Industry large model training method, device and system based on efficient fine-tuning

Abstract

The present application relates to the technical field of artificial intelligence, in particular to an industry large model training method, device and system based on efficient fine-tuning, which solves the technical problem of poor training data quality in the prior art. The method comprises: collecting multi-source heterogeneous data of each power station; the multi-source heterogeneous data comprises meteorological time series data, power generation time series data, an operation and maintenance record set and geographical environment parameters; each operation and maintenance record in the operation and maintenance record set is used for recording a fault type and a fault time of a fault occurring in the power station; performing correlation analysis on the multi-source heterogeneous data of each power station to obtain an influence sequence corresponding to each operation and maintenance record; performing similarity analysis on the influence sequence corresponding to each operation and maintenance record to determine an importance index corresponding to each operation and maintenance record; performing data processing on the influence sequence based on the importance index corresponding to each operation and maintenance record, and training an industry large model by using a training data set obtained through the data processing.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method, apparatus, and system for training large industry models based on efficient fine-tuning. Background Technology

[0002] With the breakthroughs achieved by Large Language Models (LLMs) in the field of natural language processing, applying them to vertical industries such as photovoltaics to build dedicated large models has become an important trend in improving the intelligence level of these industries. By fine-tuning general-purpose large models to enable them to understand photovoltaic operation and maintenance reports, predict power generation, or assist in fault diagnosis, industry operational efficiency can be effectively improved. However, the context window of large models has a limit on the number of tokens; inputting extremely long photovoltaic numerical time series can lead to memory overflow. This necessitates data cleaning to reduce the length of each training data set when training and fine-tuning large models.

[0003] Current technologies for processing photovoltaic data used for fine-tuning typically employ a divide-and-conquer followed by fusion strategy: first, heterogeneous data from multiple sources such as meteorology, power generation, and operation and maintenance are separated and each data is cleaned independently (e.g., outlier removal and missing value imputation); then, the cleaned data is simply concatenated for fine-tuning training of downstream models. This generic data cleaning approach severs the inherent, continuous temporal causal relationships among meteorology, power generation, and operation and maintenance, leading to the misclassification and deletion of some high-quality data with strong correlations as redundant data. This results in poor-quality training data, ultimately affecting the practical application effectiveness of the trained large model in critical tasks such as power generation prediction and operation and maintenance fault early warning. Summary of the Invention

[0004] To address the technical problem of poor training data quality in existing technologies, the present invention aims to provide a method, apparatus, and system for training large industry models based on efficient fine-tuning. The specific technical solution adopted is as follows:

[0005] This application provides a method for training large industry models based on efficient fine-tuning, including: Collect multi-source heterogeneous data from various power plants; the multi-source heterogeneous data includes meteorological time-series data, power generation time-series data, operation and maintenance record set, and geographical environmental parameters; each operation and maintenance record in the operation and maintenance record set is used to record the fault type and fault time of the power plant. Correlation analysis was performed on the multi-source heterogeneous data of each power station to obtain the impact sequence corresponding to each operation and maintenance record; the impact sequence is used to characterize the impact of the fault corresponding to the operation and maintenance record on meteorological time series data, power generation time series data and geographical environmental parameters. A similarity analysis is performed on the impact sequence corresponding to each maintenance record to determine the importance index corresponding to each maintenance record; the importance index is used to characterize the regularity and stability of the impact sequence corresponding to the maintenance record within the same fault type; The impact sequence is processed based on the importance index corresponding to each operation and maintenance record, and the industry large model is trained using the training dataset obtained from the data processing.

[0006] In one possible implementation, the method includes: Based on the meteorological time-series data and the power generation time-series data, the key geographical parameters of each power station are determined from the geographical environment parameters; For each maintenance record, related time-series data is extracted from the meteorological time-series data and the power generation time-series data based on the fault time corresponding to the maintenance record; The key geographical parameters of the power station where the operation and maintenance record is located, the associated time-series data, and the fault type corresponding to the operation and maintenance record are associated and combined to form the impact sequence corresponding to the operation and maintenance record.

[0007] In one possible implementation, the method includes: The meteorological time-series data and the power generation time-series data are divided into windows with a preset duration to determine the meteorological sequence segment and the power generation sequence segment of each power station in each time window. For each power station in each time window, a cumulative distance matrix is ​​constructed between the meteorological sequence segment and the power generation sequence segment within the time window; the cumulative distance matrix is ​​used to characterize the cumulative temporal difference between the meteorological sequence segment and the power generation sequence segment within the time window. Multiple meteorological clusters were obtained by clustering meteorological sequence segments of each power station within each time window. For each meteorological cluster, the core influence matrix of the meteorological cluster is determined based on the cumulative distance matrix of each power station in each time window within the meteorological cluster; the core influence matrix is ​​used to characterize the core influence pattern between meteorological data and power generation data within the meteorological cluster. Regression analysis was performed based on the geographical environment parameters of each power station within the meteorological cluster and the core influence matrix of the meteorological cluster to determine the key geographical parameters of each power station.

[0008] In one possible implementation, the method includes: For each meteorological cluster, a local variance analysis is performed on each cumulative distance matrix in the meteorological cluster to determine the weighted basis matrix corresponding to each cumulative distance matrix; the weighted basis matrix is ​​used to characterize the importance distribution of each local region in the cumulative distance matrix; Based on the weight basis matrix corresponding to each cumulative distance matrix in the meteorological cluster, a mask consensus graph of the meteorological cluster is constructed; the mask consensus graph is used to characterize the consensus region distribution among the cumulative distance matrices in the meteorological cluster. Based on the mask consensus graph of the meteorological cluster, the weight basis matrix corresponding to each cumulative distance matrix in the meteorological cluster is weighted and calculated to obtain the weighted weight matrix corresponding to each cumulative distance matrix in the meteorological cluster; the weighted weight matrix is ​​used to characterize the importance weight distribution of the cumulative distance matrix after fusing consensus information within the cluster; The impact of each cumulative distance matrix is ​​evaluated based on the weighted weight matrix corresponding to each cumulative distance matrix in the meteorological cluster, and the core impact matrix of the meteorological cluster is determined from each cumulative distance matrix in the meteorological cluster.

[0009] In one possible implementation, the method includes: For each cumulative distance matrix in the meteorological cluster, based on the cumulative distance matrix and the weighted weight matrix corresponding to the cumulative distance matrix, a similarity analysis is performed with other cumulative distance matrices and their corresponding weighted weight matrices in the meteorological cluster to determine the average similarity coefficient between the cumulative distance matrix and other cumulative distance matrices in the meteorological cluster. The cumulative distance matrix with the largest average similarity coefficient among the cumulative distance matrices in the meteorological cluster is taken as the core influence matrix of the meteorological cluster.

[0010] In one possible implementation, the method includes: The impact sequence corresponding to each maintenance record is divided into multiple impact sequence sets according to the fault type; For each operation and maintenance record in each set of impact sequences, a similarity analysis is performed between the impact sequence corresponding to the operation and maintenance record and the impact sequences corresponding to other operation and maintenance records in the set of impact sequences to determine the importance index corresponding to the operation and maintenance record.

[0011] In one possible implementation, the method includes: For each impact sequence corresponding to each operation and maintenance record in each impact sequence set, calculate the feature distance between the impact sequence corresponding to the operation and maintenance record and the impact sequences corresponding to other operation and maintenance records in the impact sequence set. The feature distances between the impact sequence corresponding to the operation and maintenance record and the impact sequences corresponding to other operation and maintenance records in the impact sequence set are negatively correlated and normalized to obtain the similarity corresponding to each feature distance. The mean of the similarity corresponding to each feature distance is used as the importance index corresponding to the operation and maintenance record.

[0012] In one possible implementation, the method includes: For each operation and maintenance record, the impact sequence corresponding to the operation and maintenance record is pruned according to the importance index corresponding to the operation and maintenance record, and the pruned impact sequence of each operation and maintenance record is used as the training dataset; The model parameters of the low-rank adaptation layer in the large industry model are trained based on the training dataset.

[0013] This application provides a large-scale industry model training device based on efficient fine-tuning, comprising: a processor and a communication interface; the communication interface and the processor are coupled, and the processor is used to run computer programs or instructions to implement the methods described in any of the above embodiments.

[0014] This application provides a large-scale industry model training system based on efficient fine-tuning, including: The data acquisition unit is used to collect multi-source heterogeneous data from various power plants. The multi-source heterogeneous data includes meteorological time-series data, power generation time-series data, operation and maintenance record sets, and geographical environmental parameters. Each operation and maintenance record in the operation and maintenance record set is used to record the fault type and fault time of the power plant. The sequence construction unit is used to perform correlation analysis on the multi-source heterogeneous data of each power station to obtain the impact sequence corresponding to each operation and maintenance record; the impact sequence is used to characterize the impact of the fault corresponding to the operation and maintenance record on meteorological time series data, power generation time series data and geographical environmental parameters. The importance determination unit is used to perform similarity analysis based on the impact sequence corresponding to each operation and maintenance record to determine the importance index corresponding to each operation and maintenance record; the importance index is used to characterize the regularity and stability of the impact sequence corresponding to the operation and maintenance record in the same fault type; The model training unit is used to process the impact sequence based on the importance index corresponding to each operation and maintenance record, and to train the industry-wide large model using the training dataset obtained from the data processing.

[0015] The present invention has the following beneficial effects: Based on the above technical solution, this application collects multi-source heterogeneous data from various power plants and performs correlation analysis on the multi-source heterogeneous data from various power plants to obtain the impact sequence corresponding to each operation and maintenance record, thereby uncovering the inherent temporal causal relationship among meteorology, power generation, and operation and maintenance. Subsequently, similarity analysis is performed on the impact sequence corresponding to each operation and maintenance record to determine the importance index corresponding to each operation and maintenance record, achieving a quantitative assessment of the data training value. Finally, the impact sequence is processed based on the importance index corresponding to each operation and maintenance record, and the resulting training dataset is used to train a large-scale industry model. This allows for the retention of high-value data and the removal of redundant information based on importance indicators, reducing the length of training data while ensuring data quality. This effectively solves the problem of low dataset quality caused by data cleaning fragmenting business causal relationships in existing technologies, significantly improving the accuracy of the large-scale industry model in key tasks such as power generation prediction and fault early warning. Attached Figure Description

[0016] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating a method for training large industry models based on efficient fine-tuning, as provided in one embodiment of the present invention. Figure 2 This is a system architecture diagram of an industry large model training system based on efficient fine-tuning, provided as an embodiment of the present invention. Figure 3 This is a schematic diagram of the hardware structure of an industry large model training device based on efficient fine-tuning, provided as an embodiment of the present invention. Detailed Implementation

[0018] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation methods, structures, features, and effects of the industry-wide large model training method, apparatus, and system based on efficient fine-tuning proposed by the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0020] In all division and logarithmic operations covered in this application, a smoothing mechanism is employed to prevent computer program crashes or invalid values ​​from being generated due to a zero denominator or zero input. Specifically, a positive correction factor is superimposed on the denominator term of the division operation or the argument term of the logarithmic function, thereby ensuring the robustness and feasibility of the algorithm under extreme conditions.

[0021] The specific solutions for the industry-wide large model training method, apparatus, and system based on efficient fine-tuning provided by the present invention are described in detail below with reference to the accompanying drawings.

[0022] Please see Figure 1 The diagram illustrates a flowchart of a method for training large industry models based on efficient fine-tuning, according to an embodiment of the present invention. The method includes the following steps: Step 101: Collect multi-source heterogeneous data from each power station.

[0023] Among them, the multi-source heterogeneous data includes meteorological time-series data, power generation time-series data, operation and maintenance record set, and geographical environmental parameters. Each operation and maintenance record in the operation and maintenance record set is used to record the fault type and fault time of the power plant.

[0024] It should be noted that the operation of a photovoltaic power station involves data from multiple dimensions, including meteorological data reflecting the external environmental conditions, power generation data reflecting the equipment's operating status, maintenance records reflecting fault conditions, and geographical environmental parameters reflecting the station's geographical location. These data exhibit complex causal relationships over time; changes in meteorological conditions directly affect power output, while prolonged power generation anomalies often foreshadow maintenance failures. Therefore, collecting this multi-source heterogeneous data can provide a data foundation for subsequent analysis of the temporal causal relationships between meteorology, power generation, and maintenance.

[0025] In some embodiments, meteorological time-series data can be collected through meteorological monitoring stations deployed at the power plant site, including parameters such as temperature, light intensity, cloud cover, and wind speed. Power generation time-series data can be collected through the power plant's management system, including parameters such as module power generation, module temperature, and inverter efficiency. The operation and maintenance record set can be collected from the operation and maintenance management platform, including inspection records and fault reports; each record must include a fault occurrence timestamp, fault type, and descriptive information. Geographical environmental parameters can include parameters describing the geographical location characteristics of the power plant, such as altitude, terrain type, and latitude and longitude.

[0026] For example, the sampling frequency of meteorological time-series data can be set to once every 10 minutes, and the sampling frequency of power generation time-series data can be set to once every 5 minutes, with precise timestamps recorded for subsequent time alignment. For power generation data with a higher sampling frequency, downsampling or averaging can be used to align the time nodes of all data to a uniform sampling frequency. Furthermore, this application can also normalize the meteorological and power generation time-series data to eliminate the influence of dimensions.

[0027] Step 102: Perform correlation analysis on the multi-source heterogeneous data of each power station to obtain the impact sequence corresponding to each operation and maintenance record.

[0028] Among them, the impact sequence is used to characterize the impact of the corresponding faults in the operation and maintenance records on meteorological time series data, power generation time series data, and geographical environmental parameters.

[0029] Because the operation of photovoltaic systems exhibits a distinct physical causal chain, with continuous temporal causal relationships existing between weather changes, power generation anomalies, and operation and maintenance failures, it is necessary to analyze multi-source heterogeneous data as a whole. By performing correlation analysis on multi-source heterogeneous data, an impact sequence reflecting this cross-domain temporal causal relationship can be constructed, thereby preserving high-quality data containing strong correlations.

[0030] Step 103: Perform similarity analysis based on the impact sequence corresponding to each operation and maintenance record to determine the importance index corresponding to each operation and maintenance record.

[0031] Among them, the importance index is used to characterize the regularity and stability of the impact sequence corresponding to the operation and maintenance record within the same fault type.

[0032] Under the same fault type, the impact sequences corresponding to different maintenance records may differ. Some sequences can reflect the typical patterns of the fault type and are highly representative, while others may be affected by accidental factors and have weaker regularity. Therefore, this application can quantify the typicality of each impact sequence in the same type of fault by performing similarity analysis, thereby identifying data with stable patterns and high value.

[0033] Step 104: Process the impact sequence based on the importance index corresponding to each operation and maintenance record, and train the industry large model using the training dataset obtained from the data processing.

[0034] Because large models have a limited number of terms in their context window, excessively long input time-series data can lead to memory overflow. Therefore, it is necessary to process the influential sequences to compress the data length. This application can reduce the length of training data while ensuring data quality by processing the data according to importance indicators, retaining high-value data while removing redundant information.

[0035] In one possible implementation, this application can perform data pruning on the impact sequence corresponding to each operation and maintenance record based on the importance index corresponding to the operation and maintenance record, and use the pruned impact sequence of each operation and maintenance record as a training dataset.

[0036] By performing differentiated cropping based on importance indicators, this application can retain high-value data while eliminating redundant information, achieving a balance between data compression and quality preservation.

[0037] In some embodiments, this application can achieve data pruning through a hierarchical pruning strategy: For each operation and maintenance record, if the importance index corresponding to the operation and maintenance record is greater than or equal to the first importance threshold, the impact sequence corresponding to the operation and maintenance record is used as the training data in the training dataset. If the importance index corresponding to the operation and maintenance record is greater than or equal to the second importance threshold but less than the first importance threshold, the first subsequence in the impact sequence corresponding to the operation and maintenance record shall be used as the training data in the training dataset. If the importance index corresponding to the maintenance record is less than the second importance threshold, the second subsequence in the impact sequence corresponding to the maintenance record is used as the training data in the training dataset.

[0038] Among them, the first importance threshold is greater than the second importance threshold, and the sequence range of the first subsequence is greater than the sequence range of the second subsequence.

[0039] If the importance index corresponding to an operation and maintenance record is greater than or equal to the first importance threshold, it indicates that the operation and maintenance record is a high-importance record. In this case, this application can retain the complete time-series data of the impact sequence corresponding to the operation and maintenance record. If the importance index corresponding to an operation and maintenance record is greater than or equal to the second importance threshold but less than the first importance threshold, it indicates that the operation and maintenance record is a medium-importance record. In this case, this application can partially prune the impact sequence corresponding to the operation and maintenance record, retaining a portion of the time-series data of the impact sequence. If the importance index corresponding to an operation and maintenance record is less than the second importance threshold, it indicates that the operation and maintenance record is a low-importance record. In this case, this application can further prune the impact sequence corresponding to the operation and maintenance record, retaining even less time-series data of the impact sequence.

[0040] For example, the first importance threshold, the second importance threshold, and the sequence ranges of the first and second subsequences can be determined based on statistical data. For instance, the first importance threshold can be 0.8, the second importance threshold can be 0.5, the sequence range of the first subsequence can be within 12 hours before and after the fault time in the impact sequence corresponding to the operation and maintenance record, and the sequence range of the second subsequence can be within 1 hour before and after the fault time in the impact sequence corresponding to the operation and maintenance record.

[0041] By using the above methods, the proportion of redundant data in the training data is reduced as a whole, while ensuring the integrity of high-value data.

[0042] Then, the model parameters of the low-rank adaptation layer in the large industry model are trained based on the training dataset.

[0043] Since the industry-wide models are pre-trained models, their main parameters already possess powerful general language understanding and generation capabilities, eliminating the need to retrain all parameters during domain-specific fine-tuning. Therefore, this application utilizes Low-Rank Adaptation (LoRA) technology to train only the parameters of the newly added low-rank adaptation layer, freezing the main model parameters, thereby significantly reducing computational resource consumption and training time while ensuring fine-tuning effectiveness.

[0044] For example, low-rank adaptation layers can be inserted into the attention and feedforward layers of large industry models, approximating the effect of full parameter fine-tuning through low-rank matrix factorization (e.g., decomposing the weight matrix into the product of two low-rank matrices). Training can employ standard gradient descent optimization algorithms, updating the model parameters of the low-rank adaptation layer by minimizing the loss function between the predicted output and the true label.

[0045] Based on the above technical solution, this application collects multi-source heterogeneous data from various power plants and performs correlation analysis on the multi-source heterogeneous data from various power plants to obtain the impact sequence corresponding to each operation and maintenance record, thereby uncovering the inherent temporal causal relationship among meteorology, power generation, and operation and maintenance. Subsequently, similarity analysis is performed on the impact sequence corresponding to each operation and maintenance record to determine the importance index corresponding to each operation and maintenance record, achieving a quantitative assessment of the data training value. Finally, the impact sequence is processed based on the importance index corresponding to each operation and maintenance record, and the resulting training dataset is used to train a large-scale industry model. This allows for the retention of high-value data and the removal of redundant information based on importance indicators, reducing the length of training data while ensuring data quality. This effectively solves the problem of low dataset quality caused by data cleaning fragmenting business causal relationships in existing technologies, significantly improving the accuracy of the large-scale industry model in key tasks such as power generation prediction and fault early warning.

[0046] As a possible embodiment of this application, the above-mentioned "contextual analysis of multi-source heterogeneous data from various power plants to obtain the impact sequence corresponding to each operation and maintenance record" can be achieved through the following steps: Step 201: Based on meteorological time series data and power generation time series data, determine the key geographical parameters of each power station from the geographical environment parameters.

[0047] Due to differences in geographical environment, photovoltaic power plants in different regions exhibit significant variations in meteorological change characteristics and power generation response patterns. This application analyzes the correlation between meteorological time-series data and power generation time-series data to identify the geographical environmental parameters that have the most significant impact on power generation, thereby determining the key geographical parameters for each power plant and providing geographical background characteristics for subsequent construction of impact sequences.

[0048] Step 202: For each maintenance record, extract the associated time series data from the meteorological time series data and the power generation time series data according to the fault time corresponding to the maintenance record.

[0049] The fault timestamp in the operation and maintenance records is a key node connecting meteorological, power generation, and operation and maintenance data. This application can extend a certain time range forward and backward from the fault time to extract the complete meteorological and power generation changes before and after the fault, thereby constructing correlated time-series data that reflects the entire fault evolution process.

[0050] For example, taking the fault time corresponding to the maintenance record as the center, and taking preset time periods (e.g., 24 hours) before and after, the meteorological time-series data and power generation time-series data within the corresponding time period of the power plant are extracted as associated time-series data. Taking a time-series data sampling interval of 10 minutes as an example, this time period includes 144 time points, and the associated time-series data includes the meteorological data and power generation data corresponding to the 144 time points. The meteorological data may include one or more meteorological parameters, and the power generation data may include one or more meteorological parameters.

[0051] Step 203: Combine the key geographical parameters of the power station where the operation and maintenance record is located, the associated time series data, and the fault type corresponding to the operation and maintenance record to form the impact sequence corresponding to the operation and maintenance record.

[0052] By associating and combining key geographic parameters, related time-series data, and fault types, an impact sequence that includes geographic background, meteorological change process, power generation response process, and fault results can be constructed, thus fully characterizing the temporal causal relationship between meteorology, power generation, and operation and maintenance.

[0053] Based on the above technical solution, this application determines the key geographical parameters of each power station from geographical environmental parameters based on meteorological time-series data and power generation time-series data, thereby identifying the geographical environmental parameters that have the most significant impact on power generation. Then, for each operation and maintenance record, related time-series data is extracted from meteorological time-series data and power generation time-series data according to the fault time corresponding to the operation and maintenance record, thereby obtaining the complete meteorological and power generation change process before and after the fault occurs. Finally, the key geographical parameters of the power station where the operation and maintenance record is located, the related time-series data, and the fault type corresponding to the operation and maintenance record are associated and combined to form the impact sequence corresponding to the operation and maintenance record. This can completely characterize the temporal causal relationship between geographical environment, meteorological changes, power generation response, and fault results, effectively retain high-quality data containing strong correlations, and improve the ability of the impact sequence to characterize fault patterns.

[0054] As a possible embodiment of this application, the above-mentioned "determining the key geographical parameters of each power station from geographical environmental parameters based on meteorological time-series data and power generation time-series data" can be achieved through the following steps: Step 301: Divide the meteorological time series data and power generation time series data into windows with a preset duration to determine the meteorological sequence segment and power generation sequence segment of each power station within each time window.

[0055] Taking photovoltaic systems as an example, the power generation of new energy power generation systems is usually directly affected by meteorological conditions, but this impact often has a certain time lag and persistence. By dividing continuous time series data into fixed-length windows, the complete cycle from the occurrence of meteorological processes to their impact on power generation can be captured, thereby analyzing the time series correlation pattern between meteorology and power generation.

[0056] For example, the preset duration can be set to 6 hours to ensure that the divided sequence segments can cover the complete cycle of most meteorological processes (such as a complete cloud passage or a rainfall event) from occurrence to dissipation. The window sliding step size can be set to 3 hours to ensure continuity between windows.

[0057] Step 302: For each time window of each power station, construct the cumulative distance matrix between the meteorological sequence segment and the power generation sequence segment within the time window.

[0058] The cumulative distance matrix is ​​used to characterize the cumulative temporal differences between meteorological sequence segments and power generation sequence segments within a time window.

[0059] Because there is a complex nonlinear time series relationship between meteorological changes and power generation response, direct point-to-point comparison is difficult to accurately characterize the correlation pattern between the two. Therefore, this application can construct a cumulative distance matrix to capture the optimal alignment path between the meteorological sequence and the power generation sequence, thereby quantifying the cumulative difference between the two in time series.

[0060] In some embodiments, this application may employ the Dynamic Time Warping (DTW) algorithm to construct the cumulative distance matrix.

[0061] For example, the cumulative distance matrix satisfies the following formula:

[0062] in, For the meteorological sequence segment up to the [number]th The first time point and the power generation sequence segment The cumulative distance at each point in time. The first segment in the meteorological sequence The values ​​corresponding to each time point The first segment in the power generation sequence The values ​​corresponding to each time point For the meteorological sequence segment up to the [number]th The first time point and the power generation sequence segment The cumulative distance at each point in time. For the meteorological sequence segment up to the [number]th The first time point and the power generation sequence segment The cumulative distance at each point in time. For the meteorological sequence segment up to the [number]th The first time point and the power generation sequence segment The cumulative distance value at each point in time. This is a function that takes the minimum value.

[0063] In the above formula, and These represent the sampled values ​​of meteorological and power generation sequence segments at corresponding time points, respectively. Their physical meaning is the normalized value of specific meteorological parameters (such as irradiance and temperature) and power generation parameters (such as power generation and module temperature). This means that when calculating the current cumulative distance, the path with the minimum cumulative cost among the three directions of top, left, or top left is selected for extension, and the optimal alignment between meteorological sequence segments and power generation sequence segments is found, allowing the sequence to be flexibly stretched on the time axis to match similar waveform patterns.

[0064] The element value corresponding to the lower right corner of the final cumulative distance matrix is ​​the DTW distance, which reflects the overall difference between the two sequences. The internal structure of the matrix records the temporal alignment relationship between meteorological and power generation, which can characterize the temporal influence relationship between meteorological changes and power generation output.

[0065] Step 303: Cluster the meteorological sequence segments of each power station within each time window to obtain multiple meteorological clusters.

[0066] Since the meteorological characteristics of photovoltaic power plants vary significantly in different regions, directly mixing and analyzing all the data would introduce geographical noise. Therefore, this application can eliminate the interference of geographical environment by grouping power plants with similar meteorological characteristics into the same category and conducting refined analysis within the category.

[0067] For example, for any two meteorological sequence segments and DTW is used to calculate the difference distance between them. (Only meteorological parameters, such as irradiance and temperature, are used here). Using this difference distance as the distance metric, the K-medoids clustering algorithm is employed to cluster all meteorological segments into several clusters (the optimal number of clusters can be determined using the elbow rule). This results in multiple meteorological clusters.

[0068] Step 304: For each meteorological cluster, determine the core influence matrix of the meteorological cluster based on the cumulative distance matrix of each power station in each time window within the meteorological cluster.

[0069] Among them, the core influence matrix is ​​used to characterize the core influence patterns between meteorological data and power generation data within a meteorological cluster.

[0070] Within the same meteorological feature cluster, the cumulative distance matrices of different power station windows reflect their respective meteorological-power generation impact patterns, among which some representative core patterns exist. This application can extract the most representative core impact matrix that best reflects the meteorological-power generation impact pattern of each meteorological cluster from the cumulative distance matrices corresponding to the numerous power station windows within the cluster, providing core features for subsequent geographic parameter analysis.

[0071] Step 305: Based on the geographical environment parameters of each power station within the meteorological cluster and the core influence matrix of the meteorological cluster, perform regression analysis to determine the key geographical parameters of each power station.

[0072] The differences in the core influence matrix exhibited by different cluster groups are usually related to the geographical environment of the power plant. This application can identify the key geographical parameters that cause the differences by analyzing the correlation between the geographical environment and the core influence matrix.

[0073] For example, common regional characteristics of power stations within each meteorological cluster can be collected as independent variables (e.g., altitude, terrain type, latitude and longitude), and the lower right corner value of the core influence matrix of the corresponding window of these power stations (i.e., DTW distance, representing the overall influence strength of the meteorological-power generation pair) can be used as the dependent variable to train a random forest regression model. This model outputs the importance score of each geographic parameter, and selects the top-ranked geographic parameters as key geographic parameters. For example, the three geographic parameters with the highest importance scores can be selected as key geographic parameters.

[0074] Based on the above technical solution, this application can capture the complete cycle of meteorological processes from occurrence to impact on power generation by dividing meteorological time-series data and power generation time-series data into windows with preset durations. Then, a cumulative distance matrix is ​​constructed for each time window of each power station, which can quantify the cumulative difference between meteorological and power generation sequences in time, revealing the temporal impact relationship between meteorological changes and power output. At the same time, multiple meteorological clusters are obtained by clustering meteorological sequence segments of each power station in each time window, which can group power station time periods with similar meteorological characteristics into the same category and eliminate the interference of geographical environment. On this basis, a core impact matrix is ​​determined for each meteorological cluster based on the cumulative distance matrix, which can then extract the most representative meteorological-power generation correlation features within the cluster. Finally, regression analysis is performed based on geographical environment parameters and the core impact matrix to determine the key geographical parameters of each power station, thereby identifying the geographical parameters that have the most significant impact on the meteorological-power generation correlation pattern, providing accurate geographical background features for subsequent construction of impact sequences, and effectively improving the accuracy and pertinence of geographical parameter analysis.

[0075] As a possible embodiment of this application, the above-mentioned "determining the core influence matrix of a meteorological cluster based on the cumulative distance matrix of each power station within the meteorological cluster for each time window" can be achieved through the following steps: Step 401: For each meteorological cluster, perform local variance analysis on each cumulative distance matrix in the meteorological cluster to determine the weight basis matrix corresponding to each cumulative distance matrix.

[0076] The weighted basis matrix is ​​used to characterize the distribution of importance of each local region in the cumulative distance matrix. Different regions of the cumulative distance matrix reflect the degree of difference between meteorology and power generation at different time-series alignment positions. Regions with larger variances often contain richer dynamic change information and are of greater importance for characterizing the meteorology-power generation correlation pattern. Key regions in the matrix can be identified through local variance analysis.

[0077] In some embodiments, this application can treat each cumulative distance matrix as a grayscale image, calculate the local variance using a sliding window, and normalize the local variances at each location to obtain the weighted basis matrix. For example, the sliding window size is set to 5×5. Then, the local variance within each sliding window is used as the local variance corresponding to the center position of that sliding window, and the local variances at each position are normalized to obtain the weighted basis matrix.

[0078] For example, the weighted basis matrix can be calculated in the following way:

[0079] in, For the first The cumulative distance matrix in the meteorological sequence segment is the first... The first time point and the power generation sequence segment The weight value corresponding to each time point. For the first The cumulative distance matrix in the meteorological sequence segment is the first... The first time point and the power generation sequence segment The local variance of the location corresponding to each time point. This is a normalization function (e.g., maximum / minimum normalization) used to map the calculation results to the range of 0 to 1.

[0080] in, Characterizing the first The cumulative distance matrix in the meteorological sequence segment is the first... The first time point and the power generation sequence segment The degree of numerical dispersion in the neighborhood of a given time point reflects the fluctuation of the meteorological-power generation alignment cost in that local area. The larger the variance, the richer the dynamic change information contained in the area, and the larger the weight value of the corresponding position after normalization.

[0081] Step 402: Construct a mask consensus graph for the meteorological clusters based on the weight basis matrix corresponding to each cumulative distance matrix in the meteorological clusters.

[0082] Among them, the mask consensus graph is used to characterize the distribution of consensus regions among the cumulative distance matrices in the meteorological cluster.

[0083] Within the same meteorological feature cluster, the cumulative distance matrices of different power plant windows typically exhibit a certain degree of consistency in key areas. This consistency reflects the generally applicable meteorological-power generation correlation patterns within that meteorological feature cluster. By identifying these consensus areas, it is possible to distinguish between universally significant correlation features and the specific circumstances of individual power plants.

[0084] In some embodiments, the mean and variance of all weight basis matrices within a cluster can be calculated position by position to obtain the average weight and weight consensus degree of each position (weight consensus degree is defined as the negative correlation normalization of the variance of all weight basis matrices at the corresponding position, that is, the smaller the variance, the higher the weight consensus degree. For example, the variance can be normalized to between 0 and 1 by normalizing the maximum and minimum values ​​first, and then negative correlation normalization can be achieved by subtracting the normalization value from 1).

[0085] Next, locations where the average weight is greater than the mean threshold and the weight consensus degree is greater than the consensus degree threshold can be marked as "high consensus regions," generating a masked consensus graph. In the masked consensus graph, 1 indicates that the corresponding location is a high consensus region, and 0 indicates that the corresponding location is a non-high consensus region. For example, the mean threshold and consensus degree threshold can be determined based on statistical analysis, such as setting both the mean threshold and consensus degree threshold to 0.6.

[0086] Step 403: Based on the mask consensus graph of the meteorological cluster, perform weighted calculation on the weight basis matrix corresponding to each cumulative distance matrix in the meteorological cluster to obtain the weighted weight matrix corresponding to each cumulative distance matrix in the meteorological cluster.

[0087] The weighted weight matrix is ​​used to characterize the importance weight distribution of the cumulative distance matrix after fusing consensus information within the cluster.

[0088] By combining the mask consensus graph to weight the basis matrix, the importance of the consensus region within the cluster can be strengthened while the characteristics of individual matrices are preserved, and the regions with large individual differences can be weakened, so that the weight distribution can better reflect the overall association pattern of the cluster.

[0089] For example, the weighted matrix satisfies the following formula:

[0090] in, The weighted weight matrix is ​​the first... The cumulative distance matrix in the meteorological sequence segment is the first... The first time point and the power generation sequence segment The weighted weights corresponding to the positions at each time point. For the first The cumulative distance matrix in the meteorological sequence segment is the first... The first time point and the power generation sequence segment The weight value corresponding to each time point. The weight basis matrix of all elements in the cluster is the first element in the meteorological sequence segment. The first time point and the power generation sequence segment The average weight at each time point corresponding to the location For the mask consensus graph, the first The cumulative distance matrix in the meteorological sequence segment is the first... The first time point and the power generation sequence segment The mask value corresponding to each time point. The weight basis matrix of all elements in the cluster is the first element in the meteorological sequence segment. The first time point and the power generation sequence segment The consensus degree of the weight at each point in time.

[0091] In the above formula This characterizes the general importance of the location within the cluster. The weighted average is used to represent the degree of consensus within the cluster. For regions with high consensus, the average weight is used to strengthen the region and highlight the association features that are generally applicable within the cluster. For regions with low consensus, the consensus degree is used to modulate the region, preserving individual features while considering consistency. The resulting weighted average reflects both the local features of the individual matrix and the overall pattern within the cluster.

[0092] Step 404: Based on the weighted weight matrix corresponding to each cumulative distance matrix in the meteorological cluster, conduct an impact assessment on each cumulative distance matrix and determine the core impact matrix of the meteorological cluster from each cumulative distance matrix in the meteorological cluster.

[0093] In one possible implementation, this application can perform similarity analysis on each cumulative distance matrix in the meteorological cluster, based on the cumulative distance matrix and the weighted weight matrix corresponding to the cumulative distance matrix, with other cumulative distance matrices and their corresponding weighted weight matrices in the meteorological cluster, to determine the average similarity coefficient between the cumulative distance matrix and other cumulative distance matrices in the meteorological cluster.

[0094] Within the same meteorological feature cluster, if a certain cumulative distance matrix can well represent the meteorological-power generation correlation pattern of the cluster, then the cumulative distance matrix usually has a high degree of similarity with other cumulative distance matrices in the cluster. This application can quantify the representativeness of each cumulative distance matrix by calculating the average similarity coefficient.

[0095] In some embodiments, the weighted 2D-DTW algorithm can be used to calculate the weighted difference distance matrix between two cumulative distance matrices.

[0096] For example, the weighted difference distance matrix satisfies the following formula:

[0097] in, For the cumulative distance matrix In the meteorological sequence segment, the first The first time point and the power generation sequence segment Matrix of location and cumulative distance at each time point In the meteorological sequence segment, the first The first time point and the power generation sequence segment Weighted difference distance at each time point Cumulative distance matrix Up to the meteorological sequence segment The first time point and the power generation sequence segment The cumulative distance at each point in time. Cumulative distance matrix Up to the meteorological sequence segment The first time point and the power generation sequence segment The cumulative distance at each point in time. The cumulative distance matrix in the weighted weight matrix In the meteorological sequence segment, the first The first time point and the power generation sequence segment The weighted weights corresponding to the positions at each time point. The cumulative distance matrix in the weighted weight matrix In the meteorological sequence segment, the first The first time point and the power generation sequence segment The weighted weights corresponding to the positions at each time point.

[0098] and These respectively reflect the degree of difference between meteorological and power generation at the time-series alignment position of the two cumulative distance matrices. and The importance weights for corresponding positions are used to enhance the difference calculation of key regions. When calculating the difference between two cumulative distance matrices, regions identified as important (high weight) are given a greater influence, while less important regions (low weight) are given a smaller influence, thus making the difference calculation more focused on key correlation features.

[0099] The lower right corner value of the weighted difference distance matrix constructed above is the weighted 2D-DTW distance between the two cumulative distance matrices. This application can perform negative correlation normalization on all the calculated weighted 2D-DTW distances (for example, the weighted 2D-DTW distance can be normalized to between 0 and 1 by first normalizing the maximum and minimum values, and then negative correlation normalization can be achieved by subtracting the normalization value from 1) to obtain the similarity between the two cumulative distance matrices, and then determine the average similarity coefficient between the cumulative distance matrix and other cumulative distance matrices in the meteorological cluster.

[0100] For example, the average similarity coefficient satisfies the following formula:

[0101] in, The first in the meteorological cluster The average similarity coefficient between the cumulative distance matrix and other cumulative distance matrices in the meteorological cluster. This represents the number of cumulative distance matrices in the meteorological cluster. The first in the meteorological cluster The cumulative distance matrix and the first Similarity between cumulative distance matrices.

[0102] Thus, this application can use the cumulative distance matrix with the largest average similarity coefficient among the cumulative distance matrices in the meteorological cluster as the core influence matrix of the meteorological cluster.

[0103] Among them, the cumulative distance matrix with the largest average similarity coefficient indicates that it has the highest similarity with other cumulative distance matrices within the cluster, and best represents the common meteorological-power generation correlation pattern within the meteorological cluster. Therefore, it is suitable as the core influence matrix. If there are multiple cumulative distance matrices whose average similarity coefficients are the maximum values ​​within the meteorological cluster, then any one of these cumulative distance matrices can be selected as the core influence matrix of the meteorological cluster.

[0104] Based on the above technical solution, this application determines the weighted basis matrix corresponding to each cumulative distance matrix by performing local variance analysis on each cumulative distance matrix in the meteorological cluster. This allows for the identification of the importance distribution of each local region in the cumulative distance matrix and the location of key regions containing rich dynamic information. Subsequently, a mask consensus map of the meteorological cluster is constructed based on the weighted basis matrix, which can characterize the consensus region distribution among the cumulative distance matrices in the meteorological cluster and identify common features reflecting general correlation patterns. Thus, a weighted weight matrix can be obtained by weighting the weighted basis matrix based on the mask consensus map, thereby integrating intra-cluster consensus information and making the weight distribution more reflective of the overall correlation pattern of the cluster. Finally, an impact assessment is performed on each cumulative distance matrix based on the weighted weight matrix to determine the core impact matrix, thereby extracting the most representative meteorological-power generation correlation pattern within the meteorological feature cluster and effectively improving the accuracy and representativeness of the core impact pattern extraction.

[0105] As a possible embodiment of this application, the above-mentioned "conducting similarity analysis based on the impact sequence corresponding to each operation and maintenance record to determine the importance index corresponding to each operation and maintenance record" can be achieved through the following steps: Step 501: Divide the impact sequence corresponding to each maintenance record into multiple impact sequence sets according to the fault type.

[0106] Different types of faults (such as module hotspots, inverter faults, and shading faults) typically have different causal mechanisms and evolutionary patterns, and their corresponding impact sequences will also exhibit different characteristic patterns in meteorological, power generation, and geographical dimensions. Therefore, by classifying faults according to their types, this application can group impact sequences with similar causes and patterns into one category, thereby comparing similarity within the same category and identifying typical patterns.

[0107] For example, this application can divide all impact sequences into multiple impact sequence sets based on the fault type field in the operation and maintenance records, with each impact sequence set corresponding to a fault type.

[0108] Step 502: For each operation and maintenance record in each set of impact sequences, perform a similarity analysis between the impact sequence corresponding to the operation and maintenance record and the impact sequences corresponding to other operation and maintenance records in the set of impact sequences to determine the importance indicators corresponding to the operation and maintenance record.

[0109] Within the same set of fault types, the impact sequences corresponding to different maintenance records may differ. Some impact sequences can well reflect the typical patterns of the fault type and are highly representative, while others may be affected by accidental factors or recording errors, resulting in weaker regularity. This application quantifies the typicality of each impact sequence through similarity analysis.

[0110] In one possible implementation, this application can calculate the feature distance between the influence sequence corresponding to each operation and maintenance record in each influence sequence set and the influence sequences corresponding to other operation and maintenance records in the influence sequence set.

[0111] For example, this application can flatten the multidimensional time series (meteorological and power generation parameters) in the influence sequence into a univariate time series, or use the multidimensional DTW algorithm to calculate the DTW distance between any two influence sequences as the feature distance.

[0112] Subsequently, the feature distances between the impact sequences corresponding to the operation and maintenance records and the impact sequences corresponding to other operation and maintenance records in the impact sequence set are negatively correlated and normalized to obtain the similarity corresponding to each feature distance. The mean of the similarity corresponding to each feature distance is then used as the importance index corresponding to the operation and maintenance records.

[0113] Since feature distance and similarity are negatively correlated (the greater the distance, the lower the similarity), negative correlation normalization can be used to convert distance into a similarity value. This can be achieved by first normalizing the feature distance to between 0 and 1 using maximum and minimum value normalization, and then subtracting this normalization value from 1 to achieve negative correlation normalization. The importance index is defined as the average similarity to other influential sequences within the set. A higher value indicates greater similarity between the sequence and other sequences of the same type, reflecting a more typical and stable pattern.

[0114] For example, the importance metric satisfies the following formula:

[0115] in, For the first The importance indicators corresponding to each maintenance record To influence the number of operation and maintenance records in the sequence set, For the first The similarity between the impact sequence corresponding to one maintenance record and the impact sequence corresponding to other maintenance records.

[0116] This reflects the degree of consistency in the patterns between these two similar influence sequences. For the first The average similarity of the impact sequence corresponding to each maintenance record in the set of fault types. The larger the value, the more similar the impact sequence is to other impact sequences of the same type, the more typical and stable the fault pattern it reflects, and the higher the data value.

[0117] Based on the above technical solution, this application calculates the feature distance between the impact sequence corresponding to each maintenance record in each impact sequence set and the impact sequences corresponding to other maintenance records in the impact sequence set. This allows for the quantification of the comprehensive differences between two impact sequences in dimensions such as geographical features, meteorological change patterns, and power generation response patterns. The feature distance is negatively correlated and normalized to obtain the similarity, and the mean of the similarity is used as an importance indicator. This transforms the distance metric into a comparable importance score, effectively quantifying the typicality and regularity of each maintenance record in the same type of fault. This provides a precise quantitative basis for subsequent graded data screening based on importance, improving the accuracy and reliability of data value assessment.

[0118] Please see Figure 2 The diagram illustrates a system architecture of an industry-wide large model training system based on efficient fine-tuning, according to an embodiment of the present invention. This industry-wide large model training system 20 based on efficient fine-tuning includes: The data acquisition unit 21 is used to collect multi-source heterogeneous data from various power plants. The multi-source heterogeneous data includes meteorological time-series data, power generation time-series data, operation and maintenance record set, and geographical environmental parameters. Each operation and maintenance record in the operation and maintenance record set is used to record the fault type and fault time of the power plant. Sequence construction unit 22 is used to perform correlation analysis on multi-source heterogeneous data of various power plants to obtain the impact sequence corresponding to each operation and maintenance record; the impact sequence is used to characterize the impact of the fault corresponding to the operation and maintenance record on meteorological time series data, power generation time series data and geographical environmental parameters. The importance determination unit 23 is used to perform similarity analysis based on the impact sequence corresponding to each operation and maintenance record to determine the importance index corresponding to each operation and maintenance record; the importance index is used to characterize the regularity and stability of the impact sequence corresponding to the operation and maintenance record in the same fault type; Model training unit 24 is used to process the impact sequence based on the importance index corresponding to each operation and maintenance record, and to train the industry large model using the training dataset obtained from the data processing.

[0119] It should be noted that the various embodiments of this application can be referenced or learned from each other. For example, the same or similar steps, method embodiments, system embodiments and device embodiments can be referenced from each other without limitation.

[0120] This application also provides a hardware structure diagram of an industry large model training device based on efficient fine-tuning (hereinafter referred to as industry large model training device based on efficient fine-tuning 30), see [link to relevant documentation]. Figure 3 The industry-leading large model training device 30 based on efficient fine-tuning includes a processor 31, and optionally, a memory 32 connected to the processor 31.

[0121] In the first possible implementation, see Figure 3 The industry-wide large-scale model training device 30 based on efficient fine-tuning also includes a communication interface 33. The processor 31, memory 32, and communication interface 33 are connected via a bus. The communication interface 33 is used to communicate with other devices or communication networks. Optionally, the communication interface 33 may include a transmitter and a receiver. The device in the communication interface 33 used to implement the receiving function can be considered as a receiver, which is used to perform the receiving steps in the embodiments of this application. The device in the communication interface 33 used to implement the transmitting function can be considered as a transmitter, which is used to perform the transmitting steps in the embodiments of this application.

[0122] Based on the first possible implementation method Figure 3 The schematic diagram shown can be used to illustrate the structure of the industry large model training device based on efficient fine-tuning involved in the above embodiments.

[0123] in, Figure 3 This can also be illustrated by the system chip in a large-scale industry model training device based on efficient fine-tuning. In this case, the actions performed by the aforementioned large-scale industry model training device based on efficient fine-tuning can be implemented by this system chip. The specific actions performed can be found above and will not be repeated here.

[0124] In implementation, each step of the method provided in this embodiment can be completed by integrated logic circuits in the processor or by instructions in software form. The steps of the method disclosed in the embodiments of this application can be directly manifested as being executed by a hardware processor, or being executed by a combination of hardware and software modules in the processor.

[0125] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0126] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A method for training large industry models based on efficient fine-tuning, characterized in that, include: Collect multi-source heterogeneous data from various power plants; Multi-source heterogeneous data includes meteorological time-series data, power generation time-series data, operation and maintenance record sets, and geographical environmental parameters; Each maintenance record in the maintenance record set is used to record the type of failure and the time of failure that occurred in the power plant; Correlation analysis was performed on the multi-source heterogeneous data of each power station to obtain the impact sequence corresponding to each operation and maintenance record; The impact sequence is used to characterize the impact of the corresponding faults in the operation and maintenance records on meteorological time-series data, power generation time-series data, and geographical environmental parameters. A similarity analysis is performed on the impact sequence corresponding to each operation and maintenance record to determine the importance index corresponding to each operation and maintenance record; Importance metrics are used to characterize the regularity and stability of the impact sequence corresponding to the operation and maintenance records within the same fault type; The impact sequence is processed based on the importance indicators corresponding to each operation and maintenance record, and the training dataset obtained from the data processing is used to train the industry-wide model. Specifically, based on meteorological time-series data and power generation time-series data, key geographical parameters for each power station are determined from geographical environmental parameters; for each operation and maintenance record, related time-series data are extracted from meteorological time-series data and power generation time-series data according to the fault time corresponding to the operation and maintenance record; the key geographical parameters of the power station where the operation and maintenance record is located, the related time-series data, and the fault type corresponding to the operation and maintenance record are associated and combined to form the impact sequence corresponding to the operation and maintenance record. The process involves dividing meteorological and power generation time-series data into windows of preset duration to determine the meteorological and power generation sequence segments for each power station within each time window. For each time window of each power station, a cumulative distance matrix is ​​constructed between the meteorological and power generation sequence segments within that window. This cumulative distance matrix characterizes the cumulative temporal differences between the meteorological and power generation sequence segments within the time window. Based on the meteorological sequence segments of each power station within each time window, multiple meteorological clusters are obtained. For each meteorological cluster, a core influence matrix is ​​determined based on the cumulative distance matrix of each power station within that cluster for each time window. This core influence matrix characterizes the core influence pattern between meteorological and power generation data within the meteorological cluster. Finally, regression analysis is performed based on the geographical environmental parameters of each power station within the meteorological cluster and the core influence matrix of the meteorological cluster to determine the key geographical parameters of each power station. Specifically, the impact sequence corresponding to each operation and maintenance record is divided into multiple impact sequence sets according to the fault type; for each operation and maintenance record in each impact sequence set, a similarity analysis is performed between the impact sequence corresponding to the operation and maintenance record and the impact sequences corresponding to other operation and maintenance records in the impact sequence set to determine the importance indicators corresponding to the operation and maintenance record; Specifically, for each operation and maintenance record in each set of impact sequences, the feature distance between the impact sequence corresponding to the operation and maintenance record and the impact sequences corresponding to other operation and maintenance records in the set is calculated. The feature distance between the impact sequence corresponding to the operation and maintenance record and the impact sequences corresponding to other operation and maintenance records in the set is negatively correlated and normalized to obtain the similarity corresponding to each feature distance. The mean of the similarity corresponding to each feature distance is used as the importance index corresponding to the operation and maintenance record.

2. The industry-wide large model training method based on efficient fine-tuning according to claim 1, characterized in that, For each meteorological cluster, the core influence matrix of the meteorological cluster is determined based on the cumulative distance matrix of each power station within the cluster for each time window, including: For each meteorological cluster, a local variance analysis is performed on each cumulative distance matrix in the meteorological cluster to determine the weighted basis matrix corresponding to each cumulative distance matrix; the weighted basis matrix is ​​used to characterize the importance distribution of each local region in the cumulative distance matrix; Based on the weight basis matrix corresponding to each cumulative distance matrix in the meteorological cluster, a mask consensus graph of the meteorological cluster is constructed; the mask consensus graph is used to characterize the consensus region distribution among the cumulative distance matrices in the meteorological cluster. Based on the mask consensus graph of the meteorological cluster, the weight basis matrix corresponding to each cumulative distance matrix in the meteorological cluster is weighted and calculated to obtain the weighted weight matrix corresponding to each cumulative distance matrix in the meteorological cluster; the weighted weight matrix is ​​used to characterize the importance weight distribution of the cumulative distance matrix after fusing consensus information within the cluster; The impact of each cumulative distance matrix is ​​evaluated based on the weighted weight matrix corresponding to each cumulative distance matrix in the meteorological cluster, and the core impact matrix of the meteorological cluster is determined from each cumulative distance matrix in the meteorological cluster.

3. The industry-wide large model training method based on efficient fine-tuning according to claim 2, characterized in that, An impact assessment is performed on each cumulative distance matrix based on the weighted weight matrix corresponding to each cumulative distance matrix in the meteorological cluster. The core impact matrix of the meteorological cluster is then determined from the cumulative distance matrices in the meteorological cluster, including: For each cumulative distance matrix in the meteorological cluster, based on the cumulative distance matrix and the weighted weight matrix corresponding to the cumulative distance matrix, a similarity analysis is performed with other cumulative distance matrices and their corresponding weighted weight matrices in the meteorological cluster to determine the average similarity coefficient between the cumulative distance matrix and other cumulative distance matrices in the meteorological cluster. The cumulative distance matrix with the largest average similarity coefficient among the cumulative distance matrices in the meteorological cluster is taken as the core influence matrix of the meteorological cluster.

4. The industry-wide large model training method based on efficient fine-tuning according to claim 1, characterized in that, The impact sequence is processed based on the importance indicators corresponding to each operation and maintenance record, and the resulting training dataset is used to train a large industry model, including: For each operation and maintenance record, the impact sequence corresponding to the operation and maintenance record is pruned according to the importance index corresponding to the operation and maintenance record, and the pruned impact sequence of each operation and maintenance record is used as the training dataset; The model parameters of the low-rank adaptation layer in the large industry model are trained based on the training dataset.

5. A training device for large industry models based on efficient fine-tuning, characterized in that, include: A processor and a communication interface; the communication interface is coupled to the processor, the processor being used to run computer programs or instructions to implement the industry large model training method based on efficient fine-tuning as described in any one of claims 1-4.

6. A large-scale industry model training system based on efficient fine-tuning, characterized in that, The method for training large industry models based on efficient fine-tuning as described in any one of claims 1-4 includes: The data acquisition unit is used to collect multi-source heterogeneous data from various power plants. The multi-source heterogeneous data includes meteorological time-series data, power generation time-series data, operation and maintenance record sets, and geographical environmental parameters. Each operation and maintenance record in the operation and maintenance record set is used to record the fault type and fault time of the power plant. The sequence construction unit is used to perform correlation analysis on the multi-source heterogeneous data of each power station to obtain the impact sequence corresponding to each operation and maintenance record; the impact sequence is used to characterize the impact of the fault corresponding to the operation and maintenance record on meteorological time series data, power generation time series data and geographical environmental parameters. The importance determination unit is used to perform similarity analysis based on the impact sequence corresponding to each operation and maintenance record to determine the importance index corresponding to each operation and maintenance record; the importance index is used to characterize the regularity and stability of the impact sequence corresponding to the operation and maintenance record in the same fault type; The model training unit is used to process the impact sequence based on the importance index corresponding to each operation and maintenance record, and to train the industry-wide large model using the training dataset obtained from the data processing.

Images