A new energy project risk management and control method, system, medium and product

Abstract

A method, system, medium, and product for risk management in new energy projects, relating to the field of artificial intelligence, are disclosed. The method includes: acquiring real-time operating parameters of various devices in the new energy project; responding to a query request, performing semantic parsing to determine the intent type and query entity of the query request, identifying the target device corresponding to the query entity, determining the target operating parameters, acquiring historical operating parameters corresponding to each parameter dimension of the target operating parameters, calculating the device status deviation, generating text feature vectors and device status feature vectors, and generating a fusion retrieval vector, inputting it into a preset risk knowledge base for similarity retrieval, identifying at least one candidate risk item, filtering each candidate risk item based on the intent type and device status deviation, and generating risk guidance information through a preset risk identification model. This application can improve the targeting and accuracy of risk identification.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and in particular to a method, system, medium and product for risk management of new energy projects. Background Technology

[0002] Currently, the new energy industry is rapidly expanding towards larger-scale facilities and unmanned management models. This poses a challenge to traditional safety management systems that rely on manual inspections and experience-based judgment. In actual operation, risk identification at new energy facilities mainly depends on the personal experience of operators. Meanwhile, relevant laws, regulations, and standards exist in the form of static electronic documents, with numerous entries and a lack of systematic organization. When faced with specific work scenarios, operators need to rely on their experience to find matching compliance requirements and control measures from a massive amount of documents. This not only results in low search efficiency but also makes the accuracy of risk identification highly dependent on individual professional skills. Differences in experience can easily lead to inconsistent judgments, increasing the uncertainty of on-site safety management.

[0003] To improve the situation, the industry has attempted to introduce a rule-based risk knowledge base system. By structuring laws, regulations, and standards, pre-setting standard risk items and corresponding control measures, users can search for relevant information using keywords. When operators enter keywords, they can retrieve predefined risk analysis tables from the knowledge base. This system achieves a certain degree of systematic storage of risk knowledge and improves the convenience of information retrieval.

[0004] However, in the actual application of new energy power stations, especially integrated power stations that combine solar thermal and photovoltaic operations, the operating scenarios are often complex and dynamically changing. The description of the on-site operating conditions is difficult to accurately correspond to the preset standard entries in the knowledge base. When the query keywords entered by the operators differ from the preset entries in the database, the system can only return broad related results, which increases the burden on the operators to perform secondary screening and self-judgment from the generalized results, resulting in insufficient targeting and accuracy of risk identification. Summary of the Invention

[0005] This application provides a method, system, medium, and product for risk management of new energy projects, which addresses the technical problem of how to improve the pertinence and accuracy of risk identification.

[0006] In a first aspect, embodiments of this application provide a risk management method for new energy projects, including: Obtain real-time operating parameters of each device in new energy projects; In response to a user-input query request, the query request is semantically parsed to determine the intent type of the query request and at least one query entity, wherein the query entity includes at least one of equipment name, operating condition description or risk type; Based on the query entity, at least one target device corresponding to the query entity is determined, and target operating parameters are determined from the real-time operating parameters corresponding to the target device, wherein the target operating parameters include at least one parameter dimension; Based on the target operating parameters, obtain the historical operating parameters corresponding to each of the parameter dimensions in the target operating parameters; Based on the target operating parameters and the historical operating parameters, calculate the equipment state deviation. The query request is subjected to feature calculation to obtain a text feature vector, and the deviation between the target operating parameters and the device state is subjected to feature calculation to obtain a device state feature vector. Based on the text feature vector and the device status feature vector, a fusion retrieval vector is generated; The fusion retrieval vector is input into a preset risk knowledge base for similarity retrieval to determine at least one candidate risk item that matches the fusion retrieval vector. Based on the deviation between the intent type and the device state, each candidate risk item is filtered to obtain a target candidate risk item set, and risk guidance information is generated based on the target candidate risk item set through a preset risk identification model.

[0007] Optionally, calculating the device state deviation based on the target operating parameters and the historical operating parameters includes: extracting time-series data of the target parameter dimension within a preset historical time window from the historical operating parameters, wherein the target parameter dimension is any of the parameter dimensions; calculating the historical mean and historical standard deviation of the target parameter dimension based on the time-series data; calculating the difference between the real-time operating parameters of the target parameter dimension and the historical mean; calculating the ratio of the difference to the historical standard deviation to obtain the sub-state deviation corresponding to the target parameter dimension; and performing a weighted calculation on all the sub-state deviations based on a preset weight set to obtain the device state deviation.

[0008] Optionally, generating a fused retrieval vector based on the text feature vector and the device state feature vector includes: inputting the text feature vector and the device state feature vector into a pre-trained attention layer; calculating a first attention weight of the text feature vector to the device state feature vector and a second attention weight of the device state feature vector to the text feature vector through the attention layer; and performing a weighted calculation on the text feature vector and the device state feature vector based on the first attention weight and the second attention weight to generate a fused retrieval vector.

[0009] Optionally, the step of filtering each candidate risk item based on the intent type and the device state deviation to obtain a target candidate risk item set, and generating risk guidance information based on the target candidate risk item set through a preset risk identification model, includes: selecting candidate risk items that match the deviation range of the device state deviation as intermediate candidate risk items to obtain an intermediate candidate risk item set, and obtaining a preset baseline risk level for each intermediate candidate risk item; matching the device state deviation with multiple preset deviation threshold intervals to determine the risk level adjustment coefficient corresponding to the device state deviation; and adjusting the risk level based on the risk level adjustment coefficient for each intermediate candidate risk item. The baseline risk level of each selected risk item is adjusted to generate an intermediate risk level corresponding to each intermediate candidate risk item. Real-time micro-meteorological data and operation duration data are obtained based on the operation scenario anchor point corresponding to the target equipment, and the intermediate risk levels are dynamically adjusted based on the real-time micro-meteorological data and operation duration data to obtain the dynamic risk level of each intermediate candidate risk item. Based on the intent type, a set of target candidate risk items matching the intent type is selected from the set of intermediate candidate risk items, and risk guidance information matching the intent type is generated based on the dynamic risk level of each target candidate risk item in the set of target candidate risk items.

[0010] Optionally, the step of acquiring real-time micro-meteorological data and operation duration data based on the operation scenario anchor point corresponding to the target equipment, and dynamically adjusting each of the intermediate risk levels based on the real-time micro-meteorological data and operation duration data to obtain the dynamic risk level of each of the intermediate candidate risk items, includes: acquiring real-time micro-meteorological data of the operation area of ​​the new energy project where the target equipment is located based on the operation scenario anchor point, wherein the real-time micro-meteorological data includes at least one of wind speed time series data, ultraviolet radiation index, and surface icing humidity; acquiring the operation duration data of the current operators, and inferring the physiological metabolic fatigue index based on the operation duration data and a preset heavy physical operation duty cycle; calculating the ecological stress gradient value characterizing the environmental deterioration trend based on the real-time micro-meteorological data; performing feature combination transformation on the ecological stress gradient value and the physiological metabolic fatigue index to generate a comprehensive spatiotemporal attenuation factor; and performing nonlinear dynamic adjustment of each of the intermediate risk levels through the comprehensive spatiotemporal attenuation factor to obtain the dynamic risk level of each of the intermediate candidate risk items.

[0011] Optionally, the intent type includes risk query intent and risk question-and-answer intent. The step of filtering a set of target candidate risk items matching the intent type from the intermediate candidate risk item set, and generating risk guidance information matching the intent type based on the dynamic risk level of each target candidate risk item in the target candidate risk item set, includes: when the intent type is the risk query intent, filtering intermediate candidate risk items belonging to historical risk data from the intermediate candidate risk item set to form a target candidate risk item set, wherein the historical risk data is associated with historical handling measures; extracting each target risk item from the target candidate risk item set... The historical handling measures and dynamic risk levels associated with candidate risk items are used to generate a list of risk operation guidelines sorted from high to low according to the dynamic risk level, thus obtaining risk guidance information. When the intent type is the risk question and answer intent, intermediate candidate risk items belonging to regulatory clauses are selected from the intermediate candidate risk item set to form the target candidate risk item set, and the regulatory clauses are associated with compliance basis information. The compliance basis information and dynamic risk level associated with each target candidate risk item in the target candidate risk item set are extracted to generate a list of compliance analysis reports sorted from high to low according to the dynamic risk level, thus obtaining the risk guidance information.

[0012] Optionally, the method further includes: acquiring historical trend data of the historical operating parameters within a preset historical time window; generating predicted operating parameters of the target operating parameters at a future preset time point based on the historical trend data using a preset trend prediction model; calculating the predicted state deviation based on the predicted operating parameters and the historical operating parameters; when the predicted state deviation exceeds a preset warning threshold, weighting and fusing the predicted state deviation with the equipment state deviation to generate a fused state deviation; forward-lookingly correcting the risk level and handling suggestions in the risk guidance information based on the fused state deviation to generate forward-looking risk guidance information containing future risk warnings; and outputting the forward-looking risk guidance information simultaneously with or before outputting the risk guidance information.

[0013] Optionally, the method further includes: acquiring user feedback operation data on the risk guidance information, the feedback operation data including at least one of adoption identifier, modification record, or rejection instruction; associating and storing the feedback operation data with the corresponding query request, the real-time operating parameters, and the target candidate risk item set to construct a feedback sample library; periodically extracting feedback samples that meet preset screening conditions from the feedback sample library, the preset screening conditions including an adoption rate lower than a preset adoption threshold or a modification count higher than a preset modification count threshold; and correcting all candidate risk items in the preset risk knowledge base based on the feedback samples to obtain a dynamically optimized preset risk knowledge base, the correction including at least one of updating the correlation deviation range of candidate risk items, adjusting the baseline risk level, or supplementing new candidate risk items.

[0014] Secondly, embodiments of this application provide a risk management system for new energy projects. The new energy project risk management system 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, the computer program code including computer instructions, and the one or more processors call the computer instructions to cause the new energy project risk management system to perform the method described in the first aspect and any possible implementation thereof.

[0015] Thirdly, embodiments of this application provide a computer program product containing instructions that, when the computer program product is run on a new energy project risk management system, cause the new energy project risk management system to execute the method described in the first aspect and any possible implementation thereof.

[0016] Fourthly, embodiments of this application provide a computer-readable storage medium including instructions that, when executed on a new energy project risk management system, cause the new energy project risk management system to perform the method described in the first aspect and any possible implementation thereof.

[0017] In summary, one or more technical solutions provided in this application have at least the following technical effects or advantages: 1. By semantically parsing user natural language queries to accurately capture query intent, and combining real-time operating parameters and historical data of the equipment to calculate the deviation of the equipment status, the text semantics and equipment status features are then fused for retrieval. The final risk guidance information is generated by matching and filtering from the risk knowledge base, realizing a fully intelligent closed loop from user fuzzy query to accurate risk analysis. It not only understands the user's surface language needs, but also deeply integrates the real-time health status (deviation) of the equipment. This makes the retrieved risk items not simply based on text matching, but highly consistent with the actual working conditions of the current equipment, thereby greatly improving the accuracy and pertinence of risk identification.

[0018] 2. By employing a deviation algorithm based on historical mean versus real-time value, multi-dimensional device operating parameters are transformed into standardized sub-state deviations and then weighted and fused. Simultaneously, an attention mechanism is used to dynamically weight and fuse text features and device state features to generate a fused retrieval vector. This achieves accurate quantification of device health status and deep interaction of multimodal information. On the one hand, deviation calculation based on historical statistics eliminates the influence of different dimensions between different devices and parameters, making risk quantification more scientific. On the other hand, the introduction of the attention mechanism enables the retrieval process to dynamically focus on key points in the query semantics and abnormal points in the device state, ensuring that the generated retrieval vector understands both what the user wants to ask and what the device is currently doing, laying a solid data foundation for subsequent high-precision retrieval.

[0019] 3. By adjusting the baseline risk level based on equipment status deviation and introducing real-time micro-meteorological data and personnel working hours as scenario anchors, the risk level is dynamically corrected non-linearly. Furthermore, by combining user risk queries or Q&A intents to filter and differentiate candidate items, dynamic and accurate calculation of risk levels and on-demand adaptation of output formats are achieved. This upgrades the static risk knowledge base into a dynamic risk assessment system, where risk levels change in real time with the environment (weather) and personnel status (fatigue index), avoiding the failure of static risk levels in dynamic working environments. Simultaneously, operation guidance lists or compliance analysis reports are generated for both query and Q&A intents, greatly improving the efficiency of human-computer interaction and user experience, making the output information more aligned with user decision-making needs in different scenarios.

[0020] 4. By predicting future operating parameters based on historical trends and calculating the deviation of the predicted state, forward-looking risk guidance information containing future risk warnings is generated. At the same time, user feedback on risk guidance is collected to build a sample library. Based on the feedback samples, the risk knowledge base is periodically revised and dynamically optimized, constructing a complete technology chain from passive response to proactive warning and continuous evolution. This achieves pre-emptive risk control and reserves valuable emergency response time for users. The dynamic optimization mechanism of the knowledge base based on real feedback enables the system to learn from every human-computer interaction and continuously correct the accuracy and applicability of risk entries, effectively solving the industry pain point of static and fixed knowledge base that cannot be self-iteratively updated. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the risk management method for new energy projects provided in the embodiments of this application; Figure 2 This is another flowchart illustrating the risk management method for new energy projects provided in this application embodiment; Figure 3 This is a schematic diagram of the structure of a risk management system for new energy projects provided in an embodiment of this application. Explanation of reference numerals in the attached figures: 401, Central Processing Unit; 402, Read-Only Memory; 403, Random Access Memory; 404, Bus; 405, Input / Output Interface; 406, Input Section; 407, Output Section; 408, Storage Section; 409, Communication Section; 410, Driver; 411, Removable Media. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] In the description of the embodiments of this application, words such as "illustrative," "for example," or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "illustrative," "for example," or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Rather, the use of words such as "illustrative," "for example," or "for example" is intended to present the relevant concepts in a specific manner.

[0024] In the embodiments of this application, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0025] In the implementation of this application, the collection and processing of relevant data should strictly comply with the requirements of relevant national laws and regulations, obtain the informed consent or separate consent of the personal information subject, and carry out subsequent data use and processing within the scope of laws and regulations and the authorization of the personal information subject.

[0026] Unless otherwise defined, all technical and scientific terms used in the embodiments of this application have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in the embodiments of this application is for the purpose of describing the embodiments of this application only and is not intended to limit this application.

[0027] This application provides a method, system, medium, and product for risk management of new energy projects, which addresses the technical problem of how to improve the pertinence and accuracy of risk identification.

[0028] Figure 1 This is a flowchart illustrating the risk management method for new energy projects provided in this application embodiment.

[0029] This application discloses a risk management method for new energy projects, such as... Figure 1 As shown, the steps include the following.

[0030] S101. Obtain the real-time operating parameters of each device in the new energy project.

[0031] Specifically, by connecting to the industrial Internet of Things (IoT) system deployed at the project site, real-time operating parameters of various devices are actively retrieved or passively received at a preset data acquisition frequency (e.g., every 5 seconds or every minute). These devices cover the core assets of new energy projects, such as photovoltaic arrays, wind turbine generators, inverters, box-type transformers, and energy storage systems. The acquired parameters are key indicators characterizing the health status and operating conditions of the equipment, such as the current, voltage, and backsheet temperature of photovoltaic modules, the rotational speed, blade angle, and nacelle vibration amplitude of wind turbine generators, and the DC-side input power, AC-side output power, and power module temperature of inverters. After receiving this raw data, it is usually preliminarily formatted and timestamped to transform unstructured sensor signals into structured real-time operating parameter datasets with precise time dimensions, providing an accurate data foundation for subsequent risk identification and state deviation calculation.

[0032] New energy projects refer to engineering projects that utilize new energy sources such as wind and solar power to produce electricity, such as large-scale centralized photovoltaic power plants or distributed wind farms. Equipment refers to the physical entities in new energy projects that perform core functions such as power generation, conversion, voltage boosting, and storage, such as inverters that convert direct current to alternating current, or energy storage battery clusters that store electrical energy for peak shaving. Real-time operating parameters refer to specific physical or electrical quantities that are collected and transmitted back in real time by sensors deployed on the equipment, characterizing the equipment's current instantaneous operating state. For example, the "main shaft temperature" of a wind turbine indicates the bearing lubrication status, and the "DC-side voltage" of an inverter indicates the output capacity of the photovoltaic array.

[0033] S102. In response to a query request input by the user, perform semantic parsing on the query request to determine the intent type of the query request and at least one query entity, wherein the query entity includes at least one of equipment name, operating condition description or risk type.

[0034] Specifically, when a user enters a query request in natural language on the monitoring interface, such as entering "Is the vibration of wind turbine No. 2 recently too high?" in the search box or "Historical cases of overheating risks in photovoltaic inverters" via voice input, the system will respond to this input and trigger a semantic parsing process. Relying on a pre-trained natural language processing model, the query text is first segmented, part-of-speech tagging is performed, and dependency parsing is conducted. Then, entity recognition algorithms extract key information particles strongly related to the new energy project field from the text. These identified information particles are the query entities, specifically including three categories: first, equipment names, such as wind turbine No. 2, inverter, photovoltaic array, etc., used to identify the target object; second, operating condition descriptions, such as excessive vibration, decreased power generation efficiency, abnormal noise, etc., used to characterize the abnormal state of the equipment; and third, risk types, such as overheating risk, fire hazard, reduced insulation, etc., used to directly locate the risk category. At the same time, by classifying the overall semantics of the query statement, the user's intent type is determined. For example, is the intent to query the current status of the equipment risk, or is the intent to obtain historical cases or regulatory clauses risk Q&A? The final output is a structured parsing result, which is a clear intent type (such as risk query intent) and at least one query entity (such as equipment name: No. 2 fan, operating condition description: excessive vibration), providing accurate semantic anchors for subsequent precise location of target equipment and operating parameters.

[0035] In this context, "user" refers to the operator who exchanges information with the risk management system through a human-computer interaction interface, typically an operation and maintenance engineer or safety manager of a new energy power plant. "Query request" refers to a user's instruction in natural language to express their need for information; this can be text entered via keyboard, speech-to-text, or a phrase selected via touch. "Intent type" refers to a label categorizing the user's query purpose, indicating whether the user's core demand in this interaction is to query status, seek knowledge, or perform an operation. "Query entity" refers to key information elements extracted from the user's query text and associated with the knowledge base of the new energy project domain; it serves as a semantic bridge connecting natural language and business data. "Equipment name" is a designation used to uniquely identify a specific piece of equipment or a set of equipment in a new energy project. "Operating condition description" refers to the user's natural language expression describing the equipment's operating status or abnormal phenomena, such as large current fluctuations, abnormal temperature rises, or shutdown. "Risk type" refers to category labels classifying potentially dangerous events or adverse conditions that may occur during the operation of a new energy project, such as electrical fires, blade breakage, or insulation breakdown.

[0036] S103. Based on the query entity, determine at least one target device corresponding to the query entity, and determine the target operating parameters from the real-time operating parameters corresponding to the target device. The target operating parameters include at least one parameter dimension.

[0037] Specifically, based on the parsed query entity, it is matched against a pre-built equipment ledger information database to determine the specific target equipment. This process is not a simple string comparison, but rather integrates fuzzy matching and semantic mapping mechanisms. For example, when the query entity is "Wind Turbine No. 2," it precisely identifies the wind turbine generator set numbered 2. When the query entity is a general term like "photovoltaic inverter," it may be mapped to a list of all inverter-type equipment at the project site. When the query entity contains industry abbreviations like "main transformer," it can be identified as a main transformer through a thesaurus and located to the corresponding physical equipment. After identifying the target equipment, the operating condition description or risk type in the query entity is further parsed and transformed into specific parameter dimension filtering conditions. For example, when the query entity contains the operating condition description "excessive vibration," a predefined operating condition-parameter mapping table is used to determine the parameters that need to be focused on, such as nacelle vibration amplitude and main shaft vibration frequency. When the query entity contains the risk type "overheating risk," it is mapped to temperature-related parameter dimensions such as winding temperature, bearing temperature, and power module temperature. Finally, data values ​​of specific parameter dimensions that match the above mapping results are filtered from the real-time operating parameter streams of these target devices to form a parameter set, namely the target operating parameters. This set contains at least one parameter dimension. For example, when a user only queries "current power of wind turbine No. 3", the target operating parameters only contain the single dimension of active power; while when a user queries a broad description such as "operating status of inverter No. 2", the target operating parameters may contain multiple parameter dimensions such as DC side voltage, AC side current, module temperature, and conversion efficiency, thus laying a data foundation for subsequent comprehensive status assessment.

[0038] In this context, "target equipment" refers to one or more specific physical devices identified by matching their names against equipment records in the query entity. For example, "Wind Turbine No. 2" identifies wind turbine generator set numbered 2, and "Photovoltaic Zone Inverter" identifies all inverter devices in the photovoltaic power generation area. "Target operating parameters" refers to a subset of data highly relevant to the current query intent, selected from numerous real-time operating parameters of the target equipment based on the operating condition description or risk type in the query entity. "Parameter dimension" refers to the specific indicator name used to classify equipment operating parameters, characterizing a particular aspect of the equipment's operating features. For example, "active power" represents power generation capacity, "winding temperature" represents thermal state, and "cumulative power generation" represents historical output.

[0039] S104. Based on the target operating parameters, obtain the historical operating parameters corresponding to each parameter dimension in the target operating parameters.

[0040] Specifically, after determining the target operating parameters, using the current moment as the base point, and based on a preset historical backtracking window (e.g., the past 30 days, the past 12 months, or the same period last year), historical operating parameters that are strictly aligned with each parameter dimension on the time axis are retrieved from the time-series database. Simultaneously, it is necessary to ensure a unique correspondence between parameter dimensions. That is, for each parameter dimension in the target operating parameters, such as "nacelle vibration amplitude of wind turbine No. 2," the historical database is precisely queried for the numerical sequence of the same equipment and the same parameter dimension at the same historical moment or historical time period. For example, for the nacelle vibration amplitude at the current moment, the vibration amplitude data at the same moment every day over the past 30 days will be extracted to form the historical time-series dataset for that parameter dimension. At the same time, the retrieved raw historical data will be cleaned to remove invalid data points generated during shutdown maintenance or sensor malfunctions, ensuring that the historical data used for subsequent analysis can accurately reflect the statistical characteristics of the equipment under normal operating conditions.

[0041] Historical operating parameters refer to the set of data values ​​recorded by the same device and the same parameter dimension at multiple historical time points within a historical time window before the current moment. They are usually stored in the database in the form of a time sequence.

[0042] S105. Calculate the equipment status deviation based on the target operating parameters and historical operating parameters.

[0043] Specifically, the target operating parameters are aligned with historical operating parameters collected in real time or periodically from the equipment's sensors. This comparison is not a single-dimensional numerical comparison, but involves a multi-dimensional state space composed of multiple key indicators such as temperature, pressure, vibration, and current. By applying statistical models or distance measurement algorithms (such as Euclidean distance, Mahalanobis distance, etc.), the difference between the current actual operating state point and the preset ideal target state point or state area is quantitatively calculated. This difference is then synthesized, weighted, and finally integrated into a standardized single value or index, namely the equipment state deviation. This quantitative indicator intuitively reflects the size of the gap between the current operating state of the equipment and its optimal state. The higher the deviation, the greater the risk of abnormal equipment operating state or performance degradation.

[0044] Among them, the equipment status deviation is a comprehensive quantitative indicator used to characterize the degree of deviation of the current overall operating status of the target equipment from its own historical normal status. The larger the value, the more significant the abnormality.

[0045] Through the above embodiments, the difference between the current state of the equipment and its historical normal state can be objectively and scientifically quantified, eliminating the obstacle that makes it impossible to directly compare different equipment and different parameters due to differences in units and magnitudes. This provides accurate and quantitative input basis for subsequent risk retrieval and level adjustment, making risk identification truly based on the dynamic quantitative assessment of the real-time health status of the equipment.

[0046] Based on the above embodiments, as an optional embodiment, for Figure 1 The step S105 shown can be implemented through steps S1051-S1055, which will be explained in detail below.

[0047] S1051. Extract the time series data of the target parameter dimension within the preset historical time window from the historical operating parameters. The target parameter dimension can be any parameter dimension.

[0048] Specifically, for any parameter dimension among the target operating parameters (i.e., the target parameter dimension), the time-series database storing historical data is accessed. Based on pre-configured historical backtracking rules, the corresponding parameter time-series data is extracted. The preset historical time window is a key time range parameter, and its length may vary depending on different parameter characteristics and equipment type: for parameters reflecting the slow aging process of equipment, such as insulation resistance, the time window may be set to the past 12 months; for parameters significantly affected by environmental factors, such as the backsheet temperature of photovoltaic modules, the time window may be set to the past 30 days to include the complete meteorological cycle; for frequently fluctuating operating parameters, such as wind turbine speed, the time window may be set to the past 7 days to ensure the timeliness of statistical characteristics. Strict time alignment is ensured during extraction, i.e., using the current moment as the reference point, the entire time window length is traced back, extracting the values ​​of all historical sampling moments of the target parameter dimension within the window, forming a data sequence arranged in chronological order, i.e., parameter time-series data. The completeness and continuity of this data sequence directly affect the accuracy of subsequent statistical calculations. Therefore, the data is usually pre-processed after extraction, such as removing abnormal operating data recorded during downtime and filling in temporary data gaps caused by communication interruptions.

[0049] The preset historical time window is a pre-configured time length parameter used to limit the time range for extracting historical data, such as the past 30 days, the past 12 months, or the same period last year. Its length setting needs to comprehensively consider equipment characteristics, parameter variation patterns, and analysis accuracy requirements. The target parameter dimension refers to the specific parameter dimension selected as the processing object in the current calculation process; it can be any dimension from the target set of operating parameters. Parameter time series data refers to the sequence of data values ​​arranged chronologically within the preset historical time window for a specific parameter dimension. It typically includes timestamps and corresponding numerical values ​​and is the basic data format for time series analysis and statistical calculations.

[0050] S1052. Based on the time series data of the parameters, calculate the historical mean and historical standard deviation of the target parameter dimension.

[0051] Specifically, the system calls a statistical calculation function to sum all numerical points in the parameter time series data, then divides the sum by the total number of data points to calculate the arithmetic mean of these historical data, i.e., the historical mean. This mean reflects the central trend or typical level of the parameter dimension under normal operating conditions. Next, the system calculates the difference between each data point in the parameter time series data and the historical mean, squares each difference to eliminate positive and negative cancellation, sums all squares, divides the sum by the number of data points, and then takes the square root to obtain the historical standard deviation. This standard deviation quantifies the dispersion of the historical data of the parameter dimension relative to its mean, i.e., the fluctuation range under normal conditions. For example, for equipment temperature parameters with gradual fluctuations, the historical standard deviation is usually small; while for power parameters that fluctuate frequently with operating conditions, the historical standard deviation is relatively large. Finally, the system outputs the historical mean and historical standard deviation corresponding to the target parameter dimension. These two statistics will serve as the benchmark parameters for subsequent calculations of sub-state deviation.

[0052] The historical mean refers to the arithmetic mean of all historical operating parameter values ​​for a certain parameter dimension within a preset historical time window, representing the central trend of that parameter under normal operating conditions. The historical standard deviation refers to the degree of dispersion of all historical operating parameter values ​​for a certain parameter dimension relative to its historical mean within a preset historical time window, used to quantify the normal fluctuation range of that parameter.

[0053] S1053. Calculate the difference between the real-time running parameters and the historical average of the target parameter dimension.

[0054] Specifically, the real-time operating parameter value of the target parameter dimension is read from the target operating parameters, and then subtracted from the previously calculated historical average to obtain the difference. This difference is an absolute value with dimensions, reflecting the actual deviation of the current real-time value relative to the historical average level: a positive difference indicates that the current value is higher than the historical average level; a negative difference indicates that the current value is lower than the historical average level; the larger the absolute value of the difference, the greater the deviation of the current state from the historical norm. This difference is an intermediate variable for subsequent calculations of standardized deviation. Although it cannot be directly compared across parameters due to the influence of dimensions, accurate difference calculation is a prerequisite for ensuring the accuracy of subsequent ratio calculations.

[0055] Among them, real-time operating parameters refer to specific physical or electrical quantity data that are collected and transmitted back in real time by sensors deployed on the equipment, and are used to characterize the current instantaneous operating state of the equipment.

[0056] S1054. Calculate the ratio of the difference to the historical standard deviation to obtain the sub-state deviation corresponding to the target parameter dimension.

[0057] Specifically, the difference calculated in the previous step is used as the numerator, and the historical standard deviation is used as the denominator. A division operation is performed to obtain the ratio between the two. This ratio is the sub-state deviation corresponding to the target parameter dimension. This indicator is essentially a standardized statistic, representing the degree to which the current real-time value deviates from the historical average level in units of standard deviation. This ratio follows the assumption of a standard normal distribution. Generally, an absolute value exceeding 2 is considered significant anomaly, and exceeding 3 is considered highly anomaly. Through this standardization transformation, parameter dimensions that originally had different dimensions and numerical ranges, such as temperature, voltage, and vibration amplitude, are unified to the same dimensionless scale, allowing for direct comparison and subsequent weighted fusion of the anomaly degrees between different parameters.

[0058] Among them, the sub-state deviation is a deviation index calculated for a single parameter dimension, expressed in standard deviation, reflecting the degree of abnormality of the current value of that single dimension relative to its historical mean.

[0059] S1055. The deviation of all sub-states is calculated by weighting based on a preset weight set to obtain the device state deviation.

[0060] Specifically, a pre-configured set of weights is read, which assigns a weight coefficient to each parameter dimension. These coefficients reflect the differences in the importance of different parameter dimensions in assessing the overall risk level of the equipment. For example, for wind turbine generators, the "main shaft vibration" dimension, which reflects the core mechanical health status, might be assigned a high weight, such as 0.4, while the "nacelle external temperature" dimension, which is greatly affected by the environment, might only be assigned a low weight, such as 0.05. No specific restrictions are imposed here. Then, for each parameter dimension, its sub-state deviation is multiplied by the corresponding weight coefficient to obtain the weighted contribution value of that dimension. Finally, the weighted contribution values ​​of all parameter dimensions are summed to obtain the final equipment state deviation. This comprehensive index is a dimensionless value that quantifies the overall degree to which the target equipment's current operating state deviates from its historical normal mode. The larger the value, the more abnormal the overall operating state of the equipment and the higher the potential risk, providing a key decision-making basis for subsequent risk level adjustments and risk item screening.

[0061] The preset weight set is a set of pre-configured values ​​used to assign different importance coefficients to different parameter dimensions, reflecting the differences in the contribution of each dimension in evaluating the overall status of the equipment. Its setting is usually based on equipment mechanism analysis, expert experience, or historical fault data analysis.

[0062] S106. Perform feature calculation on the query request to obtain the text feature vector, and perform feature calculation on the deviation between the target operating parameters and the device status to obtain the device status feature vector.

[0063] Specifically, two feature computation paths are executed in parallel: In the text processing path, the user's original query text is input into a pre-trained natural language encoding model (such as LLM, BERT, Word2Vec, etc.). The model's embedding and attention layers map the words, syntax, and semantic information in the text into a fixed-length dense vector, i.e., the text feature vector. This vector captures the core semantics of the query request, including the user's intent type and the entity information they are interested in. In the state processing path, the real-time values ​​of each dimension of the target operating parameters and the calculated device state deviation are combined into a multi-dimensional state data sequence. This sequence is input into a pre-trained numerical feature encoder, which, through normalization and feature transformation, maps it into a fixed-length dense vector, i.e., the device state feature vector. This vector represents the device's current real-time health status and degree of abnormality. Finally, two feature vectors with the same or aligned dimensions are output, representing the semantic information of the user query and the device's current physical state information, respectively.

[0064] Among them, the text feature vector refers to the semantic features extracted from the user's query text through a natural language processing model and represented as a fixed-length numerical vector, which is used for subsequent similarity calculation. The device status feature vector refers to the fixed-length numerical vector obtained by combining the real-time values ​​of each dimension of the target operating parameters with the device status deviation and converting it through a numerical encoder, which is used to characterize the current real-time health status of the device.

[0065] For example, given a user query "Is the temperature of inverter #5 too high?", the current device status is: module temperature 47.3℃, DC side voltage 820V, and device status deviation 1.65. This can be encoded into a 768-dimensional text feature vector using the BERT model. Simultaneously, the real-time parameters of inverter #5 and the device status deviation 1.65 are combined into a single 3D raw data set {47.3, 820, 1.65}, which is then normalized and mapped through a fully connected layer to form another 768-dimensional device status feature vector. These two vectors encapsulate the semantic information relevant to the user and the actual physical state of the device, respectively, and are prepared for input into the subsequent fusion module. No specific restrictions are imposed here.

[0066] S107. Generate a fused retrieval vector based on the text feature vector and the device status feature vector.

[0067] Specifically, the text feature vector and device status feature vector are mapped and combined into a unified high-dimensional feature space through a specialized fusion model (such as a multimodal neural network, attention mechanism, or simple vector concatenation followed by dimensionality reduction). The final product of this process is a novel, higher-information-density single vector, namely the fused retrieval vector. This vector simultaneously encodes "what the problem is" (textual semantics) and "what the device status is" (physical condition), enabling it to serve as a highly accurate "digital fingerprint" for subsequent similarity matching and retrieval in a knowledge base or historical case database.

[0068] Among them, the fusion retrieval vector refers to a new vector generated by weighted fusion of text feature vectors and device status feature vectors through an attention mechanism, which is used for unified similarity retrieval in the risk knowledge base.

[0069] Through the above embodiments, it is ensured that when similarity searches are performed in the risk knowledge base, the risk entries hit are not only semantically relevant, but also highly matched with the current real-time health status of the device, thereby greatly improving the accuracy of risk identification and context awareness.

[0070] Based on the above embodiments, as an optional embodiment, for Figure 1 The step S107 shown can be implemented through steps S1071-S1073, which will be explained in detail below.

[0071] S1071. Input the text feature vector and device state feature vector into the pre-trained attention layer.

[0072] Specifically, after obtaining the text feature vector and the device state feature vector, the initial stage of feature interaction begins. These two heterogeneous but dimensionally aligned vectors are used as input and passed to a pre-trained attention layer module. This attention layer is a neural network component built on the Transformer architecture, whose parameters have been pre-trained on a large amount of domain data. It has mastered how to identify the association patterns between semantic information and state information. The text feature vector is placed at the query position of the attention mechanism, and the device state feature vector is placed at the key and value positions, ready to start calculating the mutual attention relationship between the two. This initiates the multimodal information fusion process and establishes an interaction channel for the two originally independent feature vectors.

[0073] The attention layer refers to a neural network layer based on the attention mechanism that has been trained on a large-scale dataset. Its core function is to calculate the correlation weights between input features and realize the interaction and fusion of information.

[0074] S1072. Calculate the first attention weight of the text feature vector to the device state feature vector and the second attention weight of the device state feature vector to the text feature vector through the attention layer.

[0075] Specifically, the attention layer performs two sets of parallel calculations through a dot product attention mechanism: First, it uses the text feature vector as the query and the device state feature vector as the key, calculating the similarity score between them. This score is then normalized using the Softmax function to obtain the first attention weight. This weight distribution reflects the extent to which each component of the user query semantics should focus on each dimension of the device state features. Second, it uses the device state feature vector as the query and the text feature vector as the key, calculating the similarity score between them and normalizing it to obtain the second attention weight. This weight distribution reflects the extent to which each component of the device's current abnormal state features should focus on each word concept in the user query semantics. Both sets of weights are in probability distribution form, with values ​​between 0 and 1, and their sum is 1, accurately depicting the bidirectional interaction between semantics and state.

[0076] The first attention weight refers to the distribution of attention given to the device status feature vector by the text feature vector, reflecting the query semantics' focus on each dimension of the device status. The second attention weight refers to the distribution of attention given to the text feature vector by the device status feature vector, reflecting the emphasis required on each part of the query semantics by the device's abnormal status.

[0077] S1073. Based on the first attention weight and the second attention weight, the text feature vector and the device status feature vector are weighted and calculated to generate a fused retrieval vector.

[0078] Specifically, after calculating the first and second attention weights, the device state feature vectors are weighted and summed based on the first attention weight to obtain a state feature representation guided by the query semantics. This process highlights and enhances the dimensions of the device state most relevant to the user query. Then, based on the second attention weight, the text feature vectors are weighted and summed to obtain a semantic feature representation guided by the device state. This process highlights and enhances the semantic components of the user query most relevant to the current device anomaly. Subsequently, these two enhanced feature representations are concatenated or element-wise added to generate the final fused retrieval vector. This vector contains both the core semantic direction of the user query and the current abnormal state features of the device. Furthermore, the interaction between the two has been dynamically aligned and enhanced through attention weights, enabling the fused vector to more accurately locate risk entries in the risk knowledge base's vector space that are both semantically relevant and state-matching.

[0079] S108. Input the fused retrieval vector into the preset risk knowledge base for similarity retrieval, and determine at least one candidate risk item that matches the fused retrieval vector.

[0080] Specifically, the fused retrieval vector is used as the query vector and input into a pre-built and quantized risk knowledge base. Each risk entry in this knowledge base, such as "inverter module overheating may lead to shutdown" or "wind turbine blade icing risk management plan," has been converted into a vector representation of the same dimension offline using the same encoder, and an efficient vector index has been established. An approximate nearest neighbor search is performed in the knowledge base to calculate the similarity score between the fused retrieval vector and each risk entry vector in the base, typically using cosine similarity or Euclidean distance as the metric. Subsequently, the similarity scores are sorted from high to low, and a certain number of risk entries with scores exceeding a preset similarity threshold or ranking highly are selected as candidate risk entries that match the fused retrieval vector. These candidate risk entries are semantically relevant to the user query and match the current abnormal characteristics of the equipment in terms of status, forming the raw material pool for subsequent filtering and generation of risk guidance information.

[0081] The pre-built risk knowledge base refers to a structured database that is pre-constructed and stored, containing various risk-related entries for new energy projects. Each entry has been converted into a vector representation and indexed. Candidate risk entries refer to one or more risk information entries that match the fused search vector and are selected from the risk knowledge base after similarity retrieval, serving as candidate objects for subsequent processing.

[0082] S109. Based on the deviation between intent type and device status, each candidate risk item is screened to obtain a target candidate risk item set, and risk guidance information is generated based on the target candidate risk item set through a preset risk identification model.

[0083] Specifically, the intent type (e.g., risk intent type or risk question-and-answer type) is the primary filtering condition. A preliminary, purpose-based scope definition is performed on a large knowledge base of candidate risk items encompassing all possible scenarios. The equipment status deviation calculated in the previous step is used as a key quantitative screening threshold. If the deviation is high, risk items related to severe failures or sharp performance degradation are prioritized; if the deviation is low, items related to early wear, minor efficiency declines, or optimization suggestions are emphasized. Through this combined filtering of intent and status severity, the analysis scope is significantly narrowed, ultimately resulting in a highly relevant and easily processed set of target candidate risk items. This refined set of target candidate risk items is then submitted as input. A deep analysis is performed on a pre-defined risk identification model. This model is not a simple search and match; it may be a complex expert system, decision tree, or a trained machine learning classifier. The model comprehensively evaluates each item in the target set, combines it with lower-level equipment operation data (such as specific sensor readings, historical trends, etc.), calculates the probability and confidence level of each candidate risk, and infers the causal relationship between them. Finally, the model integrates the analysis results and transforms them into human-readable text or instructions with clear guidance, i.e., risk guidance information. This information usually clearly explains what the most likely risk is, what the possible consequences may be, and the specific countermeasures or troubleshooting steps recommended, thereby providing direct and accurate decision support for operations and maintenance personnel.

[0084] The target candidate risk item set refers to the set of risk items ultimately selected for generating risk guidance after multiple screening processes, including intent type and state deviation. The preset risk identification model is a generative model built on large language model technology, used to convert structured risk data into natural language risk guidance information. Risk guidance information refers to the readable text or list that the system ultimately outputs to the user, containing risk level, handling suggestions, compliance basis, etc.

[0085] Based on the above embodiments, as an optional embodiment, for Figure 1 Step S109 shown can be implemented through steps S1091-S1095, as explained in detail below.

[0086] S1091. Select the candidate risk items that match the deviation range of the equipment status deviation as intermediate candidate risk items, obtain the intermediate candidate risk item set, and obtain the preset benchmark risk level of each intermediate candidate risk item.

[0087] Specifically, the pre-configured applicable deviation range for each candidate risk item is read. This range defines the level of equipment anomaly under which the risk item is relevant; for example, the applicable range for the "inverter over-temperature risk" item is a deviation > 1.5, and the applicable range for the "minor voltage fluctuation" item is a deviation < 1.0. The currently calculated equipment state deviation is compared with the applicable deviation range of each candidate risk item. Only items whose current deviation falls within their applicable range are retained as intermediate candidate risk items, forming an intermediate candidate risk item set. For each retained intermediate candidate risk item, a pre-configured baseline risk level is further read from its metadata. This level is typically pre-assessed based on the probability of the risk occurring and the severity of its consequences, such as "low risk," "medium risk," "high risk," or using a numerical scale of 1-5. Finally, the intermediate candidate risk item set and its associated baseline risk level are output, laying the foundation for subsequent dynamic adjustment of risk levels.

[0088] Here, the deviation range refers to the pre-configured range of equipment status deviation values ​​for each risk item, used to filter risk information that matches the current level of anomaly. Intermediate candidate risk items are those retained after the deviation range filtering, matching the current level of equipment anomaly. The "intermediate candidate risk item set" refers to the collection of all intermediate candidate risk items. The baseline risk level is a pre-assessed level for each risk item, reflecting its inherent risk level, typically determined based on the probability of risk occurrence and the severity of its consequences.

[0089] Through the above embodiments, it is ensured that the risk information recommended subsequently is not only semantically relevant to the user's query, but also quantitatively matched with the current level of anomaly of the device. This avoids recommending risk management measures applicable to serious faults to devices with minor anomalies, and also avoids recommending routine reminders applicable only to normal operating conditions to devices with serious anomalies, thereby greatly improving the pertinence and practicality of risk guidance.

[0090] S1092. Match the equipment status deviation with multiple preset deviation threshold ranges to determine the risk level adjustment coefficient corresponding to the equipment status deviation.

[0091] Specifically, the system reads multiple pre-configured deviation threshold ranges and their corresponding risk level adjustment coefficients. These threshold ranges are typically divided based on statistical principles and operational experience. For example, equipment status deviations are divided into normal range (<1.0), watch range (1.0-2.0), warning range (2.0-3.0), and alarm range (>3.0). Each range is configured with a corresponding adjustment coefficient, such as 0.8, 1.0, 1.5, and 2.0. The currently calculated equipment status deviation is matched sequentially with these threshold ranges to determine its specific range, and the corresponding risk level adjustment coefficient is extracted. This adjustment coefficient reflects the factor by which the risk level should be amplified or reduced at the current deviation level. For example, a deviation of 1.65 falls into the watch range, with an adjustment coefficient of 1.0, meaning the baseline risk level remains unchanged. If a deviation of 2.5 falls into the warning range, with an adjustment coefficient of 1.5, the risk level needs to be increased by 50%. Finally, the determined risk level adjustment coefficients are output for subsequent unified adjustment of the baseline risk level of intermediate candidate risk items.

[0092] The preset multiple deviation threshold intervals refer to several numerical intervals that are pre-divided based on statistical principles and operational experience, mapping continuous values ​​of equipment status deviation to discrete risk levels. The risk level adjustment coefficient is a multiplier factor determined according to the threshold interval in which the equipment status deviation falls, used to amplify or reduce the baseline risk level.

[0093] S1093. Adjust the baseline risk level of each intermediate candidate risk item based on the risk level adjustment coefficient to generate the corresponding intermediate risk level for each intermediate candidate risk item.

[0094] Specifically, each item in the set of intermediate candidate risk items is traversed, and its pre-configured baseline risk level is read. This level is usually stored in numerical form for easy calculation; for example, low risk corresponds to 1, medium risk to 2, and high risk to 3. For each item, the risk level adjustment coefficient determined in the previous step is multiplied by its baseline risk level to obtain a new value, which is the intermediate risk level of that item under the current equipment status deviation. For example, if the adjustment coefficient is 1.5, the intermediate risk level of an item with an original baseline risk level of 2 (medium risk) will become 3.0; if the adjustment coefficient is 0.8, the intermediate risk level of an item with an original baseline risk level of 3 (high risk) will become 2.4. This intermediate risk level reflects the degree to which the risk level of each risk item should be adjusted up or down when considering only the single factor of equipment status deviation, providing an initial quantitative benchmark for further adjustments by introducing more dynamic factors such as environment and personnel.

[0095] The baseline risk level refers to the pre-assessed risk level of each risk item, reflecting its inherent risk level, and is usually stored in numerical form. The intermediate risk level refers to the transitional risk level obtained by applying a risk level adjustment coefficient to the baseline risk level, which initially reflects the impact of the current abnormality of the equipment.

[0096] S1094. Based on the anchor point of the operation scenario corresponding to the target equipment, obtain real-time micro-meteorological data and operation duration data, and dynamically adjust each intermediate risk level based on the real-time micro-meteorological data and operation duration data to obtain the dynamic risk level of each intermediate candidate risk item.

[0097] Specifically, based on the geographical location or work area of ​​the target equipment, a preset work scenario anchor point is invoked as an information index. This could be a precise GPS coordinate, a site number, or a specific workstation identification code. A request is proactively sent to an external data source. Through this anchor point, the specific physical environment of the equipment can be accurately located. Real-time micro-meteorological data strongly correlated with that location is retrieved from meteorological service systems, IoT environmental sensors, or on-site data collection. This data includes not only conventional temperature and humidity but may also include wind speed, air pressure, rainfall, and light intensity, which have potential impacts on equipment operation. The system incorporates refined environmental indicators and also obtains operational duration data from the equipment's own monitoring system or production management system (MES). This refers to key information reflecting work fatigue, such as the continuous operating time of the equipment since its startup or the duration under current high-load conditions. The acquired real-time micro-meteorological data and operational duration data are used as dynamic influencing factors to quantitatively correct the intermediate risk levels previously assessed. This adjustment process is based on a built-in knowledge base or dynamic risk model, which pre-defines the impact logic of different environmental conditions and work intensity on different risk types. For example, the model might define "when the ambient temperature exceeds 40°C, the 'overheating' risk level is increased by 20%"; or "when continuous uninterrupted operation exceeds 10 hours, the 'mechanical fatigue' risk level is increased by 15%"; or "when high humidity and high dust concentration occur simultaneously, the 'electrical short circuit' risk level is increased by a combined 30%." The system will evaluate each intermediate candidate risk item one by one and trigger corresponding adjustment rules or algorithms based on real-time data. Through this series of refined weighting and correction, the originally static risk assessment value is updated in real time, and finally the dynamic risk level of each intermediate candidate risk item is generated. This final level value more accurately reflects the real risk status of the equipment at this moment, in this place and in this situation, providing a more timely and accurate basis for subsequent early warning and decision-making.

[0098] Among them, the operation scenario anchor point refers to the identifier or interface bound to the geographical location of the target equipment, used to associate and obtain real-time environmental data and operation information for that location. Real-time micro-meteorological data refers to the refined meteorological parameters of the area where the target equipment is located, obtained in real time through on-site meteorological stations or meteorological service interfaces, including wind speed time series data, ultraviolet radiation index, and surface icing humidity. Operation duration data refers to the continuous working time of personnel currently working in the target equipment area. Dynamic risk level refers to the final risk level obtained after further applying a comprehensive spatiotemporal attenuation factor to the intermediate risk level for nonlinear adjustment, comprehensively reflecting the three factors of equipment status, environmental conditions, and personnel status.

[0099] Based on the above embodiments, as an optional embodiment, step S1094 can be implemented through steps S10941-S10945, which will be explained in detail below.

[0100] S10941. Based on the anchor point of the operation scenario, obtain real-time micro-meteorological data of the operation area of ​​the new energy project where the target equipment is located. The real-time micro-meteorological data includes at least one of wind speed time series data, ultraviolet sun exposure index and surface icing humidity.

[0101] Specifically, the system reads the work scenario anchor point bound to the target device. This anchor point is a pre-configured geographic location identifier or data interface address pointing to the specific work area where the device is located, such as "Inverter No. 5 - Photovoltaic Zone B - Weather Station 03". Through this anchor point, it calls the micro-weather station deployed at the project site or connects to a third-party meteorological service interface to obtain real-time micro-meteorological data for the area. Based on common environmental risk factors in new energy projects, three key parameters are collected: first, wind speed time-series data, including current instantaneous wind speed, average wind speed, and gust wind speed, used to assess the impact of strong winds on high-altitude operations and equipment stability; second, ultraviolet radiation index, used to assess accelerated equipment aging and the risk of heatstroke among personnel under strong sunlight; and third, surface icing humidity, used to assess the risks of equipment icing, reduced insulation, and slippery work surfaces under low-temperature freezing weather. One or more of these parameters may be collected, depending on the equipment type and the characteristics of the work scenario. Finally, the system outputs a real-time micro-meteorological dataset of the work area where the target device is located, which serves as the environmental dimension input for subsequent calculations of the comprehensive spatiotemporal attenuation factor.

[0102] The work area refers to the physical space where the target equipment is located, usually corresponding to a specific meteorological monitoring point or work management unit. Wind speed time-series data refers to a sequence of wind speed values ​​collected in chronological order, including instantaneous wind speed, average wind speed, and gust wind speed, used to characterize wind conditions. The ultraviolet (UV) exposure index is a quantitative indicator characterizing the intensity of solar ultraviolet radiation, typically divided into several levels, used to assess exposure risk. Surface icing humidity refers to a parameter characterizing the degree of icing or slipperiness on the equipment or work surface, used to assess the work risk in low-temperature freezing environments.

[0103] S10942. Obtain the current worker's work duration data, and based on the work duration data and the preset heavy physical labor duty cycle, deduce the physiological metabolic fatigue index.

[0104] Specifically, the system reads the work duration data of personnel currently working in the target equipment's work area from the personnel operation management module or the on-site attendance system. This data represents the continuous working time of the personnel from the start of their shift to the current moment. Based on the nature of the job, the system reads the pre-configured duty cycle for heavy physical labor operations. This parameter is a value between 0 and 1, reflecting the proportion of time that the position spends on heavy physical labor operations within a standard work cycle. For example, the duty cycle for the inspection position is 0.3, and the duty cycle for the maintenance position is 0.7. The operation time data and the duty cycle of heavy physical labor are then input into a preset physiological fatigue simulation model. This model is built based on ergonomics and fatigue accumulation theory. By calculating the effective physical exertion time and mapping it to a preset fatigue curve, the physiological metabolic fatigue index of the current worker is derived. This index is usually a value between 0 and 1. The larger the value, the higher the degree of fatigue and the greater the impact on risk. For example, 0.3 indicates mild fatigue, 0.7 indicates significant fatigue, and more than 0.9 indicates extreme fatigue. Finally, the physiological metabolic fatigue index is output as an important component of the subsequent comprehensive spatiotemporal decay factor.

[0105] The preset heavy physical labor duty cycle refers to a parameter pre-configured based on the nature of the job, reflecting the proportion of time spent on heavy physical labor within a standard work cycle, used to estimate the actual intensity of physical exertion. The physiological metabolic fatigue index is a quantitative indicator characterizing the current level of physical fatigue of the worker, calculated using a physiological fatigue extrapolation model based on the work duration and heavy physical labor duty cycle; it typically ranges from 0 to 1.

[0106] Through the above embodiments, by introducing work duration and duty cycle of heavy physical labor, a quantitative assessment of the fatigue state of workers is achieved. This enables the risk management system to perceive the real-time status changes of "people," the most critical safety element. By incorporating personnel fatigue, an important risk factor, into the comprehensive assessment system, the impact of fatigue level on risk level can be objectively quantified. This avoids underestimating the risk after personnel have been working continuously for a long time or after high-intensity physical exertion, and provides a key human factor dimension input for generating risk guidance information that is more in line with actual working conditions.

[0107] S10943. Calculate the ecological stress gradient value that characterizes the trend of environmental degradation based on real-time micro-meteorological data.

[0108] Specifically, feature extraction is performed on the collected wind speed time-series data, and its rate of change and peak intensity are calculated to characterize the abruptness and extreme nature of the wind environment; the cumulative effect of the ultraviolet radiation index is calculated to quantify its continuous stress on equipment and personnel; and phase transition trend analysis is performed on the surface icing humidity data to assess the rate of icing formation or melting. These feature-engineered environmental parameters are then input into a pre-built ecological stress calculation model. This model is constructed based on the coupling relationship between environmental stress and equipment reliability. Through weighted fusion and nonlinear transformation, the multidimensional environmental parameters are converted into a comprehensive scalar value, namely the ecological stress gradient value. This value characterizes the comprehensive stress degree of current environmental conditions on equipment operation and safety; the larger the value, the more significant the environmental degradation trend and the stronger the amplification effect on risks.

[0109] Among them, the ecological stress gradient value refers to a quantitative indicator obtained by extracting features from real-time micrometeorological data and calculating models, which is used to comprehensively characterize the current degree of environmental degradation and its amplifying effect on risks.

[0110] For example, suppose we obtain real-time micrometeorological data for the current area: wind speed time-series data shows that the average wind speed has rapidly increased from 5 m / s to 8 m / s in the past 10 minutes, with a rate of change of 0.3 m / s²; the UV index is level 4 (strong); and the surface icing humidity is 0%. Inputting this data into an ecological stress calculation model, the model assigns a weight of 0.6 to the abrupt change in wind speed, a weight of 0.4 to UV intensity, and no contribution from icing humidity. After nonlinear transformation, the output ecological stress gradient value is 0.65, indicating that the current environment has a moderate amplification effect on risk.

[0111] S10944. The ecological stress gradient value and the physiological metabolic fatigue index are combined and transformed to generate a comprehensive spatiotemporal decay factor.

[0112] Specifically, the ecological stress gradient value and the physiological metabolic fatigue index are used as inputs to a pre-defined feature combination transformation model. This model, based on risk coupling theory, does not simply perform a weighted summation of the two indicators, but rather captures their synergistic enhancement effect through a nonlinear mapping function. For example, in high-temperature environments, the risk of decreased attention due to personnel fatigue is further amplified; in windy environments, the consequences of operational errors caused by personnel fatigue are more severe. The model calculates the interaction term between the two and combines it with a pre-defined environment-fatigue coupling coefficient to generate a comprehensive spatiotemporal attenuation factor. This factor is a dimensionless value, typically greater than or equal to 1; a larger value indicates a stronger amplification effect of the current environment and personnel status on the risk level. The final output is the comprehensive spatiotemporal attenuation factor, used for dynamic adjustment of intermediate risk levels.

[0113] Among them, the comprehensive spatiotemporal attenuation factor refers to the multiplier factor generated by combining and transforming the ecological stress gradient value and the physiological metabolic fatigue index, which is used to dynamically adjust the risk level. It comprehensively reflects the amplification effect of environmental and human factors on risk.

[0114] For example, suppose the ecological stress gradient of the current work area is 0.65, and the physiological metabolic fatigue index of the current workers is 0.45. Inputting these two values ​​into the feature combination transformation model, the model not only calculates the weighted sum of the two (0.65 + 0.45 = 1.1), but also calculates the interaction term 0.65 × 0.45 = 0.2925. Combined with a preset coupling coefficient, the final comprehensive spatiotemporal attenuation factor is 1.3. This factor indicates that, under the current environmental and personnel conditions, the risk level should be amplified by 30%.

[0115] Through the above embodiments, by combining and transforming features, dynamic risk factors of both environment and personnel are integrated into a unified attenuation factor. This can quantify the synergistic enhancement effect between environmental severity and personnel fatigue, making the risk adjustment factor more consistent with the actual risk evolution law and providing a scientific basis for subsequent precise adjustment of intermediate risk levels.

[0116] S10945. The dynamic risk level of each intermediate risk level is dynamically adjusted nonlinearly by comprehensively considering the spatiotemporal attenuation factor.

[0117] Specifically, the process iterates through each item in the set of intermediate candidate risk items, reads its current intermediate risk level, and for each item, performs a nonlinear dynamic adjustment calculation between the comprehensive spatiotemporal attenuation factor and the intermediate risk level. This is achieved through a preset mapping function, which considers the impact of the risk level itself on the adjustment magnitude: for example, for a higher risk level, the amplification effect of the environment and personnel may be more significant; for a lower risk level, the amplification effect may be relatively mild. The intermediate risk level is input into the nonlinear adjustment function, with the comprehensive spatiotemporal attenuation factor as the control parameter, to calculate the final dynamic risk level of the item. This dynamic risk level is a final quantitative indicator that comprehensively reflects three factors: equipment status deviation, environmental degradation, and personnel fatigue, and is used for the subsequent sorting and generation of risk guidance information.

[0118] S1095. Based on the intent type, select a set of target candidate risk items that match the intent type from the intermediate candidate risk item set, and generate risk guidance information that matches the intent type based on the dynamic risk level of each target candidate risk item in the target candidate risk item set.

[0119] Specifically, the user's operational goal or the system's preset task mode, i.e., the intent type (e.g., "troubleshooting", "perform preventative maintenance", or "seek performance optimization suggestions"), serves as a clear instruction filter. Then, each entry in the intermediate candidate risk item set is traversed. Since each risk item is pre-labeled with its applicable intent category in the knowledge base (e.g., "bearing wear" may be associated with both "troubleshooting" and "preventative maintenance"), a matching operation is performed, retaining only those risk items whose labels perfectly match the current intent type. This filtering process greatly improves the targeting of subsequent processing, narrowing the analysis scope from all possible risks to those that the current task needs to focus on, ultimately forming a highly focused set of target candidate risk items. The dynamic risk level of each target candidate risk item in the target candidate risk item set, calculated in the previous step, is read. This level is a quantitative score that integrates multiple dimensions such as equipment status deviation, real-time environment, and operation duration, directly reflecting the real-time urgency of each risk. Based on this dynamic risk level, all risks in the target set are sorted in descending order, with the most urgent and critical risks placed at the top. Then, based on the sorting results and the current intent type, a preset solution or suggestion generation module is invoked. For example, for a high-risk "troubleshooting" intent, a strongly worded warning and clear troubleshooting steps will be generated; while for a medium-risk "preventive maintenance" intent, a suggested maintenance plan and a list of required spare parts will be generated. Finally, this sorted and intent-customized information is integrated to form risk guidance information that matches the intent type, providing users with clear, actionable, and prioritized action plans.

[0120] Based on the above embodiments, as an optional embodiment, the intent type includes risk query intent and risk question and answer intent. For step S1095, it can be implemented through steps S10951-S10954, which will be explained in detail below.

[0121] S10951. When the intent type is risk query intent, intermediate candidate risk items belonging to historical risk data are selected from the intermediate candidate risk item set to form the target candidate risk item set. The historical risk data is associated with historical handling measures.

[0122] Specifically, when the user's intent is determined to be a risk query intent, each entry in the intermediate candidate risk entry set is traversed, and its data type label is checked to see if it is historical risk data. This label indicates that the content of the entry belongs to a record of a risk event that actually occurred in history, rather than a theoretical norm or general clause. For each entry identified as historical risk data, its associated historical handling measures field is further extracted. This field records in detail the specific operational steps, resource allocation, and handling effects taken at that time for the risk. After filtering, all entries that belong to the historical risk data type and carry effective historical handling measures are gathered together to form the target candidate risk entry set. Each entry in this set carries two core pieces of information: dynamic risk level and corresponding historical handling measures, providing experience support for the subsequent generation of actionable risk guidelines.

[0123] Among these, "risk query intent" specifically refers to a user's desire to understand current risk management references. Historical risk data refers to a special type of risk entry in the risk knowledge base, containing records of actual risk events that occurred historically, rather than theoretical norms. Historical management measures refer to structured information associated with historical risk data entries, recording the specific operational steps, resource allocation, and management effects taken at the time of the risk.

[0124] S10952. Extract the historical handling measures and dynamic risk levels associated with each target candidate risk item in the target candidate risk item set, generate a risk operation guidance list that is sorted from high to low dynamic risk level, and obtain risk guidance information.

[0125] Specifically, for each item in the target candidate risk item set, two core information fields are extracted: first, the historical handling measures associated with the item, i.e., the specific operational steps and solutions previously implemented for this type of risk; second, the dynamic risk level obtained after multi-dimensional dynamic adjustment of the item by human, machine, and environment. This level comprehensively reflects the quantitative assessment of the actual threat level of the risk based on the current state of the equipment, the severity of the environment, and the fatigue level of the personnel. After extracting information from all items, all target candidate risk items are sorted in descending order using the dynamic risk level as the sorting key, placing the items with the highest and most urgent risk levels at the top. The sorted list of items and their associated historical handling measures are then input into a preset risk identification model. This model organizes the structured risk data and handling measures into coherent and clear natural language text, generating a risk operation guidance list containing risk level ranking and corresponding handling suggestions. Finally, this list is output as risk guidance information, providing maintenance personnel with priority-ranked, directly actionable plans.

[0126] The risk operation guidance list refers to an operational list generated by sorting multiple risk items from high to low according to their dynamic risk level, which includes the risk level and corresponding historical handling measures.

[0127] S10953. When the intent type is risk question and answer intent, intermediate candidate risk items belonging to the normative clauses are selected from the intermediate candidate risk item set to form the target candidate risk item set, and the normative clauses are associated with compliance basis information.

[0128] Specifically, when a user's intent is determined to be a risk-based question-and-answer intent, each entry in the intermediate candidate risk item set is traversed, and its data type tag is checked to see if it is a normative clause. This tag indicates that the content of the entry belongs to normative documents such as industry standards, operating procedures, or safety management regulations, rather than historical cases or theoretical descriptions. For each entry identified as a normative clause, its associated compliance basis information field is further extracted. This field records the source of the clause in detail, including the standard number, clause number, publication date, and the original text of the specific compliance requirements. After filtering, all entries belonging to the normative clause type and carrying valid compliance basis information are gathered to form the target candidate risk item set. Each entry in this set carries two core pieces of information: the dynamic risk level and the corresponding compliance basis information, providing a basis for generating an authoritative compliance analysis report.

[0129] Among these, "risk Q&A" specifically refers to users' desire to understand the compliance basis for risks. "Normative clauses" refers to a special type of risk entry in the risk knowledge base, containing specific clauses from industry standards, operating procedures, or safety management regulations, rather than historical cases. "Compliance basis information" refers to standardized information associated with the normative clause entry, recording the source of the clause, including the standard number, clause number, publication date, and the original text of the specific requirements.

[0130] S10954. Extract the compliance basis information and dynamic risk level associated with each target candidate risk item in the target candidate risk item set, generate a list of compliance analysis reports sorted from high to low dynamic risk level, and obtain risk guidance information.

[0131] Specifically, for each item in the target candidate risk item set, two core information fields are extracted: first, the compliance basis information associated with the item, i.e., the source and specific requirements of the clause, including the standard number, clause number, and the original text of the compliance requirement; second, the dynamic risk level obtained after multi-dimensional dynamic adjustment of the item by human, machine, and environment. This level comprehensively reflects the quantitative assessment of the actual threat level of the risk involved by the clause based on the current state of the equipment, the severity of the environment, and the degree of personnel fatigue. After extracting information from all items, the dynamic risk level is used as the sorting key to sort all target candidate risk items in descending order, placing the compliance requirements with the highest risk level and the most urgent at the top. Subsequently, the sorted list of items and their associated compliance basis information are input into a preset risk identification model. This model organizes the structured compliance data and risk levels into coherent and clear natural language text, generating a compliance analysis report list containing compliance requirements and their sources sorted by dynamic risk level. This list is output as risk guidance information, providing maintenance personnel with priority-sorted compliance analysis references with clear standard basis.

[0132] The compliance analysis report list refers to a reference list generated by sorting multiple regulatory clauses from high to low according to dynamic risk levels, and containing information on risk levels and corresponding compliance bases.

[0133] Figure 2 This is another flowchart illustrating the risk management method for new energy projects provided in this application embodiment.

[0134] Based on the above embodiments, as an optional embodiment, for Figure 1 The risk management methods for new energy projects shown may also include Figure 2 Steps S201-S206 are explained in detail below.

[0135] S201. Obtain historical trend data of historical operating parameters within a preset historical time window.

[0136] Specifically, after calculating the deviation of the current equipment status, the historical trend analysis stage begins. Using the current moment as a benchmark, and based on a preset historical trend analysis window (e.g., the past 7 days, the past 30 days, or the past 12 months), all historical data points for each parameter dimension of the target operating parameters within the complete time window are retrieved from the time series database. Unlike the parameter time series data used when calculating the historical mean and standard deviation, the historical trend data extracted here not only includes the numerical values ​​themselves, but also emphasizes the sequence characteristics of the data evolution over time, including the overall upward or downward trend, periodic fluctuation patterns, abrupt change points, and rate of change. During the extraction process, the temporal continuity and integrity of the data are ensured, and the data is smoothed to eliminate short-term random noise interference. Finally, the preprocessed historical trend data sequence for each parameter dimension within the preset historical time window is output.

[0137] The preset historical time window is a pre-configured time length parameter used to limit the time range for extracting historical trend data, such as "the past 30 days" or "the past 12 months". Its length setting needs to consider capturing the complete change cycle of the parameter. Historical trend data refers to a complete set of data extracted within the preset historical time window for a specific parameter dimension, emphasizing the time series evolution characteristics, including information such as the trend, cycle, abrupt changes, and rate of data.

[0138] S202. Based on historical trend data, generate predicted operating parameters for the target operating parameters at a future preset time point through a preset trend prediction model.

[0139] Specifically, historical trend data for each parameter dimension is input into a pre-trained trend prediction model. This model is built on time series prediction algorithms, such as long short-term memory networks or gated recurrent units, and can learn complex patterns such as periodic fluctuations, trend drifts, and seasonal changes of parameters from historical data. Based on the input time series data, the model extrapolates and calculates the possible value of the parameter dimension at a future preset time point (e.g., 15 minutes, 1 hour, or 4 hours), generating predicted operating parameters. When generating predicted values, the model not only outputs point prediction results but also usually outputs prediction intervals to characterize the uncertainty of the prediction. The above prediction process is executed in parallel for each parameter dimension in the target operating parameters, and finally, a set of predicted operating parameters for each parameter dimension at the same future time is obtained.

[0140] The preset trend prediction model refers to a pre-trained machine learning model built based on time series prediction algorithms, used to predict future parameter values ​​based on historical data. The preset future time point refers to a pre-set future moment that needs to be predicted, such as "the next 15 minutes" or "the next hour." Predicted operating parameters refer to values ​​generated by the trend prediction model that characterize the operating state the equipment may reach at the preset future time point.

[0141] For example, regarding the "module temperature" dimension of "Inverter No. 5," the system inputs the temperature change trend data of the past 30 days into a trained Long Short-Term Memory (LSTM) network prediction model. Model analysis reveals that this parameter exhibits a 24-hour periodic fluctuation pattern. Combined with the recent overall upward drift trend, the system extrapolates and calculates the predicted module temperature for 15 minutes to come to 47.8℃ and for 1 hour to come to 48.5℃. Simultaneously, the same prediction is performed on the DC-side voltage dimension, yielding a predicted voltage of 825V for the next hour. Finally, the system outputs a set of predicted operating parameters for each parameter dimension at the preset future time points.

[0142] S203. Calculate the deviation of the predicted state based on the predicted operating parameters and historical operating parameters.

[0143] Specifically, historical benchmark data matching the future preset time point is extracted from historical operating parameters. This includes historical data from the same time period as the future time or historical statistical features within a recent sliding window. Then, for each parameter dimension, the predicted operating parameters are compared and analyzed with the corresponding historical operating parameters, using the same calculation logic as for real-time state deviation: the difference between the predicted value and the historical mean is calculated, and this difference is divided by the historical standard deviation to obtain the predicted sub-state deviation for that parameter dimension. This indicator, expressed in standard deviation, characterizes the degree to which the equipment operating parameters may deviate from their historical normal levels at the future preset time point. After obtaining the predicted sub-state deviations for all parameter dimensions, the same preset weight set as in the real-time evaluation is used to perform a weighted summation of all predicted sub-state deviations, ultimately generating a comprehensive predicted state deviation. This indicator quantifies the overall degree to which the overall operating state of the equipment may deviate from its historical normal mode at a future time.

[0144] Among them, the predicted state deviation is a comprehensive quantitative indicator used to characterize the expected deviation of the overall operating state of the target equipment at a future preset time point from its own historical normal state.

[0145] S204. When the predicted state deviation exceeds the preset warning threshold, the predicted state deviation and the device state deviation are weighted and fused to generate a fused state deviation.

[0146] Specifically, the predicted state deviation is compared with a pre-configured warning threshold. This threshold is typically set based on statistical principles and operational experience, such as 2.0 or 2.5, indicating that when the predicted future deviation reaches this level, the device faces significant foreseeable risks. If the predicted state deviation does not exceed the warning threshold, the fusion process is not triggered; if it exceeds the warning threshold, the fusion calculation process begins. Pre-configured fusion weight coefficients are read, reflecting the relative importance of the current and future states in the comprehensive evaluation. For example, the current state has a weight of 0.3, and the future state has a weight of 0.7, reflecting the guidance of forward-looking warnings. Then, the device state deviation and the predicted state deviation are multiplied by their corresponding weights, and the products are summed to generate the fused state deviation. This comprehensive indicator includes both the degree of anomaly that has already occurred in the device and the expected trend of potential future deterioration, providing a quantitative decision-making basis for forward-looking risk warnings.

[0147] Among them, the fusion state deviation refers to a quantitative indicator that comprehensively reflects the current degree of abnormality and the future deterioration trend after the equipment state deviation and the predicted state deviation are weighted and fused.

[0148] S205. Based on the deviation of the fusion state, the risk level and handling suggestions in the risk guidance information are forward-lookingly revised to generate forward-looking risk guidance information that includes future risk warnings.

[0149] Specifically, the deviation from the fusion state is used as input and remapped to a preset risk level range to generate a forward-looking risk level that integrates the current state and future expectations. Based on this forward-looking risk level, the risk level labels in the original risk guidance information are replaced or supplemented, for example, by adding labels such as "expected upward trend" or "future high-risk warning" next to the original level. At the same time, a preset risk management knowledge base is invoked, and the management suggestions in the original risk guidance information are forward-lookingly revised according to the future deterioration trend represented by the deviation from the fusion state: for risks that are about to worsen, supplementary suggestions such as "it is recommended to activate backup equipment in advance" or "it is recommended to complete preventive maintenance within the current window" are added; for risks that are currently abnormal but tend to ease in the future, the suggestion is adjusted to "continuous monitoring, no active intervention is required for now". After the revision is completed, the updated risk level and management suggestions are integrated and output to generate forward-looking risk guidance information that includes future risk warnings.

[0150] Forward-looking risk guidance information refers to risk guidance information that has been forward-lookingly revised and includes warnings of future risks and suggestions for proactive measures.

[0151] Through the above embodiments, risk guidance information is proactively corrected by integrating state deviation, and future risk expectations are explicitly incorporated into current risk warnings. This enables maintenance personnel to not only understand the current state risks of the equipment, but also to know the evolution trend of risks in advance. This allows the timing of handling to be advanced from "post-event response" to "pre-event intervention", effectively avoiding equipment failures and unplanned downtime caused by unforeseen risk deterioration. This represents a substantial leap in risk management from passive response to proactive early warning.

[0152] S206. Providing forward-looking risk guidance information at the same time as or before providing risk guidance information.

[0153] Specifically, the system determines the output sequence of forward-looking risk guidance information relative to existing risk guidance information based on the urgency of the information and the proximity of the forecast time window. For forward-looking warnings with short forecast time windows and rapid risk deterioration, a priority output strategy is adopted. The forward-looking risk guidance information is immediately pushed to the user interface after the current risk guidance information is generated but before its official output, and prominently displayed with a label such as "High-Risk Warning for the Next 1 Hour" to ensure users are aware of the impending risk immediately. For forward-looking warnings with longer forecast time windows or relatively mild risk deterioration trends, the system adopts a "synchronous output" strategy, integrating forward-looking risk guidance information with current risk guidance information. This is displayed in sections on the same interface or presented side-by-side via a timeline; for example, the left side displays the current risk list, and the right side displays a future risk prediction dashboard. Regardless of the output sequence, the system ensures that forward-looking risk guidance information includes clear forecast time points, expected risk levels, and proactive handling suggestions, providing users with sufficient advance warning time.

[0154] Based on the above embodiments, as an optional embodiment, for Figure 1 The risk management methods for new energy projects shown may also include Figure 2 Steps S301-S304 are explained in detail below.

[0155] S301. Obtain user feedback data on risk guidance information. The feedback data includes at least one of the following: adoption flag, modification record, or rejection instruction.

[0156] Specifically, by monitoring interactive events on the user interface, the system captures and records user feedback actions in response to the risk guidance information in real time. Based on the type of action, three core feedback data categories are collected: first, adoption flags—when a user fully follows the advice in the risk guidance information and clicks the "Accept" or "Confirm" button on the interface, the adoption flag and adoption time are recorded; second, modification records—when a user edits, adjusts, or supplements the system's recommended actions and submits them, the differences in content before and after modification, the modification time, and the final version are captured; third, rejection instructions—when a user determines that the system's recommended risk guidance information does not match the actual situation and clicks the "Reject" or "Not Applicable" button, the rejection instruction and the possible rejection reason provided by the user are recorded. All feedback action data are associated and bound to the original query request that generated the guidance, real-time operating parameters, and the target candidate risk item set, forming a complete feedback sample record.

[0157] Feedback operation data refers to the digital record of a user's response to the risk guidance information output by the system. Adoption markers are positive feedback records generated when a user fully accepts and confirms that they are following the system's recommendations. Modification records are records of changes made by users when they edit, adjust, or supplement the system's recommended actions and submit them. Rejection instructions are negative feedback records generated when a user determines that the system's recommendations do not match the actual situation and refuses to adopt them.

[0158] S302. Associate and store the feedback operation data with the corresponding query request, real-time running parameters, and target candidate risk item set to build a feedback sample library.

[0159] Specifically, a unique feedback sample identifier is generated for this interaction. Then, the collected feedback operation data is associated and bound with the original input data that generated this risk guidance, including: the user query request that triggered this interaction, the real-time operating parameters of the target device used to calculate the device status, and the target candidate risk item set on which the final risk guidance information is based. These data, which were originally scattered in different modules, are organized according to a unified sample structure to form a complete feedback sample record. Subsequently, this sample record is written into a specially constructed feedback sample library. This library is a dedicated database for long-term accumulation and storage of user feedback data, supporting retrieval and statistical analysis according to multiple dimensions such as time, device, user, and risk type. Finally, the feedback sample library accumulates a large amount of "system recommendation-user feedback" pairing data, providing real-world validation training material for subsequent model optimization and dynamic updates of the knowledge base.

[0160] The feedback sample library refers to a structured database used for long-term accumulation and storage of user feedback samples. Each sample contains complete interaction context and user response information.

[0161] Through the above embodiments, a complete feedback sample library was constructed, so that each user feedback is not an isolated event, but a reusable data asset deeply bound to the complete technical context in which the feedback was generated. This provides a structured and high-quality data foundation for subsequent mining of optimization clues from feedback samples and periodic revision of the knowledge base.

[0162] S303. Periodically extract feedback samples that meet preset screening conditions from the feedback sample library. The preset screening conditions include an adoption rate lower than a preset adoption threshold or a number of modifications higher than a preset number of modifications threshold.

[0163] Specifically, the system automatically triggers a screening task according to a preset cycle (e.g., weekly or monthly), accesses the feedback sample library, and iterates through all feedback samples added in the most recent cycle. For each feedback sample, it calculates the proportion of risk guidance information adopted by users (adoption rate) and the number of times it has been modified by users. These statistical indicators are then compared with pre-configured screening thresholds: if the adoption rate of risk guidance information corresponding to a sample is lower than the preset adoption threshold (e.g., lower than 60%), it indicates that the system's recommendation quality in this scenario has not been recognized by users; or if the number of times risk guidance information corresponding to a sample has been modified by users is higher than the preset modification threshold (e.g., more than 3 times), it indicates that the system's recommended handling suggestions deviate significantly from the user's actual needs. Feedback samples meeting either of these conditions are considered "samples to be optimized" requiring close attention. All feedback samples meeting the screening criteria are extracted to form a subset of feedback samples to be processed, for use in subsequent knowledge base correction processes.

[0164] The preset screening criteria refer to pre-configured rules used to identify samples from the feedback sample library that need optimization and improvement. The adoption rate refers to the proportion of users who fully accept and follow the system's recommendations, usually expressed as a percentage. The preset adoption threshold is a pre-set lower limit for the adoption rate; samples below this value are considered to have poor recommendation performance. The number of modifications refers to the number of times users edit or adjust the system's recommended actions. The preset modification threshold is a pre-set upper limit for the number of modifications; samples exceeding this value are considered to deviate significantly from user needs.

[0165] S304. Based on feedback samples, all candidate risk entries in the preset risk knowledge base are corrected to obtain a dynamically optimized preset risk knowledge base. The correction includes at least one of updating the correlation deviation range of candidate risk entries, adjusting the benchmark risk level, or supplementing new candidate risk entries.

[0166] Specifically, each feedback sample undergoes in-depth analysis, analyzing the differences in content within user modification records or the reasons for rejection in rejection instructions. This identifies deviation patterns between existing risk items and user expectations. Based on these analysis results, at least one type of correction operation is performed on candidate risk items in the pre-set risk knowledge base: First, updating the correlation deviation range of candidate risk items. For example, if a user frequently rejects a recommendation for an item, and analysis reveals that the actual applicable deviation range for that item does not match the original configuration, its applicable deviation range is automatically adjusted. Second, adjusting the baseline risk level of candidate risk items. For example, if a user repeatedly modifies the handling suggestion for a risk item from "low risk" to a more severe handling measure, the baseline risk level of that item is adjusted accordingly. Third, supplementing new candidate risk items. For example, when entirely new risk patterns or handling solutions are identified in user modification records, they are structured and added to the knowledge base as new risk items. After all correction operations are completed, a dynamically optimized pre-set risk knowledge base is output. This knowledge base becomes more closely aligned with users' actual needs and the real-world situation after each iteration.

[0167] Through the above embodiments, the risk knowledge base is corrected in multiple dimensions using real user feedback, realizing the continuous evolution and dynamic optimization of the knowledge base. This can gradually eliminate deviations from the actual situation on site, continuously improve the accuracy of risk identification and the practicality of recommendation results, and form a virtuous cycle of "use-feedback-optimization".

[0168] The risk management system for new energy projects in the embodiments of this invention is described below from the perspective of hardware processing. Please refer to [link / reference needed]. Figure 3 , Figure 3 This is a schematic diagram of the structure of a new energy project risk management system provided in an embodiment of this application.

[0169] It should be noted that, Figure 3 The structure of the new energy project risk management system shown is merely an example and should not impose any limitations on the functionality and scope of application of the embodiments in this application.

[0170] like Figure 3 As shown, the new energy project risk management system includes a central processing unit 401, which can perform various appropriate actions and processes based on programs stored in read-only memory 402 or programs loaded from storage section 408 into random access memory 403, such as executing the methods described in the above embodiments. The random access memory 403 also stores various programs and data required for system operation. The central processing unit 401, read-only memory 402, and random access memory 403 are interconnected via bus 404. Input / output interface 405 is also connected to bus 404.

[0171] The following components are connected to the input / output interface 405: an input section 406 including audio input devices, push-button switches, etc.; an output section 407 including a liquid crystal display (LCD) and audio output devices, indicator lights, etc.; a storage section 408 including a hard disk, etc.; and a communication section 409 including a network interface card such as a LAN (Local Area Network) card, modem, etc. The communication section 409 performs communication processing via a network such as the Internet. A drive 410 is also connected to the input / output interface 405 as needed. A removable medium 411, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 410 as needed so that computer programs read from it can be installed into the storage section 408 as needed.

[0172] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing a computer program for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 409, and / or installed from removable medium 411. When the computer program is executed by central processing unit 401, it performs the various functions defined in the present invention. It should be noted that specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0173] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. Each block in a flowchart or block diagram may represent a module, program segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those shown in the drawings.

[0174] Specifically, the new energy project risk management system of this embodiment includes a processor and a memory. The memory stores a computer program. When the computer program is executed by the processor, it implements the new energy project risk management method provided in the above embodiment.

[0175] In another aspect, the present invention also provides a computer-readable storage medium, which may be included in the new energy project risk management system described in the above embodiments; or it may exist independently and not assembled into the new energy project risk management system. The storage medium carries one or more computer programs, which, when executed by a processor of the new energy project risk management system, enable the new energy project risk management system to implement the new energy project risk management method provided in the above embodiments.

[0176] The above description is merely an exemplary embodiment of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure. The specification and embodiments are considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

Claims

1. A risk management method for new energy projects, characterized in that, include: Obtain real-time operating parameters of each device in new energy projects; In response to a user-input query request, the query request is semantically parsed to determine the intent type of the query request and at least one query entity, the query entity including at least one of equipment name, operating condition description or risk type; Based on the query entity, at least one target device corresponding to the query entity is determined, and target operating parameters are determined from the real-time operating parameters corresponding to the target device, wherein the target operating parameters include at least one parameter dimension; Based on the target operating parameters, obtain the historical operating parameters corresponding to each of the parameter dimensions in the target operating parameters; Based on the target operating parameters and the historical operating parameters, calculate the equipment state deviation. The query request is subjected to feature calculation to obtain a text feature vector, and the deviation between the target operating parameters and the device state is subjected to feature calculation to obtain a device state feature vector. Based on the text feature vector and the device status feature vector, a fusion retrieval vector is generated; The fusion retrieval vector is input into a preset risk knowledge base for similarity retrieval to determine at least one candidate risk item that matches the fusion retrieval vector. Based on the deviation between the intent type and the device state, each candidate risk item is filtered to obtain a target candidate risk item set, and risk guidance information is generated based on the target candidate risk item set through a preset risk identification model.

2. The method according to claim 1, characterized in that, The calculation of equipment state deviation based on the target operating parameters and the historical operating parameters includes: Extract the time-series data of the target parameter dimension within a preset historical time window from the historical operating parameters, wherein the target parameter dimension is any of the parameter dimensions; Based on the time series data of the parameters, calculate the historical mean and historical standard deviation of the target parameter dimension; Calculate the difference between the real-time operating parameters and the historical average values ​​of the target parameter dimension; Calculate the ratio of the difference to the historical standard deviation to obtain the sub-state deviation corresponding to the target parameter dimension; The device state deviation is obtained by weighting all the sub-state deviations based on a preset weight set.

3. The method according to claim 1, characterized in that, The step of generating a fused retrieval vector based on the text feature vector and the device status feature vector includes: The text feature vector and the device state feature vector are input into a pre-trained attention layer; The attention layer calculates the first attention weight of the text feature vector to the device state feature vector, and the second attention weight of the device state feature vector to the text feature vector. Based on the first attention weight and the second attention weight, the text feature vector and the device state feature vector are weighted and calculated to generate a fusion retrieval vector.

4. The method according to claim 1, characterized in that, The step involves filtering each candidate risk item based on the deviation between the intent type and the device state to obtain a target candidate risk item set, and generating risk guidance information based on the target candidate risk item set using a preset risk identification model, including: The candidate risk entries that match the deviation range of the device state deviation are taken as intermediate candidate risk entries to obtain an intermediate candidate risk entry set, and the preset baseline risk level of each intermediate candidate risk entry is obtained. The deviation of the equipment status is matched with multiple preset deviation threshold ranges to determine the risk level adjustment coefficient corresponding to the deviation of the equipment status; The baseline risk level of each intermediate candidate risk item is adjusted based on the risk level adjustment coefficient to generate an intermediate risk level corresponding to each intermediate candidate risk item. Real-time micro-meteorological data and operation duration data are obtained based on the operation scenario anchor point corresponding to the target equipment, and the intermediate risk levels are dynamically adjusted based on the real-time micro-meteorological data and operation duration data to obtain the dynamic risk level of each intermediate candidate risk item. Based on the intent type, a set of target candidate risk items matching the intent type is selected from the intermediate candidate risk item set, and risk guidance information matching the intent type is generated based on the dynamic risk level of each target candidate risk item in the target candidate risk item set.

5. The method according to claim 4, characterized in that, The process involves acquiring real-time micro-meteorological data and operation duration data based on the anchor point of the operation scenario corresponding to the target equipment, and dynamically adjusting each intermediate risk level based on the real-time micro-meteorological data and operation duration data to obtain the dynamic risk level of each intermediate candidate risk item, including: Real-time micro-meteorological data of the working area of ​​the new energy project where the target equipment is located is obtained based on the work scenario anchor point. The real-time micro-meteorological data includes at least one of wind speed time series data, ultraviolet sun exposure index and surface ice humidity. Obtain the current work duration data of the workers, and based on the work duration data and the preset duty cycle of heavy physical labor operations, deduce the physiological metabolic fatigue index; The ecological stress gradient value, which characterizes the trend of environmental degradation, is calculated based on the real-time micrometeorological data. The ecological stress gradient value and the physiological metabolic fatigue index are combined and transformed to generate a comprehensive spatiotemporal decay factor. The dynamic risk level of each intermediate candidate risk item is obtained by nonlinearly and dynamically adjusting each intermediate risk level using the comprehensive spatiotemporal attenuation factor.

6. The method according to claim 4, characterized in that, The intent types include risk query intent and risk question-and-answer intent. Based on the intent type, a set of target candidate risk items matching the intent type is selected from the intermediate candidate risk item set. Based on the dynamic risk level of each target candidate risk item in the target candidate risk item set, risk guidance information matching the intent type is generated, including: When the intent type is the risk query intent, the intermediate candidate risk entries belonging to the historical risk data are filtered from the intermediate candidate risk entry set to form the target candidate risk entry set, and the historical risk data is associated with historical handling measures. Extract the historical handling measures and dynamic risk levels associated with each target candidate risk item in the target candidate risk item set, and generate a risk operation guidance list that is sorted from high to low according to the dynamic risk level to obtain risk guidance information; When the intent type is the risk question and answer intent, the intermediate candidate risk items belonging to the normative clauses are selected from the intermediate candidate risk item set to form the target candidate risk item set, and the normative clauses are associated with compliance basis information; Extract the compliance basis information and dynamic risk level associated with each target candidate risk item in the target candidate risk item set, generate a compliance analysis report list that is sorted from high to low according to the dynamic risk level, and obtain the risk guidance information.

7. The method according to claim 1, characterized in that, The method further includes: Obtain historical trend data of the historical operating parameters within a preset historical time window; Based on the historical trend data, a preset trend prediction model is used to generate predicted operating parameters for the target operating parameters at a future preset time point. Based on the predicted operating parameters and the historical operating parameters, the deviation of the predicted state is calculated; When the predicted state deviation exceeds the preset warning threshold, the predicted state deviation and the device state deviation are weighted and fused to generate a fused state deviation. Based on the deviation of the fusion state, the risk level and handling suggestions in the risk guidance information are forward-lookingly revised to generate forward-looking risk guidance information that includes future risk warnings; The forward-looking risk guidance information is output at the same time as or before the output of the risk guidance information.

8. A risk management and control system for new energy projects, characterized in that, The new energy project risk management system 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 new energy project risk management system to perform the method as described in any one of claims 1-7.

9. A computer-readable storage medium comprising instructions, characterized in that, When the instruction is run on the new energy project risk management system, the new energy project risk management system performs the method as described in any one of claims 1-7.

10. A computer program product, characterized in that, When the computer program product is run on the new energy project risk management system, the new energy project risk management system performs the method as described in any one of claims 1-7.

Images