weighting of quality and uncertainty, semantic constraint injection, and physical security filtering, demonstrates higher state estimation accuracy and anomaly identification performance compared to baseline methods such as fixed-weight fusion, EKF+pseudo-measurement, and DS evidence fusion under conditions of missing measurements, communication delays, and semantic conflicts. At the same time, by verifying the execution flow, topology, boundary violations, and action executability of candidate operation suggestions generated by the large language model, the operational risks caused by text-based decision-making without physical constraints can be reduced. Attached Figure Description

[0031] Figure 1 This is a flowchart of a method for multi-source data fusion in a power distribution network with large language model enhancement according to the present invention. Figure 2 This is a flowchart of the overall method of the present invention.

[0032] Figure 3 This is a flowchart of the collaborative decision-making process of "search enhancement - physical constraints - security filtering" in this invention.

[0033] Figure 4 This is a comparison chart of the RMSE of voltage amplitudes for different methods under different measurement missing rates.

[0034] Figure 5 This is a comparison chart of coverage and interval width for different uncertainty methods.

[0035] Figure 6 This is a comparison chart of Precision, Recall, and F1 scores for different anomaly detection methods.

[0036] Figure 7 A comparison chart showing the voltage limit elimination rate, average computation time delay, and safe executability rate of different auxiliary decision-making methods.

[0037] Figure 8 This is a diagram showing the ablation experiment results of this invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0039] like Figure 1 As shown, this invention proposes a method for multi-source data fusion in power distribution networks enhanced by a large language model, specifically including the following steps: Acquire multi-source heterogeneous data of the power distribution network and establish a spatiotemporal index of the multi-source heterogeneous data, which includes structured measurement data and unstructured operation and maintenance text data; The structured measurement data includes at least three of the following: SCADA measurement data, synchronous phasor measurement data, AMI data, online equipment monitoring data, meteorological data, and GIS / topology data. The unstructured operation and maintenance text data includes at least one of the following: work order text, alarm text, defect record, scheduling record, and maintenance record.

[0040] Based on the spatiotemporal index, the integrity score, accuracy score, timeliness score, and semantic consistency score of each data source are calculated respectively. Based on the integrity score, accuracy score, timeliness score, and semantic consistency score, the comprehensive quality score of each data source at the current moment is calculated. The integrity score is calculated by dividing the number of data points actually received within a preset time window by the number of data points that should have been received within that time window. The accuracy score is calculated as follows: the absolute value of the difference between each observation and the corresponding reference prediction value is divided by a preset accuracy threshold to obtain a ratio result, and the smaller value between 1 and the ratio result is taken; the mean of the smaller values ​​corresponding to all observations is calculated, and the accuracy score is obtained by subtracting the mean from 1. The timeliness score is calculated as follows: the product of the delay attenuation coefficient and the current time delay is calculated, and the negative of the product is taken as the input of the exponential function to calculate the timeliness score. The semantic consistency score is calculated as follows: calculate the cosine similarity between the current semantic vector and the predicted semantic vector of the normal pattern, add 1 to the cosine similarity and divide by 2 to obtain the semantic consistency score.

[0041] The semantic consistency score is calculated as follows: extract the device number, topology node, timestamp, switch status and over-limit alarm corresponding to the measurement data, and cross-match it with the domain knowledge base and other data sources; if the match is consistent, it is assigned a value of 1.0; if there is a clear conflict, it is assigned a value of 0; if there is partial field conflict, it is assigned a value by linear interpolation based on the proportion of the number of conflicting fields to the total number of matched fields.

[0042] The method for calculating the overall quality score is as follows: a completeness weight, an accuracy weight, a timeliness weight, and a semantic consistency weight are preset. The completeness score is multiplied by the completeness weight, the accuracy score is multiplied by the accuracy weight, the timeliness score is multiplied by the timeliness weight, and the semantic consistency score is multiplied by the semantic consistency weight, and then the results are added together to obtain the overall quality score; wherein, the sum of the completeness weight, accuracy weight, timeliness weight, and semantic consistency weight is 1.

[0043] A structured state representation is obtained based on the structured measurement data; a textual semantic representation is obtained based on the unstructured operation and maintenance text data. The calculation method for the structured state representation is as follows: feature extraction is performed on the structured measurement data through a numerical coding network to obtain the structured state representation; The text semantic representation calculation method is as follows: entity extraction and event extraction are performed on the unstructured operation and maintenance text data to obtain intermediate results containing device objects, event types, event times, topology status, operation intentions and text confidence; the intermediate results are semantically encoded through a retrieval-enhanced large language model to obtain the text semantic representation. The structured state representation and the text semantic representation are fused through an attention mechanism to obtain a fused representation vector. The fused representation vector is used for anomaly detection, retrieval query and generation of auxiliary decision context, and is not directly used as the observation input of the physical constraint state estimation model.

[0044] The semantic representation of the text is verified based on a large language model with retrieval enhancement, and the verified results are converted into semantic constraints, topological corrections or pseudo-observations for state estimation. The computational method for knowledge verification of the text semantic representation using the retrieval-enhanced large language model is as follows: Retrieve topology constraints, equipment ledgers, protection settings, scheduling rules, and historical similar cases that match the current moment from the domain knowledge base. Input the retrieval results and the text semantic representation into the large language model, and output a structured text result that has been verified by knowledge constraints. The structured text result includes equipment object field, event type field, time constraint field, topology status field, operation intention field, and text confidence field.

[0045] The method for converting the verified result into semantic constraint terms for state estimation is as follows: When the structured text result represents node injection power, device operating status, or the time of occurrence of abnormal events, it is mapped to a semantic constraint term for solving the constraint state estimation. The method for converting the verified result into a topology correction term for state estimation is as follows: When the structured text result represents the status of switch opening / closing, line outage, equipment maintenance, or equipment commissioning / decommissioning, it is mapped to a topology correction term used to correct the topology connection relationship in the state estimation model. The method for converting the verified results into pseudo-observations for state estimation is as follows: When the structured text results represent missing measurements, low-confidence measurements, or load changes confirmed by work orders, they are mapped to pseudo-observations used to supplement the state estimation input.

[0046] Based on the comprehensive quality score of each data source and the prediction uncertainty of the corresponding observation, a joint weighted matrix of quality and uncertainty is constructed. The structured state representation is used as the observation input, and the semantic constraint term, topology correction term, or pseudo-observation term is used as the constraint term or observation term and introduced into the physical constraint state estimation model to obtain the distribution network state estimation result. The specific method for obtaining the distribution network state estimation results is as follows: The overall quality scores of each data source are truncated to upper and lower limits to obtain the truncated reliability weights; the standard deviation of the prediction error corresponding to each observation is calculated based on the historical effective prediction residuals of each observation before the current time; the square of the standard deviation of the prediction error is summed with the corresponding lower limit of variance, the reliability weight is divided by the summation result, and the quotient is used as the diagonal element to construct a quality-uncertainty joint weighting matrix composed of the diagonal elements; The structured state representation is input into the observation function of the physical constraint state estimation model to obtain the expected observation value; Construct a physical constraint equation, and add the physical constraint equation, the semantic constraint term, the topology correction term, and the pseudo-observation term as penalty terms into the optimization objective function of the physical constraint state estimation model; The state variable vector of the distribution network is obtained by minimizing the optimization objective function, and is used as the state estimation result of the distribution network.

[0047] The comprehensive anomaly score is calculated based on the power distribution network state estimation results, and anomaly discrimination information is output by combining the uncertainty quantification results. The specific method is as follows: Calculate the measurement residual vector corresponding to the power distribution network state estimation result, and calculate the product of the measurement residual vector and the inverse of the residual covariance matrix to obtain the weighted residual term; calculate the cosine similarity between the current text embedding vector and the normal pattern predicted text embedding vector, and subtract the cosine similarity from 1 to obtain the text semantic bias term; obtain the anomaly isolation score; assign weight coefficients to the weighted residual term, the text semantic bias term, and the anomaly isolation score respectively, and then perform a weighted summation to obtain the comprehensive anomaly score; The mean and variance of the state estimation results are calculated through multiple state estimation sub-models. The conformal calibration quantiles are calculated based on the historical calibration dataset. The prediction interval at a preset confidence level is generated based on the mean, variance and conformal calibration quantiles as the uncertainty quantification result. The comprehensive anomaly score is compared with a preset anomaly judgment threshold, and anomaly discrimination information including anomaly label, anomaly type, scope of impact, and confidence level is generated based on the comparison result.

[0048] The method for obtaining the isolated anomaly score includes: Based on historical operation records, select fusion representation vectors that are in normal operation to form a historical normal training sample set, and use the historical normal training sample set to train an isolated forest model containing multiple isolated trees. Input the fused representation vector at the current moment into each isolated tree to obtain the path length of the fused representation vector from the root node to the terminal node in each isolated tree, and calculate the average path length in all isolated trees. The expected path length adjustment coefficient is calculated based on the number of historical normal training samples randomly selected during the training of each isolated tree. The outlier score is calculated based on the average path length and the expected path length adjustment coefficient. Specifically, the outlier score is calculated by dividing the average path length by the negative value of the expected path length adjustment coefficient by 2 as the base, and the resulting power value is the outlier score.

[0049] Based on the power distribution network state estimation results and anomaly identification information, a large language model with retrieval enhancement is used to generate auxiliary decision candidate schemes; physical security verification is performed on the auxiliary decision candidate schemes, and only auxiliary decision schemes that pass the verification are output.

[0050] This invention also proposes a large language model-enhanced distribution network multi-source data fusion system, including a data access and alignment module, a data quality assessment module, a structured coding module, a text retrieval and large language model module, a semantic constraint conversion module, a physical constraint state estimation module, an anomaly detection module, and an auxiliary decision-making and safety verification module. The data access and alignment module is used to acquire multi-source heterogeneous data of the distribution network and establish a spatiotemporal index of the multi-source heterogeneous data, which includes structured measurement data and unstructured operation and maintenance text data. The data quality assessment module is used to calculate the integrity score, accuracy score, timeliness score and semantic consistency score of each data source based on the spatiotemporal index, and to calculate the comprehensive quality score of each data source at the current time based on the integrity score, accuracy score, timeliness score and semantic consistency score. A structured coding module is used to obtain a structured state representation based on the structured measurement data; The text retrieval and large language model module is used to obtain text semantic representations based on the unstructured operation and maintenance text data, and to perform knowledge verification on the text semantic representations based on the retrieval-enhanced large language model. The semantic constraint transformation module is used to convert the verified results into semantic constraint terms, topology correction terms, or pseudo-observation terms for state estimation. The physical constraint state estimation module is used to construct a joint weighted matrix of quality and uncertainty based on the comprehensive quality score of each data source and the prediction uncertainty of the corresponding observation. The structured state representation is used as the observation input, and the semantic constraint term, topology correction term or pseudo-observation term is used as the constraint term or observation term and introduced into the physical constraint state estimation model to obtain the distribution network state estimation result. An anomaly detection module is used to calculate a comprehensive anomaly score based on the power distribution network state estimation results, and output anomaly discrimination information in combination with the uncertainty quantification results; The auxiliary decision-making and safety verification module is used to generate auxiliary decision-making candidate schemes based on the power distribution network state estimation results and anomaly identification information using a retrieval-enhanced large language model, perform physical safety verification on the auxiliary decision-making candidate schemes, and output only the auxiliary decision-making schemes that pass the verification.

[0051] The present invention also proposes a computer device, including a memory, a processor, and a computer program stored in the memory and executable by the processor, wherein the processor executes the computer program to implement the above-described method.

[0052] The present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0053] Example 1 The technical solution of this application is described in detail below with reference to a specific implementation scenario. This embodiment takes a typical medium-voltage radial distribution network as the application object and aims to solve the problems of insufficient measurement coverage, communication quality fluctuations, difficulty in utilizing unstructured operation and maintenance semantics, and lack of physical security closed loop for auxiliary decision-making under the background of high proportion of distributed power source access. This embodiment sets the unified fusion cycle to 1 minute and adopts multi-source heterogeneous data streams including SCADA, synchronous phasor measurement device (PMU), advanced measurement system (AMI), online equipment monitoring, meteorological information, GIS / topology data, and work order text. The following steps are used to achieve accurate perception and safe decision-making of the distribution network operation status.

[0054] like Figure 2 As shown, this invention proposes a method for multi-source data fusion in power distribution networks enhanced by a large language model, specifically including the following steps: Step S1: Obtain multi-source heterogeneous data of the distribution network and establish a unified spatiotemporal index.

[0055] The multi-source heterogeneous data includes structured measurement data and unstructured operation and maintenance text data; wherein, the structured measurement data includes at least three of the following: SCADA measurement data, synchronous phasor measurement data, AMI data, equipment online monitoring data, meteorological data, and GIS / topology data; and the unstructured operation and maintenance text data includes at least one of the following: work order text, defect record, alarm text, scheduling record, and maintenance record, and is time-aligned, resampled, and marked for missing data according to a unified main time axis.

[0056] Let the set of multi-source observations at time t be...

[0057] In the formula, Represents the set of multi-source observations at time t; Let m represent the original input vector of the m-th data source at time t; M represents the total number of data sources.

[0058] After alignment and encoding, there is

[0059] In the formula, This represents the encoded representation vector of the m-th data source at time t; This represents the encoding function corresponding to the m-th data source; This represents the original input vector of the m-th data source at time t; This represents the set of parameters for the m-th encoding function.

[0060] This embodiment assumes the target object is a medium-voltage radial distribution network and uses a 1-minute main time axis for unified fusion. The data sources accessed include SCADA, synchronization phasors, AMI, online equipment monitoring, meteorology, GIS / topology, and work order text. For synchronization phasor data, the mean, standard deviation, maximum frequency change rate, and distortion summary features are first calculated within a 1-minute window, and then aligned to the main time axis. For AMI data, forward hold and time stamp interpolation are used. For work order text, knowledge retrieval and entity extraction are used to generate equipment objects, operation types, fault types, and confidence labels.

[0061] By establishing the unified spatiotemporal index, the heterogeneity of multi-source data in the temporal and spatial dimensions is eliminated, providing a standardized input basis for subsequent dynamic quality assessment of different data sources. This enables the system to accurately identify the reliability status of each data source at a specific moment based on a consistent spatiotemporal benchmark.

[0062] Step S2: Perform dynamic data quality assessment.

[0063] To differentiate the reliability differences of each data source at the current moment and to provide a basis for subsequent fusion weight allocation, this step quantifies and scores each data source from four dimensions: completeness, accuracy, timeliness, and semantic consistency, and then synthesizes a comprehensive reliability score based on these scores.

[0064] First, define an integrity score to measure the completeness of data reception:

[0065] In the formula, This represents the integrity score of the m-th data source at time t. This represents the number of valid and deduplicated data points actually received by the m-th data source at time t within a preset time window. This represents the number of data points that the m-th data source should receive within a preset time window at time t. 1 indicates taking the smaller value from the input parameters; 1 indicates the upper limit of the integrity score.

[0066] Before counting, the actual number of received data points is deduplicated according to the data source identifier, measurement channel identifier, and sampling timestamp to avoid abnormally high integrity scores caused by retransmission of data.

[0067] The start and end times of the preset time window are aligned with the nominal sampling period of the corresponding measurement channel; for sliding time windows that are not completely aligned with the nominal sampling time, the number of data points to be received is calculated based on the planned sampling time registered in the data acquisition configuration table.

[0068] The number of data points that the m-th data source should receive within the preset time window at time t is determined based on the measurement channel configuration, measurement channel operating status, nominal sampling period, and preset time window length of the data source.

[0069] Specifically, the number of data points to be received is calculated according to the following formula:

[0070] In the formula, This represents the number of data points that the m-th data source should receive within the preset time window at time t; t represents the current data quality assessment time; m represents the data source number; Km represents the total number of measurement channels contained in the m-th data source; and k represents the measurement channel number. This indicates the valid status identifier of the k-th measurement channel in the m-th data source at time t; Indicates the preset time window length; This represents the nominal sampling period of the k-th measurement channel in the m-th data source; This indicates the floor function.

[0071] The valid status identifier is determined according to the following formula:

[0072] In the formula, This represents the valid status identifier of the k-th measurement channel in the m-th data source at time t; t represents the current data quality assessment time; m represents the data source number; and k represents the measurement channel number. When the measurement channel is in normal operating condition... When the measurement channel is under planned maintenance, equipment shutdown, communication isolation, or manually withdrawn from data acquisition, .

[0073] When all valid measurement channels of the m-th data source have the same nominal sampling period, the number of data points that the m-th data source should receive within the preset time window at time t can be simplified as:

[0074] This represents the number of valid measurement channels that should be uploading data normally in the m-th data source at time t.

[0075] The number of effective measurement channels is calculated according to the following formula:

[0076] For example, preset time window length The sampling period is 60 seconds. The m-th data source has 20 measurement channels, of which 18 measurement channels are in normal operation and 2 measurement channels are in planned maintenance. The nominal sampling period of each normally operating measurement channel is... If the time interval is 5 seconds, then the number of data points that the m-th data source should receive within the preset time window at time t is:

[0077] The total number of measurement channels, nominal sampling period, and measurement channel operating status can be obtained from the data acquisition configuration table, equipment ledger, communication configuration, maintenance plan, and equipment commissioning / decommissioning records. By excluding measurement channels that are under planned maintenance, equipment outage, communication isolation, or manually deactivated, it is possible to avoid misclassifying normal planned data failures as data missing.

[0078] Secondly, an accuracy score is defined to measure the degree of deviation between the measured value and the expected value:

[0079] In the formula, This represents the accuracy score of the m-th data source at time t; This represents the observation dimension of the m-th data source; This represents the k-th observation from the m-th data source at time t; This represents the reference predicted value of the k-th observation from the m-th data source at time t; This represents the accuracy threshold for the k-th observation from the m-th data source.

[0080] Wherein, the reference predicted value The method employs a single exponential smoothing approach, calculated based on historical observation data that have been confirmed to be normal up to the current time.

[0081] For the k-th observation from the m-th data source at time t, set a valid historical observation identifier:

[0082] In the formula, Represents time t 1. The historical observation valid identifier of the k-th observation from the m-th data source; t represents the current data quality assessment time; t 1 represents the time preceding the current data quality assessment; m represents the data source number; k represents the observation number; Represents time t 1. The m-th data source and the k-th observation.

[0083] When the historical observation data is complete, the timestamp is valid, it does not exceed the reasonable upper and lower limits of the corresponding physical quantity, and it is not judged as abnormal by the bad data detection program, it is determined that... When the historical observations contain missing data, communication errors, abnormal timestamps, physical limitations, or confirmed anomalies, the following conditions are determined: .

[0084] The reference prediction value of the k-th observation from the m-th data source at time t is calculated using the following formula:

[0085] In the formula, This represents the reference predicted value of the k-th observation from the m-th data source at time t; t represents the current data quality assessment time; m represents the data source number; and k represents the observation number. This represents the smoothing coefficient corresponding to the k-th observation from the m-th data source; Represents time t 1. The actual observed value of the k-th observation from the m-th data source; Represents time t 1. Reference predicted value of the k-th observation from the m-th data source; Represents time t 1. The historical observation valid identifier of the m-th data source and the k-th observation.

[0086] The smoothness coefficient satisfy:

[0087] The smoothness coefficient The preferred value range is 0.3 to 0.8. In this embodiment, the smoothing coefficient... We set the value to 0.6. The larger the smoothing coefficient, the more sensitive the reference forecast is to changes in the most recent valid observation; the smaller the smoothing coefficient, the more stable the changes in the reference forecast.

[0088] The initial value of the reference prediction value is determined according to the following formula:

[0089] This indicates the number of normal historical observations selected during the initialization phase; n represents the sequence number of the normal historical observation. This represents the nth normal historical observation during the initialization phase.

[0090] In this embodiment, the initialization phase selects the 30 most recent historical observations that have been confirmed to be normal, namely:

[0091] In the formula, This represents the number of normal historical observations selected for the k-th observation from the m-th data source during the initialization phase; m represents the data source number; k represents the observation number; and 30 represents the number of normal historical observations selected. When the number of normal historical observations obtainable during the initialization phase is less than 30 but greater than zero, the initial reference prediction is calculated using all obtainable normal historical observations, and the actual number of normal historical observations is used as the reference value. When there are no normal historical observation values, the rated operating value registered in the equipment ledger, the effective value of the same historical period, or the initial value configured by the acquisition device shall be used as the initial reference prediction value.

[0092] For example, let time t 1. The k-th observation from the m-th data source is a valid observation, and its value is:

[0093] Let the reference prediction value at the previous time step be:

[0094] When the smoothing coefficient At that time, the reference prediction value is:

[0095] At time t When an observation of value 1 is missing or is judged as outlier, the following applies:

[0096] At this point, the reference prediction value for the current moment remains the same as the reference prediction value for the previous moment:

[0097] The reference prediction value is the actual observed value at the current moment. The actual observations at the current moment are generated before they arrive and participate in the accuracy evaluation. Not included in the reference forecast value This allows for more precise calculations, thus avoiding cyclical calculations.

[0098] Define a timeliness score to measure the impact of time delay in data:

[0099] In the formula, This represents the timeliness score of the m-th data source at time t; This represents the latency attenuation coefficient of the m-th data source; This represents the time delay of the m-th data source at time t.

[0100] Specifically, regarding semantic consistency, cosine similarity is used for quantification of unstructured operational text data; for structured measurement data, cross-source cross-matching is employed.

[0101] In the formula, The semantic consistency score of the m-th data source at time t is represented. This represents the current semantic vector of the m-th data source at time t; This represents the predicted semantic vector of the m-th data source at time t in normal mode; This represents the cosine similarity.

[0102] Specifically, for structured measurement data, the aforementioned semantic consistency score... The acquisition process specifically involves: extracting the device number, topology node ID, timestamp, switch position signal, and over-limit alarm label from the measurement data; cross-matching this data with the device ledger, topology connection relationships, planned maintenance schedules, and real-time status of other data sources from the same source in the domain knowledge base; if multiple matching results are consistent and there are no logical conflicts, the semantic consistency score approaches the upper limit of 1.0; if there is a clear conflict (e.g., an unstructured text record shows the switch is open but the SCADA remote signaling is closed, or the device's online monitoring temperature exceeds the limit but the measurement value is normal), then the data source is determined to be semantically inconsistent at this moment, and the semantic consistency score is [not specified]. The lower limit value of 0 is assigned, indicating that it is completely unreliable; for the intermediate state where some fields are matched successfully but there are individual attribute conflicts, the corresponding low score between 0 and 0.3 is assigned by linear interpolation based on the proportion of the number of conflicting fields to the total number of matched fields.

[0103] Finally, the scores from the four dimensions mentioned above are weighted and merged according to preset weights to obtain the overall quality score:

[0104] In the formula, This represents the overall quality score of the m-th data source at time t; , , , These represent the weighting coefficients for completeness, accuracy, timeliness, and semantic consistency, respectively. Indicates accuracy score; Indicates timeliness score; This represents the semantic consistency score.

[0105] The resulting overall quality score This not only reflects the real-time health status of each data source at the current moment, but also serves as a core weighting criterion in the subsequent multi-source fusion representation process. It distinguishes between high-confidence and low-confidence data at the feature level, thereby suppressing the interference of low-quality data on the state perception results. This comprehensive quality score... This will serve as the core input for attention weight calculation and state estimation weighting matrix construction in subsequent steps, enabling data quality-driven adaptive fusion.

[0106] Step S3: Construct a structured-unstructured dual-channel fusion representation.

[0107] After completing the quality assessment of each data source, in order to overcome the shortcomings of traditional fixed-weight fusion in adapting to the dynamic changes in data source quality, this step adopts a dual-channel architecture to extract features from structured measurement data and unstructured operation and maintenance text respectively, and adaptively aggregates the representation vectors of each channel based on the attention mechanism. The attention weight is determined jointly based on the encoded representation vector of each data source, the comprehensive quality score and the data source-level quality uncertainty description vector.

[0108] For structured measurement data, a numerical coding network is used to extract state representation vectors; for unstructured operation and maintenance text data, a retrieval-enhanced large language model is used to extract textual semantic representations containing device objects, event types, topology status, and operational intentions. After alignment, the two types of representations are fused using the following formula to generate a unified fused representation vector:

[0109] In the formula, The fusion representation vector at time t represents the total number of data sources; M represents the total number of data sources. This represents the attention weight of the m-th data source at time t; This represents the encoded representation vector of the m-th data source at time t.

[0110] To overcome the limitation of traditional fixed-weight fusion methods in adapting to dynamic changes in data source quality, this embodiment employs an attention mechanism for adaptive aggregation of multi-source encoded representations. Specifically, the encoded representation vectors of each data source are... Overall quality score and uncertainty description vector Using all data sources as input, the contribution of each data source in the final fused representation is dynamically adjusted. For unstructured operational text data, to further enhance the accuracy and domain adaptability of semantic understanding, this embodiment introduces a retrieval-enhanced generation (RAG) mechanism to retrieve relevant context from the domain knowledge base to assist the large language model in reasoning.

[0111] Attention weights are calculated as follows:

[0112] In the formula, This represents the attention weight of the m-th data source at time t; The attention score for the m-th data source at time t is represented. Represents the attention scoring parameter vector; This indicates vector concatenation; This represents the encoded representation vector of the m-th data source at time t; This represents the overall quality score of the m-th data source at time t; This represents the uncertainty description vector of the m-th data source at time t.

[0113] Wherein, the uncertainty description vector This is a data source-level quality uncertainty description vector, used to describe the degree of uncertainty of the m-th data source at time t in four dimensions: data integrity, data accuracy, data timeliness, and semantic consistency.

[0114] The uncertainty description vector is determined based on the completeness score, accuracy score, timeliness score, and semantic consistency score obtained in step S2, and is specifically expressed as follows:

[0115] In the formula, This represents the data source-level quality uncertainty description vector for the m-th data source at time t; t represents the current fusion time; and m represents the data source number. This represents the integrity uncertainty component of the m-th data source at time t; This represents the accuracy uncertainty component of the m-th data source at time t; This represents the timeliness uncertainty component of the m-th data source at time t; Represents the semantic uncertainty component of the m-th data source at time t; This represents the integrity score of the m-th data source at time t. This represents the accuracy score of the m-th data source at time t; This represents the timeliness score of the m-th data source at time t; Represents the semantic consistency score of the m-th data source at time t; superscript The corresponding column vector form represents the arrangement of each uncertainty component in a preset order.

[0116] Due to integrity score Accuracy score Timeliness score and semantic consistency score The range of values ​​for all are:

[0117] In the formula, This represents the integrity score of the m-th data source at time t. This represents the accuracy score of the m-th data source at time t; This represents the timeliness score of the m-th data source at time t; This represents the semantic consistency score of the m-th data source at time t; 0 represents the lowest score for the corresponding quality dimension; and 1 represents the highest score for the corresponding quality dimension.

[0118] Therefore, the uncertainty components satisfy:

[0119] In the formula, Represents the integrity uncertainty component; Indicates the accuracy uncertainty component; Indicates the time-related uncertainty component; The value represents the semantic uncertainty component; 0 indicates the lowest degree of uncertainty in the corresponding quality dimension; 1 indicates the highest degree of uncertainty in the corresponding quality dimension.

[0120] The higher the completeness score, accuracy score, timeliness score, or semantic consistency score, the smaller the corresponding uncertainty component; conversely, the lower the corresponding quality score, the larger the corresponding uncertainty component.

[0121] Overall quality score The weighted summary of completeness score, accuracy score, timeliness score, and semantic consistency score is used to represent the overall reliability of the m-th data source; uncertainty description vector. The uncertainty of each of the four quality dimensions is retained to enable the attention scoring model to distinguish different types of data quality problems.

[0122] For example, when the integrity score of the m-th data source at time t is 0.95, the accuracy score is 0.80, the timeliness score is 0.90, and the semantic consistency score is 0.70, its uncertainty description vector is:

[0123] This indicates that the data source exhibits the highest degree of uncertainty in the semantic consistency dimension and the lowest degree of uncertainty in the integrity dimension. The attention scoring model can adjust the weights of data sources with different types of quality defects based on the differences in each component.

[0124] It should be noted that the uncertainty description vector... This is a data source-level quality descriptor, not a statistically significant prediction standard deviation. (Step S5) The standard deviation of the prediction for the k-th observation from the m-th data source is used to construct the joint weighted matrix of quality and uncertainty. The uncertainty quantification result in step S6 represents the prediction interval of the distribution network state estimation result. The three differ in their calculation objects, application stages, and technical roles.

[0125] Given the current query vector and knowledge base vector set The search result set is defined as follows:

[0126] In the formula, Represents the set of search results at time t; The query vector at time t; This represents the vector representation of the i-th record in the knowledge base; Indicates the total number of records in the knowledge base; Represents the similarity function; This means returning the top K records with the highest similarity.

[0127] The LLM generation distribution is defined as:

[0128] In the formula, The parameter is The probability of generating large language models; This represents the output sequence at time t; The structured context input represents time t; Represents the set of search results at time t; Represents the set of constraints at time t; The sequence length at time t; Indicates the first One output token; This indicates the preceding output token sequence.

[0129] If low-rank adaptation is used, the weight update is written as:

[0130] In the formula, This represents the updated weight matrix; Represents the original weight matrix that is frozen; Represents a low-rank increment matrix; Represents the projection matrix; Let r denote the lower projection matrix; r denotes the low-rank fitting rank.

[0131] The structured state representation and textual semantic representation obtained through the dual-channel fusion already contain rich operational information, but they still need to be further transformed into mathematical constraints that the state estimation model can parse. Therefore, the structured text results output by the retrieval-enhanced large language model will be used, combined with topological and device rules from the domain knowledge base, to generate constraints for correcting model biases.

[0132] The resulting fusion characterization It simultaneously contains both high-quality numerical features of measurements and semantic features of operational text, but it still exists in the form of latent vectors and cannot be directly parsed by the state estimation model. Therefore, subsequent steps will further utilize a large language model to transform the text semantic representation into an explicit mathematical constraint form.

[0133] It should be noted that the fusion representation vector This is a fusion latent vector in the feature space. Unlike observations with definite physical dimensions such as voltage, current, active power, and reactive power, it is not directly used as the observation input for weighted least squares state estimation. The basic observation input for physical constraint state estimation is a structured measurement vector aligned with a unified spatiotemporal index; the fusion representation vector... It is used for anomaly identification in isolated forests, knowledge retrieval query construction, and auxiliary decision context generation.

[0134] Step S4: Perform retrieval enhancement and knowledge constraint injection.

[0135] Since the text semantic representation output in step S3 is still in implicit vector form and cannot be directly used for state estimation based on physical equations, this step introduces a domain knowledge base and retrieval enhancement generation mechanism to perform knowledge verification and structured parsing on the text semantic representation, and explicitly convert it into topological correction terms, semantic constraint terms or pseudo-observation terms that can be recognized by the state estimation model.

[0136] The topology constraints, equipment ledger, protection settings, scheduling rules, and historical similar cases corresponding to the current moment are written into the domain knowledge base. Vector retrieval is performed on the current query to obtain a set of candidate knowledge fragments, which are then fed into a large language model to form a knowledge-constrained structured text result. The structured text result is not directly used as the final running conclusion, but is converted into semantic constraint terms, topology correction terms, or pseudo-observation terms for the state estimation model. Among them, the topology correction term is used to correct the switch state, line connection relationship, or equipment commissioning / decommissioning state; the semantic constraint term is used to constrain node injected power, equipment operating state, abnormal event occurrence time, or fault type; and the pseudo-observation term is used to supplement missing or low-reliability measurements.

[0137] In the text semantic constraint conversion process, work order texts, alarm texts, defect records, or scheduling records are first parsed into structured text results. These structured text results include equipment objects, event types, event times, topology status, operational intentions, and text confidence levels. Subsequently, based on the mapping relationship between equipment objects and nodes, lines, switches, transformers, or reactive power compensation devices in the GIS / topology data, the structured text results are converted into constraint forms that can be invoked by the state estimation model.

[0138] Specifically, when structured text results represent switch opening / closing, line outage, equipment maintenance, or equipment commissioning / decommissioning status, they are converted into topology correction terms; when structured text results represent fault type, anomaly occurrence time, equipment operating status, or operational intent, they are converted into semantic constraint terms; when structured text results represent missing measurements, low-reliability measurements, or load changes confirmed by work orders, they are converted into pseudo-observations. The aforementioned topology correction terms, semantic constraint terms, and pseudo-observations are uniformly represented as... and with electrical physical constraints Joint participation in state estimation.

[0139] For example, when the structured text result is "Switch K12 is open, text confidence level is 0.92", the system determines the branch number corresponding to switch K12 based on the equipment ledger and GIS / topology mapping relationship, and constrains the conduction state variable of the branch to 0, forming a topology correction term; when the structured text result is "Transformer area T3 load suddenly increases, text confidence level is 0.88", the system adds the range of change of active power injection of the corresponding node as a pseudo-observation term to the state estimation model based on the mapping relationship between transformer area T3 and topology node number; when the structured text result is "Feeder F2 has a grounding alarm, event time is t", the system adds the abnormal time constraint, residual weight adjustment amount or abnormal event label of the related node or branch of feeder F2 as a semantic constraint term to the anomaly discrimination model.

[0140] On the text side, using a retrieval-enhanced generation framework, the conditional probability can be written as:

[0141] In the formula, The parameter is The probability of generating large language models; The text output sequence representing time t; The structured context input represents time t; Represents the set of search results at time t; Represents the set of physical and operational constraints at time t; Indicate the length of the output sequence at time t; Indicates the output sequence at time t. One token; Indicates the first The output before each token.

[0142] Through the above transformation, the operational semantic information that originally existed in isolation in the text is systematically mapped into topology correction terms in the state estimation model. (Used to correct switch states and connection relationships), semantic constraints (used to constrain power variation range and abnormal time), and spurious observations (used to supplement missing measurements). These constraints will be combined with traditional electrophysical constraints. Together, they constitute an augmented optimization problem, thereby significantly improving the accuracy and robustness of state estimation in weakly observable scenarios under the dual guidance of physical laws and operational experience.

[0143] The structured text results generated by the aforementioned retrieval-enhanced large language model are not directly output as the final result. Instead, they serve as a bridge connecting unstructured semantics and the structured physical model, participating sequentially in subsequent state estimation optimization and security decision filtering. The overall logic and information flow of the "retrieval enhancement—physical constraint—security filtering" collaborative decision-making mechanism proposed in this application can be found in the appendix to the specification. Figure 3 .like Figure 3 As shown, the operation and maintenance semantics after knowledge verification are first converted into topology correction terms, semantic constraint terms, or pseudo-observation terms, and then incorporated into the physical constraint state estimation model as augmented constraints to achieve a highly robust perception of the distribution network operation status. Subsequently, based on the state estimation results and anomaly identification conclusions, the large language model generates auxiliary decision candidate schemes, and before output, it enforces multiple physical security filters such as power flow verification, topology feasibility verification, and protection setting verification, thereby forming a complete closed loop of "semantic understanding - physical optimization - security adjudication", fundamentally avoiding the operational risks caused by unconstrained text generation.

[0144] Step S5: Perform quality-aware physical constraint state estimation.

[0145] Based on the comprehensive quality score obtained in step S2 and the semantic constraint term obtained in step S4, this step constructs a quality-aware physical constraint state estimation model. This model, building upon the traditional weighted least squares framework, introduces data source quality weights and textual semantic constraints, thus maintaining high-accuracy estimation even under conditions of missing measurements, sudden noise increases, or inter-source conflicts.

[0146] Assume the hidden state of the distribution network is Fusion observations The state estimation problem can then be formalized into the following optimization objective:

[0147] In the formula, the first term is the standard WLS objective function; the second term enforces physical constraints such as Kirchhoff's laws; and the third term incorporates the semantic constraints, topology correction constraints, or pseudo-observation constraints generated in step S4 into the optimization process. Specifically, Represents the estimated state vector at time t; The vector representing the state to be estimated at time t; M represents the total number of data sources; This represents the observation vector of the m-th data source at time t; This represents the observation function corresponding to the m-th data source; This represents the joint weighted matrix of quality and uncertainty for the m-th data source; Indicates the physical constraint penalty coefficient; The physical constraint matrix at time t; This represents the physical constraint constant vector at time t. This represents semantic constraints, topological correction constraints, or pseudo-observation constraints obtained from the transformation of structured text results; This represents the semantic constraint penalty coefficient.

[0148] To suppress the interference of low-quality data on the estimation results, the weighting matrix is ​​jointly determined by the overall quality score and the prediction uncertainty, and the overall quality score is truncated with upper and lower limits to avoid weight degradation or abnormal amplification.

[0149]

[0150] In the formula, This represents the joint weighted matrix of quality and uncertainty of the m-th data source at time t; This represents the overall quality score of the m-th data source at time t; This represents the reliability weight after truncation by upper and lower limits; and These represent the preset lower reliability limit and the upper reliability limit, respectively; This represents the standard deviation of the prediction for the k-th observation from the m-th data source at time t. This represents the observation dimension of the m-th data source; This represents the lower bound of the positive variance corresponding to the k-th observation from the m-th data source, and its dimensions are the same as... Similarly, this is used to prevent excessively small prediction standard deviations from causing abnormally large weights in the joint weighting matrix. Among them, Indicates when Time to take ,when Time to take In other cases, take .

[0151] Wherein, the prediction standard deviation is the standard deviation of the prediction error at the observation level, used to describe the degree of fluctuation in the historical prediction error of the k-th observation from the m-th data source before time t.

[0152] The predicted standard deviation is calculated based on the historical valid observations prior to the current time and their corresponding reference predicted values ​​before the state estimation begins. The actual observations to be included in the state estimation at the current time are not included in the calculation of the predicted standard deviation.

[0153] Specifically, for the k-th observation from the m-th data source at time t, the most recent observation before the current time is selected. There are 10 valid historical moments, which are represented sequentially as follows:

[0154] In the formula, This represents the i-th valid historical time selected for calculating the predicted standard deviation of the k-th observation from the m-th data source at time t; i represents the sequence number of the valid historical time; t represents the current state estimation time; m represents the data source number; and k represents the observation number. This indicates the number of valid historical moments selected.

[0155] For the i-th valid historical moment, its prediction residual is calculated according to the following formula:

[0156] In the formula, This represents the i-th historical prediction residual used to calculate the predicted standard deviation of the k-th observation from the m-th data source at time t; i represents the sequence number of the valid historical time; t represents the current state estimation time; m represents the data source number; and k represents the observation number. This represents the i-th valid historical moment; This represents the actual observed value of the k-th observation from the m-th data source at the i-th valid historical moment; This represents the corresponding reference prediction value obtained at the i-th valid historical moment, before the actual observations are used for accuracy evaluation.

[0157] The average value of the historical prediction residuals is calculated according to the following formula:

[0158] In the formula, This represents the average prediction residual of the k-th observation from the m-th data source at time t in the selected historical samples; t represents the current state estimation time; m represents the data source number; and k represents the observation number. This indicates the number of valid historical forecast residuals; i represents the sequence number of the valid historical forecast residuals. Let represent the residual of the i-th historical prediction.

[0159] The standard deviation of the prediction for the k-th observation from the m-th data source at time t is calculated using the following formula:

[0160] In the formula, The standard deviation of the prediction for the kth observation from the m-th data source at time t is given; t represents the current state estimation time; m represents the data source number; and k represents the observation number. This represents the number of effective historical forecast residuals, and ; i represents the sequence number of the effective historical forecast residual; This represents the residual of the i-th historical prediction. This represents the average value of historical forecast residuals.

[0161] The valid historical moment refers to a historical moment when both the actual observation and its corresponding reference prediction are valid. If the historical actual observation or its corresponding reference prediction has missing data, communication errors, abnormal timestamps, physical limitations, or has been identified as abnormal by the bad data detection program, the prediction residual corresponding to that historical moment will not be included in the calculation of the prediction standard deviation.

[0162] The number of effective historical prediction residuals The value range is from 10 to 200, and 60 is preferred in this embodiment. When the number of valid prediction residuals in the most recent historical time window is less than the preset number, the most recent valid historical prediction residuals are selected again until the preset number is reached or the maximum historical backtracking range is reached.

[0163] When the number of effective historical forecast residuals meets the following condition:

[0164] The current forecast standard deviation is the same as the forecast standard deviation of the previous time step.

[0165] In the formula, This represents the standard deviation of the prediction for the k-th observation from the m-th data source at time t. Represents time t 1. The predicted standard deviation of the k-th observation from the m-th data source; t represents the current state estimation time; m represents the data source number; k represents the observation number.

[0166] When the system is in the initialization phase and there is no previous prediction standard deviation, the nominal measurement standard deviation of the acquisition device corresponding to the m-th data source and the k-th observation is used as the initial prediction standard deviation.

[0167] For example, if the prediction residuals of a certain observation for the most recent 10 valid historical moments are as follows:

[0168] The units of the above prediction residuals are per unit values, so the mean prediction residual is 0, and the prediction standard deviation is:

[0169] The prediction standard deviation is updated on a rolling basis with the historical time window. When the historical prediction residual fluctuation of a certain observation increases, its prediction standard deviation increases, which reduces the weight of the observation in the joint weighting matrix of quality and uncertainty; when the historical prediction residual fluctuation decreases, its prediction standard deviation decreases, which increases the weight of the observation.

[0170] It should be noted that the predicted standard deviation The variance and prediction interval calculated in step S6 based on multiple state estimation sub-models belong to the observation-level prediction error statistics and are used to construct the joint weighted matrix of quality and uncertainty in step S5. The variance and prediction interval calculated in step S6 based on multiple state estimation sub-models belong to the state estimation result level uncertainty. The two have different calculation objects, calculation stages, and uses.

[0171] The unified observation model is written as:

[0172] In the formula, Represents the joint observation vector at time t; This represents the system state vector at time t; Represents the joint observation function; Represents the network topology at time t; This represents the observation noise vector at time t.

[0173] Under the linearized DistFlow approximation, it can be written as:

[0174] In the formula, The node voltage vector at time t; Represents the reference voltage vector at the slack node; The sensitivity matrix representing the active power to voltage ratio; Represents the active power injection vector at time t; The sensitivity matrix representing reactive power to voltage; This represents the reactive power injection vector at time t.

[0175] Through the aforementioned quality-aware physical constraint state estimation, a distribution network state vector that takes into account data quality, physical laws, and operational semantics is obtained. However, in actual operation, data sources may contain hidden faults or malicious attacks that are not fully captured by the quality assessment. Therefore, based on the high-precision state estimation results mentioned above, it is necessary to further conduct uncertainty quantification and deep anomaly detection to achieve early warning of potential risks.

[0176] Step S6: Perform uncertainty quantification and anomaly detection.

[0177] Although step S5 provides point results for state estimation, point estimation alone is insufficient to support high-risk decision-making in environments with high proportions of renewable energy access and strong uncertainty. Therefore, this step further quantifies the prediction uncertainty based on point estimation and integrates measurement residuals, semantic consistency bias, and data density characteristics to construct a comprehensive anomaly score, thereby achieving early warning of potential faults or network attacks.

[0178] For uncertainty quantification, a deeply integrated mean and variance calculation is employed, and a statistically guaranteed prediction interval is generated based on conformal calibration.

[0179] In the formula, B represents the ensemble mean output at time t; B represents the number of ensemble models. This represents the output of the b-th sub-model at time t; Let represent the covariance matrix of cognitive uncertainty at time t.

[0180] Then, combine conformal calibration to form the range:

[0181] In the formula, Indicates the confidence level as The prediction range; Indicates the significance level; This represents the point prediction result at time t; Indicates the calibration quantile; This represents the standard deviation estimate vector at time t.

[0182] Before calculating the overall anomaly score, an isolation score for time t is calculated using an isolation forest model. This isolation score measures the degree of deviation of the multi-source fusion representation at time t from the historical normal multi-source fusion representation distribution.

[0183] First, based on historical operation records, select fusion representation vectors that are confirmed to be in normal operation to form a historical normal training sample set:

[0184] In the formula, This represents the historical normal training sample set; This represents the multi-source fusion representation vector corresponding to the i-th historical normal moment; i represents the number of the historical normal training sample. This represents the total number of historical normal training samples; This indicates a fusion representation.

[0185] The historical normal training samples do not include data during confirmed equipment failures, communication anomalies, bad measurement data, malicious attack data, planned maintenance data, or data during distribution network topology switching transitions.

[0186] An isolated forest model containing B isolated trees is trained using the aforementioned historical normal training sample set. For each isolated tree, a sample is randomly selected from the historical normal training sample set. The training samples are recursively divided by randomly selecting feature dimensions and random selecting feature segmentation values.

[0187] Among them, the number of isolated trees B and the number of training samples for each isolated tree satisfy:

[0188] In the formula, B represents the number of isolated trees included in the isolated forest model; This represents the number of historical normal training samples randomly selected when constructing each isolated tree; This represents the total number of historical normal training samples.

[0189] In this embodiment, the number of isolated trees B ranges from 50 to 300, preferably 100; the number of training samples for each isolated tree... The value of is between 64 and 512, with 256 being the preferred value. When the total number of historical normal training samples is less than 256, let . .

[0190] At time t, the fused representation vector obtained in step S3 is... Input the B isolated trees respectively, obtain the corresponding path lengths, and calculate the average path length:

[0191] In the formula, represents the average path length of the fused representation vector in the isolated forest model at time t; t represents the current anomaly detection time; B represents the number of isolated trees in the isolated forest model; b represents the isolated tree number. The fusion representation vector at time t The length of the path from the root node of the b-th isolated tree to the terminal node.

[0192] Calculate the number of training samples for each isolated tree. Corresponding adjustment coefficients:

[0193] In the formula, This represents the number of training samples for each isolated tree. The corresponding expected path length adjustment coefficient; This represents the number of historical normal training samples randomly selected when constructing each isolated tree; Indicates the first Harmonic number.

[0194] The first The harmonic number is calculated according to the following formula:

[0195] In the formula, Indicates the first Harmonic number; represents the number of historical normal training samples randomly selected when constructing each isolated tree; j represents the harmonic summation index.

[0196] The isolated anomaly fraction at time t is calculated using the following formula:

[0197] In the formula, The score represents the number of isolated anomalies at time t; t represents the current time of anomaly detection. This represents the average path length of the fused representation vector at time t in the isolated forest model; This represents the number of training samples for each isolated tree. The corresponding expected path length adjustment coefficient.

[0198] The isolated outlier score satisfies:

[0199] In the formula, represents the isolated fraction at time t; 0 represents the theoretical lower limit of the isolated fraction; 1 represents the upper limit of the isolated fraction.

[0200] When the average path length of the fused representation vector at time t in each isolated tree is short, it indicates that the fused representation vector can be separated individually with fewer node partitions, and its abnormal isolation score is low. The large difference indicates a significant discrepancy between the current multi-source data distribution and the historical normal data distribution.

[0201] When the average path length of the fused representation vector in each isolated tree at time t is relatively long, it indicates that the fused representation vector needs to undergo a large number of node partitions to be separated individually, resulting in an abnormal isolation score. The smaller value indicates that the current distribution of multi-source data is closer to the historical normal data distribution.

[0202] For example, when the number of training samples for each isolated tree is:

[0203] The corresponding expected path length adjustment factor is approximately:

[0204] When the average path length of the fused representation vector in the 100 isolated trees at time t is 4.5, the outlier isolation score is:

[0205] The outlier isolation score is calculated on a rolling basis along with the input data of the isolation forest model. The isolation forest model can be periodically updated based on newly added historical fusion representation vectors that have been confirmed to be in normal operating condition.

[0206] In terms of anomaly detection, to simultaneously capture numerical and semantic anomalies, an anomaly score is constructed by combining residuals, text consistency, and isolated forest scores:

[0207] In the formula, The overall anomaly score at time t; , , Indicates the weighting coefficient of the components of the outlier score; This represents the measurement residual vector at time t; Represents the residual covariance matrix; This represents the current text embedding vector at time t; This represents the text embedding vector predicted by the normal pattern at time t; Indicates cosine similarity; This represents the isolated fraction of anomalies at time t.

[0208] The first term of the anomaly score uses weighted residuals to characterize measurement consistency; the second term uses textual semantic bias to capture discrepancies between maintenance records and operational status; and the third term uses isolated forest anomaly scores to identify the offset of the multi-source fusion representation relative to the historical normal data distribution. After weighted fusion of these three terms, anomaly labels are output based on a threshold.

[0209] In the formula, The anomaly label representing time t; Indicates an indicator function; Indicates the threshold for anomaly detection; This represents the overall abnormal score.

[0210] At this point, a comprehensive perception and risk assessment of the distribution network's operational status has been completed, outputting integrated situational information including point estimates, uncertainty ranges, and anomaly tags. However, simple status perception and anomaly alarms are insufficient to support efficient operation and maintenance. It is still necessary to generate auxiliary decision-making schemes that combine technical feasibility and physical security on this basis, thereby forming a complete closed loop of "perception-diagnosis-decision".

[0211] Step S7: Perform auxiliary decision generation and security filtering.

[0212] Based on the uncertain state estimation results and anomaly diagnosis conclusions obtained in step S6, this step aims to generate a safe, executable, and interpretable auxiliary decision-making scheme. To prevent the large language model from generating "illusion" or illegal operation suggestions under the condition of no physical constraints, this step strictly follows the principle of "model generates suggestions, physical rules adjudicate," and constructs a two-stage decision-making mechanism of "candidate generation - physical filtering."

[0213] In the first stage, the large language model generates auxiliary decision-making candidate schemes, including on-load voltage regulation, reactive power compensation switching, distributed power supply regulation, or network reconfiguration, under the constraints of the domain knowledge base and the current state estimation results. These auxiliary decision-making candidate schemes are not directly output to the operators, but are instead sequentially processed through power flow verification, node voltage upper and lower limit verification, branch power flow limit verification, topology connectivity verification, radial operation constraint verification, equipment action number constraint, action interval constraint, and protection setting constraint. If a candidate scheme does not meet any physical safety constraint, it is eliminated or regenerated. Only the candidate operation schemes that meet all physical safety constraints, along with the operation explanations and risk warnings, are output as the final recommendations.

[0214] For decision support, the large language model is only used to generate candidate solutions based on the current state estimation results, anomaly identification results, and domain knowledge base; it does not serve as the final control action arbiter. In the second stage, multiple physical safety checks are performed on the candidate solutions, including but not limited to: power flow check, node voltage upper and lower limit check, branch power flow limit check, topology connectivity check, radial operation constraint check, equipment action count constraint, action interval constraint, and protection setting constraint. A candidate solution is only output as the final recommendation if it passes all checks; otherwise, it is discarded or returned for regeneration.

[0215] The above screening process can be formalized into the following optimization problem: Let the control variable be... The physical screening target is:

[0216] In the formula, This represents the optimal executable control scheme at time t; Represents the set of candidate controls at time t; Represents the control cost function; This represents the estimated state at time t; This represents the voltage deviation penalty coefficient; This represents the node voltage vector at the next time step. Represents the reference voltage vector; This represents the penalty coefficient for action complexity; This represents the complexity function of the action.

[0217] It is worth emphasizing that this embodiment strictly follows the principle of "generating suggestions using a large language model and making decisions based on physical rules". By placing candidate solutions under strict physical constraints such as power flow, topology, equipment actions, and protection settings for screening, it effectively avoids the "illusion" problem or unsafe control commands that may arise from relying solely on data-driven models, ensuring the reliability and executability of the assisted decision-making.

[0218] Step S8: Output the final result.

[0219] The output includes fused state estimation results, node voltages and branch power, anomalous event labels, uncertainty ranges, candidate operational recommendations, and interpretable text descriptions for operators.

[0220] It should be noted that steps S1 to S8 above describe the online operation process of the technical solution of this application after actual deployment in the distribution network dispatch center. Before the system is officially launched, the model parameters need to be trained offline using historical distribution network operation datasets to ensure the performance of each submodule in practical applications. The specific training method is as follows: Before the above online operation steps, the coding network, attention weight parameters, LoRA parameters of the large language model, and parameters of the deep ensemble sub-model in this scheme need to be jointly trained using historical distribution network operation datasets. The training objective is to minimize the comprehensive loss function.

[0221] In the formula, Represents the total loss function; Indicates the weights of the state estimation loss; Indicates the state estimation loss; This represents the token loss weights of the language model; This represents the token-level cross-entropy loss; This represents the weight of the negative log-likelihood loss; This represents the negative log-likelihood loss due to uncertainty. Indicates the weight of the anomaly detection loss; This indicates loss due to anomaly detection; Indicates the weight of physical constraint loss; This represents the loss due to physical constraints.

[0222] in,

[0223] In the formula, The loss represents the state estimation loss; T represents the number of training time steps. This represents the estimated state at time t; It represents the true state at time t.

[0224]

[0225] In the formula, This represents the token cross-entropy loss; N represents the number of training samples; This indicates the output length of the nth sample; Indicates the nth sample. One real token; Indicates the nth sample. The sequence of real tokens preceding each position; The parameter is The generation probability; This represents the structured input of the nth sample; This represents the set of search results for the nth sample. Let n represent the set of constraints for the nth sample.

[0226] After training is complete, all model parameters are fixed, and the trained model can be deployed to the distribution network dispatch center to execute the online operation process described in S1 to S8. Once training converges and parameters are fixed, the model is applied to a real-world test scenario. To verify the superiority of this application over existing technologies, the technical effects are further illustrated below through several specific embodiments and comparative experiments.

[0227] As a preferred option, in step S3, a retrieval-enhanced generation framework is used for the text, and low-rank adaptation fine-tuning is used for the model parameters; in step S5, the state estimation uses linearized DistFlow constraints or AC power flow constraints; in step S6, the uncertainty results are represented by 90%, 95%, or 99% confidence intervals; and in step S7, the operational suggestions include at least one of on-load tap regulation, reactive power compensation switching, distributed power reactive power regulation, flexible interconnection regulation, and network reconfiguration suggestions.

[0228] Example 2 This invention also provides a large language model-enhanced distribution network multi-source data fusion system, comprising: a data access and alignment module, a data quality assessment module, a structured coding module, a text retrieval and large language model module, a semantic constraint conversion module, a physical constraint state estimation module, an anomaly detection module, and an auxiliary decision-making and safety verification module; the modules work together to implement the above method steps.

[0229] The data access and alignment module is used to acquire multi-source heterogeneous data of the distribution network and establish a spatiotemporal index of the multi-source heterogeneous data, which includes structured measurement data and unstructured operation and maintenance text data. The data quality assessment module is used to calculate the integrity score, accuracy score, timeliness score and semantic consistency score of each data source based on the spatiotemporal index, and to calculate the comprehensive quality score of each data source at the current time based on the integrity score, accuracy score, timeliness score and semantic consistency score. A structured coding module is used to obtain a structured state representation based on the structured measurement data; The text retrieval and large language model module is used to obtain text semantic representations based on the unstructured operation and maintenance text data, and to perform knowledge verification on the text semantic representations based on the retrieval-enhanced large language model. The semantic constraint transformation module is used to convert the verified results into semantic constraint terms, topology correction terms, or pseudo-observation terms for state estimation. The physical constraint state estimation module is used to construct a joint weighted matrix of quality and uncertainty based on the comprehensive quality score of each data source and the prediction uncertainty of the corresponding observation. The structured state representation is used as the observation input, and the semantic constraint term, topology correction term or pseudo-observation term is used as the constraint term or observation term and introduced into the physical constraint state estimation model to obtain the distribution network state estimation result. An anomaly detection module is used to calculate a comprehensive anomaly score based on the power distribution network state estimation results, and output anomaly discrimination information in combination with the uncertainty quantification results; The auxiliary decision-making and safety verification module is used to generate auxiliary decision-making candidate schemes based on the power distribution network state estimation results and anomaly identification information using a retrieval-enhanced large language model, perform physical safety verification on the auxiliary decision-making candidate schemes, and output only the auxiliary decision-making schemes that pass the verification.

[0230] Example 3 The present invention also provides a computer device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described method steps.

[0231] Example 4 The present invention also provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the above-described method steps.

[0232] Example 5 This embodiment uses a 33-node radial distribution topology, a 1-minute fusion cycle, and 14 days of continuous data. The total number of steps on the main time axis is 20,160. The overall quality weight is set as follows: , , , The large language model employs a 32-layer decoder architecture with 4096 hidden dimensions, 32 attention heads, LoRA fine-tuning, rank r=16, scaling factor 32, dropout 0.05, and a top-K retrieval value of 6. The significance level is... The number of deep integration models B=5.

[0233] Example 6 The difference between this embodiment and Embodiment 5 is that it sets up a scenario with a high missing rate, strong latency, and severe semantic conflict. The SCADA missing rate is 30%, the AMI random arrival delay is 5 minutes, and 12% of ambiguous expressions are set in the work order text. All other parameters remain unchanged. This embodiment is used to verify that under conditions of weak observability and high conflict data, the quality-aware weights, adaptive state estimation, and semantic retrieval constraints of this invention can still maintain stable output.

[0234] Example 7 The difference between this embodiment and Embodiment 5 is that a concentrated set of abnormal scenarios is introduced, including 36 persistent missing events, 24 bias drift events, 18 events where the switch status is inconsistent with the text record, and 18 events where the equipment temperature rises abnormally. All other parameters remain unchanged. This embodiment is used to verify the effectiveness of the comprehensive anomaly score model and the auxiliary decision-making safety filtering mechanism.

[0235] The baselines for comparison are fixed into four categories. The first category is fixed-weight fusion + WLS; the second category is EKF + pseudo-measurement; the third category is DS evidence fusion + WLS; and the fourth category is a text baseline that "directly generates suggestions using LLM on the same base but does not perform retrieval enhancement and physical filtering". The first three categories correspond to the classic approaches in existing structured multi-source estimation, while the last category is used to demonstrate that the innovation of this application lies not in "using a large model", but in the complete closed loop of "LLM + retrieval + quality awareness + physical constraints + uncertainty + security filtering". Public power and LLM research also shows that for the basic model to work stably in power grid tasks, it must be coordinated with tools, knowledge, and constraints.

[0236] The fixed hyperparameters are shown in Table 1: Table 1 Examples of Hyperparameters

[0237] Evaluation indicators include: State estimation uses voltage amplitude RMSE and branch active power MAE; uncertainty uses PICP (Prediction Interval Coverage Probability) and MPIW (Mean Prediction Interval Width); anomaly detection uses Precision, Recall, F1, and AUC; and decision support uses voltage over-limit elimination rate, average time delay, and safety executability rate.

[0238] Results Table 2: Accuracy of State Estimation

[0239] Table 2 shows that in the three missing scenarios of 10%–30%, such as Figure 4 As shown, the average reduction in voltage RMSE of this invention relative to the optimal structured baseline EKF is 42.73%, and the average reduction relative to fixed-weight fusion is 51.41%; the average reduction in branch active power MAE relative to the optimal structured baseline is 40.45%. When the missing rate rises to 30%, this invention still maintains a voltage RMSE of 0.0143 pu, indicating that quality-aware weights and textual semantic constraints are particularly effective for weakly observable conditions.

[0240] As shown in Table 2, the estimation error of traditional methods increases nonlinearly with the increase of measurement missing rate. However, this application, by introducing a quality-aware weighting mechanism and text semantic constraints, can effectively use high-confidence text information to make up for the lack of measurement data, thereby controlling the voltage amplitude RMSE at a low level, which verifies the significant advantages of this application under weak observable conditions.

[0241] Results Table 3: Uncertainty Calibration

[0242] Combination Figure 5 As shown in Table 3, the present invention calibrates the nominal coverage from 95% to 95.6% while maintaining a narrower interval width, indicating that the combination of "deep integration + conformal calibration" is superior to simple MC-Dropout or uncalibrated deep integration.

[0243] The results in Table 3 show that the "deep integration + conformal calibration" strategy adopted in this application significantly improves the accuracy of interval coverage (lowest ECE) while ensuring the width of the prediction interval. This indicates that this application can not only provide point results for state estimation, but also provide statistically significant confidence guarantees, greatly enhancing operators' confidence in the decision-making results.

[0244] Results Table 4: Anomaly Detection Performance

[0245] Combination Figure 6 As can be seen from the data in Table 4, by deeply fusing residual information, semantic consistency bias and data density features, the comprehensive anomaly score constructed in this application can more sensitively capture complex anomalies that are difficult to detect by a single method, achieving the best balance between precision and recall, and proving the synergistic gain effect of multi-source information fusion in anomaly identification.

[0246] Table 4 shows that the present invention improves the F1 score by 0.054 compared to the best non-inventive baseline, representing a relative increase of approximately 5.97%. The core reason for this is not simply replacing the classifier, but rather that the three types of information—residual, semantic conflict, and anomaly density—are unified into a single anomaly score.

[0247] Results Table 5: Decision Support Effect

[0248] Combination Figure 7 As shown in Table 5, if only the LLM directly provides operational suggestions, although the latency is very low, the safe execution rate is only 78.0%. However, this invention improves the safe execution rate to 100% through physical screening and limit constraints, while maintaining an average response speed of 29 seconds, which is much faster than a purely manual rule base.

[0249] Results Table 6: Ablation Experiment

[0250] Combination Figure 8 The ablation results shown in Table 6 demonstrate that this invention does not rely on a single component. Removing the text semantic channel significantly reduced the anomaly detection F1 score; removing retrieval enhancements decreased knowledge matching and executability; removing uncertainty calibration resulted in severe distortion of interval coverage; and removing physical security filtering, although some numerical indicators did not appear to be the worst, resulted in a sharp drop in security executability. In other words, the focus of this invention is "end-to-end collaboration," rather than a single isolated model module.

[0251] Ablation experiments further confirmed that the superior performance of this application does not stem from a single component, but is the result of the synergistic effect of the entire chain of "quality awareness—semantic enhancement—uncertainty quantification and physical filtering". The absence of any key link will lead to a significant decline in overall performance. In particular, the removal of the physical security filtering link will directly lead to serious operational safety hazards, highlighting the systematic and rigorous design of the technical solution of this application.

[0252] In summary, this application effectively solves a series of problems, such as the difficulty in utilizing unstructured semantic information, the difficulty in dynamically adapting data quality, the difficulty in quantifying operational risks, and the difficulty in verifying the security of auxiliary decision-making, by constructing a multi-source data fusion framework for distribution networks that deeply integrates large language models and physical constraints. This provides a practical and feasible technical solution for the intelligent operation of distribution networks under the new power system.

[0253] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0254] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0255] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0256] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0257] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for multi-source data fusion in power distribution networks enhanced by a large language model, characterized in that, Includes the following steps: Acquire multi-source heterogeneous data of the power distribution network and establish a spatiotemporal index of the multi-source heterogeneous data, which includes structured measurement data and unstructured operation and maintenance text data; Based on the spatiotemporal index, the integrity score, accuracy score, timeliness score, and semantic consistency score of each data source are calculated respectively. Based on the integrity score, accuracy score, timeliness score, and semantic consistency score, the comprehensive quality score of each data source at the current moment is calculated. A structured state characterization is obtained based on the structured measurement data; Text semantic representation is obtained based on the unstructured operation and maintenance text data; The semantic representation of the text is verified based on a large language model with retrieval enhancement, and the verified results are converted into semantic constraints, topological corrections or pseudo-observations for state estimation. Based on the comprehensive quality score of each data source and the prediction uncertainty of the corresponding observation, a joint weighted matrix of quality and uncertainty is constructed. The structured state representation is used as the observation input, and the semantic constraint term, topology correction term, or pseudo-observation term is used as the constraint term or observation term and introduced into the physical constraint state estimation model to obtain the distribution network state estimation result. The comprehensive anomaly score is calculated based on the power distribution network state estimation results, and anomaly discrimination information is output by combining the uncertainty quantification results. Based on the power distribution network state estimation results and anomaly identification information, a large language model with retrieval enhancement is used to generate auxiliary decision candidate schemes; physical security verification is performed on the auxiliary decision candidate schemes, and only auxiliary decision schemes that pass the verification are output.

2. The method for multi-source data fusion in distribution networks with large language model enhancement according to claim 1, characterized in that: The structured measurement data includes at least three of the following: SCADA measurement data, synchronous phasor measurement data, AMI data, equipment online monitoring data, meteorological data, and GIS / topology data. The unstructured operation and maintenance text data includes at least one of the following: work order text, alarm text, defect record, scheduling record, and maintenance record.

3. The method for multi-source data fusion in distribution networks with large language model enhancement according to claim 1, characterized in that: The integrity score is calculated by dividing the number of data points actually received within a preset time window by the number of data points that should have been received within that time window. The accuracy score is calculated as follows: the absolute value of the difference between each observation and the corresponding reference prediction value is divided by a preset accuracy threshold to obtain a ratio result, and the smaller value between 1 and the ratio result is taken; the mean of the smaller values ​​corresponding to all observations is calculated, and the accuracy score is obtained by subtracting the mean from 1. The timeliness score is calculated as follows: the product of the delay attenuation coefficient and the current time delay is calculated, and the negative of the product is taken as the input of the exponential function to calculate the timeliness score. The semantic consistency score is calculated as follows: calculate the cosine similarity between the current semantic vector and the predicted semantic vector of the normal pattern, add 1 to the cosine similarity and divide by 2 to obtain the semantic consistency score.

4. The method for multi-source data fusion in distribution networks with large language model enhancement according to claim 3, characterized in that: The semantic consistency score is calculated as follows: extract the device number, topology node, timestamp, switch status and over-limit alarm corresponding to the measurement data, and cross-match it with the domain knowledge base and other data sources; if the match is consistent, it is assigned a value of 1.0; if there is a clear conflict, it is assigned a value of 0; if there is partial field conflict, it is assigned a value by linear interpolation based on the proportion of the number of conflicting fields to the total number of matched fields.

5. The method for multi-source data fusion in distribution networks with large language model enhancement according to claim 1, characterized in that: The method for calculating the overall quality score is as follows: a completeness weight, an accuracy weight, a timeliness weight, and a semantic consistency weight are preset. The completeness score is multiplied by the completeness weight, the accuracy score is multiplied by the accuracy weight, the timeliness score is multiplied by the timeliness weight, and the semantic consistency score is multiplied by the semantic consistency weight, and then the results are added together to obtain the overall quality score; wherein, the sum of the completeness weight, accuracy weight, timeliness weight, and semantic consistency weight is 1.

6. The method for multi-source data fusion in distribution networks with large language model enhancement according to claim 1, characterized in that: The calculation method for the structured state representation is as follows: feature extraction is performed on the structured measurement data through a numerical coding network to obtain the structured state representation; The text semantic representation calculation method is as follows: entity extraction and event extraction are performed on the unstructured operation and maintenance text data to obtain intermediate results containing device objects, event types, event times, topology status, operation intentions and text confidence; the intermediate results are semantically encoded using a retrieval-enhanced large language model to obtain the text semantic representation. The structured state representation and the text semantic representation are fused through an attention mechanism to obtain a fused representation vector. The fused representation vector is used for anomaly detection, retrieval query and generation of auxiliary decision context, and is not directly used as the observation input of the physical constraint state estimation model.

7. The method for multi-source data fusion in distribution networks with large language model enhancement according to claim 1, characterized in that: The computational method for knowledge verification of the text semantic representation using the retrieval-enhanced large language model is as follows: Retrieve topology constraints, equipment ledgers, protection settings, scheduling rules, and historical similar cases that match the current moment from the domain knowledge base. Input the retrieval results and the text semantic representation into the large language model, and output a structured text result that has been verified by knowledge constraints. The structured text result includes equipment object field, event type field, time constraint field, topology status field, operation intention field, and text confidence field.

8. The method for multi-source data fusion in distribution networks with large language model enhancement according to claim 1, characterized in that: The method for converting the verified result into semantic constraint terms for state estimation is as follows: When the structured text result represents node injection power, device operating status, or the time of occurrence of abnormal events, it is mapped to a semantic constraint term for solving the constraint state estimation. The method for calculating the topology correction term used for state estimation by converting the verified result is as follows: When the structured text result represents the status of switch opening / closing, line outage, equipment maintenance, or equipment commissioning / decommissioning, it is mapped to a topology correction term used to correct the topology connection relationship in the state estimation model. The method for converting the verified results into pseudo-observations for state estimation is as follows: When the structured text results represent missing measurements, low-confidence measurements, or load changes confirmed by work orders, they are mapped to pseudo-observations used to supplement the state estimation input.

9. The method for multi-source data fusion in distribution networks with large language model enhancement according to claim 1, characterized in that: The specific method for obtaining the distribution network state estimation results is as follows: The overall quality scores of each data source are truncated to upper and lower limits to obtain the truncated reliability weights; the standard deviation of the prediction error corresponding to each observation is calculated based on the historical effective prediction residuals of each observation before the current time; the square of the standard deviation of the prediction error is summed with the corresponding lower limit of variance, the reliability weight is divided by the summation result, and the quotient is used as the diagonal element to construct a quality-uncertainty joint weighting matrix composed of the diagonal elements; The structured state representation is input into the observation function of the physical constraint state estimation model to obtain the expected observation value; Construct a physical constraint equation, and add the physical constraint equation, the semantic constraint term, the topology correction term, and the pseudo-observation term as penalty terms into the optimization objective function of the physical constraint state estimation model; The state variable vector of the distribution network is obtained by minimizing the optimization objective function, and is used as the state estimation result of the distribution network.

10. The method for multi-source data fusion in distribution networks with large language model enhancement according to claim 1, characterized in that: The comprehensive anomaly score is calculated based on the power distribution network state estimation results, and anomaly discrimination information is output by combining the uncertainty quantification results. The specific method is as follows: Calculate the measurement residual vector corresponding to the power distribution network state estimation result, and calculate the product of the measurement residual vector and the inverse of the residual covariance matrix to obtain the weighted residual term; calculate the cosine similarity between the current text embedding vector and the normal pattern predicted text embedding vector, and subtract the cosine similarity from 1 to obtain the text semantic bias term; obtain the anomaly isolation score; assign weight coefficients to the weighted residual term, the text semantic bias term, and the anomaly isolation score respectively, and then perform a weighted summation to obtain the comprehensive anomaly score; The mean and variance of the state estimation results are calculated through multiple state estimation sub-models. The conformal calibration quantiles are calculated based on the historical calibration dataset. The prediction interval at a preset confidence level is generated based on the mean, variance and conformal calibration quantiles as the uncertainty quantification result. The comprehensive anomaly score is compared with a preset anomaly judgment threshold, and anomaly discrimination information including anomaly label, anomaly type, scope of impact, and confidence level is generated based on the comparison result.

11. The method for multi-source data fusion in a distribution network with large language model enhancement according to claim 10, characterized in that: The method for obtaining the isolated anomaly score includes: Based on historical operation records, select fusion representation vectors that are in normal operation to form a historical normal training sample set, and use the historical normal training sample set to train an isolated forest model containing multiple isolated trees. Input the fused representation vector at the current moment into each isolated tree to obtain the path length of the fused representation vector from the root node to the terminal node in each isolated tree, and calculate the average path length in all isolated trees. The expected path length adjustment coefficient is calculated based on the number of historical normal training samples randomly selected during the training of each isolated tree. The outlier score is calculated based on the average path length and the expected path length adjustment coefficient. Specifically, the outlier score is calculated by dividing the average path length by the negative value of the expected path length adjustment coefficient by 2 as the base, and the resulting power value is the outlier score.

12. The method for multi-source data fusion in distribution networks with large language model enhancement according to claim 1, characterized in that: Based on the power distribution network state estimation results and anomaly identification information, a large language model with enhanced retrieval is used to generate auxiliary decision-making candidate schemes; physical security checks are performed on the auxiliary decision-making candidate schemes, and only auxiliary decision-making schemes that pass the checks are output. The specific method is as follows: The distribution network state estimation results, anomaly identification information, scheduling rules in the domain knowledge base, and historical similar cases are input into the retrieval-enhanced big language model to generate auxiliary decision-making candidate schemes that include at least one of on-load voltage regulation, parallel capacitor switching, distributed power reactive power regulation, flexible interconnection power regulation, and network reconfiguration suggestions. The multiple physical security checks include: Based on power flow calculation, it is determined whether the node voltage after the candidate scheme is executed meets the upper and lower voltage limits, whether the branch power flow meets the line capacity constraints, whether the switch action meets the topology connectivity and radial operation constraints, and whether the candidate scheme meets the constraints of equipment action number, action interval and protection setting. The system filters candidate solutions that simultaneously satisfy all constraints in multiple physical security checks. The filtered candidate solutions are output as the final auxiliary decision solution. If no candidate solution is found, the system returns a message indicating that no feasible solution can be generated.

13. A distribution network multi-source data fusion system enhanced by a large language model using the method described in any one of claims 1-12, comprising a data access and alignment module, a data quality assessment module, a structured coding module, a text retrieval and large language model module, a semantic constraint conversion module, a physical constraint state estimation module, an anomaly detection module, and an auxiliary decision-making and safety verification module, characterized in that: The data access and alignment module is used to acquire multi-source heterogeneous data of the distribution network and establish a spatiotemporal index of the multi-source heterogeneous data, which includes structured measurement data and unstructured operation and maintenance text data. The data quality assessment module is used to calculate the integrity score, accuracy score, timeliness score and semantic consistency score of each data source based on the spatiotemporal index, and to calculate the comprehensive quality score of each data source at the current time based on the integrity score, accuracy score, timeliness score and semantic consistency score. A structured coding module is used to obtain a structured state representation based on the structured measurement data; The text retrieval and large language model module is used to obtain text semantic representations based on the unstructured operation and maintenance text data, and to perform knowledge verification on the text semantic representations based on the retrieval-enhanced large language model. The semantic constraint transformation module is used to convert the verified results into semantic constraint terms, topology correction terms, or pseudo-observation terms for state estimation. The physical constraint state estimation module is used to construct a joint weighted matrix of quality and uncertainty based on the comprehensive quality score of each data source and the prediction uncertainty of the corresponding observation. The structured state representation is used as the observation input, and the semantic constraint term, topology correction term or pseudo-observation term is used as the constraint term or observation term and introduced into the physical constraint state estimation model to obtain the distribution network state estimation result. An anomaly detection module is used to calculate a comprehensive anomaly score based on the power distribution network state estimation results, and output anomaly discrimination information in combination with the uncertainty quantification results; The auxiliary decision-making and safety verification module is used to generate auxiliary decision-making candidate schemes based on the power distribution network state estimation results and anomaly identification information using a retrieval-enhanced large language model, perform physical safety verification on the auxiliary decision-making candidate schemes, and output only the auxiliary decision-making schemes that pass the verification.

14. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable by the processor, characterized in that, When the processor executes the computer program, it implements the method according to any one of claims 1 to 12.

15. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method according to any one of claims 1 to 12.