A causal and credible reasoning method and system based on grid instruction semantic uncertainty guidance

Abstract

The application provides a kind of based on power grid instruction semantic uncertainty guidance causal credible reasoning method and system, it is related to processing natural language data technical field.The semantic analysis is carried out to power grid instruction text and the associated structure is constructed, which provides the structured logic basis for subsequent all determination.On this basis, the calculation and grading of the uncertainty of power grid business semantic elements are introduced, which enables the perception of the inherent risk of the input text, and can identify which are the high-risk areas that need external knowledge intervention.Furthermore, the uncertainty level is used to dynamically guide the subsequent search, mapping, causal reasoning and correction links, ensuring that the computing resources can be put into the semantic elements that need governance the most, improving the adaptability of power dispatching instruction processing, and thus reducing resource consumption and improving instruction accuracy.

Description

Technical Field

[0001] This application relates to the field of natural language data processing technology, and in particular to a causal and reliable reasoning method and system based on the semantic uncertainty of power grid commands. Background Technology

[0002] With the rapid development of natural language processing technology, large-scale language models, with their powerful semantic understanding and generation capabilities, are gradually penetrating from general domains into vertical industries with high security and reliability, such as power grid dispatching. In these scenarios, the models need to process a large amount of power grid command text. However, the inherent illusion problem of large-scale language models—that is, generating content that is inconsistent with the facts or logically contradictory—has become a bottleneck for their in-depth application in the power grid field. Any erroneous command parsing caused by illusion may lead to equipment damage or even power grid accidents.

[0003] To improve the accuracy of power grid command parsing, a related technique involves pre-verification using a power grid knowledge base before the model processes the input commands. The specific implementation logic of this scheme is as follows: First, preliminary semantic element extraction is performed on the input power grid command text. Then, a search of the external knowledge base is initiated based on a fixed trigger rule. For example, when a device number not registered in the local device dictionary appears in the command, the system will compare and verify the information in the external knowledge base. The retrieved information will serve as enhanced context to assist the model in subsequent understanding and generation.

[0004] However, in actual use, the semantic complexity of business command flows in power grid dispatch centers is unevenly distributed. The strategies of the aforementioned related technologies are rigid and blind, ultimately falling into a dilemma: either wasting resources by excessively searching for massive amounts of simple commands, or affecting the reliability of decision-making by insufficiently searching for key complex commands. Summary of the Invention

[0005] This application provides a causal and reliable reasoning method and system based on the semantic uncertainty of power grid commands, which can improve the adaptability to power grid dispatch command processing, thereby reducing resource consumption and improving command accuracy.

[0006] Firstly, this application provides a causal and reliable reasoning method guided by the semantic uncertainty of power grid instructions, comprising: performing semantic parsing on the power grid instruction text to determine the association structure containing power grid business semantic elements; calculating the uncertainty of the power grid business semantic elements based on the association structure; distinguishing the uncertainty according to the uncertainty level and selecting a target element set based on the uncertainty level; performing dynamic retrieval and mode switching on the target element set in the corresponding candidate retrieval sources based on a priority scheduling strategy to obtain corresponding evidence information; mapping the evidence information to the target element set, calculating the evidence matching degree and updating the evidence reliability, generating an extended element set and a corresponding evidence structure; selecting target association paths based on the association quantity in the evidence structure, constructing a causal chain, and calculating the corresponding reasoning reliability factor; locating the problematic power grid business semantic elements in the power grid instruction text based on the reasoning reliability factor, selecting the corresponding strategy to perform text correction, and outputting the corrected power grid text.

[0007] By employing the aforementioned technical solution, semantic parsing of power grid command text and the construction of associated structures provide a structured logical foundation for all subsequent judgments. Building upon this, the calculation and classification of uncertainties in power grid business semantic elements are introduced, enabling the perception of inherent risks in the input text and the identification of high-risk areas requiring external knowledge intervention. Furthermore, the uncertainty level is used to dynamically guide subsequent retrieval, mapping, causal reasoning, and correction processes, ensuring that computational resources are allocated to the semantic elements most in need of governance. This transforms power grid command parsing from a black-box, probabilistic generation process into a reliable and efficient processing flow where every decision is verifiable, logically traceable, and self-correcting, improving adaptability to power grid dispatch command processing, thereby reducing resource consumption and increasing command accuracy.

[0008] In conjunction with some embodiments of the first aspect, in some embodiments, the step of semantically parsing the power grid instruction text to determine the association structure containing power grid business semantic elements includes: extracting and segmenting the power grid instruction text according to a preset length threshold to obtain multiple power grid business semantic elements; taking any two power grid business semantic elements as a first power grid business semantic element and a second power grid business semantic element, and calculating the association quantity between the first power grid business semantic element and the second power grid business semantic element; the association quantity is determined by weighting vector similarity and co-occurrence statistics; calculating the difference between the positive association quantity from the first power grid business semantic element to the second power grid business semantic element and the negative association quantity from the second power grid business semantic element to the first power grid business semantic element; if the difference is greater than a preset positive threshold, establishing a dependency edge from the first power grid business semantic element to the second power grid business semantic element; if the difference is less than a preset negative threshold, establishing a dependency edge from the second power grid business semantic element to the first power grid business semantic element; and determining the association structure based on the power grid business semantic elements and the corresponding dependency edges.

[0009] By employing the aforementioned technical solution, the correlation quantity is calculated through a combination of vector similarity and co-occurrence statistics. This approach considers both the deeper meaning of the words themselves and incorporates collocation habits from the domain, resulting in a more comprehensive and accurate calculation of correlation strength. By calculating the difference between positive and negative correlation quantities and establishing unidirectional dependency edges based on thresholds, the originally ambiguous textual narrative can be transformed into a topological network with clear primary and secondary, sequential, and causal orientations. This avoids logical confusion caused by semantic comprehension biases. Ultimately, the constructed correlation structure is no longer a simple model, but a logical graph that accurately reflects the command flow, constraint relationships, and execution prerequisites within power grid instructions, improving the model's depth and accuracy in structurally understanding complex instructions.

[0010] In conjunction with some embodiments of the first aspect, after determining the association structure based on the power grid business semantic elements and their corresponding dependent edges, the method further includes: traversing the topological paths of the association structure, extracting a set of closed paths containing directed loops, and bidirectional dependent edges with mutually opposite directions; calculating the average association quantity corresponding to the closed path set and the bidirectional dependent edges respectively, and determining that the corresponding dependent edge has a logical conflict if the average association quantity is greater than a preset conflict threshold; adjusting the association quantity in the association structure using a preset interpolation function for the dependent edges with logical conflicts, and then jumping to the step of calculating the difference between the positive association quantity from the first power grid business semantic element to the second power grid business semantic element and the negative association quantity from the second power grid business semantic element to the first power grid business semantic element, so as to update the association structure; after updating the association structure, if the original closed path set or the original bidirectional dependent edge still has logical conflicts, then splitting or merging the corresponding power grid business semantic elements, and then jumping to the step of taking any two power grid business semantic elements as the first power grid business semantic element and the second power grid business semantic element, and calculating the association quantity between the first power grid business semantic element and the second power grid business semantic element.

[0011] By employing the aforementioned technical solution, potential logical deadlocks such as directed cycles or bidirectional dependencies are extracted through topological path traversal. A comparison mechanism between average correlation quantity and conflict threshold is introduced, ensuring that conflict determination is no longer based on fuzzy rules but on quantifiable mathematical indicators, thus improving diagnostic accuracy. Building upon this, a two-stage progressive reconstruction strategy of interpolation adjustment and node splitting / merging is implemented. The first stage performs soft logical guidance by adjusting correlation quantity. If this method fails, the second stage initiates structural reorganization of the semantic elements themselves. Even when faced with unclear or contradictory original instructions, it can self-iterate and correct itself, ultimately outputting a logically consistent and unambiguous correlation structure, providing a reliable data foundation for subsequent trustworthy reasoning.

[0012] In conjunction with some embodiments of the first aspect, in some embodiments, the step of performing dynamic retrieval and mode switching on the target element set in the corresponding alternative retrieval sources based on a priority scheduling strategy to obtain corresponding evidence information specifically includes: calculating the initial priority score of each alternative retrieval source based on historical matching quality, availability, and current task relevance, and establishing or updating a scheduling table; extracting the element type of the target element in the target element set and calculating the matching score between it and the capability feature vector of the alternative retrieval sources; updating the initial priority score in the scheduling table with weighted average based on the matching score; performing retrieval in the alternative retrieval sources according to the order of the scheduling table and obtaining feedback scores; calculating retrieval consistency reflecting the alignment degree of a single retrieval based on the feedback scores; determining whether the retrieval consistency is greater than a preset quality threshold and whether the change in priority score is within the allowable range; if yes, then determining the information obtained from the current retrieval as evidence information; if no, then changing the order of the scheduling table and performing retrieval in the alternative retrieval sources according to the order of the scheduling table and obtaining feedback scores.

[0013] By adopting the above technical solution, and through initial priority scoring of candidate search sources, combined with weighted updates based on the matching scores of target element types and search source capability characteristics, a shift from blind general screening to on-demand selection is achieved. This ensures that each search hits the knowledge source most relevant to the current task with a high probability. A closed-loop quality inspection mechanism based on feedback scoring and search consistency is introduced. This feature transforms the search from a one-off action into a continuous self-evaluation and dynamic adjustment process. When the search quality is substandard, the scheduling strategy can be proactively changed and the search can be re-executed until satisfactory evidence is obtained. This self-correcting capability ensures that high-quality external evidence is obtained with optimal resource consumption, balancing search efficiency, coverage, and the reliability of the final content.

[0014] In conjunction with some embodiments of the first aspect, in some embodiments, the steps of mapping evidence information to a target element set, calculating the evidence matching degree and updating the evidence credibility, and generating an extended element set and a corresponding evidence structure include: calculating the initial matching degree between the corresponding target element and the corresponding evidence information, wherein the initial matching degree is determined by a semantic similarity item, a structural consistency item, and a contextual association item; obtaining the integrity score of the evidence information, performing an adjustment calculation on the initial matching degree using the integrity score to obtain the adjusted evidence matching degree; performing a weighted correction on the original evidence credibility based on the adjusted evidence matching degree to obtain a new evidence credibility; determining the extended element set based on the new evidence credibility, and determining the corresponding evidence structure based on the extended element set.

[0015] By employing the aforementioned technical solution and integrating multi-dimensional indicators such as semantic similarity, structural consistency, and contextual relevance to calculate the initial matching degree, the degree of matching between evidence and target elements can be measured. Based on this, an adjustment is made by introducing an inherent integrity score for the evidence itself, avoiding the risk of secondary decision-making due to the acceptance of partially matching, incomplete evidence. Finally, by weighting and correcting the credibility of the original evidence, the credit rating of each evidence source can be dynamically evaluated and updated. This mechanism allows high-quality, highly matching evidence to play a more important role in subsequent reasoning, while low-quality noise information is naturally weakened. Ultimately, the generated evidence structure is a credible knowledge network that has undergone layer-by-layer screening and has clearly defined weights, providing solid and reliable argumentative support for constructing causal chains.

[0016] In conjunction with some embodiments of the first aspect, in some embodiments, the steps of filtering target association paths based on the association quantity in the evidence structure, constructing a causal chain, and calculating the corresponding reasoning credibility factor specifically include: extracting the connection lines of power grid business semantic elements in the evidence structure whose association quantity is greater than a preset association quantity threshold as target association paths, and generating a set of target association paths; performing a graph traversal algorithm on the set of target association paths to construct an initial causal chain with directed association paths; detecting the direction corresponding to the dependency relationship in the initial causal chain, and extracting the corresponding power grid business semantic elements as conflicting power grid business semantic elements when the dependency relationship is a contradictory dependency relationship; obtaining the association quantity corresponding to the candidate branch path containing the conflicting power grid business semantic elements, sorting the candidate branch paths according to the association quantity, and extracting the optimal branch path; reconstructing the initial causal chain based on the optimal branch path to perform conflict resolution, and confirming the final causal chain under the condition that the reconstructed causal chain satisfies the constraints of temporal monotonicity and consistency; and calculating the reasoning credibility factor of the corresponding power grid instruction text based on the final causal chain.

[0017] By employing the aforementioned technical solution and filtering out strongly correlated paths using correlation thresholds, the core logical backbone can be extracted from complex evidence networks, eliminating interference from weakly related information. Furthermore, the use of a graph traversal algorithm ensures that the constructed initial causal chain is structurally connected. Crucially, a conflict node identification and resolution mechanism for contradictory dependencies is introduced. By sorting candidate branch paths and selecting the optimal branch, logical paradoxes caused by multi-source evidence or ambiguous instructions can be proactively resolved. The finally confirmed causal chain, under the constraints of temporal monotonicity and consistency, transforms the originally disordered text elements into an irreversible reasoning path from cause to effect, providing clear, consistent logical guidance and a basis for credibility calculation for subsequent text correction.

[0018] In conjunction with some embodiments of the first aspect, in some embodiments, the steps of locating problematic power grid business semantic elements in power grid instruction text based on inference credibility factors, selecting corresponding strategies to perform text correction, and outputting corrected power grid text specifically include: locating target corrected power grid business semantic elements in power grid instruction text based on causal chains and inference credibility factors; extracting inference credibility factors corresponding to the target corrected power grid business semantic elements, constructing a power grid business semantic element credibility feature vector; constructing multiple sets of strategy feature vectors corresponding to different correction methods, the strategy feature vectors containing adaptation parameters, risk coefficients, and multi-dimensional factors of strategy depth; calculating the similarity between the power grid business semantic element credibility feature vector and each set of strategy feature vectors as strategy preference, selecting the correction method with the largest preference value as the optimal correction method; and calling the optimal correction method to correct the target corrected power grid business semantic elements in the power grid instruction text. The system performs replacement, insertion, or expansion operations on network service semantic elements to generate local correction results. It then calculates the matching degree between these local correction results and the evidence structure and causal chain, and weights the results to obtain a consistency score. If the consistency score is greater than a set consistency threshold, the local correction results are merged into the power grid instruction text; otherwise, the correction is abandoned. For the merged power grid instruction text, a confidence decay function is called to perform a smoothing penalty calculation on the inference credibility factor, resulting in a smoothed credibility factor. If the smoothed credibility factor is lower than a preset target threshold, the system jumps to a step of dynamically searching and switching methods on the target element set in the corresponding alternative search sources based on a priority scheduling strategy to obtain the corresponding evidence information. If the smoothed credibility factor is not lower than the target threshold, the merged power grid instruction text is determined as the power grid corrected text and output.

[0019] By adopting the above technical solution, the optimal correction method is selected by calculating the similarity between the credible feature vector and the policy feature vector, thus achieving an upgrade from rule-based encoding modification to data-driven adaptive correction. After the correction is performed, a consistency score verification between the local correction result and the evidence structure and causal chain is introduced. This feature sets up a safety barrier for the correction operation, preventing the introduction of new logical errors due to improper modification. A smoothing penalty and secondary retrieval jump mechanism based on the confidence decay function are designed. When there is insufficient confidence in the final output, it is better to backtrack and re-examine the evidence than to output a low-quality result. This closed-loop control logic with feedback callback, combined with the final structured audit record, ensures the high reliability, security, and post-event traceability of the power grid correction text.

[0020] In a second aspect, this application provides a causal and reliable reasoning system guided by the semantic uncertainty of power grid commands. The causal and reliable reasoning system guided by the semantic uncertainty of power grid commands includes: one or more processors and a memory; the memory is coupled to one or more processors, and the memory is used to store computer program code, which includes computer instructions. One or more processors call the computer instructions to cause the causal and reliable reasoning system guided by the semantic uncertainty of power grid commands to perform the method described in the first aspect and any possible implementation thereof.

[0021] Thirdly, this application provides a computer program product containing instructions that, when the computer program product is run on a causal and reliable reasoning system guided by the semantic uncertainty of power grid instructions, causes the causal and reliable reasoning system guided by the semantic uncertainty of power grid instructions to execute the method described in the first aspect and any possible implementation thereof.

[0022] Fourthly, this application provides a computer-readable storage medium including instructions that, when executed on a causal and reliable reasoning system guided by the semantic uncertainty of power grid instructions, cause the causal and reliable reasoning system guided by the semantic uncertainty of power grid instructions to perform the method described in the first aspect and any possible implementation thereof.

[0023] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0024] By semantically parsing the power grid command text and constructing a relational structure, a structured logical foundation is provided for all subsequent judgments. Based on this, the calculation and classification of uncertainties in the semantic elements of power grid operations are introduced, enabling the perception of inherent risks in the input text and the identification of high-risk areas requiring external knowledge intervention. Furthermore, the uncertainty level is used to dynamically guide subsequent retrieval, mapping, causal reasoning, and correction processes, ensuring that computational resources are allocated to the semantic elements most in need of governance. This transforms the parsing of power grid commands from a black-box, probabilistic generation process into a reliable and efficient processing flow where every decision is verifiable, logically traceable, and self-correcting, improving adaptability to power grid dispatch command processing, thereby reducing resource consumption and improving command accuracy. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating a causal and reliable reasoning method based on the semantic uncertainty of power grid commands in an embodiment of this application.

[0026] Figure 2 This is a flowchart illustrating step S101 in an embodiment of this application;

[0027] Figure 3 This is a flowchart illustrating step S104 in an embodiment of this application;

[0028] Figure 4 This is a flowchart illustrating step S105 in an embodiment of this application;

[0029] Figure 5 This is an exemplary hardware structure diagram of a causal and reliable reasoning system guided by the semantic uncertainty of power grid commands in this application embodiment. Detailed Implementation

[0030] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification and appended claims of this application, the singular expressions a, an, said, the above, and this are intended to also include the plural expressions unless the context clearly indicates otherwise. It should also be understood that the terms used in this application refer to and / or include any or all possible combinations of one or more of the listed items.

[0031] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0032] Please see Figure 1 , Figure 1 This is a flowchart illustrating a causal and reliable reasoning method based on the semantic uncertainty of power grid commands in an embodiment of this application.

[0033] A causal and reliable reasoning method based on the semantic uncertainty of power grid commands includes:

[0034] S101. Perform semantic parsing on the power grid instruction text to determine the association structure containing power grid business semantic elements.

[0035] Among them, power grid instruction text refers to natural language text issued by the dispatch center that contains descriptions of power grid operations or status; power grid business semantic elements refer to the smallest semantic units with independent business meanings extracted from power grid instruction text, such as entities, events, attributes, or constraints; and association structure refers to the data structure used to represent the logical dependencies between various power grid business semantic elements, usually a directed graph or network topology.

[0036] Specifically, the system receives raw, unstructured power grid command text and transforms it into a machine-understandable and computable format through a series of natural language processing operations. This process includes not only text segmentation and entity recognition, but also analyzing and establishing logical relationships between these entities, such as causal, modifying, or parallel relationships. Ultimately, it outputs a stable graph-like relational structure that reflects the internal logical flow of the command, providing a structured data foundation for the calculation and judgment of all subsequent steps.

[0037] In some embodiments, a deep learning-based language model can be used first to segment and tag the power grid instruction text, then the subject-verb-object structure of the sentence can be identified using syntactic analysis tree technology, and the dependency direction and connection relationship between the semantic elements of each power grid business can be determined based on grammatical relations, thereby constructing an association structure; no limitation is made here.

[0038] In other embodiments, a rule-based template matching approach can be adopted, in which common instruction sentence templates in the field of power grid dispatching are predefined, the input power grid instruction text is matched with the template, and the structured power grid business semantic elements and their preset association relationships are directly extracted from the matching results, thereby quickly constructing the association structure. This approach is not limited here.

[0039] Please see Figure 2 , Figure 2 This is a flowchart illustrating step S101 in an embodiment of this application;

[0040] In some specific embodiments, step S101 specifically includes:

[0041] S1011. Extract and segment the power grid instruction text according to the preset length threshold to obtain multiple power grid business semantic elements.

[0042] This step is the initial stage of semantic parsing, responsible for dividing a complete and continuous power grid instruction text into several independent semantic segments that are easy to calculate later, according to predetermined granularity rules.

[0043] S1012. Take any two power grid business semantic elements as the first power grid business semantic element and the second power grid business semantic element, and calculate the correlation between the first power grid business semantic element and the second power grid business semantic element; the correlation is determined by weighting vector similarity and co-occurrence statistics.

[0044] Among them, the correlation quantity is a numerical value used to quantify the degree of logical correlation between two semantic elements of power grid business.

[0045] Specifically, a quantitative relationship model is established between semantic elements of power grid business. This involves iterating through all element combinations obtained in the previous step and scoring each pair. This score is not based on a single dimension but rather considers two aspects: firstly, vector similarity assesses the semantic closeness between the two elements; secondly, co-occurrence statistics assess the frequency of their co-occurrence in historical power grid business data. By weighted fusion of these two dimensions, a comprehensive correlation quantity that reflects both semantic logic and domain experience is obtained, providing a mathematical basis for subsequently determining whether a dependency exists between the two elements.

[0046] In some embodiments, a pre-trained word embedding model can be used to convert the semantic elements of power grid business into high-dimensional vectors, and the cosine similarity between the two vectors can be calculated as a vector similarity term. At the same time, the frequency of the two elements appearing in the same window is counted in a large-scale corpus in the background, and normalization is performed to obtain a co-occurrence statistical term. Finally, the two items are added together according to a preset weight to obtain the association quantity, which is not limited here.

[0047] In other embodiments, the texts of two power grid business semantic elements can be concatenated and input into a fine-tuned twin network, which directly outputs a correlation score representing the strength of the correlation between the two elements end-to-end. This is not limited here.

[0048] In some specific embodiments, the function for calculating the correlation quantity is:

[0049]

[0050] In the formula, To obtain semantic elements of the first power grid business To the second power grid business semantic elements The correlation quantity; is; , As weighting coefficients, in some embodiments ; Semantic elements for the first power grid business Normalized vector representation; Semantic elements for the second power grid business Normalized vector representation; Semantic elements for the first power grid business Second power grid business semantic elements The frequency of co-occurrence in the historical corpus is used to calculate co-occurrence statistics. It is a very small positive number (to prevent division by zero constant) to avoid the denominator being zero.

[0051] S1013. Calculate the difference between the positive correlation between the first power grid business semantic element and the second power grid business semantic element, and the negative correlation between the second power grid business semantic element and the first power grid business semantic element.

[0052] This step determines whether there is an asymmetric, primary-secondary potential dependency relationship between the two power grid business semantic elements by calculating the difference in correlation strength in two directions.

[0053] S1014. If the difference is greater than a preset positive direction threshold, establish a dependency edge from the first power grid business semantic element to the second power grid business semantic element; if the difference is less than a preset negative direction threshold, establish a dependency edge from the second power grid business semantic element to the first power grid business semantic element.

[0054] In this context, a dependency edge refers to a directed connection in a graph structure that represents a logical dependence of one power grid business semantic element on another power grid business semantic element.

[0055] Specifically, this step is a decision-making process that makes potential dependencies explicit and directional based on the difference calculation results from the previous step. Two decision thresholds are set for positive and negative directions. Only when the influence of one element on another is sufficiently large—that is, the difference in their correlation exceeds this boundary—is a unidirectional dependency relationship recognized, and a directed dependency edge is established between the nodes representing these two elements. Through this step, the originally undirected and ambiguous relationships are transformed into directed relationships with a clear logical flow. This is a crucial step in constructing the final relationship structure, ensuring that the entire structure reflects the execution order and constraints of the instructions.

[0056] In some embodiments, the establishment of dependent edges can be achieved in a variety of ways: in other embodiments, a positive direction threshold and a negative direction threshold with the same absolute value (i.e., the opposite direction threshold) can be set. If the difference falls outside the two thresholds, a one-way dependent edge is established. If it falls between the two thresholds, the two are considered to be equal or have no explicit dependence, and no dependent edge is established. This is not limited here.

[0057] S1015. Determine the association structure based on the semantic elements of power grid business and the corresponding dependency edges.

[0058] This step is the final step in the semantic parsing phase. It assembles all processed power grid business semantic elements (as nodes) and determined dependency edges (as directed edges) to form a complete, structured graph-like association structure, which is then output to subsequent modules.

[0059] It is evident that calculating association strength by combining vector similarity with co-occurrence statistics considers both the deeper meaning of the words themselves and incorporates collocation habits from the domain, making the calculation of association strength more comprehensive and accurate. By calculating the difference between positive and negative associations and establishing unidirectional dependency edges based on thresholds, the originally ambiguous textual narrative can be transformed into a topological network with clear primary and secondary, sequential, and causal orientations. This avoids logical confusion caused by semantic comprehension biases. Ultimately, the constructed association structure is no longer a simple model, but a logical graph that accurately reflects the command flow, constraint relationships, and execution prerequisites within power grid instructions, improving the model's depth and accuracy in structurally understanding complex instructions.

[0060] In some preferred embodiments, after step S1015, the method further includes:

[0061] S1016. Traverse the topological paths of the associated structure, extract the set of closed paths with directed cycles, and the bidirectional dependent edges with mutually inverse directions.

[0062] In this context, a directed loop refers to a closed loop path formed by connecting three or more power grid business semantic elements end to end through dependency edges in an associated structure; a bidirectional dependency edge refers to two dependency edges with opposite directions existing simultaneously between two power grid business semantic elements.

[0063] This step is a quality inspection step after the associated structure is generated. It is responsible for searching and identifying all structural patterns in the entire logical network that may lead to logical dead loops or contradictions.

[0064] S1017. Calculate the average correlation between the closed path set and the bidirectional dependency edge respectively. If the average correlation is greater than the preset conflict threshold, determine that the corresponding dependency edge has a logical conflict.

[0065] Among them, the average correlation quantity refers to the arithmetic mean of the correlation quantities of all dependent edges that constitute a closed path set or bidirectional dependent edges; logical conflict refers to the existence of one or more combinations of dependencies, which makes it impossible for the correlation structure to form an acyclic, unidirectional logical sequence.

[0066] Specifically, this step doesn't simply treat all found circular or bidirectional structures as errors. Instead, it introduces a quantitative set of criteria to distinguish between harmless structural redundancy and fatal logical conflicts. The strength of these suspicious structures—that is, the average correlation of the dependent edges that constitute them—is calculated. Only when this strength exceeds a preset conflict threshold are these high-strength dependencies officially classified as logical conflicts. This step avoids the pitfall of classifying all closed paths as conflicts, improving the effectiveness of logical judgment and allowing the system to focus on handling contradictory dependencies with higher deviations.

[0067] In some embodiments, a fixed global conflict threshold may be set, and the average correlation of all detected loop or bidirectional structures is compared with this threshold; no limitation is made here.

[0068] In other embodiments, a dynamic conflict threshold may be used depending on the type or context of the power grid instruction text. For example, when processing operation ticket instructions with extremely high security requirements, a lower conflict threshold may be used to perform more stringent logical conflict review, while a relatively higher threshold may be used when processing routine work reports. This is not limited here.

[0069] In some embodiments, the formula for calculating the average correlation coefficient is:

[0070]

[0071] In the formula, The average association value of a closed set of paths (directed cycle); The number of dependent edges that constitute this closed path set; for to The correlation quantity.

[0072] S1018. For dependency edges with logical conflicts, use a preset interpolation function to adjust the associated quantities in the associated structure, and then jump to step S1013.

[0073] The interpolation function is a predefined mathematical function used to calculate a new output value based on the input value (such as the original correlation quantity and the external adjustment factor).

[0074] Specifically, upon identifying a dependency edge with a logical conflict, it invokes a preset interpolation function, such as linear interpolation or a weighted average function, to numerically adjust the original correlation values ​​on the conflicting dependency edge, typically by attenuating or reweighting. After adjustment, it does not immediately proceed to the next step but instead uses a jump mechanism to redirect the process back to step S1013, where the difference between the positive and negative correlation values ​​is recalculated based on the adjusted correlation values. This adjustment-reassessment loop aims to break the logical loop by fine-tuning the correlation strength, representing a low-computational-cost repair attempt.

[0075] In some embodiments, a linear decay method can be used, multiplying the association amount of logically conflicting dependency edges by a fixed decay coefficient less than 1, which is not limited here.

[0076] In some embodiments, the interpolation function is:

[0077]

[0078] In the formula, This is the adjusted correlation value. This represents the original association quantity; These are the weighting coefficients; These are external feedback adjustment items or preset benchmark values ​​used to assist in adjustments.

[0079] S1019. After updating the association structure, if the original closed path set or the original bidirectional dependency edge still has logical conflicts, the corresponding power grid business semantic elements are split or merged, and the process jumps to step S1012.

[0080] Specifically, if logical conflicts still exist after reconstructing the association structure, intervention is applied to the corresponding power grid business semantic elements. After adjusting the association quantities and updating the association structure in step S1018, conflict detection is performed again. If the paths or dependency edges that were previously identified as logical conflicts still exist and their strength still exceeds the limit, the soft adjustment is deemed ineffective. At this point, structural reconstruction is initiated, directly operating on the power grid business semantic elements that caused the conflict: for elements that are too general and cause ambiguity, a splitting process is performed, decomposing them into more specific and precise sub-elements; for semantically overlapping or actually synonymous elements, a merging process is performed, merging them into a single, unambiguous node. After the splitting or merging is completed, the process jumps back to step S1012, and the association quantities are recalculated based on the processed elements to eliminate logical contradictions in the structure.

[0081] In some embodiments, the splitting can be based on a preset domain ontology library. For example, when a power element causes a conflict, it can be split into two sub-elements, active power and reactive power, according to the hierarchical relationship in the ontology library. This is not limited here.

[0082] In other embodiments, merging can be performed by calculating the word embedding vector similarity between elements. When the vector similarity between two different elements exceeds a merging threshold, they are considered synonyms and merged into a standardized representative element. This is not limited to this approach.

[0083] It is evident that by traversing topological paths to extract potential logical deadlocks such as directed cycles or bidirectional dependencies, and introducing a comparison mechanism between average correlation quantity and conflict threshold, conflict determination is no longer based on fuzzy rules but on quantifiable mathematical indicators, thus improving diagnostic accuracy. Building upon this, a two-stage progressive reconstruction strategy based on interpolation adjustment and node splitting / merging is employed. The first stage performs soft logical guidance by adjusting correlation quantity. If this method fails, the second stage initiates a structural reorganization of the semantic elements themselves. Even when faced with unclear or even contradictory original instructions, it can iteratively correct itself, ultimately outputting a logically consistent and unambiguous correlation structure, providing a reliable data foundation for subsequent trustworthy reasoning.

[0084] S102. Calculate the uncertainty of the semantic elements of power grid business based on the associated structure.

[0085] Uncertainty refers to a comprehensive numerical indicator used to quantify the completeness of information, clarity of meaning, and logical consistency of a power grid business semantic element in the current context.

[0086] Specifically, a risk assessment is performed on each node (i.e., the semantic element of power grid business) in the association structure generated in the previous step. Each node is examined to determine if it presents any issues requiring external knowledge intervention. This calculation process serves as the basis for all subsequent dynamic strategy decisions. It transforms abstract semantic ambiguities into quantifiable and comparable values, enabling the identification of weaknesses in the instruction text. A higher calculated uncertainty value indicates a greater risk associated with that power grid business semantic element, thus requiring more focused attention and enhanced processing in subsequent steps.

[0087] In some embodiments, a multi-dimensional index weighted fusion approach can be adopted. First, a missing measure reflecting the completeness of information, an ambiguity measure reflecting the degree of semantic interpretation divergence, and a conflict measure reflecting the degree of logical contradiction with adjacent nodes are calculated respectively. Then, the three measures are linearly combined using preset weighting coefficients to obtain the final uncertainty score. No limitation is imposed here.

[0088] In other embodiments, a machine learning model-based prediction approach can be adopted, in which the feature vector of each power grid business semantic element and its adjacency information in the association structure are input into a pre-trained classification or regression model, and the model directly outputs the uncertainty score or risk level of the element. This is not limited here.

[0089] In some specific embodiments, step S102 specifically includes:

[0090] S1021. For any current power grid business semantic element in the association structure, calculate the missing measure reflecting the completeness of semantic information and the ambiguity measure reflecting the discreteness of semantic interpretation, and extract the association weight between the power grid business semantic element and the adjacent element to calculate the conflict measure reflecting the degree of contradiction in the semantic structure.

[0091] Among them, the missing measure is used to indicate the degree of lack of information required for the current power grid business semantic elements to constitute complete business logic; the ambiguity measure is used to indicate the possibility and degree of divergence of multiple different reasonable interpretations of the current power grid business semantic elements; and the conflict measure is used to indicate whether there is a logical contradiction or inconsistency between the current power grid business semantic elements and their surrounding adjacent elements in the association structure.

[0092] Specifically, this step is the core of multi-dimensional uncertainty quantification, aiming to comprehensively diagnose the inherent "health status" of a power grid business semantic element from three different and complementary perspectives. First, the system checks the completeness of the element's own information; for example, does an element representing "voltage adjustment" explicitly include a specific voltage value? If not, its missing value metric is high. Second, the system assesses the semantic clarity of the element; for example, expressions like "depending on the situation" have a large potential interpretive space, thus their ambiguity metric is high. Most importantly, the system utilizes the pre-constructed association structure to extract the association weights between the current element and its directly connected adjacent elements. By analyzing these weights (e.g., the difference between positive and negative associations), it calculates whether contradictory dependencies exist, thus obtaining a conflict metric. The calculation of these three metrics provides the foundational data for subsequent comprehensive uncertainty assessment.

[0093] In some embodiments, the calculation of these three metrics can be implemented in several ways: Optionally, the system can predefine an attribute completeness template for various power grid business semantic elements, and calculate the missing metric by comparing the current element with the template and counting the number of missing necessary attributes; simultaneously, the system can compare the word embedding vector of the current element with a preset semantic clustering center, and calculate the ambiguity metric based on its weighted distance to multiple centers; finally, the system extracts the association quantities of all edges pointing to and being pointed to by the current element from the association structure, and calculates the conflict metric based on the symmetry or difference of these association quantities; optionally, these three metrics can also be calculated in parallel by three independent neural network models fine-tuned for specific tasks, with each model taking into account the context information of the current power grid business semantic element and directly outputting the corresponding missing metric, ambiguity metric, or conflict metric score. It is understood that other methods can also be used to achieve the calculation of these three metrics, which are not limited here.

[0094] The function for calculating the missing metric is:

[0095]

[0096] In the formula, For missing data; This represents the number of valid pieces of information matched in the knowledge base or context; theoretically, the upper limit for the number of matched pieces of information is [value missing].

[0097] The function for calculating the ambiguity measure is:

[0098]

[0099] In the formula, For ambiguous measurement; The vector representing the cluster centers; This is the normalization factor.

[0100] The function for calculating the conflict metric is:

[0101]

[0102] In the formula, For conflict measurement; It is a set of semantic elements of adjacent power grid services; This is a positive correlation quantity. This represents the reverse correlation quantity.

[0103] S1022. Substitute the missing measure, ambiguity measure, and conflict measure into the calculation model with preset weighting coefficients to calculate the uncertainty.

[0104] This step involves integrating the three basic metrics obtained in the previous step and combining them into a final uncertainty score that comprehensively reflects the risk level of the semantic elements of the current power grid business through a weighted formula.

[0105]

[0106] In the formula, The uncertainty is... , , These are the weighting coefficients.

[0107] S103. Distinguish uncertainties according to uncertainty levels and select the target element set based on uncertainty levels.

[0108] Among them, uncertainty level refers to the discrete risk category that maps continuous uncertainty values, such as low, medium, high and extremely high; target element set refers to the set of all power grid business semantic elements that are determined to require external knowledge enhancement.

[0109] Based on the uncertainty score calculated in the previous step, all power grid business semantic elements are classified and screened. Elements with uncertainties exceeding the preset standard are screened to form a list to be processed for targeted operation in subsequent steps.

[0110] S104. Based on the priority scheduling strategy, perform dynamic retrieval and mode switching on the target element set in the corresponding alternative retrieval sources to obtain the corresponding evidence information.

[0111] Evidence information refers to text fragments or data records retrieved from external knowledge bases that can be used to clarify uncertainties, supplement missing information, or resolve logical conflicts.

[0112] Specifically, unlike traditional fixed retrieval, this method dynamically plans an optimal retrieval path based on the characteristics of each element in the target element set (such as uncertainty level and element type) and the status of alternative retrieval sources (such as historical performance and availability). This determines the dynamic polling order in which retrieval sources perform retrieval operations, and proactively changes the retrieval method (mode switching) when retrieval results are unsatisfactory, for example, switching from fast keyword matching to more time-consuming but more accurate deep semantic search. The entire process is a closed loop with feedback adjustment, aiming to obtain the highest quality and most relevant evidence information with the least computational cost and the highest efficiency.

[0113] In some embodiments, a scheduling strategy based on score ranking can be adopted. First, the historical matching quality, real-time availability and content relevance of each candidate retrieval source are comprehensively calculated to obtain an initial priority score. Then, the priority is adjusted by weighting according to the type of the target element, and the retrieval is performed in order of score. The score is dynamically updated according to the retrieval feedback. This is not limited here.

[0114] In other embodiments, a reinforcement learning-based agent strategy can be adopted to model the retrieval process as a Markov decision process. The agent learns an optimal retrieval strategy function through continuous trial and error (performing retrieval and obtaining reward signals). This function can directly output the best combination of retrieval source and retrieval method based on the current state (target element, load, etc.), which is not limited here.

[0115] Please see Figure 3 , Figure 3 This is a flowchart illustrating step S104 in an embodiment of this application;

[0116] In some specific embodiments, step S104 specifically includes:

[0117] S1041. Calculate the initial priority score of each candidate retrieval source based on historical matching quality, availability and relevance to the current task, and establish or update the scheduling table.

[0118] Among them, historical matching quality refers to the performance score of the candidate search source in providing accurate and relevant information in past tasks; availability refers to the service status indicators of the candidate search source such as current network connectivity and response speed; current task relevance refers to the overall relevance assessment of the content topic of the candidate search source to the set of target elements to be processed this time; the scheduling table is a dynamic sorted list used to record each candidate search source and its corresponding priority score.

[0119] By comprehensively evaluating the historical performance and current status of the candidate search sources, an initial priority score is calculated. Taking into account their historical performance, current status, and matching degree with the task, an initial search order suggestion is derived.

[0120] In some embodiments, the calculation function is:

[0121]

[0122] In the formula, Priority score; Match quality to historical records; For availability; Relevance to the current task; , , These are the weighting coefficients.

[0123] S1042. Extract the feature types of the target features in the target feature set and calculate the matching score between them and the capability feature vectors of the alternative retrieval sources; update the initial priority scores in the scheduling table based on the matching scores.

[0124] Among them, the element type refers to the classification of the domain attributes of the target element, such as entity, event, or background knowledge; the capability feature vector is a pre-defined high-dimensional vector used to represent the domain of expertise of each candidate retrieval source.

[0125] Specifically, instead of applying a uniform retrieval order to all target elements, the search is now customized for each individual target element. The type of the target element is analyzed—for example, is it a query for the physical parameters of a device or a market transaction rule? Then, the matching degree between this element type and the expertise tags (capability feature vectors) of each candidate search source is calculated. Candidate search sources whose expertise highly matches the current element type receive a significant weighted increase in their initial priority score. This dynamic adjustment ensures that for specific types of questions, the most specialized knowledge base is always prioritized, improving the hit rate and accuracy of the search.

[0126] In some embodiments, a fixed weighting coefficient matrix can be pre-defined for combinations of different feature types and capability feature vectors. After calculating the matching score, the corresponding coefficients are found in the matrix to update the initial priority score by multiplication. This is not limited here.

[0127] In other embodiments, a small neural network model can be used, taking the type vector of the target element and the capability feature vectors of the candidate retrieval sources as input. The network directly outputs a dynamic weighting adjustment factor, which is then used to adjust the initial priority score. It is understood that other methods can also be used to achieve the weighted update, and this is not limited here.

[0128] S1043. Perform the search in the candidate search sources according to the order of the schedule and obtain feedback scores.

[0129] The feedback score is a numerical value returned by the alternative search sources after executing this search task, used to indicate whether the search process itself was successful and the strength of the initial relevance between the search results and the query content.

[0130] This step is the execution phase of the retrieval action. Based on the dynamically updated scheduling table from the previous step, the actual query operation is initiated starting from the highest priority candidate retrieval source, and preliminary execution result feedback is received from that source.

[0131] S1044. Calculate the retrieval consistency, which reflects the alignment degree of a single retrieval, based on the feedback score.

[0132] Among them, retrieval consistency is an indicator used to quantify whether the information retrieved from one or more sources is mutually supportive in terms of content and whether there are any contradictions.

[0133] This step is a preliminary quality check of the retrieved content. By calculating a consistency score, it assesses whether the retrieved information has internal conflicts or whether the answers are not unique.

[0134] S1045. Determine whether the retrieval consistency is greater than the preset quality threshold and whether the change in priority score is within the allowable range.

[0135] This step is a decision-making checkpoint. It checks whether the consistency of the retrieved content meets the standards and whether there are drastic fluctuations in the priority ranking of the search sources to decide whether to accept the current search results or reschedule.

[0136] S1046. If so, the information obtained from the current retrieval will be determined as evidence information.

[0137] When the judgment conditions are met, the search is confirmed to be successful, and the obtained information is formally adopted as credible evidence and passed to subsequent modules for processing.

[0138] S1047. If not, change the order of the scheduling table and execute step S1043.

[0139] Specifically, when the consistency of the retrieved data does not meet the requirements (e.g., multiple pieces of information retrieved from a single source contradict each other), or when the actual performance of a high-priority source is found to be far below expectations (causing a sharp drop in its priority score), it is determined that the current search has failed to obtain high-quality evidence. In this case, the process does not terminate; instead, the order of the scheduling table is proactively changed. For example, the poorly performing search source is downgraded, or a candidate source is promoted. After the order change, a jump mechanism returns to step S1043, initiating a new round of searching according to the new order. This search-evaluation-adjustment-retry loop ensures that, through self-regulation, evidence information that meets the quality requirements can ultimately be obtained.

[0140] In some embodiments, a simple de-weighting strategy can be adopted, which involves directly multiplying the priority score of the search source with the lowest feedback score in this search by a penalty coefficient less than 1, and then reordering the scheduling table. This is not limited here.

[0141] In other embodiments, a switch to a retrieval method can be triggered. For example, if a single-source deep search fails, the strategy is changed to multi-source parallel retrieval, that is, queries are simultaneously initiated to several retrieval sources with the highest priority, and the returned results are cross-validated to increase the probability of obtaining high-quality evidence. It is understood that other methods can also be used to achieve the purpose of changing the order of the scheduling table, which is not limited here.

[0142] As can be seen, by initially prioritizing candidate search sources and then weighting the updates based on the matching scores between target element types and search source capability characteristics, a shift from blind general screening to on-demand selection is achieved, ensuring that each search hits the knowledge source most relevant to the current task with a high probability. A closed-loop quality control mechanism based on feedback scoring and search consistency is introduced. This feature transforms retrieval from a one-off action into a continuous self-evaluation and dynamic adjustment process. When search quality is substandard, the scheduling strategy can be proactively changed and the process re-executed until satisfactory evidence is obtained. This self-correcting capability ensures that high-quality external evidence is obtained with optimal resource consumption, balancing search efficiency, coverage, and the reliability of the final content.

[0143] S105. Map the evidence information to the target element set, calculate the evidence matching degree and update the evidence credibility, and generate an extended element set and the corresponding evidence structure.

[0144] Among them, evidence matching degree is an indicator that measures the degree of relevance between the retrieved evidence information and the target element to be processed in terms of content and structure; evidence credibility is a score that represents the degree of trust in the authenticity and reliability of a piece of evidence information or a source of evidence; the extended element set is a new set formed by incorporating verified evidence information on the basis of the original power grid business semantic elements; and the evidence structure is an enhanced association structure used to represent the internal structure of the extended element set and its logical relationship with the original elements.

[0145] The preliminary evidence obtained in the previous step is subjected to quality inspection and screening. High-quality and highly relevant evidence is integrated with the original semantic elements, and the data structure is updated to prepare enhanced and verified raw materials for subsequent causal inference.

[0146] Please see Figure 4 , Figure 4 This is a flowchart illustrating step S105 in an embodiment of this application;

[0147] In some specific embodiments, step S105 specifically includes:

[0148] S1051. Calculate the initial matching degree between the corresponding target element and the corresponding evidence information, wherein the initial matching degree is determined by the semantic similarity term, the structural consistency term and the contextual association term.

[0149] Among them, the semantic similarity term refers to the numerical value used to measure the degree of semantic similarity between the target element and the evidence information; the structural consistency term refers to the numerical value used to measure whether the data structure of the evidence information conforms to the format or type expected by the target element; and the contextual relevance term refers to the numerical value used to measure whether the context in which the evidence information is located is consistent with the context in which the target element is located.

[0150] Specifically, this step does not simply involve keyword comparison, but rather a comprehensive scoring process across three levels: first, semantic similarity to determine if the two pieces of evidence are equivalent in their core meaning; second, structural consistency to determine if the format of the evidence meets the requirements—for example, if the target element requires a numerical value, does the evidence provide the corresponding quantitative data? Finally, contextual relevance to determine if the context in which the evidence is generated matches the context of the current instruction. Integrating these three scores yields a more reliable initial matching degree than a single relevance score, providing a solid foundation for subsequent evidence screening and credibility assessment.

[0151] In some embodiments, the normalized scores of the three indicators can be calculated separately, and then the three scores can be linearly weighted and summed using preset weighting coefficients to obtain the final initial matching degree. This is not limited here.

[0152] In other embodiments, the feature vectors of the target element and the feature vectors of the evidence information can be concatenated and input into a pre-trained regression model, which directly outputs a comprehensive score representing the initial matching degree end-to-end. This is not limited here.

[0153]

[0154] In the formula, For evidence information With target elements The degree of matching; For semantic similarity terms; For structural consistency terms; For context-related items; , , These are the weighting coefficients.

[0155] S1052. Obtain the integrity score of the evidence information, and perform adjustment calculation on the initial matching degree through the integrity score to obtain the adjusted evidence matching degree.

[0156] The integrity score is a preliminary quantitative assessment of whether the evidence information itself is complete, without missing parts, and without truncation.

[0157] This step is a correction to the initial matching score obtained in the previous step. Its main purpose is to punish evidence with incomplete content and prevent misjudgments caused by accepting information taken out of context.

[0158] S1053. The credibility of the original evidence is obtained by performing a weighted correction based on the adjusted evidence matching degree.

[0159] The credibility of original evidence refers to the internally maintained benchmark score on the reliability of a piece of evidence or its source in historical tasks; the credibility of new evidence is the updated credibility score of evidence after incorporating the performance in the current task.

[0160] Specifically, the system internally maintains a long-term original evidence credibility score for each piece of evidence or its source, representing its historical reliability. In each task, this long-term credibility score is dynamically adjusted based on the actual performance of the retrieved evidence (i.e., the adjusted evidence match rate). Specifically, the original evidence credibility score is weighted and summed or exponentially smoothed with the adjusted match rate for the current task. If the performance is good, the credibility of new evidence increases; if the performance is poor, it decreases accordingly. This mechanism ensures that the credibility assessment of evidence is not static, but rather adaptively learns and evolves based on continuous interactive feedback.

[0161] In some embodiments, the exponential moving average (EMA) method can be used to calculate the credibility of new evidence as a weighted average of the credibility of the original evidence and the matching degree of the adjusted evidence, wherein the weighting coefficient determines the degree of impact of the performance of this task on long-term credit, which is not limited here.

[0162] In other embodiments, a framework based on Bayesian updates can be designed, where the original evidence credibility is regarded as the prior probability, the adjusted evidence matching degree is regarded as the likelihood, and the posterior probability is calculated using Bayes' formula as the new evidence credibility. This is not limited here.

[0163] In some specific embodiments, the calculation function is:

[0164]

[0165] In the formula, To assess the credibility of new evidence; The adjusted evidence matching score; The adjusted evidence matching score; To enhance the threshold; This is a step function; its value is 1 when its parameter is greater than zero, and 0 otherwise. This is the gain coefficient.

[0166] S1054. Determine the extended set of elements based on the credibility of the new evidence, and determine the corresponding evidence structure based on the extended set of elements.

[0167] This step is the final step in this embodiment. Based on the final credibility score updated in the previous step, all evidence information that meets the acceptance criteria is selected and merged with the original target elements to form an extended set of elements containing more and more reliable information. On this basis, the evidence structure for subsequent reasoning is constructed or updated.

[0168] It is evident that by integrating multi-dimensional indicators such as semantic similarity, structural consistency, and contextual relevance to calculate the initial matching degree, the degree of matching between evidence and target elements can be measured. Based on this, an adjustment is made by introducing an inherent completeness score for the evidence itself, avoiding the risk of secondary decision-making due to the acceptance of partially matching, incomplete evidence. Finally, by weighting and correcting the credibility of the original evidence, the credit rating of each evidence source can be dynamically evaluated and updated. This mechanism allows high-quality, highly matching evidence to play a more important role in subsequent reasoning, while low-quality noise information is naturally weakened. Ultimately, the generated evidence structure is a credible knowledge network that has undergone layer-by-layer screening and has clearly defined weights, providing solid and reliable argumentative support for constructing causal chains.

[0169] S106. Based on the correlation quantity in the evidence structure, filter the target correlation path, construct the causal chain, and calculate the corresponding reasoning credibility factor.

[0170] Among them, the target association path refers to the core logical pathway in the evidence structure where the association strength between nodes exceeds a preset threshold; the causal chain refers to a reasoning sequence consisting of a series of target association paths connected together, with a clear unidirectional, temporal, and causal logical relationship; the reasoning credibility factor is a quantitative confidence score obtained by comprehensively evaluating the evidence support strength, logical coherence, and consistency of the entire causal chain.

[0171] Specifically, the process goes beyond simply listing evidence. It involves identifying one or more of the strongest and most logically consistent chains of evidence within the complex network of the enhanced evidence structure. First, weakly related connections are filtered out, retaining only strongly related paths with sufficient correlation. Then, algorithms such as graph traversal are used to connect these paths, forming a clear causal chain from cause to effect. During construction, potential logical conflicts within the chain are actively detected and resolved to ensure the final causal chain is unidirectional and contradiction-free. Finally, this hard-won causal chain is scored overall, outputting a reasoning credibility factor that directly reflects the reliability of subsequent corrective decisions.

[0172] In some embodiments, the construction of causal chains and conflict resolution can be achieved in a variety of ways: In other embodiments, a threshold-based filtering and depth-first search method can be used. First, edges with a correlation value lower than the threshold in the evidence structure are removed. Then, a depth-first search is performed on the remaining graph starting from the starting node to form a path, which is the causal chain. When a fork is encountered, the path is selected based on the weight. This is not limited here.

[0173] S1061. Extract the connection lines of power grid business semantic elements in the evidence structure whose correlation quantity is greater than the preset correlation quantity threshold as target correlation paths, and generate a set of target correlation paths.

[0174] Among them, the target association path refers to the edge that is determined to have a strong logical relationship and connects two power grid business semantic elements in the evidence-enhanced association network.

[0175] This step involves pruning a complex network of evidence. By filtering through a threshold of correlation, only the most closely related and core logical connections are retained to simplify the network structure and highlight the main lines of reasoning.

[0176] S1062. Perform a graph traversal algorithm on the target associated path set to construct an initial causal chain with directed associated paths.

[0177] The initial causal chain refers to a reasoning sequence that is initially linked together based on strong correlations, reflecting a certain logical flow but possibly still containing local contradictions.

[0178] This step combines the scattered, strongly correlated paths selected in the previous step into one or more continuous logical chains with preliminary directionality, according to the connection rules of graph theory.

[0179] S1063. Detect the direction of the dependency relationship in the initial causal chain. If the dependency relationship is a contradictory dependency relationship, extract the corresponding power grid business semantic element as a conflicting power grid business semantic element.

[0180] This step involves performing an internal logical consistency check on the initially constructed causal chain, aiming to locate nodes in the chain that exhibit self-contradictions or logical dead loops.

[0181] S1064. Obtain the correlation quantity corresponding to the candidate branch path containing the semantic elements of the conflicting power grid business, sort the candidate branch paths according to the correlation quantity, and extract the optimal branch path.

[0182] Among them, candidate branch paths refer to multiple logical paths that can be selected at the point of conflict, pointing in different directions; the optimal branch path refers to the path that is determined to be the most reasonable and credible among all candidate branches according to preset rules.

[0183] Specifically, when a conflict point (i.e., a conflicting electrical grid business semantic element) is located in step S1063, it usually means that there are more than one logically mutually exclusive reasoning path starting from this node. At this point, all these candidate branch paths are extracted, and the correlation values ​​of the edges constituting these paths are obtained. By comparing these correlation values ​​(e.g., selecting the path with the largest sum of correlation values ​​or the highest average value), these candidate paths can be ranked. Finally, the highest-ranked path is extracted as the optimal branch path, providing a basis for decision-making in the next step of path reconstruction.

[0184] In some embodiments, a maximum weight path priority strategy can be adopted, which calculates the sum of all associated quantities on each candidate branch path and determines the path with the largest sum as the optimal branch path. This is not limited here.

[0185] In other embodiments, a prior knowledge rule base in the power grid business domain can be combined to perform a semantic-level rationality assessment of candidate branch paths. The assessment scores are then weighted and combined with the correlation quantity for ranking, thereby extracting the optimal branch path that is both supported by correlation strength and conforms to business logic. This is not limited here.

[0186] S1065. Reconstruct the initial causal chain based on the optimal branch path to perform conflict resolution. Under the condition that the reconstructed causal chain satisfies the constraints of temporal monotonicity and consistency, confirm the final causal chain.

[0187] Among them, temporal monotonicity means that in a causal chain, the power grid business semantic element that is the cause must be earlier than or equal to the power grid business semantic element that is the result in time; consistency means that there are no logically contradictory dependencies in the causal chain.

[0188] Specifically, based on the unique optimal branch path extracted in the previous step, the initial causal chain containing conflicts is processed. Specifically, the connection relationship of the optimal branch path is preserved, and all other contradictory candidate branch paths originating from the same conflict point are disconnected or deleted. After reconstruction, the result is not directly confirmed; instead, a final quality check is performed to verify whether the reconstructed causal chain strictly satisfies two core constraints: first, whether the logical order of all nodes conforms to chronological order (temporal monotonicity); and second, whether there are no longer any logical contradictions in the chain (consistency). Only when both conditions are met simultaneously will this conflict-free reasoning path, after resolution and verification, be confirmed as the final causal chain.

[0189] In some embodiments, a timestamp attribute can be attached to each power grid business semantic element. By traversing the final causal chain, it can be checked whether the timestamps of the preceding nodes in all dependencies are less than or equal to the timestamps of the following nodes to verify the temporal monotonicity. This is not limited here.

[0190] In other embodiments, the conflict detection algorithm can be executed again to confirm that there are no more directed cycles or strong bidirectional dependencies in the reconstructed causal chain, in order to verify its consistency; this is not limited here.

[0191] S1066. Calculate the reasoning credibility factor of the corresponding power grid instruction text based on the final causal chain.

[0192] This step is the final output stage of this embodiment. It performs an overall evaluation of the confirmed final causal chain, and calculates a total confidence score, i.e., the reasoning credibility factor, by taking into account its path length, the credibility of the evidence contained therein, and other information, and then outputs it.

[0193] As can be seen, filtering strongly correlated paths using correlation thresholds allows for the extraction of the core logical framework from complex evidence networks, eliminating interference from weakly correlated information. Furthermore, the use of graph traversal algorithms ensures that the constructed initial causal chain is structurally connected. Crucially, a conflict node identification and resolution mechanism for contradictory dependencies is introduced. By sorting candidate branch paths and selecting the optimal branch, logical paradoxes arising from multi-source evidence or ambiguous instructions can be proactively resolved. The final confirmed causal chain, under the constraints of temporal monotonicity and consistency, transforms the originally disordered text elements into an irreversible reasoning path from cause to effect, providing clear, consistent logical guidance and a basis for credibility calculations for subsequent text revisions.

[0194] S107. Based on the reasoning credibility factor, locate the problematic power grid business semantic elements in the power grid instruction text, select the corresponding strategy to perform text correction, and output the power grid corrected text.

[0195] Among them, problematic power grid business semantic elements refer to key nodes that are identified as having low credibility, weak evidence support, or being at the center of logical conflict during causal reasoning; power grid corrected text refers to the final text that is more accurate, logically clear, and credible after a series of processing and corrections based on the original power grid instruction text.

[0196] Based on the reasoning results, the nodes to be corrected in the original instructions are located and the corresponding modification strategies are executed. At the same time, constraints are constructed based on consistency checks to ensure the standardization of the correction.

[0197] It is evident that semantic parsing of power grid command text and the construction of its relational structure provide a structured logical foundation for all subsequent judgments. Building upon this, the calculation and classification of uncertainties in the semantic elements of power grid operations are introduced, enabling the perception of inherent risks in the input text and the identification of high-risk areas requiring external knowledge intervention. Furthermore, the uncertainty level is used to dynamically guide subsequent retrieval, mapping, causal reasoning, and correction processes, ensuring that computational resources are allocated to the semantic elements most in need of governance. This transforms power grid command parsing from a black-box, probabilistic generation process into a reliable and efficient processing flow where every decision is verifiable, logically traceable, and self-correcting, improving adaptability to power grid dispatch command processing, thereby reducing resource consumption and increasing command accuracy.

[0198] In some specific embodiments, step S107 specifically includes:

[0199] S1071. Based on causal chains and reasoning credibility factors, locate the target correction power grid business semantic elements in the power grid instruction text.

[0200] This step is the starting point for text correction. Based on the reasoning results of the previous steps, the target correction elements that are judged to have the lowest credibility and most in need of modification are accurately found and marked in the original power grid instruction text.

[0201] S1072. Extract the reasoning credibility factor corresponding to the target modified power grid business semantic element, construct the power grid business semantic element credibility feature vector; and construct multiple sets of strategy feature vectors corresponding to different modification methods. The strategy feature vector includes the fit parameter, risk coefficient and strategy depth multi-dimensional factor.

[0202] This step involves preparing data for selecting an appropriate correction strategy. On the one hand, the current credibility status of the element to be corrected is vectorized. On the other hand, all alternative correction strategies (such as replacement, insertion, etc.) are also vectorized into multi-dimensional features.

[0203] S1073. Calculate the similarity between the credible feature vector of the semantic elements of power grid business and the feature vector of each group of strategies as the strategy preference, and select the correction method with the largest preference value as the optimal correction method.

[0204] Specifically, the similarity (e.g., cosine similarity or Euclidean distance) is calculated one by one between the credible feature vector representing the severity of the current problem and the policy feature vector representing the characteristics of various correction strategies. The calculated similarity value is the policy preference, which quantifies the applicability of a certain correction strategy to the current problem. Finally, the correction method with the highest policy preference value is selected as the optimal correction method to be executed in this operation, thereby realizing the automation and personalized selection of correction strategies and avoiding the use of a single, fixed correction rule.

[0205] In some embodiments, all calculated similarity values ​​can be directly compared, and the correction method corresponding to the maximum value can be selected as the optimal one; this is not limited here.

[0206] In other embodiments, a Softmax function with a temperature coefficient can be used to normalize all similarity values ​​to obtain a probability distribution of the selection of each correction method. Then, sampling or selection of the method with the highest probability based on this probability distribution is performed to introduce a certain degree of randomness and avoid getting trapped in local optima. It is understood that other methods can also be used to achieve the purpose of selecting the optimal correction method, which is not limited here.

[0207] S1074. Invoke the optimal correction method to perform replacement, insertion, or expansion operations on the corresponding target correction power grid business semantic elements in the power grid instruction text, and generate local correction results.

[0208] Among them, the local correction result refers to the temporary text fragment generated after modifying only the target power grid business semantic elements and their adjacent context.

[0209] This step is the execution phase of the correction action. Based on the optimal strategy selected in the previous step, specific textual or data-level modifications are made to the problematic parts of the original text, resulting in a preliminary draft of the correction.

[0210] S1075. Calculate the matching degree between the local correction results and the evidence structure and causal chain respectively, and combine the calculation results with weight to obtain the consistency score.

[0211] The consistency score is a numerical value used to quantitatively assess whether the results of local corrections are consistent with the previously established evidence system and causal logic.

[0212] This step involves a fact-checking and logical verification of the generated draft amendment to ensure that the revised content does not deviate from the existing evidence and reasoning.

[0213] S1076. If the consistency score is greater than the set consistency threshold, the local correction results are merged into the power grid instruction text.

[0214] Once the draft amendment has passed verification, a merging operation is performed, and the modified content is formally applied to the complete power grid instruction text.

[0215] S1077. If the value is less than the consistency threshold, then this correction is abandoned.

[0216] If the verification results show that the draft amendment conflicts with the evidence or logic, the modification will be reversed to prevent the introduction of new errors.

[0217] S1078. For the merged power grid command text, call the confidence decay function to perform smoothing penalty calculation on the inference confidence factor to obtain the smoothed confidence factor.

[0218] The confidence decay function is a preset function that can smoothly lower the confidence score based on the depth or magnitude of the correction action; the smoothed confidence factor is a conservative assessment value used to reflect the overall confidence of the final output text after the decay penalty calculation.

[0219] This step involves making a slight downward adjustment after a successful correction, in order to reflect the principle that any modification may introduce potential risks.

[0220] In some embodiments, the calculation function is:

[0221]

[0222]

[0223] In the formula, This is the confidence decay function, whose value varies with the parameter. It decreases as it increases; Variables such as the depth, magnitude, or number of corrections; This is the attenuation rate coefficient; This is the smoothing confidence factor obtained after smoothing penalty calculation; As a factor of reasoning credibility; This represents the confidence decay function value. This is the smoothing coefficient.

[0224] S1079. If the smoothing credibility factor is lower than the preset target threshold, proceed to the step of performing dynamic retrieval and mode switching on the target element set in the corresponding alternative retrieval sources based on the priority scheduling strategy to obtain the corresponding evidence information.

[0225] Specifically, after a series of corrections to the text, a final smoothed confidence factor is obtained. This factor is compared with a target threshold representing the lowest acceptable confidence level. If the smoothed confidence factor is lower than this threshold, it indicates that even after self-correction, confidence in the final result is still insufficient, and outputting the result carries a high risk. At this point, the output process is proactively halted, and a jump mechanism is used to redirect the entire processing flow back to the evidence retrieval stage (i.e., the fourth major step in claim 1). This callback will trigger a deeper secondary search targeting the current low-confidence region, aiming to obtain stronger evidence to support or refute the previous conclusion.

[0226] In some embodiments, a secondary search flag can be attached to the target element during the jump. After the scheduling module receives the flag, it will automatically increase the search depth, for example, by switching from single-source search to multi-source cross-validation search. This is not limited here.

[0227] In other embodiments, an updated set of target elements may be passed during the jump, which may include new points of doubt that have arisen during the correction process, thereby making the secondary search more targeted, which is not limited here.

[0228] S10710. If the smoothing confidence factor is not lower than the target threshold, determine the fused power grid command text as the power grid correction text and output it.

[0229] Once the final evaluation confirms that the overall credibility of the text, after all revisions and verifications, meets the standards, it is officially finalized as the final power grid revised text and output.

[0230] As can be seen, by calculating the similarity between the credible feature vector and the policy feature vector to select the optimal correction method, an upgrade from rule-based encoding modification to data-driven adaptive correction is achieved. After the correction is performed, a consistency score verification between the local correction result and the evidence structure and causal chain is introduced. This feature sets up a safety barrier for the correction operation, preventing the introduction of new logical errors due to improper modification. A smoothing penalty and secondary retrieval jump mechanism based on the confidence decay function are designed. When the smoothing confidence factor does not meet the output requirements, the system will trigger a callback to the retrieval stage for supplementary verification, effectively limiting the abnormal results caused by potential data bias. This closed-loop control logic with feedback callback, combined with the final structured audit record, ensures the high reliability, security, and post-event traceability of the power grid correction text.

[0231] The following describes an exemplary causal and reliable reasoning system 500 based on the semantic uncertainty of power grid commands, provided in an embodiment of this application. Figure 5 This is an exemplary hardware structure diagram of a causal and reliable reasoning system 500 based on the semantic uncertainty of power grid commands provided in this application embodiment.

[0232] In some embodiments, the causal reliable reasoning system 500 guided by the semantic uncertainty of power grid commands is a computer device or includes a computer device. The computer device includes a processor, memory, and a network interface connected via a system bus. The processor of the computer device provides computational and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the computer device stores data. The network interface of the computer device is used to communicate with other external terminals or servers via a network connection. In some embodiments, the network interface can be a wired network interface; in some embodiments, the network interface can be a wireless network interface. When the computer program is executed by the processor, it implements the methods in the embodiments of this application.

[0233] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0234] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0235] As used in the above embodiments, depending on the context, the term "when..." can be interpreted as meaning if... or after... or in response to determining... or in response to detecting... Similarly, depending on the context, the phrase "when determining... or if (the stated condition or event) is interpreted as meaning if determining... or in response to determining... or in response to detecting (the stated condition or event)" or in response to detecting (the stated condition or event).

[0236] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive), etc.

[0237] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A causal and reliable reasoning method based on the semantic uncertainty of power grid commands, characterized in that, include: Semantic parsing is performed on the power grid instruction text to determine the association structure containing power grid business semantic elements; The uncertainty of the semantic elements of power grid business is calculated based on the aforementioned association structure; The uncertainties are classified according to their uncertainty levels, and a set of target elements is selected based on these uncertainty levels. Based on a priority scheduling strategy, the target element set is dynamically searched and the method is switched among the corresponding alternative search sources to obtain the corresponding evidence information. The evidence information is mapped to the target element set, the evidence matching degree is calculated and the evidence credibility is updated, and an extended element set and the corresponding evidence structure are generated. Based on the correlation quantity in the evidence structure, target correlation paths are selected, causal chains are constructed, and the corresponding reasoning credibility factors are calculated. Based on the reasoning credibility factor, locate the problematic power grid business semantic elements in the power grid instruction text, select the corresponding strategy to perform text correction, and output the corrected power grid text.

2. The method according to claim 1, characterized in that, The step of performing semantic parsing based on the power grid instruction text to determine the association structure containing power grid business semantic elements includes: The power grid instruction text is extracted and segmented according to a preset length threshold to obtain multiple power grid business semantic elements; Take any two of the aforementioned power grid business semantic elements as the first power grid business semantic element and the second power grid business semantic element, and calculate the correlation between the first power grid business semantic element and the second power grid business semantic element; the correlation is determined by a weighted average of vector similarity and co-occurrence statistics. Calculate the difference between the positive correlation between the first power grid business semantic element and the second power grid business semantic element, and the negative correlation between the second power grid business semantic element and the first power grid business semantic element; If the difference is greater than a preset positive threshold, a dependency edge is established between the first power grid business semantic element and the second power grid business semantic element; if the difference is less than a preset negative threshold, a dependency edge is established between the second power grid business semantic element and the first power grid business semantic element. The association structure is determined based on the semantic elements of the power grid business and the corresponding dependency edges.

3. The method according to claim 2, characterized in that, After determining the association structure based on the power grid business semantic elements and the corresponding dependency edges, the method further includes: Traverse the topological path of the associated structure, extract the set of closed paths with directed loops, and the bidirectional dependent edges with mutually inverse directions; Calculate the average correlation between the closed path set and the bidirectional dependency edge respectively. If the average correlation is greater than a preset conflict threshold, determine that the corresponding dependency edge has a logical conflict. For dependency edges with logical conflicts, a preset interpolation function is used to adjust the correlation quantity in the correlation structure, and the step of calculating the difference between the positive correlation quantity from the first power grid business semantic element to the second power grid business semantic element and the negative correlation quantity from the second power grid business semantic element to the first power grid business semantic element is skipped to update the correlation structure. After updating the association structure, if the original closed path set or the original bidirectional dependency edge still has logical conflicts, the corresponding power grid business semantic elements are split or merged, and the process jumps to the step of taking any two of the power grid business semantic elements as the first power grid business semantic element and the second power grid business semantic element, and calculating the association amount between the first power grid business semantic element and the second power grid business semantic element.

4. The method according to claim 1, characterized in that, The step of performing dynamic retrieval and mode switching on the target element set in the corresponding candidate retrieval sources based on the priority scheduling strategy to obtain the corresponding evidence information specifically includes: Calculate the initial priority score of each candidate retrieval source based on historical matching quality, availability, and relevance to the current task, and establish or update the scheduling table. Extract the element types of the target elements in the target element set and calculate the matching score between them and the capability feature vectors of the alternative retrieval sources; update the initial priority scores in the scheduling table based on the matching scores. The search is performed in the candidate search sources according to the order of the scheduling table, and feedback scores are obtained. The retrieval consistency, reflecting the alignment degree of a single retrieval, is calculated based on the feedback score. Determine whether the retrieval consistency is greater than a preset quality threshold and whether the change in the priority score is within an allowable range; If so, the information obtained from the current retrieval will be determined as the evidence information; If not, then change the order of the scheduling table and execute the steps of performing a search in the alternative search sources according to the order of the scheduling table and obtaining feedback scores.

5. The method according to claim 1, characterized in that, The steps of mapping the evidence information to the target element set, calculating the evidence matching degree and updating the evidence credibility, and generating an extended element set and corresponding evidence structure include: Calculate the initial matching degree between the corresponding target element and the corresponding evidence information, wherein the initial matching degree is determined by a semantic similarity term, a structural consistency term, and a contextual association term; Obtain the integrity score of the evidence information, and perform an adjustment calculation on the initial matching degree based on the integrity score to obtain the adjusted evidence matching degree; The new evidence credibility is obtained by performing a weighted correction on the original evidence credibility based on the adjusted evidence matching degree; The extended set of elements is determined based on the credibility of the new evidence, and the corresponding evidence structure is determined based on the extended set of elements.

6. The method according to claim 1, characterized in that, The steps of filtering target association paths based on the correlation quantity in the evidence structure, constructing causal chains, and calculating the corresponding reasoning credibility factors specifically include: Extract the connections of power grid business semantic elements in the evidence structure whose correlation quantity is greater than a preset correlation quantity threshold as target correlation paths, and generate a set of target correlation paths; A graph traversal algorithm is performed on the target set of associated paths to construct an initial causal chain with directed associated paths; Detect the direction of the dependency relationship in the initial causal chain. If the dependency relationship is a contradictory dependency relationship, extract the corresponding power grid business semantic element as a conflicting power grid business semantic element. Obtain the correlation quantity corresponding to the candidate branch path containing the semantic elements of the conflicting power grid business, sort the candidate branch paths according to the correlation quantity, and extract the optimal branch path; The initial causal chain is reconstructed based on the optimal branch path to perform conflict resolution. The final causal chain is confirmed under the condition that the reconstructed causal chain satisfies the constraints of temporal monotonicity and consistency. The reasoning credibility factor of the corresponding power grid instruction text is calculated based on the final causal chain.

7. The method according to claim 1, characterized in that, The steps of locating problematic power grid business semantic elements in the power grid instruction text based on the reasoning credibility factor, selecting the corresponding strategy to perform text correction, and outputting the corrected power grid text specifically include: Based on the causal chain and the reasoning credibility factor, locate the target corrected power grid business semantic elements in the power grid instruction text; Extract the reasoning credibility factor corresponding to the target modified power grid business semantic element, construct the power grid business semantic element credibility feature vector; and construct multiple sets of strategy feature vectors corresponding to different modification methods. The strategy feature vector includes a fitness parameter, a risk coefficient, and a strategy depth multidimensional factor. The similarity between the credible feature vector of the power grid business semantic elements and the feature vector of each group of strategies is calculated as the strategy preference, and the correction method with the largest preference value is selected as the optimal correction method. The optimal correction method is invoked to perform replacement, insertion, or expansion operations on the target corrected power grid business semantic elements in the power grid instruction text, generating local correction results. The matching degree is calculated by comparing the local correction results with the evidence structure and the causal chain respectively, and the calculation results are weighted and combined to obtain the consistency score. If the consistency score is greater than the set consistency threshold, the local correction result will be integrated into the power grid instruction text. If the value is less than the aforementioned consistency threshold, then this correction is abandoned; For the fused power grid command text, the confidence decay function is invoked to perform a smoothing penalty calculation on the inference confidence factor to obtain the smoothed confidence factor; If the smoothing credibility factor is lower than the preset target threshold, the process jumps to the step of performing dynamic retrieval and mode switching on the target element set in the corresponding alternative retrieval sources based on the priority scheduling strategy to obtain the corresponding evidence information. If the smoothing reliability factor is not lower than the target threshold, the fused power grid command text is determined as the power grid correction text and output.

8. A causal and reliable reasoning system based on the semantic uncertainty of power grid commands, characterized in that, The causal and reliable reasoning system guided by the semantic uncertainty of power grid instructions includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code includes computer instructions, and the one or more processors call the computer instructions to cause the causal and reliable reasoning system guided by the semantic uncertainty of power grid instructions to perform the method as described in any one of claims 1-7.

9. A computer program product containing instructions, characterized in that, When the computer program product is run on a causal and reliable reasoning system guided by the semantic uncertainty of power grid commands, the causal and reliable reasoning system guided by the semantic uncertainty of power grid commands performs the method as described in any one of claims 1-7.

10. A computer-readable storage medium comprising instructions, characterized in that, When the instruction is executed on a causal and reliable reasoning system guided by the semantic uncertainty of power grid instructions, the causal and reliable reasoning system guided by the semantic uncertainty of power grid instructions performs the method as described in any one of claims 1-7.

Images