Power generation system fault auxiliary decision generation method and system based on large model

Abstract

The application relates to the technical field of power systems, in particular to a power generation system fault auxiliary decision generation method and system based on a large model. The method and system successively acquire current detection information set of a power generation system, pre-process the current detection information set to generate a time sequence data set, generate first prediction information and a residual sequence based on a preset ARIMA model according to the time sequence data set, input the residual sequence into a preset LSTM network to generate second prediction information, combine the first prediction information and the second prediction information to generate preliminary fault information, input the preliminary fault detection information into a pre-trained Transformer model to perform deep feature extraction and context correlation analysis and generate auxiliary decision information, and send the auxiliary decision information to managers, so that the accuracy of power generation system fault detection is improved, the intelligent level of auxiliary decision is enhanced, and the operation stability and management efficiency of the system are significantly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power systems, in particular to a power generation system fault auxiliary decision generation method and system based on a large model. BACKGROUND

[0002] With the continuous expansion of the scale of photovoltaic power generation systems and the increasing complexity of operating environments, the stability and reliability of photovoltaic systems are particularly important. In the actual operation process of photovoltaic power generation systems, timely detection and accurate positioning of faults are of key significance to ensure the continuity of power supply and the safety of the system. However, the existing fault detection and decision support schemes have the following problems in photovoltaic power generation systems.

[0003] Firstly, the existing technology lacks the ability to process data and extract features. Traditional photovoltaic fault detection methods often rely on simple statistical analysis or rule-based discrimination, which is difficult to effectively process a large amount of real-time monitoring data from photovoltaic systems. These methods have limited ability to extract complex time series features and capture potential patterns in data, resulting in low accuracy and reliability of fault prediction.

[0004] Secondly, the existing decision support system is imperfect. Even if the existing method can detect faults, the existing auxiliary decision system often lacks the ability of deep feature extraction and context association analysis when generating maintenance strategies and processing recommendations. This results in the lack of pertinence and effectiveness of the generated decision information, which cannot provide reliable decision support for managers, thereby affecting the efficiency and effectiveness of fault handling.

[0005] Therefore, it is necessary to improve the existing technology to solve the above problems. SUMMARY

[0006] Therefore, it is necessary to improve the existing technology to solve the above problems.

[0007] The present application provides a power generation system fault auxiliary decision generation method based on a large model, comprising:

[0008] Obtaining a current detection information set of a power generation system, and preprocessing the current detection information set to generate a time series data set;

[0009] Generating first prediction information and a residual sequence based on a preset ARIMA model according to the time series data set;

[0010] Inputting the residual sequence into a preset LSTM network to generate second prediction information;

[0011] generate preliminary fault information according to the first prediction information and the second prediction information;

[0012] input the preliminary fault detection information into a pre-trained Transformer model for deep feature extraction and context association analysis and generate auxiliary decision information, and send the auxiliary decision information to the management personnel.

[0013] As a preferred solution, the first prediction information and the residual sequence are generated based on a preset ARIMA model according to the time series data set, comprising:

[0014] select target parameters according to a preset historical fault database and an expert experience library;

[0015] acquire a running parameter sequence related to the target parameters by traversing the time series data set, and generate a target running parameter sequence according to each related running parameter sequence;

[0016] align and combine each target running parameter sequence according to a time stamp to generate a target sequence, and input the target sequence into a preset ARIMA model to generate first prediction information;

[0017] perform prediction error calculation according to the target sequence and the first prediction information, and generate a residual sequence.

[0018] As a preferred solution, the residual sequence is input into a preset LSTM network to generate second prediction information, comprising:

[0019] extract time-dependent features from the residual sequence by a sliding window algorithm, and generate an input feature sequence based on a set time step according to the residual sequence;

[0020] perform feature extraction on the input feature sequence to generate statistical features, periodic features and frequency domain features, fuse the time-dependent features, the statistical features, the periodic features and the frequency domain features, and generate an enhanced feature sequence;

[0021] input the enhanced feature sequence into a preset LSTM network for deep modeling to generate first nonlinear prediction information;

[0022] perform short-term and medium-term nonlinear prediction and trend analysis on the first nonlinear prediction information based on a multi-step prediction algorithm to generate second nonlinear prediction information;

[0023] filter and correct the second nonlinear prediction information based on the expert experience library, and generate second prediction information.

[0024] As a preferred solution, preliminary fault information is generated according to the first prediction information and the second prediction information, comprising:

[0025] generating an aligned prediction information set according to the timestamps of the first prediction information and the second prediction information being aligned;

[0026] performing a difference evaluation analysis on the first prediction information and the second prediction information in the same time interval based on a mutual information metric, and generating a difference analysis index;

[0027] generating a weight parameter according to the difference analysis index, and generating comprehensive prediction information according to the first prediction information, the second prediction information and the weight parameter;

[0028] generating fault matching information based on a pattern matching algorithm according to the comprehensive prediction information, the historical fault database and the expert experience library, wherein the fault matching information includes a historical fault identification and an expert rule identification;

[0029] generating preliminary fault information based on a Naive Bayes algorithm according to the fault matching information, wherein the preliminary fault information includes a historical fault matching identification, a fault type, a fault probability and a fault severity.

[0030] As a preferred scheme, the preliminary fault detection information is input into a pre-trained Transformer model for deep feature extraction and context correlation analysis to generate auxiliary decision information, and the auxiliary decision information is sent to the management personnel, including:

[0031] retrieving corresponding fault case text information in the historical fault database according to the historical fault matching identification, and querying corresponding processing suggestion text in the expert experience library according to the expert rule identification, and performing vectorization representation on the fault case text information and the processing suggestion text based on a text encoding algorithm to generate context embedding data;

[0032] performing feature splicing and weighted integration according to the preliminary fault detection information and the context embedding data to generate fusion feature representation;

[0033] inputting the fusion feature representation into a pre-trained Transformer model, performing deep feature extraction and context correlation analysis on the fusion feature through a multi-head attention mechanism, and generating deep correlation features;

[0034] generating auxiliary decision candidate information through strategy inference according to the deep correlation features and the expert experience library;

[0035] performing result formatting and preferred screening on the auxiliary decision candidate information to generate auxiliary decision information, wherein the auxiliary decision information includes executable maintenance strategies, maintenance opportunity suggestions and resource allocation schemes.

[0036] The auxiliary decision-making information is sent to the management personnel and feedback information is obtained, and the feedback information is stored in the historical fault database and the expert experience database.

[0037] As a preferred solution, the ARIMA model is established based on the following steps:

[0038] A set of routine detection information of the power generation system is obtained, a routine time series data set is generated according to the set of routine detection information, and the routine time series data set is subjected to multi-granularity partitioning to generate a modeling sequence and a verification sequence;

[0039] A plurality of intrinsic mode components are generated by decomposing the modeling sequence by means of variational mode decomposition, a corresponding initial ARIMA sub-model is established according to each intrinsic mode component, and an ARIMA sub-model is generated by iteratively searching and updating the parameters of the initial ARIMA sub-model by means of a Bayesian optimization algorithm;

[0040] A first prediction result is generated by performing a rolling prediction on the verification sequence based on the ARIMA sub-model using a fixed-length sliding window algorithm;

[0041] A residual sequence is generated by comparing the first prediction result with the true value of the verification sequence and refining the residual, and an optimized ARIMA sub-model is generated by optimizing the ARIMA sub-model according to the residual sequence;

[0042] The performance of the optimized ARIMA sub-model is quantified according to the residual sequence, and an error value is generated, which is obtained based on the following formula:

[0043]

[0044] wherein E is the error value, N is the number of time steps of the verification sequence, t is the ordinal number of the time step, c t is the weight factor corresponding to the time step t, y t is the true value of the verification sequence at the time step t, y′ t is the model prediction value of the verification sequence at the time step t, and ε is a predetermined positive constant;

[0045] A sub-model weight parameter is generated according to the error value, the historical fault database and the expert experience database, and the ARIMA model is constructed according to the ARIMA sub-model and the sub-model weight parameter.

[0046] As a preferred solution, a set of current detection information of the power generation system is obtained, and a time series data set is generated by preprocessing the set of current detection information, including:

[0047] Acquire a current detection information set of a power generation system, perform data fusion preprocessing on the current detection information set to generate a fusion information set, wherein the detection information set includes environmental detection information, system operation information and state detection information;

[0048] Adopt a locally weighted regression interpolation algorithm on the fusion information set to perform adaptive data completion on missing values, and generate a completed information set, fit and correct the completed information set through a random forest algorithm, and generate a corrected information set;

[0049] Perform abnormal screening on the corrected information set based on a z-score method to generate an abnormal value set, and perform abnormal screening on the abnormal value set through an isolation forest algorithm to generate a confirmed abnormal value set, and generate a screening information set based on the locally weighted regression interpolation algorithm according to the completed detection information set and the confirmed abnormal value set;

[0050] Acquire time series data of each detection feature in the screening information set, calculate the median and median absolute deviation value of each feature, and perform normalization processing on the screening information set according to the absolute deviation value to generate a normalized information set;

[0051] Perform standardization on the normalized information set based on a z-score method to generate a standard information set, and perform feature engineering processing on the standard information set to generate a feature extraction information set, wherein the feature extraction information set includes statistical features, periodic features and frequency domain features.

[0052] The application also provides a power generation system fault auxiliary decision generation system based on a large model, comprising:

[0053] A time series data generation module is configured to acquire a current detection information set of a power generation system, and perform preprocessing on the current detection information set to generate a time series data set;

[0054] A first prediction information generation module is configured to generate first prediction information and a residual sequence based on a preset ARIMA model according to the time series data set;

[0055] A second prediction information generation module is configured to input the residual sequence into a preset LSTM network to generate second prediction information;

[0056] A preliminary fault information generation module is configured to generate preliminary fault information by combining the first prediction information and the second prediction information;

[0057] An auxiliary decision information generation module is configured to input the preliminary fault detection information into a pre-trained Transformer model to perform deep feature extraction and context association analysis and generate auxiliary decision information, and send the auxiliary decision information to a manager.

[0058] The application further provides a computer device comprising a memory and a processor, the memory storing a computer program, and the processor implementing the steps of the above-mentioned method for generating an auxiliary decision of a power generation system fault based on a large model when executing the computer program.

[0059] The application further provides a computer readable storage medium having a computer program stored thereon, the computer program implementing the steps of the method for generating an auxiliary decision of a power generation system fault based on a large model when executed by a processor.

[0060] The application achieves the following technical effects:

[0061] The above-mentioned method and system for generating an auxiliary decision of a power generation system fault based on a large model sequentially acquire a current detection information set of the power generation system, pre-process the current detection information set to generate a time series data set, generate first prediction information and a residual sequence based on a preset ARIMA model according to the time series data set, input the residual sequence into a preset LSTM network to generate second prediction information, combine the first prediction information and the second prediction information to generate preliminary fault information, input the preliminary fault detection information into a pre-trained Transformer model to perform deep feature extraction and context association analysis and generate auxiliary decision information, and send the auxiliary decision information to a manager, thereby improving the accuracy of power generation system fault detection, enhancing the intelligent level of auxiliary decision, significantly improving the operation stability and management efficiency of the system, and solving the key problems of low prediction accuracy and limited decision support in the prior art. BRIEF DESCRIPTION OF DRAWINGS

[0062] Figure 1 FIG. 1 is a flowchart of a method for generating an auxiliary decision of a power generation system fault based on a large model according to an embodiment of the application;

[0063] Figure 2 FIG. 2 is a structural block diagram of a system for generating an auxiliary decision of a power generation system fault based on a large model according to an embodiment of the application. DETAILED DESCRIPTION

[0064] In order to make the purposes, technical solutions and advantages of the application clearer, the application is further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the application and do not limit the application.

[0065] In one embodiment, a terminal is provided, which acquires a current detection information set of a power generation system, pre-processes the current detection information set to generate a time series data set, generates first prediction information and a residual sequence based on a preset ARIMA model according to the time series data set, inputs the residual sequence into a preset LSTM network to generate second prediction information, combines the first prediction information and the second prediction information to generate preliminary fault information, inputs the preliminary fault detection information into a pre-trained Transformer model for deep feature extraction and context association analysis and generates auxiliary decision information, and sends the auxiliary decision information to a management personnel.

[0066] The terminal can be, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers and portable wearable devices.

[0067] In this embodiment, as shown in the accompanying drawings, a large model-based power generation system fault auxiliary decision generation method is provided, which is especially suitable for a photovoltaic power generation system, and the method comprises the following steps: Figure 1

[0068] Step S100: Acquire a current detection information set of a power generation system, and pre-process the current detection information set to generate a time series data set.

[0069] In this step, the photovoltaic power generation system monitors the photovoltaic components, inverters, transformers, junction boxes and other equipment in real time, collects the operating parameters such as voltage, current, power, irradiance, temperature and frequency, collects and summarizes these parameters into a current detection information set to reflect the operating state and environmental conditions of the photovoltaic system at different time points. After pre-processing the current detection information set, it is organized in chronological order to form a time series data set suitable for time series analysis.

[0070] Step S200: Generate first prediction information and a residual sequence based on a preset ARIMA model according to the time series data set.

[0071] In this step, a pre-configured ARIMA (Autoregressive Integrated Moving Average) model is used to model and analyze the time series data set. The ARIMA model identifies and captures the time dependence and trend in the operating parameters of the photovoltaic power generation system, generates first prediction information, and predicts the operating state of the photovoltaic power generation system at future time points. At the same time, the difference between the actual monitoring data and the ARIMA model prediction value is calculated to form a residual sequence.

[0072] Step S300: Input the residual sequence into a preset LSTM network to generate second prediction information.

[0073] ​In this step, the LSTM (Long Short-Term Memory) network, a deep learning model capable of capturing long-term dependencies in time series, is suitable for processing and predicting data with complex temporal characteristics. The residual sequence is fed as input to the pre-configured LSTM network. By analyzing the residual sequence, the LSTM network can identify and learn nonlinear patterns and underlying trends that the ARIMA model failed to capture, thereby generating more accurate secondary prediction information.

[0074] Step S400: Generate preliminary fault information by combining the first prediction information and the second prediction information.

[0075] In this step, the first prediction information generated based on the ARIMA model and the second prediction information generated based on the LSTM network are comprehensively analyzed and fused. By combining the results of these two prediction methods, the advantages of the ARIMA model in capturing linear trends and time dependencies, as well as the ability of the LSTM network in identifying nonlinear patterns and long-term dependencies, can be fully utilized to improve the accuracy and reliability of fault prediction.

[0076] Step S500: Input the preliminary fault detection information into the pre-trained Transformer model for deep feature extraction and contextual analysis to generate auxiliary decision information, and send the auxiliary decision information to the management personnel.

[0077] In this step, the self-attention mechanism of the Transformer model is used to deeply extract key features from the initial fault information and analyze their temporal and spatial contextual relationships. Through deep feature extraction and contextual analysis, the model can accurately diagnose the cause of the fault, predict its development trend, and assess its potential impact on the entire photovoltaic power generation system. Based on these analytical results, the Transformer model is used to generate detailed auxiliary decision-making information, including fault type, occurrence time, impact range, emergency handling suggestions, and preventative maintenance measures. This auxiliary decision-making information is promptly sent to relevant management personnel through appropriate communication channels (such as management system interfaces, mobile application notifications, or email) to ensure they can quickly and accurately understand the system's operating status and take necessary countermeasures, thereby improving the overall operating efficiency and reliability of the photovoltaic power generation system.

[0078] In this embodiment, through the integration of data collection, preprocessing, multi-model prediction fusion and deep learning analysis, efficient and accurate fault prediction and decision support is achieved. First, the key equipment of the photovoltaic power generation system is monitored in real time and the operating parameters are collected. These data are organized into time series data sets for further analysis. The time dependence and trend in the operating parameters of the photovoltaic power generation system are identified and captured by the ARIMA model. Then, the LSTM network is used to identify and learn the nonlinear patterns and potential trends that the ARIMA model fails to capture. By combining the prediction information of the ARIMA and LSTM models, the accuracy and reliability of fault prediction are significantly improved. Then, the preliminary fault detection information is extracted and analyzed by the Transformer model to accurately diagnose the fault cause, predict the fault development trend, and evaluate its potential impact on the entire photovoltaic power generation system. Based on these deep analysis results, the system generates detailed decision support information, including fault type, occurrence time, impact range, emergency handling suggestions and preventive maintenance measures, etc. These decision support information is timely transmitted to relevant management personnel, so that the management personnel can accurately understand the fault status of the photovoltaic power generation system, and then send appropriate maintenance personnel with corresponding tools and materials to handle the fault, avoiding the problem of low fault handling rate caused by multiple trips of personnel due to insufficient personnel capacity and maintenance tools and materials.

[0079] In one embodiment, the step S100 of acquiring a current detection information set of the power generation system includes:

[0080] Step S110: Acquire a current detection information set of the power generation system, and perform data fusion preprocessing on the current detection information set to generate a fusion information set, wherein the detection information set includes environmental detection information, system operation information and state detection information.

[0081] In this step, first, the environmental detection information, system operation information and state detection information of the photovoltaic power generation system are collected. These detection information covers various parameters required for system operation, such as environmental conditions, equipment operating status, etc. Data fusion technology is used to preprocess these multi-source data, eliminate redundancy, improve data integrity and consistency, and generate a comprehensive fusion information set.

[0082] Step S120: Perform adaptive data completion on the missing values in the fusion information set using a locally weighted regression interpolation algorithm, and generate a completed information set. The completed information set is fitted and corrected by a random forest algorithm, and a corrected information set is generated.

[0083] Step S130: Based on the correction information set, abnormal screening is performed based on the z-score method to generate an abnormal value set, the abnormal value set is screened by the isolation forest algorithm to generate a confirmed abnormal value set, and based on the local weighted regression interpolation algorithm, the screening information set is generated based on the completion detection information set and the confirmed abnormal value set.

[0084] In this step, based on the z-score method, the deviation of the data point from the mean value is calculated to identify the abnormal values in the correction information set that deviate significantly from the normal range, and these abnormal values are combined to form an abnormal value set. The isolation forest algorithm is used to further confirm the preliminary screened abnormal value set to generate the final confirmed abnormal value set. The screening information set is generated by combining the completion detection information set and the confirmed abnormal value set using the local weighted regression interpolation algorithm to ensure the accuracy and integrity of the data.

[0085] Step S140: Obtain the time series data of each detection feature in the screening information set, calculate the median and median absolute deviation value of each feature, and normalize the screening information set based on the absolute deviation value to generate a normalized information set.

[0086] Step S150: Based on the z-score method, the normalized information set is standardized to generate a standard information set, and the standard information set is subjected to feature engineering processing to generate a feature extraction information set, wherein the feature extraction information set includes statistical features, periodic features and frequency domain features.

[0087] In this step, the z-score method is used to standardize the normalized information set to further improve the consistency and comparability of the data. The standard information set is subjected to feature engineering processing to generate a feature extraction information set. Through the extracted statistical features, periodic features and frequency domain features, important patterns and rules in the data can be captured to provide feature support for subsequent fault prediction and analysis.

[0088] In this embodiment, the environmental detection information, system operation information and state detection information are integrated into a comprehensive fusion information set through data fusion preprocessing, eliminating redundancy and improving data integrity and consistency. Then, the locally weighted regression interpolation algorithm is used to adaptively complete the missing values, and the random forest algorithm is used to fit and correct the completed data, ensuring the accuracy of the data. Subsequently, the z-score method and the isolation forest algorithm are used to screen the outliers, generate a confirmed outlier set, and generate a screening information set combined with the completion information, thereby eliminating noise data and improving the reliability of the data. Then, the median and median absolute deviation value of each detection feature are calculated to normalize the screening information set, eliminate the dimensional differences between different features, and ensure that the data are compared on the same scale. Finally, the z-score method is used to standardize the normalized information set, and feature engineering processing is performed to extract statistical features, periodic features and frequency domain features, fully capturing important patterns and rules in the data. Through the above data preprocessing, the integrity, accuracy and consistency of the data are improved, the influence of outliers and noise is reduced, and the feature expression ability is enhanced.

[0089] In one embodiment, the step S200 of generating first prediction information and residual sequence based on a preset ARIMA model according to the time series data set comprises:

[0090] Step S210: selecting target parameters according to a preset historical fault database and expert experience library.

[0091] In this step, according to the operation characteristics and failure modes of the photovoltaic power generation system, as well as the historical fault database and expert experience library, the operation parameters that need to be monitored are determined. For example, voltage and current abnormalities may have a high possibility of causing system failure, so the voltage parameter and the current parameter are selected as the target parameters to improve the pertinence and efficiency of fault detection and concentrate resources on monitoring the parameters that have the greatest impact on the system.

[0092] Step S220: traversing the time series data set to obtain operation parameter sequences related to the target parameters, and generating target operation parameter sequences according to each related operation parameter sequence.

[0093] In this step, there may be correlations between different operation parameters. By traversing the entire time series data set, the operation parameter sequences related to the selected target parameters are extracted and the target operation parameter sequences specifically used for prediction are generated, which helps to build a more comprehensive prediction model and improve the accuracy and reliability of the prediction.

[0094] Step S230: aligning and combining each target operation parameter sequence according to the time stamp to generate a target sequence, and inputting the target sequence into a preset ARIMA model to generate first prediction information.

[0095] Step S240: performing prediction error calculation according to the target sequence and the first prediction information, and generating a residual sequence.

[0096] In this embodiment, the accuracy and reliability of fault detection and prediction are significantly improved through a systematic data processing procedure. First, according to the operating characteristics and known fault modes of the photovoltaic system, combined with the historical fault database and expert experience library, key operating parameters such as voltage and current are selected for monitoring to ensure the pertinence and efficiency of monitoring. Then, the time series data set is traversed, and the operating parameter sequence related to the selected target parameter is extracted to generate a target operating parameter sequence dedicated for prediction, fully utilizing the correlation of multi-dimensional data and enhancing the input information of the prediction model. Next, the target operating parameter sequence is aligned and combined into a unified target sequence according to the time stamp, and input into the preset ARIMA model to generate the first prediction information, ensuring the consistency of the data and the accuracy of the model prediction. By calculating the error between the actual value and the predicted value of the ARIMA model, a residual sequence is generated, which is used to evaluate the prediction performance of the ARIMA model and provide a basis for subsequent fault detection and model optimization. Overall, this embodiment improves the accuracy and efficiency of photovoltaic system fault detection by selectively choosing key monitoring parameters, enhancing prediction input using multi-parameter correlation, ensuring data consistency, and optimizing the model through residual analysis.

[0097] In one embodiment, the step S300 of inputting the residual sequence into a preset LSTM network to generate second prediction information comprises:

[0098] Step S310: extracting time-dependent features from the residual sequence through a sliding window algorithm, and generating an input feature sequence based on a set time step for the residual sequence.

[0099] In this step, the sliding window method can effectively capture the time sequence dependence in the time series data. By sliding the window within the time step to extract continuous data segments, the model can learn the dynamic change pattern and potential time dependence of the data.

[0100] Step S320: performing feature extraction on the input feature sequence to generate statistical features, periodic features, and frequency domain features, and fusing the time-dependent features, the statistical features, the periodic features, and the frequency domain features to generate an enhanced feature sequence.

[0101] In this step, the extraction of statistical, periodic, and frequency domain features comprehensively describes the different characteristics of the data and enhances the feature expression ability. Fusing multiple features enables the model to understand the data from multiple angles, improves the recognition ability of complex patterns, and improves the accuracy and reliability of nonlinear prediction.

[0102] Step S330: input the enhanced feature sequence into a preset LSTM network for deep modeling to generate first nonlinear prediction information.

[0103] Step S340: perform short-term and medium-term nonlinear prediction and trend analysis on the first nonlinear prediction information based on a multi-step prediction algorithm to generate second nonlinear prediction information.

[0104] Step S350: filter and correct the second nonlinear prediction information based on the expert experience library, and generate second prediction information.

[0105] In this step, through the filtering and correction of expert knowledge, false positives and false negatives are reduced to ensure the accuracy and reliability of the prediction information. At the same time, the prediction ability of the machine learning model is combined with the expert experience to improve the professionalism and pertinence of decision support.

[0106] In this embodiment, the accuracy and reliability of fault detection and prediction are significantly improved through a systematic data processing flow. First, the sliding window algorithm is used to extract time-dependent features from the residual sequence to generate input feature sequences, ensuring that the model can capture the dynamic changes and time-dependent of the data. Then, multi-dimensional feature extraction is performed on these input feature sequences to generate statistical features, periodic features and frequency domain features, and these features are fused with the time-dependent features to form an enhanced feature sequence, which fully represents the complex patterns of the data. Next, the enhanced feature sequence is input into a preset LSTM network for deep modeling to generate a first set of nonlinear prediction information, fully utilizing the LSTM's ability to capture long-term dependencies and nonlinear relationships. Based on the multi-step prediction algorithm, the first nonlinear prediction information is analyzed for short-term and medium-term trends to generate a second set of nonlinear prediction information, expanding the time range and warning ability of the prediction. Finally, the second nonlinear prediction information is filtered and corrected in combination with the expert experience library to generate second prediction information optimized by professional knowledge, improving the reliability and practicality of the prediction results. Overall, this embodiment, through deep feature extraction, advanced prediction model, multi-step prediction and expert knowledge fusion, not only improves the prediction accuracy and robustness of key operating parameters of photovoltaic systems, but also enhances the warning ability and decision support effect of the system, reduces false positives and false negatives, and ensures the stable operation of photovoltaic power generation systems in complex environments.

[0107] In one embodiment, the step S400 of generating preliminary fault information based on the first prediction information and the second prediction information includes:

[0108] Step S410: align the timestamps of the first prediction information and the second prediction information to generate an aligned prediction information set.

[0109] In this step, since different models may generate prediction results at different time steps, unaligned data will lead to inaccurate evaluation. By time alignment, the fairness and accuracy of comparison can be ensured, effectively eliminating errors caused by time bias, making subsequent difference analysis and comprehensive prediction more accurate.

[0110] Step S420: Perform difference evaluation analysis of the first prediction information and the second prediction information in the same time interval based on the mutual information metric, and generate a difference analysis index.

[0111] In this step, the mutual information metric is used to capture nonlinear and complex dependencies, and to deeply understand the differences in the description of the system state by different prediction models in the same time period, thereby providing a basis for generating weight parameters.

[0112] Step S430: Generate a weight parameter according to the difference analysis index, and generate comprehensive prediction information according to the first prediction information, the second prediction information, and the weight parameter.

[0113] In this step, the contribution proportion of each prediction information in the comprehensive result is dynamically adjusted according to the difference degree between the prediction results. The embodiment adopts a difference degree-based weighted fusion strategy, which improves the flexibility and adaptability of comprehensive prediction, and can dynamically adjust the weight distribution of prediction results according to real-time data, thereby enhancing the accuracy and robustness of overall prediction.

[0114] Step S440: Generate fault matching information based on the comprehensive prediction information, the historical fault database, and the expert experience library based on a pattern matching algorithm, wherein the fault matching information includes historical fault identification and expert rule identification.

[0115] In this step, the pattern matching algorithm is used to compare the current prediction result with past fault cases, identify potential fault patterns, and provide targeted processing recommendations based on historical data and expert knowledge, enhancing the depth and breadth of decision support.

[0116] Step S450: Generate preliminary fault information based on a Naive Bayes algorithm according to the fault matching information, wherein the preliminary fault information includes historical fault matching identification, fault type, fault probability, and fault severity.

[0117] In this step, based on the fault matching information, the Naive Bayes algorithm is used to generate preliminary fault information, including historical fault matching identification, fault type, fault probability, and fault severity. The Naive Bayes algorithm calculates the posterior probability of different fault types, combines historical matching and expert rules, and can effectively classify and evaluate the fault state of the current system.

[0118] In this embodiment, the accuracy and reliability of fault detection and prediction are significantly improved through a systematic data processing procedure. First, the first prediction information and the second prediction information are aligned according to their timestamps to generate an aligned prediction information set, ensuring the consistency and comparability of the two sets of prediction data in time. Then, the difference between the two sets of prediction information in the same time interval is evaluated and analyzed based on mutual information metrics to generate a difference analysis index, effectively quantifying the relevance and difference between the prediction results. Next, a weight parameter is generated based on the difference analysis index, and the first prediction information and the second prediction information are weighted and fused using the weight parameter to generate comprehensive prediction information, thereby improving the accuracy and robustness of the prediction results. Then, the comprehensive prediction information is matched with the historical fault database and the expert experience library using a pattern matching algorithm to generate fault matching information containing historical fault identification and expert rule identification, ensuring the comprehensiveness and professionalism of fault identification. Finally, based on the fault matching information, a naive Bayes algorithm is used to generate preliminary fault information, including historical fault matching identification, fault type, fault probability, and fault severity, providing detailed and reliable fault diagnosis results. This embodiment improves the accuracy of photovoltaic power generation system fault detection and the effectiveness of decision support through time alignment, difference evaluation, weighted fusion, pattern matching, and probability analysis.

[0119] In one embodiment, the step S500 of inputting the preliminary fault detection information into a pre-trained Transformer model for deep feature extraction and context association analysis and generating auxiliary decision information, and sending the auxiliary decision information to the management personnel, comprises:

[0120] Step S510: According to the historical fault matching identification, the corresponding fault case text information is retrieved from the historical fault database, and according to the expert rule identification, the corresponding processing suggestion text is queried from the expert experience library. The fault case text information and the processing suggestion text are vectorized based on a text encoding algorithm to generate context embedding data.

[0121] In this step, the historical fault cases and expert suggestions are converted into numerical forms that can be processed by machine learning models, providing rich semantic information for subsequent deep feature extraction and decision analysis. By combining historical data and expert knowledge, the system can better understand fault patterns, improve the accuracy of fault identification, and enhance the relevance of decision recommendations, thereby enhancing the overall fault detection and response capability.

[0122] Step S520: Feature splicing and weighted integration are performed according to the preliminary fault detection information and the context embedding data to generate fusion feature representation.

[0123] In this step, the real-time operating state of the photovoltaic power generation system is combined with historical fault cases and expert recommendations to form a comprehensive feature set for subsequent deep learning model analysis. Through feature splicing and weighted integration, multiple sources of information can be considered simultaneously, improving the richness and diversity of feature expression and enhancing the model's ability to identify complex fault patterns.

[0124] Step S530: input the fusion feature representation into a pre-trained Transformer model, perform deep feature extraction and context association analysis on the fusion feature through a multi-head attention mechanism, and generate a deep association feature.

[0125] Step S540: generate auxiliary decision candidate information based on the deep association feature and the expert experience library.

[0126] In this step, the complex features extracted by the deep learning model are combined with expert knowledge to provide specific fault handling recommendations and maintenance strategies through a strategy inference mechanism. By utilizing the expert experience library, the system can integrate professional judgment on a data-driven basis to generate more targeted and practical decision information.

[0127] Step S550: format and filter the auxiliary decision candidate information to generate auxiliary decision information, which includes executable maintenance strategies, maintenance timing recommendations, and resource allocation schemes.

[0128] In this step, the initially generated decision information is organized into an easy-to-understand and implementable form, ensuring that management personnel can quickly obtain and apply these recommendations. Through formatting and filtering, the system can eliminate redundant or irrelevant information, highlight key decision points, and improve the effectiveness and operability of decision support.

[0129] Step S560: send the auxiliary decision information to the management personnel and obtain feedback information, and store the feedback information in the historical fault database and the expert experience library.

[0130] In this step, the generated auxiliary decision information is sent to the management personnel and feedback information is obtained, and these feedback information is stored in the historical fault database and the expert experience library, realizing the dynamic updating of knowledge and the continuous optimization of the system. A closed-loop feedback mechanism is established, and through the actual operation feedback of the management personnel, the fault database and the expert experience library of the system are continuously enriched and improved. By storing and analyzing feedback information, the system can learn and adapt to new fault patterns and handling methods, improving the accuracy and effectiveness of future fault detection and decision support.

[0131] In this embodiment, by integrating historical fault cases and expert experience, the accuracy and reliability of fault detection and decision-making are significantly improved. First, the system retrieves and vectorizes fault cases and processing suggestion texts from the historical fault database and expert experience library according to the historical fault matching identifier and expert rule identifier, and generates context embedding data. Then, the preliminary fault detection information and these context embedding data are spliced and weighted integrated to form a fusion feature representation, which is input into a pre-trained Transformer model to extract features in depth through a multi-head attention mechanism to generate deep correlation features. Based on these deep correlation features, the system infers strategies in combination with the expert experience library to generate auxiliary decision-making candidate information, and through result formatting and optimal screening, finally provides auxiliary decision-making information including executable maintenance strategies, maintenance opportunity suggestions and resource allocation schemes. Finally, the system sends these decision-making information to the management personnel and collects feedback information to continuously update the historical fault database and expert experience library, realizing dynamic optimization of knowledge.

[0132] In one embodiment, step S430: generating a weight parameter according to the difference analysis indicator, generating comprehensive prediction information according to the first prediction information, the second prediction information and the weight parameter, comprises:

[0133] Step S431: generating a weight parameter based on the following formula according to the difference analysis indicator:

[0134]

[0135] Wherein, ω(t) is the weight parameter, t is the ordinal number of the time step, α is a preset change sensitivity hyperparameter, β is a preset expected difference threshold, and D(t) is the difference analysis indicator.

[0136] Step S432: generating a basic fusion result based on the following formula according to the first prediction information and the second prediction information:

[0137] P'(t) = ω(t) μ1(t) + [1- ω(t)] μ2(t);

[0138] Wherein, P'(t) is the basic fusion result, t is the ordinal number of the time step, ω(t) is the weight parameter, μ1(t) is the first prediction information, and μ2(t) is the second prediction information.

[0139] Step S433: generating a relative difference degree based on the following formula according to the first prediction information and the second prediction information:

[0140]

[0141] wherein R(t) is the relative difference degree, t is the ordinal number of the time step, μ1(t) is the first prediction information, μ2(t) is the second prediction information, and γ is a preset positive constant.

[0142] Step S434: correcting the basic fusion result according to the relative difference degree based on the following formula to generate a comprehensive prediction result:

[0143] P(t) = P'(t) - η·R(t)·[P'(t) - ξ];

[0144] wherein P(t) is the comprehensive prediction result, t is the ordinal number of the time step, P'(t) is the basic fusion result, η is a preset correction amplitude adjustment parameter, R(t) is the relative difference degree, and ξ is a preset historical mean value.

[0145] In this embodiment, the contribution proportion of the two kinds of prediction information is flexibly adjusted by using the weight parameter, and the deficiency of the model at a specific time step is compensated by the difference degree correction, the accuracy and stability of the prediction result are ensured by combining the historical mean value and the correction amplitude adjustment mechanism. The advantage of the scheme lies in strong dynamic adaptability, the fusion strategy can be adjusted according to the real-time difference index, the complementarity of multi-source information is fully utilized, and the prediction accuracy is significantly improved. In addition, by introducing the normalization and smoothing factor, the robustness of the model to noise and extreme values is effectively enhanced, so that it can adapt to the variable environment of the photovoltaic power generation system.

[0146] In one embodiment, the ARIMA model is established based on the following steps:

[0147] Step S610: acquiring a set of conventional detection information of the power generation system, generating a set of conventional time series data based on the set of conventional detection information, and generating a modeling sequence and a verification sequence by performing multi-granularity partitioning on the set of conventional time series data.

[0148] In this step, first, various types of conventional detection information of the power generation system are collected, including environmental data (such as temperature, humidity, and light intensity), system operation data (such as voltage, current, and power), and equipment status data (such as equipment operating status and fault information). These data are usually time series data, which are converted into standard time series data sets through preprocessing. In order to ensure the generalization ability and robustness of the model, these time series data sets are further divided into modeling sequences and verification sequences. This multi-granularity partitioning method can model and verify the data through different time scales (such as hours, days, and weeks), helping the system to capture different features from short-term changes to long-term trends.

[0149] Step S620: decompose the modeling sequence by variational modal decomposition to generate a plurality of intrinsic modal components, establish a corresponding initial ARIMA sub-model according to each intrinsic modal component, and generate an ARIMA sub-model by iteratively searching and updating the initial ARIMA sub-model parameters through a Bayesian optimization algorithm.

[0150] Step S630: based on the ARIMA sub-model, a fixed-length sliding window algorithm is used to perform rolling prediction on the verification sequence to generate a first prediction result.

[0151] In this step, a fixed-length sliding window algorithm is used to perform rolling prediction on the verification sequence. Each time a fixed-size time window (such as historical data of the past few days or hours) is used to predict the system state in the future. This allows the model to update based on the latest data at each time step, avoiding the influence of outdated data and maintaining the timeliness of the prediction.

[0152] Step S640: compare the first prediction result with the true value of the verification sequence to extract residuals and generate a residual sequence, and optimize the ARIMA sub-model according to the residual sequence to generate an optimized ARIMA sub-model.

[0153] In this step, through residual extraction, possible bias or error patterns in the model can be identified, and then the ARIMA sub-model is optimized according to these residual sequences, thereby improving the fitting accuracy of the model and reducing the prediction error.

[0154] Step S650: performance quantification of the optimized ARIMA sub-model according to the residual sequence and generation of an error value, the error value being obtained based on the following formula:

[0155]

[0156] where E is the error value, N is the number of time steps of the verification sequence, t is the ordinal number of the time step, c t is the weight factor corresponding to the time step t, y t is the true value of the verification sequence at time step t, y′ t is the model prediction value of the verification sequence at time step t, and ε is a predetermined positive constant.

[0157] Step S670: generate a sub-model weight parameter according to the error value, the historical fault database and the expert experience database, and construct the ARIMA model according to the ARIMA sub-model and the sub-model weight parameter.

[0158] In this embodiment, through systematic data partitioning, variational modal decomposition and Bayesian optimization, the multi-level linear trend and dynamic change pattern of system operation parameters are accurately captured and modeled. First, the conventional detection data is divided into modeling and validation sequences to ensure the accuracy and generalization ability of model training. Then, the complex time series is decomposed into multiple intrinsic modal components through variational modal decomposition, and the parameters of each ARIMA sub-model are iteratively optimized using Bayesian optimization to improve prediction accuracy. Then, a fixed-length sliding window is used to perform rolling prediction on the validation sequence to generate preliminary prediction results, and the model is further optimized through residual refinement to reduce prediction error. Finally, the model performance is quantified by error formula to ensure the reliability of the prediction results. This embodiment combines error quantification with historical fault database and expert experience library to generate a weighted optimized ARIMA model, realizes reasonable distribution of multi-model weights, improves the prediction accuracy of key parameters of photovoltaic system and the robustness of the model, and through continuous optimization and historical data fusion, enhances the sensitivity and response ability to abnormality and failure.

[0159] In one embodiment, as shown in Figure 2 A large model-based power generation system fault auxiliary decision generation system is provided, including a time series data generation module, a first prediction information generation module, a second prediction information generation module, a preliminary fault information generation module, and an auxiliary decision information generation module.

[0160] The time series data generation module is configured to obtain a current detection information set of the power generation system, and generate a time series data set by preprocessing the current detection information set. The first prediction information generation module is configured to generate first prediction information and a residual sequence based on a preset ARIMA model according to the time series data set. The second prediction information generation module is configured to input the residual sequence into a preset LSTM network to generate second prediction information. The preliminary fault information generation module is configured to combine the first prediction information and the second prediction information to generate preliminary fault information. The auxiliary decision information generation module is configured to input the preliminary fault detection information into a pre-trained Transformer model for deep feature extraction and context association analysis, and generate auxiliary decision information. The auxiliary decision information is sent to the management personnel.

[0161] In one embodiment, the time series data generation module is further configured to:

[0162] Obtain a current detection information set of a power generation system, perform data fusion preprocessing on the current detection information set to generate a fusion information set, wherein the detection information set includes environmental detection information, system operation information, and state detection information; perform adaptive data completion on missing values in the fusion information set using a local weighted regression interpolation algorithm to generate a completed information set, perform fitting correction on the completed information set using a random forest algorithm to generate a corrected information set; perform abnormality screening on the corrected information set based on a z-score method to generate an abnormal value set, perform abnormality screening on the abnormal value set using an isolation forest algorithm to generate a confirmed abnormal value set, and generate a screening information set based on the local weighted regression interpolation algorithm according to the completed detection information set and the confirmed abnormal value set; obtain time series data of each detection feature in the screening information set, calculate the median and median absolute deviation value of each feature, perform normalization processing on the screening information set according to the absolute deviation value to generate a normalized information set, perform standardization on the normalized information set based on a z-score method to generate a standard information set, and perform feature engineering processing on the standard information set to generate a feature extraction information set, wherein the feature extraction information set includes statistical features, periodic features, and frequency domain features.

[0163] In one embodiment, the first prediction information generation module is further configured to:

[0164] According to a preset historical fault database and an expert experience library, a target parameter is selected; a running parameter sequence related to the target parameter is obtained by traversing the time series data set, a target running parameter sequence is generated according to each related running parameter sequence; each target running parameter sequence is aligned and combined according to a time stamp to generate a target sequence, the target sequence is input into a preset ARIMA model to generate first prediction information; and a prediction error calculation is performed according to the target sequence and the first prediction information, and a residual sequence is generated.

[0165] In one embodiment, the first prediction information generation module is further configured to:

[0166] A routine monitoring information set of a power generation system is acquired. A routine time series dataset is generated based on this set. This dataset is then partitioned into multi-granularity partitions to generate a modeling sequence and a validation sequence. The modeling sequence is decomposed using variational mode decomposition to generate multiple intrinsic mode components (IMCs). An initial ARIMA sub-model is established for each IMC. The parameters of the initial ARIMA sub-model are iteratively searched and updated using a Bayesian optimization algorithm to generate an ARIMA sub-model. Based on the ARIMA sub-model, a fixed-length sliding window algorithm is used to perform rolling prediction on the validation sequence, generating a first prediction result. The first prediction result is compared with the true values ​​of the validation sequence to extract residuals and generate a residual sequence. The ARIMA sub-model is then optimized based on the residual sequence to generate an optimized ARIMA sub-model. Finally, the optimized ARIMA sub-model's performance is quantified based on the residual sequence, and an error value is generated. This error value is obtained using the following formula:

[0167]

[0168] Where E is the error value, N is the number of time steps in the validation sequence, t is the ordinal number of the time step, and c t y is the weighting factor corresponding to time step t. t To verify the true value of the sequence at time step t, y′ t To verify the model prediction value of the sequence at time step t, ε is a preset positive constant; sub-model weight parameters are generated based on the error value, the historical fault database, and the expert experience database, and the ARIMA model is constructed based on the ARIMA sub-model and the sub-model weight parameters.

[0169] In one embodiment, the second prediction information generation module is further configured to:

[0170] The time-dependent features are extracted from the residual sequence using a sliding window algorithm. An input feature sequence is generated based on the residual sequence and a set time step. The input feature sequence is then used to extract statistical, periodic, and frequency domain features. These features are fused to generate an enhanced feature sequence. The enhanced feature sequence is then input into a preset LSTM network for deep modeling to generate first nonlinear prediction information. This first nonlinear prediction information is then used for short- and medium-term nonlinear prediction and trend analysis using a multi-step prediction algorithm to generate second nonlinear prediction information. Finally, the second nonlinear prediction information is filtered and corrected based on the expert experience base to generate a second prediction information.

[0171] In one embodiment, the preliminary fault information generation module is further configured to:

[0172] Align the timestamps of the first and second prediction information to generate an aligned prediction information set; perform a difference assessment analysis on the first and second prediction information within the same time interval based on mutual information metric, and generate a difference analysis index; generate weight parameters based on the difference analysis index, and generate comprehensive prediction information based on the first prediction information, the second prediction information, and the weight parameters; generate fault matching information based on the comprehensive prediction information, the historical fault database, and the expert experience base using a pattern matching algorithm, wherein the fault matching information includes historical fault identification identifiers and expert rule identifiers; generate preliminary fault information based on the fault matching information using a Naive Bayes algorithm, wherein the preliminary fault information includes historical fault matching identifiers, fault type, fault probability, and fault severity.

[0173] In one embodiment, the preliminary fault information generation module is further configured to: generate weight parameters based on the difference analysis index according to the following formula:

[0174]

[0175] Where ω(t) is the weight parameter, t is the ordinal number of the time step, α is the preset change sensitivity hyperparameter, β is the preset expected difference threshold, and D(t) is the difference analysis index.

[0176] Based on the weight parameters, a basic fusion result is generated from the first prediction information and the second prediction information using the following formula:

[0177] P′(t)=ω(t)μ1(t)+[1-ω(t)]μ2(t);

[0178] Where P′(t) is the basic fusion result, t is the ordinal number of the time step, ω(t) is the weight parameter, μ1(t) is the first prediction information, and μ2(t) is the second prediction information;

[0179] The relative difference is generated based on the first and second prediction information using the following formula:

[0180]

[0181] Where R(t) is the relative difference degree, t is the ordinal number of the time step, μ1(t) is the first prediction information, μ2(t) is the second prediction information, and γ is a preset positive constant;

[0182] The basic fusion result is corrected based on the relative difference using the following formula to generate a comprehensive prediction result:

[0183] P(t) = P'(t) - η · R(t) · [P'(t) - ξ];

[0184] P(t) = P'(t) - η · R(t) · [P'(t) - ξ]; wherein, P(t) is a comprehensive prediction result, t is the ordinal number of a time step, P'(t) is a basic fusion result, η is a preset correction amplitude adjustment parameter, R(t) is a relative difference degree, and ξ is a preset historical mean value.

[0185] In one embodiment, the auxiliary decision information generation module is further configured to:

[0186] According to the historical fault matching identifier, corresponding fault case text information is retrieved from the historical fault database, and according to the expert rule identifier, corresponding processing suggestion text is queried from the expert experience library. The fault case text information and the processing suggestion text are vectorized based on a text coding algorithm to generate context embedding data. The preliminary fault detection information and the context embedding data are combined and weighted to generate fusion feature representation. The fusion feature representation is input into a pre-trained Transformer model, and deep feature extraction and context correlation analysis are performed on the fusion feature through a multi-head attention mechanism to generate deep correlation features. The deep correlation features and the expert experience library are used to infer strategies to generate auxiliary decision candidate information. The auxiliary decision candidate information is formatted and selected to generate auxiliary decision information, wherein the auxiliary decision information includes executable maintenance strategies, maintenance opportunity suggestions, and resource allocation schemes. The auxiliary decision information is sent to the management personnel and feedback information is obtained, and the feedback information is stored in the historical fault database and the expert experience library.

[0187] In one embodiment, a computer device is provided, including a memory and a processor, the memory stores a computer program, and the processor implements the steps of the above-mentioned power generation system fault auxiliary decision generation method based on a large model when executing the computer program.

[0188] In one embodiment, a computer readable storage medium is provided, which stores a computer program, and the computer program is executed by a processor to implement the steps of the above-mentioned power generation system fault auxiliary decision generation method based on a large model.

[0189] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when the computer program is executed, the processes of the above-mentioned embodiments of the methods can be included. Any reference to memory, storage, databases, or other media in the embodiments provided by the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0190] The technical features of the above embodiments can be combined in any way. In order to make the description simple, not all possible combinations of the technical features in the above embodiments are described, but as long as the combinations of the technical features do not exist, they should be considered as the scope of the present application.

[0191] The above embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the present application, some modifications and improvements can be made, which are all within the scope of the present application. Therefore, the scope of the patent of the present application should be subject to the appended claims.

Claims

1. A method for generating fault-assisted decision-making in power generation systems based on a large model, characterized in that, include: Obtain the current detection information set of the power generation system, and preprocess the current detection information set to generate a time series dataset; Based on the time series dataset, generate first prediction information and residual sequences using a preset ARIMA model; The residual sequence is input into a preset LSTM network to generate second prediction information; Preliminary fault information is generated by combining the first prediction information and the second prediction information; The preliminary fault detection information is input into a pre-trained Transformer model for deep feature extraction and contextual analysis to generate auxiliary decision-making information, which is then sent to the management personnel. Preliminary fault information includes historical fault matching identifiers, fault type, fault probability, and fault severity; The preliminary fault detection information is input into a pre-trained Transformer model for deep feature extraction and contextual analysis to generate auxiliary decision-making information, including: Based on the historical fault matching identifier, the corresponding fault case text information is retrieved from the historical fault database. Based on the expert rule identifier, the corresponding processing suggestion text is retrieved from the expert experience base. The fault case text information and the processing suggestion text are vectorized based on the text encoding algorithm to generate context embedding data. Based on the preliminary fault detection information and the context embedding data, feature concatenation and weighted integration are performed to generate a fused feature representation; The fused feature representation is input into a pre-trained Transformer model, and deep feature extraction and contextual association analysis are performed on the fused feature through a multi-head attention mechanism to generate deep association features; Based on the deep correlation features and the expert experience base, strategy inference is performed to generate auxiliary decision-making candidate information; The auxiliary decision-making candidate information is formatted and optimized to generate auxiliary decision-making information, which includes executable maintenance strategies, maintenance timing suggestions, and resource allocation schemes.

2. The method for generating auxiliary decision-making for power generation systems based on a large model according to claim 1, characterized in that, Based on the time series dataset, a first prediction information and residual sequence are generated using a preset ARIMA model, including: Select target parameters based on a pre-set historical fault database and expert experience database; The time series dataset is traversed to obtain the sequence of operating parameters related to the target parameter, and the target operating parameter sequence is generated based on each of the related operating parameter sequences; The target operating parameter sequences are aligned by timestamp and combined to generate a target sequence. The target sequence is then input into a preset ARIMA model to generate first prediction information. The prediction error is calculated based on the target sequence and the first prediction information, and a residual sequence is generated.

3. The method for generating auxiliary decision-making for power generation systems based on a large model according to claim 2, characterized in that, The residual sequence is input into a preset LSTM network to generate second prediction information, including: Based on the residual sequence, time-dependent features are extracted using a sliding window algorithm, and an input feature sequence is generated from the residual sequence based on a set time step. The input feature sequence is subjected to feature extraction to generate statistical features, periodic features and frequency domain features. The time-dependent features, statistical features, periodic features and frequency domain features are then fused to generate an enhanced feature sequence. The enhanced feature sequence is input into a preset LSTM network for deep modeling to generate first nonlinear prediction information; The first nonlinear prediction information is used to perform short-term and medium-term nonlinear prediction and trend analysis based on a multi-step prediction algorithm to generate the second nonlinear prediction information. The second nonlinear prediction information is filtered and corrected based on the expert experience base, and then the second prediction information is generated.

4. The method for generating auxiliary decision-making for power generation systems based on a large model according to claim 3, characterized in that, Preliminary fault information is generated by combining the first prediction information and the second prediction information, including: Align the timestamps of the first prediction information and the second prediction information to generate an aligned prediction information set; Based on mutual information measurement, the difference between the first prediction information and the second prediction information within the same time interval is evaluated and analyzed, and difference analysis indicators are generated. Weight parameters are generated based on the difference analysis indicators, and comprehensive prediction information is generated based on the first prediction information, the second prediction information, and the weight parameters. Based on the pattern matching algorithm, fault matching information is generated according to the comprehensive prediction information, the historical fault database and the expert experience base, wherein the fault matching information includes historical fault identification identifiers and expert rule identifiers; Preliminary fault information is generated based on the fault matching information using the Naive Bayes algorithm. The preliminary fault information includes historical fault matching identifiers, fault types, fault probabilities, and fault severity.

5. The method for generating auxiliary decision-making for power generation systems based on a large model according to claim 4, characterized in that, Sending the decision support information to management personnel includes: The auxiliary decision-making information is sent to the management personnel and feedback information is obtained. The feedback information is stored in the historical fault database and the expert experience database.

6. The method for generating auxiliary decision-making for power generation systems based on large models according to any one of claims 2-5, characterized in that, The ARIMA model is established based on the following steps: Obtain a set of routine testing information for the power generation system, generate a routine time series dataset based on the routine testing information set, and partition the routine time series dataset into multi-granularity partitions to generate modeling sequences and validation sequences; The modeling sequence is decomposed by variational mode decomposition to generate multiple intrinsic mode components. An initial ARIMA sub-model is established based on each intrinsic mode component. The parameters of the initial ARIMA sub-model are iteratively searched and updated using a Bayesian optimization algorithm to generate an ARIMA sub-model. Based on the ARIMA sub-model, a fixed-length sliding window algorithm is used to perform rolling prediction on the verification sequence to generate a first prediction result; The first prediction result is compared with the true value of the validation sequence to extract residuals and generate a residual sequence. The ARIMA sub-model is then optimized based on the residual sequence to generate an optimized ARIMA sub-model. The optimized ARIMA sub-model is quantized based on the residual sequence to generate an error value, which is obtained based on the following formula: ； Where E is the error value, N is the number of time steps in the validation sequence, t is the ordinal number of the time step, and c t y is the weighting factor corresponding to time step t. t To verify the true value of the sequence at time step t, y' t To verify the model prediction of the sequence at time step t, ε is a preset positive constant; Sub-model weight parameters are generated based on the error value, the historical fault database, and the expert experience database. The ARIMA model is then constructed based on the ARIMA sub-model and the sub-model weight parameters.

7. The method for generating auxiliary decision-making for power generation systems based on large models according to any one of claims 1-5, characterized in that, Acquire the current detection information set of the power generation system, and preprocess the current detection information set to generate a time series dataset, including: The current detection information set of the power generation system is obtained, and the current detection information set is subjected to data fusion preprocessing to generate a fused information set, wherein the detection information set includes environmental detection information, system operation information and status detection information; The fused information set is adaptively filled with missing values ​​using a local weighted regression interpolation algorithm, and a filled information set is generated. The filled information set is then fitted and corrected using a random forest algorithm, and a corrected information set is generated. Based on the correction information set, anomalies are filtered using the z-score method to generate an outlier set. The outlier set is then filtered using the isolated forest algorithm to generate a confirmed outlier set. Finally, a filtering information set is generated based on the completion detection information set and the confirmed outlier set using the local weighted regression interpolation algorithm. Obtain time series data for each detection feature in the filter information set, calculate the median and absolute deviation of the median for each feature, and normalize the filter information set based on the absolute deviation to generate a normalized information set. The normalized information set is standardized using the z-score method to generate a standard information set. The standard information set is then subjected to feature engineering to generate a feature extraction information set, which includes statistical features, periodic features, and frequency domain features.

8. A power generation system fault auxiliary decision generation system based on a large model, employing the power generation system fault auxiliary decision generation method based on a large model as described in any one of claims 1-7, characterized in that, include: The time series data generation module is used to acquire the current detection information set of the power generation system and preprocess the current detection information set to generate a time series dataset. The first prediction information generation module is used to generate first prediction information and residual sequences based on the time series dataset and a preset ARIMA model. The second prediction information generation module is used to input the residual sequence into a preset LSTM network to generate second prediction information; A preliminary fault information generation module is used to generate preliminary fault information based on the combination of the first prediction information and the second prediction information; The auxiliary decision-making information generation module is used to input the preliminary fault detection information into a pre-trained Transformer model for deep feature extraction and contextual analysis, generate auxiliary decision-making information, and send the auxiliary decision-making information to the management personnel.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images