An automated report generation and data processing method

Abstract

The present application relates to the technical field of photovoltaic power generation operation and maintenance, and particularly relates to an automatic report generation and data processing method, comprising: responding to the start of a report generation task; calling a large language model; based on a pre-constructed photovoltaic equipment nameplate parameter knowledge graph; filling the calibrated numerical value into the report template to form a report draft. The present application constructs a dual verification mechanism of physical rationality verification and semantic logic confrontation verification working together, performs nameplate parameter boundary verification and amplitude limiting calibration before numerical filling, eliminates physically absurd data from entering the report from the source, establishes the correlation between numerical values and description texts to quantify the degree of contradiction, triggers a reverse interrogation to drive the large model to generate operation and maintenance suggestions when a logical conflict is detected, realizes the functional transition from passive error correction to active value-added, significantly reduces the manual review cost, and guarantees the physical compliance and semantic consistency of the report content.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation operation and maintenance technology, and in particular to an automated report generation and data processing method. Background Technology

[0002] With the large-scale development of distributed photovoltaic (PV) operations, operating companies are facing a new normal characterized by numerous plant areas, diverse customer needs, and a wide variety of report types. Individual customers have different requirements for the content, format, and statistical cycle of reports for inspections, quality checks, cleaning, and maintenance. Traditional manual report generation methods are insufficient to meet the needs of large-scale operations. To improve report generation efficiency, the industry is gradually adopting automated report generation technology based on template matching and automatic data filling. A typical implementation involves automatically retrieving business data from the maintenance platform database, filling the data into the corresponding positions according to a preset template, and generating a standard-format report document. Furthermore, with the deepening application of large language models in text generation, some solutions are beginning to introduce large language models to assist in generating descriptive analysis paragraphs in reports, aiming to reduce the burden of manual writing and achieve intelligent generation of report content.

[0003] However, existing report generation methods based on large language models have certain shortcomings: First, they lack verification of the physical compliance of the generated numerical content. During the autoregressive generation process, the large model may output values ​​that violate the physical limits of the photovoltaic equipment nameplate parameters, leading to reports containing physically absurd content such as inverter efficiency exceeding limits or voltage exceeding limits. Second, they lack the ability to detect the logical consistency between numerical data and generated text. When numerical values ​​are abnormal but the text description remains normal, the system cannot identify the semantic contradiction between the two. For example, Chinese patent CN121766277A discloses a report generation method that obtains key report reference information from a pre-generated database and inputs it into a large model. It imposes constraints on the generated content at the input stage to mitigate the illusion problem, but this constraint only applies to the input stage and cannot affect the autonomous generation of the large model during the generation stage. Real-time verification and correction of absurd physical values ​​generated; for example, Chinese patent CN121365699A discloses a large-model illusion suppression method based on dynamic knowledge base and multimodal consistency constraints. This method calculates the similarity between generated content and retrieved knowledge in the feature space during the model decoding stage and adjusts the generation probability in real time to suppress illusionary content inconsistent with the knowledge base facts. However, this method focuses on semantic-level factual truth judgment and does not establish cross-modal association between numerical data and descriptive text. It cannot effectively detect logical contradictions between numerical values ​​and text. In summary, existing technologies struggle to simultaneously ensure the physical compliance of numerical values ​​and the semantic consistency between numerical values ​​and descriptive text during automated report generation. There is an urgent need for an automated report generation method capable of performing both physical rationality verification and semantic logic verification of the generated content. Summary of the Invention

[0004] The technical problem to be solved by this invention is that the existing technology cannot simultaneously realize the physical compliance verification of the generated values ​​and the logical contradiction detection of the semantic consistency between the numerical data and the generated text during the automated report generation process. To this end, we propose an automated report generation and data processing method.

[0005] To achieve the above objectives, this application adopts the following technical solution: an automated report generation and data processing method, comprising: in response to the initiation of a report generation task, retrieving business data of the target plant area within a specified period from the operation and maintenance platform database, and matching the corresponding report template according to the report type and report period; calling a large language model, organizing the retrieved business data into a data context according to the template structure and inputting it into the large language model to generate a draft report, the draft report including descriptive text and numerical filling results; based on a pre-built photovoltaic equipment nameplate parameter knowledge graph, performing physical rationality verification on each value in the numerical filling results, and performing physical limiting calibration on values ​​that fail the verification to obtain calibrated values; The calibrated values ​​are filled into the corresponding positions in the report template to form a draft report. The numerical and semantic relationships between each value in the draft report and the corresponding descriptive text are constructed. For each set of numerical and semantic relationships, a semantic tendency score is calculated to quantify the degree of deviation between the text description and the actual state of the values, and a numerical anomaly score is calculated to quantify the degree of deviation of the values ​​themselves from the normal reference range. The semantic tendency score and the numerical anomaly score are combined to calculate the contradiction score. When the contradiction score exceeds a preset threshold, a reverse interrogation instruction is constructed to trigger the large language model to regenerate the analysis text for the abnormal data to replace the original descriptive text. After processing all numerical and semantic relationships, the final report document is generated.

[0006] Preferably, the photovoltaic equipment nameplate parameter knowledge graph is stored in the form of a graph database. The nodes include equipment type nodes, equipment model nodes, and parameter nodes. The edges represent the subordinate relationship between the equipment and the parameters, as well as the constraint relationship between the parameters. The photovoltaic equipment nameplate parameter knowledge graph stores at least the maximum DC input voltage, maximum conversion efficiency, and European efficiency curve data for each model of inverter, and stores at least the open-circuit voltage temperature coefficient and short-circuit current temperature coefficient for photovoltaic strings.

[0007] Preferably, the physical rationality verification includes: absolute boundary verification, determining whether the value exceeds the absolute physical limit specified by the corresponding equipment nameplate parameters; relative constraint verification, determining whether the value satisfies the physical constraint relationship between parameters stored in the photovoltaic equipment nameplate parameter knowledge graph; and environmental correlation verification, determining whether the value matches the environmental parameters. For the open-circuit voltage value of the photovoltaic string, the theoretical open-circuit voltage value is calculated based on the open-circuit voltage temperature coefficient stored in the photovoltaic equipment nameplate parameter knowledge graph and the current ambient temperature. If the relative deviation between the generated value and the theoretical value exceeds a preset threshold, it is determined to be questionable and awaiting review.

[0008] Preferably, the physical limiting calibration includes: for values ​​that fail the absolute boundary check or relative constraint check, determining the limiting target value according to the check failure type, replacing the original generated value with the limiting target value, and adding a calibration mark at the location of the value.

[0009] Preferably, the construction of the numerical and semantic association includes: traversing each block in the report template, extracting the value, unit, and physical meaning tags of each numerical value in the data block, extracting assertion statements related to the numerical value from the text description block; associating the numerical value and the assertion statement according to the rules of device identifier matching, time range matching, and physical meaning matching, to form a numerical and semantic association that includes the actual value of the numerical value, the upper and lower limits of the normal reference range, and the associated assertion text.

[0010] Preferably, the semantic tendency score is calculated as follows: a preset semantic tendency mapping table is used to map keywords in the text description to corresponding expected numerical ranges; keywords in the assertion text are extracted and matched with the semantic tendency mapping table to obtain the expected value range; the semantic tendency score is obtained by linear interpolation based on the deviation between the actual numerical value and the expected value range, and the score range is from 0 to 1.

[0011] Preferably, the calculation method for the numerical anomaly score is as follows: determine the lower and upper limits of the normal reference range for the numerical value; if there is concurrent data for the device under similar operating conditions in the historical database, then the mean of the historical data plus or minus twice the standard deviation is used as the reference range boundary; otherwise, the recommended operating range of the device's nameplate parameters in the photovoltaic device nameplate parameter knowledge graph is used as the reference range; calculate the first anomaly contribution based on the degree to which the actual numerical value is lower than the lower limit of the reference range, calculate the second anomaly contribution based on the degree to which the actual numerical value is higher than the upper limit of the reference range, and add the first anomaly contribution and the second anomaly contribution to obtain the numerical anomaly score, with the score ranging from 0 to 2.

[0012] Preferably, the contradiction score is calculated as follows: an adjustment factor is set, the numerical anomaly score is multiplied by the adjustment factor, and then input together with the semantic tendency score into a nonlinear fusion function in the form of a product to obtain a contradiction score with a value range of 0 to 1; when the semantic tendency score and the numerical anomaly score are both high, the increase in the contradiction score is greater than the sum of their independent contributions.

[0013] Preferably, the construction of the reverse interrogation instruction triggers the large language model to regenerate the analysis text for the abnormal data, including: extracting the abnormal values, numerical physical meaning labels, normal reference ranges, and original description text from the numerical and semantic correlation of the contradiction scores exceeding a preset threshold; generating an instruction containing the device name, time period, indicator name and actual value, abnormal feature description, and original contradiction explanation; requiring the large language model to generate analysis text containing possible causes of the anomaly, suggested investigation directions, and recommended handling measures; after obtaining the newly generated analysis text, replacing the corresponding original description text in the report draft, and retaining the original description text as a historical version.

[0014] Preferably, the calibrated value output by the physical rationality verification is used as the actual value used when calculating the numerical anomaly score; when multiple values ​​of the same device pass the physical rationality verification and the number of times the description text corresponding to the device triggers reverse challenge exceeds a preset number, the preset threshold used for each verification item in the subsequent physical rationality verification of the device is reduced.

[0015] The technical effects and advantages of this invention: Compared with the prior art, this invention constructs a hierarchical and bidirectional collaborative dual verification mechanism, which solves the technical problem of the difficulty in balancing physical absurdity and semantic contradiction when generating photovoltaic operation and maintenance reports from large models. First, by introducing physical rationality verification based on the knowledge graph of photovoltaic equipment nameplate parameters, the AI-generated content is subjected to triple verification of absolute boundary, relative constraint and environmental correlation before the values ​​are filled into the template, and forced amplitude limiting calibration is implemented for out-of-bounds values. This prevents absurd data that violates physical laws, such as inverter efficiency exceeding limits and voltage exceeding limits, from entering the report from the source, filling the gap in the generation stage where the input constraint scheme cannot intervene in the illusion.

[0016] Based on the above, the semantic and logical adversarial verification module further constructed does not stop at the traditional fact consistency comparison, but quantifies the degree of contradiction between the numerical value and the descriptive text by establishing the correlation between the two. When cross-modal logical conflicts such as zero power generation being described as stable operation are detected, the system does not simply report an error and interrupt the process, but automatically constructs a reverse interrogation instruction to drive the large model to transform the original generation defects into professional operation and maintenance troubleshooting suggestions for abnormal data, thereby realizing the functional leap from passive error correction to proactive value-added.

[0017] The reliable values ​​output by physical verification provide a reliable input benchmark for semantic verification, while contradictory signals discovered by semantic verification can in turn increase the rigor of physical verification. Together, they form a closed loop of cross-verification of the physical compliance and semantic consistency of the report content. This enables the report generation process to be constrained by physical common sense and has the ability to conduct cross-modal logical self-consistency review, significantly reducing the cost of manual review. While greatly improving the efficiency of large-scale report production, it ensures the professional credibility of the delivered content and effectively supports the dual stringent requirements of report quality and generation efficiency in distributed photovoltaic operation scenarios. Attached Figure Description

[0018] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts:

[0019] Figure 1 Generate an overall flowchart for the report of this invention; Figure 2 This is a detailed flowchart of the physical constraint verification process of the present invention; Figure 3 This is a detailed flowchart of the semantic logic adversarial verification process of the present invention; Figure 4 This is a schematic diagram of the dual verification and collaborative relationship of the present invention; Figure 5 This is a diagram of the overall system architecture of the present invention. Detailed Implementation

[0020] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0021] Example 1: This example provides an automated report generation and data processing method. The method is executed by an automated report generation system deployed on a server. The system establishes a data connection with the operation and maintenance platform database and calls a large language model through an application programming interface to generate descriptive content in the report.

[0022] The technical solution of this embodiment is to establish a dual verification mechanism that coordinates physical constraint verification and semantic logic adversarial verification. The physical constraint verification module is used to verify whether the numerical values ​​generated by the large language model conform to the physical laws specified by the parameters on the photovoltaic equipment nameplate. The semantic logic adversarial verification module is used to verify whether there is a logical contradiction between the numerical values ​​that have passed the physical verification and the descriptive text generated by the large language model. The two modules support each other in function: the numerical values ​​that pass the physical verification are directly used as the input benchmark for the semantic verification, and the contradictions found by the semantic verification trigger the large language model to regenerate professional analysis content for abnormal data.

[0023] The key steps of this embodiment will be described in detail below according to the overall processing flow of report generation.

[0024] like Figure 1 As shown, the method includes the following steps: Step S1: Task scheduling and data acquisition. The system automatically starts the report generation task through the timed task scheduling module according to the preset trigger cycle, and also supports users to manually submit tasks.

[0025] After the task is initiated, the system iterates through the list of target plant areas for which reports are to be generated. For each target plant area, the system matches the corresponding report template from the template library according to the report type and reporting period. The report types include, but are not limited to, inspection reports, operation reports, and water washing reports, and the reporting periods include, but are not limited to, monthly reports, weekly reports, and quarterly reports. After the template is successfully matched, the system automatically retrieves business data for the specified period from the operation and maintenance platform database. The retrieved data types include: daily power generation, cumulative power generation, peak power, and equivalent utilization hours of each inverter; current and voltage data of each combiner box; irradiance, ambient temperature, and component temperature data collected by the environmental monitoring instrument; start and end time, alarm level, and processing status of fault alarm records; and cleaning date, cleaning method, and power generation comparison data before and after cleaning in the cleaning records.

[0026] Data capture is performed using an application programming interface (API) and supports breakpoint resume functionality. The captured raw data is temporarily stored in a temporary table in the system database. Each data record includes a collection timestamp, data source identifier, and hash check value for subsequent source tracing and verification.

[0027] Step S2: Construct a knowledge graph of photovoltaic equipment nameplate parameters. Before performing data filling and report generation, the system needs to construct a knowledge graph of photovoltaic equipment nameplate parameters in advance. The knowledge graph is stored in the form of a graph database. The nodes in the graph include equipment type nodes, equipment model nodes and parameter nodes. The edges are used to represent the subordinate relationship between equipment and parameters and the constraint relationship between parameters.

[0028] The knowledge graph's data sources include technical specifications and rated parameter tables in product manuals provided by equipment manufacturers, as well as efficiency curve data specified in IEC standards. For each inverter model, the knowledge graph stores the following parameters: maximum DC input voltage, maximum conversion efficiency, efficiency value sequences corresponding to different load rates in the European efficiency curve, and ambient temperature derating curve data. For photovoltaic strings, the knowledge graph stores the open-circuit voltage temperature coefficient. and short-circuit current temperature coefficient .

[0029] In addition, the knowledge graph also stores the physical constraints between various parameters; for example, there is a negative temperature coefficient relationship between the open-circuit voltage of the photovoltaic string and the ambient temperature. These constraints are stored in the form of mathematical expressions and parameter boundaries, which are called by the physical constraint verification module at runtime.

[0030] Step S3: Call the large language model to generate the initial draft of the report. The system organizes the captured operation and maintenance data into a data context according to the template structure, and inputs it into the large language model along with the report generation prompts. Call the large language model to generate the text of each block of text in the report that requires descriptive content. This descriptive content includes, but is not limited to: monthly power generation analysis paragraphs, equipment operation status review, abnormal situation descriptions, and operation and maintenance suggestions for the next stage.

[0031] At the same time, the system calls the large language model to intelligently fill in the numerical blocks of the table in the template; for example, when generating the monthly power generation statistics table for each inverter, the large language model infers the power generation value that should be filled in for each inverter based on the data context.

[0032] This embodiment does not limit the specific type of large language model. A general large language model or a domain-specific large language model that has been fine-tuned for the power industry can be used. In this process, the large language model generates descriptive text and numerical filling suggestions that conform to the report specifications based on operation and maintenance data and prompt words. The verification, calibration and correction of these generated contents are completed by subsequent steps S4 and S5.

[0033] Step S4: Physical constraint verification and numerical limiting calibration.

[0034] like Figure 2 As shown, the steps include: Step S4.1: Physical rationality verification based on nameplate parameter knowledge graph. After the large language model completes the numerical filling, the system first sends the numerical values ​​of each template to be filled to the physical constraint verification module for rationality verification.

[0035] For numerical items matching the corresponding device model from the knowledge graph, the physical constraint verification module performs the following three checks: First, absolute boundary verification, determining whether the value exceeds the absolute physical limits specified in the device's nameplate parameters. Taking inverter conversion efficiency as an example, if the efficiency value generated by the large language model is greater than the maximum conversion efficiency of the inverter model, then the value is deemed physically absurd. Second, relative constraint verification, determining whether the value satisfies the physical constraint relationships between parameters stored in the knowledge graph. For example, when filling in the inverter output power, the large language model generates an output power value; when filling in the input power, it generates an input power value. The system verifies whether the output power is less than or equal to the product of the conversion efficiency and the input power. If this inequality does not hold, it indicates a physical contradiction between the two related values ​​generated by the large language model, and the verification is deemed a failure. Third, environmental correlation verification, determining whether the value matches environmental parameters. Specifically, the string open-circuit voltage value generated by the large language model must match the current ambient temperature. The system matches the open-circuit voltage temperature coefficient stored in the knowledge graph. Calculate the theoretical open-circuit voltage at the current temperature. The calculation formula is: ;in, The open-circuit voltage under standard test conditions is a constant. The standard test condition temperature is a constant. Temperature coefficient of open-circuit voltage; The current ambient temperature; if the string open-circuit voltage value generated by the large language model is... If the relative deviation exceeds the preset threshold, the value is judged to have physical irrationality and is marked as pending review.

[0036] Step S4.2: Categorization and processing of verification results. The system categorizes the values ​​into three types based on the verification results: First type: Verification passed. The value simultaneously meets the absolute boundary verification, relative constraint verification, and environmental correlation verification, and proceeds directly to the next processing stage. Second type: Questionable and awaiting review. The value passes the absolute boundary verification and relative constraint verification, but the deviation from the theoretical value in the environmental correlation verification exceeds the threshold but does not exceed the higher warning line. The value can still be entered into the template, but the system adds a "awaiting review" mark next to the cell or paragraph containing the value. Third type: Verification failed. The value fails the absolute boundary verification or relative constraint verification, triggering forced amplitude limiting calibration.

[0037] Step S4.3: Physical Limiting Calibration. When a value triggers forced limiting calibration, the system performs the following operations: First, it determines the limiting boundary value. For out-of-bounds values, the limiting target value is determined based on the specific type of verification failure. For example, if the large language model generates an inverter efficiency value of 99%, while the nameplate parameters show a maximum conversion efficiency of 98.2% for this inverter model, then the limiting target value is 98.2%. Here, 98.2% is an example of the actual nameplate parameters for a specific inverter model (such as Sungrow SG125HV). In actual applications, the limiting boundary value is determined by the parameters of the corresponding device model in the knowledge graph. Second, it performs value replacement, replacing the original value generated by the large language model with the calibrated value. Then, it generates a calibration label, automatically adding a label to the location of the value and appending a record to the data calibration description area at the bottom of the page. Finally, it records the calibration log, recording the calibration status and the original generated value in the system log for auditing and traceability.

[0038] Step S5: Semantic logic adversarial verification and redirection generation.

[0039] like Figure 3 As shown, this step includes: Step S5.1: Constructing numerical semantic relationships. After the physical constraint verification is completed, the system fills the verified or calibrated values ​​into the corresponding positions in the report template, forming a draft report with determined values ​​and text to be verified. Based on this, the system starts the semantic logic adversarial verification module. The core of semantic logic adversarial verification is to detect whether there is a logical contradiction between the numerical data in the draft report and the descriptive text generated by the large language model. To this end, the system first constructs numerical semantic relationships. Specifically, the system traverses all blocks in the report template. For each data block, it extracts the value, unit, and physical meaning label of each numerical value. For each text description block, it uses a combination of template matching and dependency parsing to extract assertions related to the numerical values. When words indicating numerical states or trends appear in the text, the system marks the statement as an assertion to be verified and extracts the numerical expectations implied in the assertion. Subsequently, the system associates the numerical values ​​in the data blocks with the assertion statements in the text blocks. The association rules include: the device identifier to which the numerical value belongs matches the device name described in the assertion statement; the statistical period of the numerical value matches the time range described in the assertion statement; and the physical meaning of the numerical value matches the phenomenon described in the assertion statement. After the above association, the system constructs a set of numerical and semantic associations. Each numerical and semantic association includes: a physically verified numerical value and its actual value. This value is the lower limit of the reference range under historical normal operating conditions. and upper limit and a section related to numbers Related assertion text.

[0040] Step S5.2: Semantic tendency scoring. For each numerical value and semantic association, the system calculates a semantic tendency score. It is used to quantify the degree of deviation between textual description and actual numerical state, and is a semantic tendency score. The value range is from 0 to 1. Higher numerical values ​​indicate greater deviation. The system uses a pre-defined semantic tendency mapping table to map common descriptive expressions in the text to corresponding expected numerical ranges. For example, "stable equipment operation" maps to a value that should be within the reference range and with fluctuations not exceeding a preset percentage; "power generation meets expectations" maps to a value that should be within a certain percentage range plus or minus the predicted value; and "abnormal equipment" maps to a value that should exceed the reference range or exhibit sudden changes. Semantic Tendency Score The calculation method is as follows: the system extracts keywords from the assertion text and matches them with the semantic tendency mapping table to obtain the expected value range of the text; if the actual value is... If the price falls within the expected range and the fluctuation range matches the expected description, then... ;like If the price deviates completely from the expected range or the fluctuation pattern is contrary to the description, then... The intermediate states are calculated using linear interpolation based on the degree of deviation.

[0041] Step S5.3: Numerical anomaly scoring; the system synchronously calculates numerical anomaly scores. It is used to quantify the degree to which a numerical value deviates from the normal reference range; numerical anomaly scoring. The value ranges from 0 to 2. First, determine the value... Lower limit of normal reference range and upper limit If historical databases contain data on the device under similar operating conditions during the same period, let the mean of the historical data be... The standard deviation is Then take , If no historical data is available, the recommended operating range based on the device's nameplate parameters in the knowledge graph is used as the reference range. Using twice the standard deviation as the reference range boundary is a common practice in statistics for identifying outliers; those skilled in the art can also choose other multiples (such as...) according to actual needs. or This does not affect the scope of protection of this invention.

[0042] Set reference range width Then the numerical anomaly score The calculation formula is: ;in, This refers to the actual value. This is the lower limit of the reference range; This is the upper limit of the reference range; The reference range width; the first term of the formula Used to calculate the outlier contribution when the value is below the lower limit of the reference range, when When, the value of this item is 0; when When, the value of this item is... And not exceeding 1; the second term of the formula Used to calculate the outlier contribution when the value exceeds the upper limit of the reference range, when When, the value of this item is 0; when When, the value of this item is... And not exceeding 1; if Within range If both are 0, then... ;like The further away from the reference range, The larger the value, the upper limit is 2.

[0043] Step S5.4: Calculation of Contradiction Score. System-wide comprehensive semantic tendency score. Sum of numerical anomaly scores Calculate the final contradiction score It is used to determine whether there is a logical contradiction between numerical values ​​and semantics; contradiction score. The value range is from 0 to 1. The higher the value, the greater the degree of contradiction.

[0044] Let the adjustment factor be , Used to balance the contribution weight of numerical anomalies in contradiction detection; preferably, A value of 0.5 is chosen to maintain a moderate balance between the contributions of semantic bias and numerical anomalies to the final contradiction score in typical scenarios; those skilled in the art can adjust this value according to the sensitivity requirements for numerical anomalies in practical applications. The value can be appropriately increased, for example, in scenarios where higher data precision is required. value.

[0045] Contradiction rating The calculation formula is: ;in, The semantic tendency score ranges from 0 to 1. For numerical anomalies, a score is given, with a value ranging from 0 to 2; The adjustment factor is used; the formula employs a product-based nonlinear fusion function when... The higher, or The higher, the better The higher; especially when and At the same time, when it is higher The growth of has a non-linear amplification effect, which is due to the formula in and Both factors approach a smaller value simultaneously, causing the product to decrease significantly, thus making... Significantly increased.

[0046] Step S5.5: Contradiction detection and reverse questioning. The system sets a low threshold. and high threshold Two decision thresholds; preferably, Take 0.35, The threshold value of 0.65 is an optimal value determined by maximizing the weighted harmonic mean of contradiction detection accuracy and recall on the labeled test dataset. Those skilled in the art can adjust the tolerance for false positives and false negatives according to the actual application scenario.

[0047] when When the numerical value and semantic meaning are basically consistent, the system marks the relationship between the numerical value and semantic meaning as consistent, and the corresponding description text does not need to be modified and can be directly retained.

[0048] when If the system determines that there is a potential contradiction between the numerical value and the semantics, but the evidence is insufficient, it will add a warning mark next to the description text and prompt the user in the report preview interface that the description may deviate from the actual data and suggest manual review.

[0049] when When a significant logical contradiction is found between the numerical value and its semantic meaning, the system triggers a reverse challenge mechanism. The reverse challenge process includes the following steps: First, construct a reverse challenge instruction. The system extracts abnormal values, numerical physical meaning labels, and normal reference ranges from the correlation between numerical values ​​with contradiction scores exceeding a preset threshold and semantic meaning. A new instruction is automatically constructed from the original description text; the instruction content includes: specifying the device name, time period, indicator name, and actual value. The first step involves explaining the abnormal characteristics of the value, pointing out the contradictions with the original description, and requiring the generation of a new professional operation and maintenance analysis text based on the abnormal data. The analysis should include possible causes of the anomaly, suggested troubleshooting directions, and recommended remedial measures. The second step involves calling a large language model to regenerate the text. The system sends the above reverse challenge command to the large language model to obtain the regenerated text content based on the abnormal data. The third step involves performing text replacement. The system replaces the corresponding descriptive text in the original draft report with the regenerated text content, adds a revision mark at the replacement location, and retains the original descriptive text as a historical version. The fourth step involves logging the challenge. The system records the trigger time of this reverse challenge, the numerical and semantic relationships involved, the original descriptive text content, the newly generated text content, and the contradiction score in the log. This is for subsequent quality assessment and prompt word optimization.

[0050] Step S5.6: Collaborative processing of relationships. Figure 4 This diagram illustrates the collaborative relationship between the physical constraint verification module and the semantic and logical adversarial verification module.

[0051] In this embodiment, the physical constraint verification module and the semantic logic adversarial verification module have a clear collaborative relationship.

[0052] The calibrated values ​​output by the physical constraint verification module are directly used as the input basis for the semantic logic adversarial verification module in calculating numerical anomaly scores. At that time, the system prioritizes using the values ​​that have passed physical verification. If a value has undergone physical amplitude limiting calibration, the semantic verification module uses the calibrated value as... Perform a consistency check.

[0053] The contradictions discovered by the semantic verification module can provide supplementary signals for physical verification. When multiple values ​​of a device pass physical verification, but the semantic verification module repeatedly detects contradictions in the descriptive text related to the device, the system marks the device as potentially having abnormal data sources or model cognitive biases, and increases the verification rigor of subsequent content generated by the device.

[0054] Through the aforementioned collaborative mechanism, physical verification ensures the physical compliance of the reported values, while semantic verification ensures the semantic consistency between the values ​​and the descriptions. Together, they constitute a dual verification defense for report quality.

[0055] Step S6: Report Preview and Manual Confirmation. After physical constraint verification and semantic logic adversarial verification, the system presents the verified and properly calibrated draft report in the front-end preview interface.

[0056] The preview interface displays the entire report content according to the original chapter structure of the template. Each data block is marked with its data source as either a database or AI-generated, each value that has undergone physical amplitude limiting calibration is marked with a superscript, and each descriptive paragraph that has undergone reverse challenge replacement is marked with a revision mark. Users can edit and modify any content in the preview interface, and modifications are saved in real time.

[0057] After the user confirms that the report content is correct, they can click the "Generate Final Report" button. The system will then output a formal report document in Word or PDF format according to the template.

[0058] Step S7: Report Archiving and Review. The generated official report is automatically saved to the historical report library. The system supports content and format reviews of the report; reviewers can mark issues and enter suggested modifications online. The review process adopts a closed-loop management model, with review comments immediately synchronized to the report submitter.

[0059] This embodiment achieves physical rationality verification and semantic logic consistency verification in the automated report generation process through steps S1 to S7, ensuring the accuracy and credibility of the report content.

[0060] Example 2: This example provides an automated report generation and data processing system for implementing the method described in Example 1, such as... Figure 5 As shown, the system includes the following functional modules: a task scheduling module, used to automatically initiate report generation tasks according to a preset trigger cycle, traversing the target plant area and matching the corresponding report template according to the report type and report cycle; a data acquisition module, used to automatically capture business data from the operation and maintenance platform database, including power generation, equipment status, fault records, and cleaning records, supporting breakpoint resume and data integrity verification; a template management module, used to store and manage report templates, supporting template uploading, version management, and trigger condition configuration, establishing a mapping relationship between templates and plant area, report type, and report cycle; a knowledge graph construction module, used to construct a knowledge graph of photovoltaic equipment nameplate parameters, storing equipment models, nameplate parameters, physical constraints between parameters, and IEC standard efficiency curve data in the form of a graph database; and a large model invocation module, used to invoke a large language model to generate descriptive text and numerical filling suggestions in the report based on the captured operation and maintenance data and prompt words. The physical constraint verification module is used to perform physical rationality verification on the numerical values ​​generated by the large language model based on the nameplate parameter knowledge graph, including absolute boundary verification, relative constraint verification and environmental correlation verification; for numerical values ​​that fail verification, physical limit calibration is performed to force out-of-bounds values ​​to be corrected to the physical limit boundary.

[0061] The semantic logic adversarial verification module receives the numerical values ​​verified by the physical constraint verification module and their corresponding text descriptions generated by the large language model, constructs the semantic association between the numerical values, and calculates the semantic tendency score. Sum of numerical anomaly scores And calculate the conflict score based on the conflict score calculation formula. ;when When the threshold is exceeded, a reverse challenge instruction is constructed to trigger the large language model to regenerate professional analysis text for the abnormal data, and the original description content is replaced with the new text.

[0062] The report preview module is used to display the draft report after verification on the front-end interface, allowing users to edit and confirm it online.

[0063] The report output module is used to output formal report documents in Word or PDF format according to the template format, and supports downloading and archiving.

[0064] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. An automated report generation and data processing method, characterized in that, include: In response to the start of the report generation task, the system retrieves business data of the target plant area within a specified period from the operation and maintenance platform database, and matches the corresponding report template according to the report type and report period. The system invokes a large language model, organizes the captured business data into a data context according to a template structure, and inputs it into the large language model to generate a draft report. The draft report includes descriptive text and numerical filling results. Based on a pre-built knowledge graph of photovoltaic equipment nameplate parameters, the physical rationality of each value in the numerical filling results is verified. For values ​​that fail the verification, physical limiting calibration is performed to obtain calibrated values. The calibrated values ​​are filled into the corresponding positions in the report template to form a draft report, and the numerical and semantic association between each value in the draft report and the corresponding descriptive text is constructed. For each set of numerical and semantic association, a semantic tendency score is calculated to quantify the degree of deviation between the text description and the actual state of the value, and a numerical anomaly score is calculated to quantify the degree of deviation of the value itself from the normal reference range. By combining semantic tendency score and numerical anomaly score, a contradiction score is calculated. When the contradiction score exceeds a preset threshold, a reverse interrogation command is constructed to trigger the large language model to regenerate the analysis text for the abnormal data to replace the original descriptive text. After processing all numerical and semantic relationships, the final report document is generated.

2. The automated report generation and data processing method according to claim 1, characterized in that: The photovoltaic equipment nameplate parameter knowledge graph is stored in the form of a graph database. The nodes include equipment type nodes, equipment model nodes, and parameter nodes. The edges represent the subordinate relationship between the equipment and the parameter, as well as the constraint relationship between the parameters. The photovoltaic equipment nameplate parameter knowledge graph stores at least the maximum DC input voltage, maximum conversion efficiency, and European efficiency curve data for each inverter model, and at least the open-circuit voltage temperature coefficient and short-circuit current temperature coefficient for photovoltaic strings.

3. The automated report generation and data processing method according to claim 1, characterized in that: The physical rationality verification includes: absolute boundary verification, which determines whether the value exceeds the absolute physical limit specified by the corresponding equipment nameplate parameters; relative constraint verification, which determines whether the value satisfies the physical constraint relationship between parameters stored in the photovoltaic equipment nameplate parameter knowledge graph; and environmental correlation verification, which determines whether the value matches the environmental parameters. For the open-circuit voltage value of the photovoltaic string, the theoretical open-circuit voltage value is calculated based on the open-circuit voltage temperature coefficient stored in the photovoltaic equipment nameplate parameter knowledge graph and the current ambient temperature. If the relative deviation between the generated value and the theoretical value exceeds a preset threshold, it is determined to be questionable and awaiting review.

4. The automated report generation and data processing method according to claim 3, characterized in that: The physical limiting calibration includes: for values ​​that fail the absolute boundary check or relative constraint check, determining the limiting target value according to the check failure type, replacing the original generated value with the limiting target value, and adding a calibration mark at the location of the value.

5. The automated report generation and data processing method according to claim 1, characterized in that: The constructed numerical and semantic associations include: Traverse each block in the report template, extract the value, unit, and physical meaning labels of each value in the data block, and extract the assertion statements related to the values ​​from the text description block; The numerical values ​​and assertion statements are associated according to the rules of device identification matching, time range matching, and physical meaning matching, forming a numerical and semantic association relationship that includes the actual value of the numerical value, the upper and lower limits of the normal reference range, and the associated assertion text.

6. The automated report generation and data processing method according to claim 1, characterized in that: The semantic tendency score is calculated as follows: a semantic tendency mapping table is preset to map keywords in the text description to corresponding expected numerical ranges; keywords in the assertion text are extracted and matched with the semantic tendency mapping table to obtain the expected value range; the semantic tendency score is obtained by linear interpolation based on the deviation between the actual numerical value and the expected value range, and the score range is from 0 to 1.

7. The automated report generation and data processing method according to claim 1, characterized in that: The calculation method for the numerical anomaly score is as follows: determine the lower limit and upper limit of the normal reference range of the value. If there is data of the same period under similar operating conditions for the device in the historical database, the mean of the historical data plus or minus twice the standard deviation is used as the reference range boundary. Otherwise, the recommended operating range of the nameplate parameters of the device in the photovoltaic device nameplate parameter knowledge graph is used as the reference range. The first anomalous contribution is calculated based on the degree to which the actual value is lower than the lower limit of the reference range, and the second anomalous contribution is calculated based on the degree to which the actual value is higher than the upper limit of the reference range. The first anomalous contribution and the second anomalous contribution are added together to obtain the numerical anomalous score, which ranges from 0 to 2.

8. The automated report generation and data processing method according to claim 1, characterized in that: The contradiction score is calculated as follows: a moderating factor is set, the numerical anomaly score is multiplied by the moderating factor, and then input together with the semantic tendency score into a nonlinear fusion function in the form of a product to obtain a contradiction score with a value range of 0 to 1; when both the semantic tendency score and the numerical anomaly score are high, the increase in the contradiction score is greater than the sum of their independent contributions.

9. The automated report generation and data processing method according to claim 1, characterized in that: The constructed reverse interrogation instruction triggers the large language model to regenerate the analysis text for abnormal data, including: extracting abnormal values, numerical physical meaning labels, normal reference ranges, and original description text from the numerical and semantic correlation of the contradiction scores exceeding a preset threshold; generating an instruction containing device name, time period, indicator name and actual value, abnormal feature description, and original contradiction explanation; requiring the large language model to generate analysis text containing possible causes of the anomaly, suggested investigation directions, and recommended handling measures; after obtaining the newly generated analysis text, replacing the corresponding original description text in the report draft, and retaining the original description text as a historical version.

10. The automated report generation and data processing method according to any one of claims 1-9, characterized in that: The calibrated values ​​output by the physical rationality verification are used as the actual values ​​used when calculating the numerical anomaly score. When multiple values ​​of the same device pass the physical rationality verification and the number of times the description text corresponding to the device triggers reverse challenges exceeds the preset number, the preset threshold used for each verification item in the subsequent physical rationality verification of the device is reduced.

Images