A model fusion-based intelligent mapping method for multi-source heterogeneous data of a wind farm

Abstract

The application provides a kind of model fusion-based wind farm multi-source heterogeneous data intelligent mapping method, it is related to power system automation and data processing technical field, including the following steps: constructing training set, verification set and corpus;Construct semantic similarity function, train the preset linear mapping parameter in semantic similarity function, obtain the similarity calculation model based on semantics;Construct the similarity calculation model based on statistical characteristics;Calculate fusion weight;The similarity calculation model based on semantics and the similarity calculation model based on statistical characteristics are weightedly fused;The similarity value of the source data to be mapped and each mappable data is calculated by the fused similarity calculation model, select several data with the highest similarity value, and output to the preset configuration tool interface through the output interface.The application solves the practical pain points in the current intelligent development of wind farm.

Description

Technical Field

[0001] This invention relates to the field of power system automation and data processing technology, and more specifically, to a method for intelligent mapping of multi-source heterogeneous data from wind farms based on model fusion. Background Technology

[0002] With the rapid development of the wind power industry, the level of informatization and intelligence in wind farms is constantly improving. In actual production, the central control room of a wind farm typically deploys multiple systems, including monitoring systems, operation and maintenance systems, wind turbine manufacturer-supported systems, online monitoring systems, and video monitoring systems. Furthermore, a single wind farm or multiple wind farms within a region often contain wind turbine equipment from multiple manufacturers (such as Mingyang, Haizhuang, Goldwind, Envision Energy, etc.) and various models of wind turbine units with different rated power.

[0003] This complex environment involving multiple systems, manufacturers, and models leads to a prominent problem: the same physical quantity (such as "generator active power") has different data names and measurement point names in different systems and wind turbines from different manufacturers. For example, generator active power may be named in multiple ways. This naming heterogeneity makes the data in the central control room resemble a "foreign expo" or a group of "information islands," creating huge obstacles to data integration and sharing.

[0004] On the other hand, the current number of signal connection points at wind farms is enormous. Taking a 300MW wind farm monitoring system as an example, although the number varies depending on the type of wind turbine, 50,000 to 60,000 connection points are common. As offshore wind farms expand into deeper waters, the installed capacity of single wind farms continues to increase. Meanwhile, the new energy sector currently lacks standards similar to IEC 61850 in the power transmission and transformation field, which can be parsed and configured using a unified SCD file.

[0005] Therefore, throughout the entire lifecycle of a wind farm, in actual engineering projects involving pre-commissioning system configuration, intelligent model reuse, system upgrades, and system migration or replacement, data mapping and measurement point association often rely on manual configuration by engineers. A common approach is for the system vendor to develop configuration tools for engineers, who then manually perform data mapping and association by searching through a tree of nodes, dragging and dropping, or clicking, supplemented by shortcuts such as keyword search and serial number filling. This traditional method is labor-intensive, error-prone, and inefficient.

[0006] Meanwhile, with the rapid development of technologies such as big data and artificial intelligence, the application of natural language processing (NLP) technology is also expanding, for example: extracting opinions from comments, writing news, named text recognition, and determining semantic similarity. According to published literature, the application of NLP technology in the power sector is also gradually developing, and there is some use of semantic models such as word vectors to calculate text similarity. However, existing methods have limited effectiveness in specific power sectors, and even fewer applications in engineering configuration.

[0007] Currently, in complex environments with multiple systems, manufacturers, and models, inconsistencies between data names and measurement point names lead to difficulties in data mapping. Furthermore, in practical engineering applications, manual configuration and association are primarily used. This situation presents a series of problems, including:

[0008] (1) Difficulty in reusing intelligent models; The application of intelligent technology in wind farms is becoming increasingly rich, including fault early warning models, image recognition models, sub-component life prediction models, meteorological prediction models, wind power prediction models, etc. When these models are reused among wind turbines from different manufacturers and of different models, the inconsistent naming of measurement points requires a large amount of manual configuration of data mapping relationships, which is a huge workload and prone to errors.

[0009] (2) High system upgrade and migration costs; Throughout the entire life cycle of a wind farm, monitoring systems, information systems, and advanced application systems require system upgrades, migrations, or even replacements. Each system change requires reconfiguring data mapping relationships, which consumes a lot of manpower and time, and it is difficult to guarantee accuracy.

[0010] Secondly, across multiple scenarios throughout the wind farm's lifecycle, the engineering configuration required for data mapping due to inconsistencies between data and measurement point names is enormous. Current intelligent technologies still face numerous challenges in mapping and associating heterogeneous data from wind farms.

[0011] (1) Poor adaptability of general semantic models; pre-trained models such as BERT are trained on general corpora, and their output vector space is not optimized for wind power measurement point name matching tasks. When such vectors are used as input to calculate cosine similarity, they cannot effectively distinguish the degree difference between "severe wear" and "slight wear", and they are also difficult to handle the situation where "P_Gen" and "ActivePower" have completely different word meanings but the same meaning.

[0012] (2) Semantic information and statistical information are separated; pure semantic models cannot capture statistical features such as word frequency and keyword importance; pure statistical models (such as TF-IDF) cannot understand semantic relationships. Both have limitations when used alone, resulting in limited accuracy of similarity judgment.

[0013] (3) Lack of adaptive capability; a single model uses fixed parameters and cannot make adaptive adjustments according to the data characteristics of different wind fields, resulting in large differences in performance at different stations.

[0014] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0015] In view of this, the present invention provides a model fusion-based intelligent mapping method for multi-source heterogeneous data of wind farms to solve the aforementioned problems.

[0016] To solve the above problems, the specific technical solution adopted by the present invention is as follows:

[0017] According to a first aspect of the present invention, a method for intelligent mapping of multi-source heterogeneous data from wind farms based on model fusion is provided, comprising the following steps:

[0018] S1. Collect historical monitoring data of wind farms, and construct a training set for parameter learning, a validation set for hyperparameter tuning and fusion weight calculation, and a corpus for term distribution statistics.

[0019] S2. Use the pre-trained BERT model to perform text encoding processing on the training set to obtain high-dimensional semantic vectors. Combine the cosine similarity function and the preset linear mapping parameters to construct a semantic similarity function. Use the triplet loss function that combines the triplet construction rules of wind power scenario to train the preset linear mapping parameters in the semantic similarity function to obtain a semantic similarity calculation model.

[0020] S3. Based on the corpus, construct statistical feature vectors using the entropy weight method, and construct a similarity calculation model based on statistical features based on the statistical feature vectors.

[0021] S4. Using the validation set, evaluate the performance of the semantic-based similarity calculation model and the statistical feature-based similarity calculation model respectively, and calculate the fusion weight of the semantic-based similarity calculation model and the statistical feature-based similarity calculation model according to the performance evaluation results.

[0022] S5. Based on the fusion weight, the semantic-based similarity calculation model and the statistical feature-based similarity calculation model are weighted and fused to obtain a fusion similarity calculation model for calculating fusion similarity.

[0023] S6. Encapsulate the fusion similarity calculation model into a standardized service plugin, and configure the input and output interfaces for the service plugin;

[0024] S7. Obtain the source data to be mapped and input it into the fusion similarity calculation model through the input interface. Calculate the similarity value between the source data to be mapped and each mappable data through the fusion similarity calculation model. Select the data with the highest similarity value and output it to the preset configuration tool interface through the output interface.

[0025] Preferably, the step of performing text encoding processing on the training set using a pre-trained BERT model to obtain high-dimensional semantic vectors, and constructing a semantic similarity function by combining a cosine similarity function and preset linear mapping parameters, and training the preset linear mapping parameters in the semantic similarity function using a triplet loss function that combines the triplet construction rules of wind power scenarios, to obtain a semantic-based similarity calculation model includes the following steps:

[0026] S21. Use a word segmenter to segment and encode the training set to obtain the input tensor;

[0027] S22. Use the pre-trained BERT model as the base encoder, and input the input tensor into the base encoder. Through text encoding processing, obtain the high-dimensional semantic vector.

[0028] S23. Add linear mapping parameters that can be obtained through training to the high-dimensional semantic vector, and use the linear mapping parameters to optimize the cosine similarity function to obtain the semantic similarity function.

[0029] S24. Based on the engineering characteristics of wind farms, construct the triplet construction rules for wind power scenarios, and combine them with preset training parameters. Through the triplet loss function, train the preset linear mapping parameters to obtain a semantic similarity calculation model.

[0030] Preferably, the semantic-based similarity calculation model is used to calculate the semantic similarity value; the formula for calculating the semantic similarity value is:

[0031] ;

[0032] In the formula, h_A and h_B are high-dimensional semantic vectors output by the BERT model, W and b are linear mapping parameters obtained during training, and Sim_sem(A,B) is the semantic similarity value.

[0033] Preferably, the step of constructing a statistical feature vector using the entropy weight method based on the corpus, and constructing a similarity calculation model based on statistical features based on the statistical feature vector, includes the following steps:

[0034] S31. Perform statistics on each term in the corpus and construct a term-document matrix based on the statistical results;

[0035] S32. Based on the term-document matrix, calculate the information entropy of each term and the weight of each term based on the information entropy;

[0036] S33. Construct a statistical feature vector based on entropy weight according to the weight of each term;

[0037] S34. Combine the entropy-weighted statistical feature vector with the cosine similarity formula to form a similarity calculation model based on statistical features.

[0038] Preferably, the statistical feature-based similarity calculation model is used to calculate the statistical feature similarity value, and the formula for calculating the statistical feature similarity value is:

[0039] ;

[0040] In the formula, Sim_sta(A,B) represents the statistical feature similarity value, and V A V is the statistical characteristic vector of the measurement point name A. B Let B be the statistical characteristic vector of the measurement point.

[0041] Preferably, the step of using a validation set to evaluate the performance of the semantic-based similarity calculation model and the statistical feature-based similarity calculation model, and calculating the fusion weights of the semantic-based similarity calculation model and the statistical feature-based similarity calculation model based on the performance evaluation results, includes the following steps:

[0042] S41. Based on preset complementary indicators, calculate the performance data of the semantic similarity calculation model and the statistical feature similarity calculation model on the validation set respectively.

[0043] S42. Based on the preset weight adjustment factor, calculate the comprehensive evaluation performance score of the semantic similarity calculation model and the statistical feature-based similarity calculation model respectively.

[0044] S43. Normalize the comprehensive evaluation performance scores of the semantic-based similarity calculation model and the statistical feature-based similarity calculation model to obtain the fusion weight of the semantic-based similarity calculation model and the fusion weight of the statistical feature-based similarity calculation model.

[0045] Preferably, the expression for the fusion similarity calculation model is:

[0046] ;

[0047] In the formula, α is the fusion weight value of the semantic similarity calculation model, β is the fusion weight value of the statistical feature similarity calculation model, Sim_sem(x,y) is the semantic similarity calculation value, Sim_sta(x,y) is the statistical feature similarity calculation value, and Sim_fusion(x,y) is the fusion similarity calculation value.

[0048] Preferably, the preset complementary metrics include accuracy and recall.

[0049] The accuracy rate is the ratio of the number of source measurement points that are correctly predicted to the total number of source measurement points in the validation set, where the candidate target with the highest similarity is selected as the prediction mapping point for each source measurement point data in the validation set.

[0050] The recall rate is calculated as the proportion of all source measurement data that should be mapped to the target and correctly predicted for each real target measurement point data in the validation set, and the average value of all targets is taken.

[0051] Preferably, the wind power scenario triplet construction rules include: anchor point selection rules, positive sample construction rules, and negative sample construction rules.

[0052] According to a second aspect of the present invention, a model fusion-based intelligent mapping system for multi-source heterogeneous data of wind farms is provided, the system comprising:

[0053] The data preprocessing module is used to collect historical monitoring data of wind farms and construct a training set for parameter learning, a validation set for hyperparameter tuning and fusion weight calculation, and a corpus for term distribution statistics.

[0054] The first model building module is used to perform text encoding processing on the training set using the pre-trained BERT model to obtain high-dimensional semantic vectors. It then combines the cosine similarity function and preset linear mapping parameters to construct a semantic similarity function. Finally, it uses the triplet loss function, which combines the triplet construction rules of wind power scenarios, to train the preset linear mapping parameters in the semantic similarity function, thereby obtaining a semantic similarity calculation model.

[0055] The second model building module is used to construct statistical feature vectors based on the corpus using the entropy weight method, and to construct a similarity calculation model based on statistical features based on the statistical feature vectors.

[0056] The weight calculation module is used to evaluate the performance of the semantic-based similarity calculation model and the statistical feature-based similarity calculation model using the validation set, and calculate the fusion weight of the semantic-based similarity calculation model and the statistical feature-based similarity calculation model according to the performance evaluation results.

[0057] The model fusion module is used to perform weighted fusion of the semantic-based similarity calculation model and the statistical feature-based similarity calculation model according to the fusion weight, so as to obtain a fusion similarity calculation model for calculating fusion similarity.

[0058] The system interface module includes input and output interfaces encapsulated as standardized service plugins. The input interface is used to obtain the source data to be mapped and input it into the fusion similarity calculation model. The output interface is used to output the source data to be mapped and the data with the highest similarity values ​​among the various mappable data, so as to realize the interaction with the information interface of the preset configuration tool.

[0059] The beneficial effects of this invention are as follows:

[0060] 1. This invention is the first to systematically apply intelligent mapping methods to the configuration scenario of multi-source heterogeneous data mapping in wind farms. It creatively transforms the traditional configuration operation, which relies on manual association and mapping, into an AI-based intelligent recommendation mode. Engineers can select recommended mapping points through the configuration tool interface, realizing a paradigm shift from purely manual to human-machine collaboration, significantly improving engineering configuration efficiency, and solving the practical pain points in the current intelligent development of wind farms.

[0061] 2. This invention innovatively adopts a "BERT + linear mapping optimized cosine similarity" architecture in the semantic channel and optimizes parameters through a triplet network, making the original general semantic representation capability specifically adapted to the wind farm data mapping scenario, thus solving the problem of poor performance of general models in specific domains. In the statistical channel, it innovatively introduces the entropy weight method into the calculation of wind farm measurement point data name similarity, using information entropy to measure the distribution characteristics of terms in the corpus, automatically identifying keywords with domain discrimination, and overcoming the defect of traditional TF-IDF-like methods where weights fail in similar corpora.

[0062] 3. This invention employs a comprehensive weight allocation method based on multiple performance indicators, achieving complementary advantages between semantic and statistical models. Furthermore, a weight adjustment factor is designed to automatically adjust the fusion weights under different wind farm and data characteristics scenarios. Optional auxiliary evaluation indicators are also designed to achieve adaptive optimization tailored to each wind farm, ensuring the mapping method maintains excellent performance in complex real-world environments with mixed manufacturers and different turbine models. This adaptive mechanism is an innovation that distinguishes this invention from static configuration schemes.

[0063] 4. Based on the algorithm model, this invention further designs a standardized unified interface encapsulation scheme, which enables the intelligent mapping method to be easily embedded into different wind farm system configuration tools and data governance platforms, achieving plug-and-play functionality. It also enables the technological achievements to be truly applied and can be quickly deployed in different system upgrade, migration, and replacement scenarios, significantly improving the practical value and promotion potential of the technology. Attached Figure Description

[0064] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described 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. In the drawings:

[0065] Figure 1 This is an overall flowchart of a model fusion-based intelligent mapping method for multi-source heterogeneous data in wind farms according to an embodiment of the present invention.

[0066] Figure 2 This is a schematic diagram of the semantic similarity model training principle in a model fusion-based intelligent mapping method for multi-source heterogeneous data of wind farms according to an embodiment of the present invention.

[0067] Figure 3 This is an example diagram illustrating the principle of calculating term weights using the entropy weight method in a model fusion-based intelligent mapping method for multi-source heterogeneous data of wind farms according to an embodiment of the present invention.

[0068] Figure 4 This is a comparison diagram of the effects of a model fusion-based intelligent mapping method for multi-source heterogeneous data in wind farms with other methods according to an embodiment of the present invention;

[0069] Figure 5 This is an example diagram of interface encapsulation and plug-and-play in a model fusion-based intelligent mapping method for multi-source heterogeneous data in wind farms according to an embodiment of the present invention.

[0070] Figure 6 This is a hardware structure block diagram of the host device according to an embodiment of the present invention. Detailed Implementation

[0071] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0072] The methods and embodiments provided in this application can be executed on a host device or a similar computing device. Taking running on a host device as an example, such as... Figure 6 As shown, the host device may include one or more ( Figure 6 Only one is shown in the diagram. The processor (which may include, but is not limited to, a microprocessor (MCU) or programmable logic device (FPGA), etc.) and storage for storing data are also shown. The host device may further include transmission devices for communication functions and input / output devices. Those skilled in the art will understand that... Figure 6 The structure shown is for illustrative purposes only and does not limit the structure of the host device described above. For example, the host device may also include components that are larger than... Figure 6 The more or fewer components shown, or having the same Figure 6 The different configurations shown.

[0073] The memory can be used to store computer programs, such as application software programs and modules, like the computer program corresponding to the exception handling method in this embodiment. The processor executes various functional applications and data processing by running the computer program stored in the memory, thus implementing the above-described method. The memory may include high-speed random access memory (RAM) and non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to the host device via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks (LANs), mobile communication networks, and combinations thereof.

[0074] Transmission devices are used to receive or send data over a network. Specific examples of the network described above may include a wireless network provided by a communication provider for the host device. In one example, the transmission device includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device may be a Radio Frequency (RF) module used for wireless communication with the Internet.

[0075] According to an embodiment of the present invention, a method for intelligent mapping of multi-source heterogeneous data of wind farms based on model fusion is provided.

[0076] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, according to a first embodiment of the present invention, a method for intelligent mapping of multi-source heterogeneous data from wind farms based on model fusion is provided, comprising the following steps:

[0077] S1. Collect historical monitoring data of wind farms, and construct a training set for parameter learning, a validation set for hyperparameter tuning and fusion weight calculation, and a corpus for term distribution statistics.

[0078] It should be noted that historical monitoring data from wind farms were collected to obtain the data names, measurement point names, and their corresponding physical meanings for each manufacturer and model of wind turbine. A training set (for semantic model parameter learning), a validation set (for hyperparameter tuning and fusion weight calculation), and a corpus (for statistical analysis of model term distribution) were constructed.

[0079] Specifically, historical operational data from wind farms is collected to obtain data names, measurement point names, and the meanings of corresponding physical quantities for each manufacturer and model of wind turbine. Three types of datasets are constructed:

[0080] (1) Training set: used for parameter learning of the semantic model. Construct triplet samples, each triplet containing: anchor point measurement name, positive sample measurement name, and negative sample measurement name. The size of the training set is determined according to the actual amount of data, generally not less than 5000 triplets.

[0081] (2) Validation set: Used for hyperparameter tuning during semantic model training and for calculating subsequent fusion weights. The validation set contains several source test points and their corresponding real target test points. The validation set does not overlap with the training set and its size is set to 10%-20% of the training set.

[0082] (3) Corpus: used for the word distribution statistics of the entropy weight method in the statistical model, containing the text of all data names and measurement point names.

[0083] S2. Use the pre-trained BERT model to perform text encoding processing on the training set to obtain high-dimensional semantic vectors. Combine the cosine similarity function and the preset linear mapping parameters to construct a semantic similarity function. Use the triplet loss function that combines the triplet construction rules of wind power scenario to train the preset linear mapping parameters in the semantic similarity function to obtain a semantic similarity calculation model.

[0084] This study employs a pre-trained BERT (Bidirectional Encoder Representations from Transformers) model as the basic encoder to encode data name text into high-dimensional semantic vectors. A trainable linear mapping layer is added between the BERT output layer and the cosine similarity calculation to project the original semantic vectors into a low-dimensional space optimized for the data mapping task. A triplet construction rule for the wind power scenario is designed, and the parameters of the linear mapping layer are trained using a triplet loss function to obtain a semantic-based similarity calculation model.

[0085] In a preferred embodiment, the step of performing text encoding processing on the training set using a pre-trained BERT model to obtain high-dimensional semantic vectors, and constructing a semantic similarity function by combining a cosine similarity function and preset linear mapping parameters, and training the preset linear mapping parameters in the semantic similarity function using a triplet loss function that combines the triplet construction rules of wind power scenarios, to obtain a semantic-based similarity calculation model includes the following steps:

[0086] S21. Use a word segmenter to segment and encode the training set to obtain the input tensor;

[0087] S22. Use the pre-trained BERT model as the base encoder, and input the input tensor into the base encoder. Through text encoding processing, obtain the high-dimensional semantic vector.

[0088] S23. Add linear mapping parameters that can be obtained through training to the high-dimensional semantic vector, and use the linear mapping parameters to optimize the cosine similarity function to obtain the semantic similarity function.

[0089] S24. Based on the engineering characteristics of wind farms, construct the triplet construction rules for wind power scenarios, and combine them with preset training parameters. Through the triplet loss function, train the preset linear mapping parameters to obtain a semantic similarity calculation model.

[0090] Specifically, a pre-trained BERT-base-chinese model is used as the basic encoder. The input data name and measurement point name text are segmented and encoded by a tokenizer to obtain the input tensor of the model. Then, the high-dimensional semantic vector h is obtained by the basic encoder model.

[0091] The specific parameters of the pre-trained BERT-base-chinese model are a 12-layer Transformer encoder, 768 hidden units, and 12 attention heads.

[0092] For the obtained high-dimensional semantic vector h, there is room for optimization in the application of concept and measurement point name similarity judgment in the new energy field. For example, consider two completely different concepts, "voltage level 10kV" and "voltage level 35kV," but the vector directions of these two features are quite similar. Meanwhile, cosine similarity, as a fixed, parameter-free metric function, cannot solve the problem of "metric space distortion." Therefore, this invention improves the cosine similarity calculation, obtaining a semantic similarity calculation formula. The calculation formula is:

[0093] ;

[0094] In the formula, h_A and h_B are high-dimensional semantic vectors output by the BERT model, W and b are linear mapping parameters obtained during training, and Sim_sem(A,B) is the semantic similarity value.

[0095] This improvement is equivalent to adding a trainable linear mapping layer (i.e., linear mapping layer z=W·h+b) to the high-dimensional semantic vector h output by BERT. This not only reduces the dimensionality and noise of the original semantic vector, but also increases its adaptability to the scenario. Through the trainable W, the model can learn to actively pull apart vectors with similar semantic structures but completely different meanings in the new energy field in the projection space.

[0096] The parameters W and b are trained using the triplet loss function. For each triplet (anchor a, positive sample p, negative sample n), the loss function L is defined as:

[0097] ;

[0098] In the formula, d(z) a ,z p )=1-cos(z a ,z p ) represents the cosine distance, margin is the preset boundary value, and z a It is the anchor point, z p It is a positive sample, z n It is a negative sample.

[0099] As a preferred embodiment, the wind power scenario triplet construction rules include: anchor point selection rules, positive sample construction rules, and negative sample construction rules.

[0100] Specifically, by configuring triplet construction rules for wind power scenarios, it can be ensured that the samples effectively reflect the naming differences in actual engineering projects.

[0101] (1) Anchor selection rules: Anchors are selected from the measurement point names of a specific physical quantity in the actual operation of the wind farm, covering a variety of naming styles, including but not limited to: English abbreviation type, English full name type, Chinese description type, and mixed type.

[0102] (2) Positive sample construction rules: Positive samples are measurement point names that have the same semantics and physical meaning as anchor points. The focus is on constructing three types of positive sample pairs: cross-manufacturer naming differences, cross-system naming differences, and synonymous heterogeneous expressions.

[0103] (3) Negative sample construction rules: Negative samples are measurement point names that are different from anchor points in semantics and physical meaning. The focus is on constructing three types of negative sample pairs: different physical quantities of the same equipment, easily confused similar names, and cross-equipment interference items.

[0104] Training is performed using triplet functions to bring anchor points closer to positive samples in the projection space and further away from negative samples, ensuring that the distance difference between positive and negative samples is at least greater than the margin. After training, a semantic-based similarity calculation model can be obtained.

[0105] It should be noted that while semantic-based similarity calculation models can determine similarity from a semantic perspective, they cannot address issues such as keyword weighting, which is crucial for determining the similarity of data names and measurement point names. Therefore, this invention designs a similarity calculation model based on statistical features to complement the semantic calculation model and enhance its robustness.

[0106] S3. Based on the corpus, construct statistical feature vectors using the entropy weight method, and construct a similarity calculation model based on statistical features based on the statistical feature vectors.

[0107] It should be noted that, considering the terminology distribution characteristics of wind farm data names, an entropy weight method is used to construct statistical feature vectors. The information entropy of each term in the corpus is calculated; the lower the information entropy, the higher the weight. The feature vectors based on entropy weights are combined with the cosine similarity formula to obtain a similarity calculation model based on statistical features.

[0108] In a preferred embodiment, the step of constructing a statistical feature vector using the entropy weight method based on the corpus, and constructing a similarity calculation model based on the statistical feature vector, includes the following steps:

[0109] S31. Perform statistics on each term in the corpus and construct a term-document matrix based on the statistical results;

[0110] Specifically, when constructing the term-document matrix, it can be done by statistically analyzing the occurrence of each term in the corpus, including how many documents the term appears in, i.e., the number of documents in which it appears and the total number of times the term appears.

[0111] S32. Based on the term-document matrix, calculate the information entropy of each term and the weight of each term based on the information entropy;

[0112] The formula for calculating the information entropy of each term based on the term-document matrix is ​​as follows:

[0113] ;

[0114] In the formula, p ij Let be the probability of the i-th term appearing in the j-th test point name, and n be the total number of test point names in the corpus. i Let be the information entropy of the i-th term.

[0115] The formula for calculating the weight of each term based on information entropy is as follows:

[0116] ;

[0117] In the formula, i represents the index of the current term, k represents the index of the traversed terms, and m represents the total size of the vocabulary. The lower the information entropy, the more uneven the distribution of the term, that is, it only appears in specific types of test points, and its weight is higher; the higher the information entropy, the more even the distribution of the term, that is, it appears in all types of test points, and its weight is lower.

[0118] S33. Construct a statistical feature vector based on entropy weight according to the weight of each term;

[0119] Specifically, for any data name, construct a statistical feature vector V based on entropy weights, with the vector dimension being the vocabulary size m, and the value of each dimension being:

[0120] ;

[0121] In the formula, tf j Weight represents the number of times term j appears in the current test point name. j For tf j The weight value, V[j] represents the statistical feature vector of term j.

[0122] S34. Combine the entropy-weighted statistical feature vector with the cosine similarity formula to form a similarity calculation model based on statistical features.

[0123] In a preferred embodiment, the statistical feature-based similarity calculation model is used to calculate the statistical feature similarity value, and the formula for calculating the statistical feature similarity value is as follows:

[0124] ;

[0125] In the formula, Sim_sta(A,B) represents the statistical feature similarity value, and V A V is the statistical characteristic vector of the measurement point name A. B Let B be the statistical characteristic vector of the measurement point.

[0126] As a preferred implementation, a similarity calculation model based on domain-adaptive statistical features is constructed. Alternatively, a multi-dimensional statistical feature vector can be constructed based on a corpus using a domain-constrained improved entropy weight method. Then, based on the statistical feature vector, a similarity calculation model based on domain-adaptive statistical features is constructed, specifically including the following steps:

[0127] Step 1: Preprocess the corpus, including word segmentation, terminology normalization, stop word filtering, and mapping of homonyms based on the naming conventions for wind power measurement points and a professional terminology dictionary, in order to eliminate feature noise caused by non-standard naming and obtain a standardized term set.

[0128] It should be noted that the dictionary includes standard measurement point names, equipment type terms, monitoring type terms, parameter dimension terms, and a mapping table of common homonyms. Based on the terminology dictionary, the original text in the corpus is segmented, and non-standard terms are normalized to standard terms according to the homonym mapping table. A stop word list is constructed, which includes general terms that do not contribute to actual semantics and irrelevant function words in the wind power scenario. The stop word list is used to filter stop words in the segmentation results. A standardized term set is output.

[0129] Step 2: Based on the hierarchical naming rules of wind power measurement point names, the standardized term set is divided into a core feature term set and an auxiliary feature term set, namely, equipment type-measurement point location-monitoring type-parameter dimension. The core feature term set includes equipment type, parameter dimension, measurement point location, and monitoring type, while the auxiliary feature term set includes serial number, general modifiers, and non-key suffixes.

[0130] It should be noted that the hierarchical naming rules for wind power measurement points are as follows, from left to right: equipment type, measurement point location, monitoring type, and parameter dimensions. Specifically: equipment type includes, but is not limited to: blades, main shaft, gearbox, generator, converter, hydraulic system, and yaw system; measurement point location includes, but is not limited to: front bearing, rear bearing, inlet, outlet, winding, A phase, B phase, and C phase; monitoring type includes, but is not limited to: temperature, vibration, speed, current, voltage, pressure, and angle; parameter dimensions include, but are not limited to: kW, rpm, mm / s, ℃, bar, V, and A.

[0131] Specifically, each term in the standardized term set is matched with the four categories of the hierarchical naming rule: if a term belongs to any core category, such as equipment type, measurement point location, monitoring type, or parameter dimension, it is classified into the core feature term set; if a term is a serial number, a general modifier, or a non-key suffix, such as "original" or "filtered", it is classified into the auxiliary feature term set.

[0132] Step 3: Based on the core feature term set and the auxiliary feature term set, and combined with the measurement point device level attribution of the documents in the corpus, construct a three-dimensional association matrix of term-level-document. Based on the three-dimensional association matrix of term-level-document, calculate the term frequency distribution and document proportion of the core feature term and the auxiliary feature term at the corresponding level to obtain the initial probability matrix of each term.

[0133] It should be noted that the three dimensions of the term-level-document three-dimensional association matrix are term dimension, level dimension, and document dimension, respectively. Among them, the term dimension corresponds to all unique terms in the standardized term set obtained in step one, and the dimension length is the total number of terms; the level dimension corresponds to the four core feature levels and one auxiliary feature level divided in step two, and the dimension length is fixed at 5; the document dimension corresponds to each wind power measurement point name text in the corpus, and the dimension length is the total number of documents in the corpus.

[0134] Specifically, based on hierarchical word frequency distribution and document proportion, the probability of each word appearing in its respective level is calculated to construct an initial probability matrix; the rows of the initial probability matrix correspond to words, the columns correspond to levels, and the matrix elements are the initial probability of the corresponding word appearing in the corresponding level.

[0135] Step 4: Based on the initial probability matrix, calculate the basic information entropy of each term, and correct the basic information entropy based on the weighted correction method of hierarchical prior to obtain the corrected information entropy;

[0136] It should be noted that, based on the initial probability matrix corresponding to the three-dimensional association matrix of terms-levels-documents, and using the probability of occurrence of each term under its respective wind power business level as the calculation basis, the basic information entropy of each term is solved using the standard formula for calculating information entropy with normalized coefficients. Then, based on the core feature term set and the auxiliary feature term set as the hierarchical prior, gradient positive weighted correction coefficients are set for the four core feature terms—equipment type, parameter dimensions, monitoring type, and measurement point location—according to the business priority of wind power measurement point matching. Furthermore, for the sequence number and general modifiers... Auxiliary feature terms such as words and non-key suffixes are given a reverse penalty weighting coefficient. Then, the weighting correction coefficient corresponding to each term is used to perform directional calculation with the basic information entropy of the term. The information entropy of the core feature term is reduced by the positive weighting correction coefficient to improve its subsequent weighting weight. The information entropy of the auxiliary feature term is increased by the reverse penalty weighting coefficient to suppress its invalid weight interference. At the same time, the boundary limit and compliance verification of the [0,1] interval of the initial corrected information entropy obtained by the calculation are performed. After removing extreme outliers, the corrected information entropy of each term can be directly used for subsequent weight calculation.

[0137] Step 5: Based on the corrected information entropy, calculate the comprehensive weight of each term using the jump step size of the cosine function, and generate the basic entropy weight statistical feature vector based on the comprehensive weight of each term.

[0138] Specifically, the comprehensive weight of each term is calculated using the jump step size of the cosine function based on the corrected information entropy, including the following steps:

[0139] Traverse all terms in the corpus and adjust the information entropy for each term. Calculate its basic weight value All terms are sorted from largest to smallest according to their basic weight values, and a serial number is assigned to each term.

[0140] The ordinal number of each term is normalized to obtain the normalized ranking. ;

[0141] Constructing a jump step size adjustment factor based on the cosine function Its expression is:

[0142] ;

[0143] In the formula, and These are the preset maximum and minimum values ​​of the adjustment factor, respectively, and satisfy the following conditions: ;

[0144] The basic weight value of each term is multiplied by its corresponding cosine jump step size adjustment factor, and then normalized to obtain the comprehensive weight of each term.

[0145] Step 6: Combine the basic entropy weight statistical feature vector with the cosine similarity formula to form a similarity calculation model based on statistical features.

[0146] S4. Using the validation set, evaluate the performance of the semantic-based similarity calculation model and the statistical feature-based similarity calculation model respectively, and calculate the fusion weight of the semantic-based similarity calculation model and the statistical feature-based similarity calculation model according to the performance evaluation results.

[0147] It should be noted that the performance of the two sub-models is evaluated on an independent validation set, using accuracy (defined in this invention) and recall (defined in the standard) as evaluation metrics. The combined score of each sub-model is calculated and normalized to obtain the fusion weight.

[0148] In a preferred embodiment, the step of using a validation set to evaluate the performance of the semantic-based similarity calculation model and the statistical feature-based similarity calculation model, and calculating the fusion weights of the semantic-based similarity calculation model and the statistical feature-based similarity calculation model based on the performance evaluation results, includes the following steps:

[0149] S41. Based on preset complementary indicators, calculate the performance data of the semantic similarity calculation model and the statistical feature similarity calculation model on the validation set respectively.

[0150] Specifically, the performance of the two sub-models is evaluated on independent validation sets using two complementary metrics.

[0151] In a preferred embodiment, the preset complementary metrics include accuracy and recall.

[0152] The accuracy rate is the ratio of the number of source measurement points that are correctly predicted to the total number of source measurement points in the validation set, where the candidate target with the highest similarity is selected as the prediction mapping point for each source measurement point data in the validation set.

[0153] The recall rate is calculated as the proportion of all source measurement data that should be mapped to the target and correctly predicted for each real target measurement point data in the validation set, and the average value of all targets is taken.

[0154] S42. Based on the preset weight adjustment factor, calculate the comprehensive evaluation performance score of the semantic similarity calculation model and the statistical feature-based similarity calculation model respectively.

[0155] It should be noted that by introducing a weight adjustment factor λ (0≤λ≤1), the overall score of each sub-model can be calculated:

[0156] ;

[0157] In the formula, Score sem Score is a comprehensive performance score for semantic similarity calculation models. sta The score represents the overall performance evaluation of the similarity calculation model based on statistical features. Acc represents accuracy, and Recall represents recall. λ is a weighting adjustment factor, which can be determined according to the needs of data mapping in different application scenarios: if more emphasis is placed on the correctness of the mapping results, λ > 0.5 can be chosen; if more emphasis is placed on the completeness of the data, λ < 0.5 can be chosen; if both are equally important, λ = 0.5 can be chosen.

[0158] S43. Normalize the comprehensive evaluation performance scores of the semantic-based similarity calculation model and the statistical feature-based similarity calculation model to obtain the fusion weight of the semantic-based similarity calculation model and the fusion weight of the statistical feature-based similarity calculation model.

[0159] Specifically, the fusion weights are calculated through normalization:

[0160] ;

[0161] ;

[0162] Where α is the fusion weight value of the semantic similarity calculation model, β is the fusion weight value of the statistical feature-based similarity calculation model, and α+β=1.

[0163] S5. Based on the fusion weight, the semantic-based similarity calculation model and the statistical feature-based similarity calculation model are weighted and fused to obtain a fusion similarity calculation model for calculating fusion similarity.

[0164] In a preferred embodiment, the expression for the fusion similarity calculation model is:

[0165] ;

[0166] In the formula, α is the fusion weight value of the semantic similarity calculation model, β is the fusion weight value of the statistical feature similarity calculation model, Sim_sem(x,y) is the semantic similarity calculation value, Sim_sta(x,y) is the statistical feature similarity calculation value, and Sim_fusion(x,y) is the fusion similarity calculation value.

[0167] This model fusion method can calculate unique fusion weights based on the data characteristics of different wind farms, such as naming style, terminology distribution, and abbreviation habits. It can evaluate the accuracy and recall of semantic similarity calculation models and statistical similarity calculation models on the data of each wind farm, enabling the semantic model and the statistical model to achieve optimal complementarity in each wind farm and realizing dynamic optimization of "one policy per farm".

[0168] It should be noted that this invention can be used for further performance verification of the fusion model. After completing the fusion similarity calculation, in order to further evaluate the model's generalization ability in different field scenarios and to provide a quality reference for the auxiliary configuration decision in step 7, this invention, in an optional but preferred embodiment, introduces the following auxiliary evaluation indicators. These indicators do not participate in the recalculation of fusion weights and are only used for model quality verification and engineering decision reference.

[0169] (a) Site Stability Index:

[0170] ;

[0171] In the formula, S represents the stability index value. For K wind farms in the same area, Acc is the stability index value. k This represents the mapping accuracy of the fusion model (k=1,2,…,K).

[0172] The significance of this indicator: The closer S is to 1, the more stable the model's performance and the stronger its generalization ability across different wind farms. This indicator is used to verify the effectiveness of the "one farm, one policy" adaptive mechanism of this invention when deployed across multiple wind farms. When the farm stability indicator S is low (e.g., <0.7), it indicates a significant performance difference between different wind farms. In this case, in addition to the already implemented "one farm, one policy" specific weight allocation, this invention further recommends adopting advanced strategies specific to each wind farm, such as: resampling the validation set and mining difficult samples for wind farms with poor stability, optimizing the triplet construction rules, or appropriately reducing the recommendation confidence level, etc.

[0173] (b) Confidence Reliability Indicator:

[0174] The recommendation results derived from the fused similarity scores are divided into three confidence intervals: high confidence, medium confidence, and low confidence. The actual hit rate R is calculated for each of the three intervals. high R mid R low Define the reliability index R conf :

[0175] ;

[0176] The significance of this indicator: Rconf The higher the similarity score, the better the fusion similarity model can effectively distinguish between "reliable recommendations" and "unreliable recommendations," which has guiding value for assisting configuration decisions and improving the trust and efficiency of human-machine collaboration.

[0177] S6. Encapsulate the fusion similarity calculation model into a standardized service plugin, and configure the input and output interfaces for the service plugin;

[0178] It should be noted that the above algorithm and fusion similarity calculation model are encapsulated into standardized service modules (i.e., service plugins), and a unified input and output interface is designed.

[0179] The input interface includes: the name of the source measurement point data to be mapped, a list of candidate measurement point data names that can be associated, a model fusion adjustment factor, and the returned candidate results. The input interface supports integration into software systems such as wind farm system configuration tools or data governance platforms via REST API, SDK, or configuration files.

[0180] The output interface includes: the name of the input source measurement point (for easy verification), the recommended best mapping target, the fusion similarity score of the recommended target (value range [0,1]), the confidence level (high, middle, low), and the list of alternative targets (sorted in descending order of similarity).

[0181] The candidate target objects include the candidate target name and the candidate target fusion similarity score.

[0182] This service module supports multiple integration methods: integration via REST API, SDK toolkits (mainstream programming languages), and configuration files (generating configuration file templates, supporting static configuration). The module can be integrated into software systems such as wind farm system configuration tools or data governance platforms, achieving a "plug-and-play" effect.

[0183] S7. Obtain the source data to be mapped and input it into the fusion similarity calculation model through the input interface. Calculate the similarity value between the source data to be mapped and each mappable data through the fusion similarity calculation model. Select the data with the highest similarity value and output it to the preset configuration tool interface through the output interface.

[0184] It should be noted that for engineering configurations that require data mapping, the similarity value between the source data to be mapped and each mappable data is calculated through the input interface. The N data with the highest similarity values ​​are taken, output and displayed in a specified position on the configuration tool software interface, and recommended to engineering configuration personnel to assist them in making data mapping decisions.

[0185] Specifically, for engineering configuration scenarios that require data mapping, such as intelligent model reuse, system upgrades, and system migrations, the corresponding configuration tool software calculates the similarity value between the source data to be mapped and each mappable data through an input interface, and obtains N return values ​​based on the set number of return results.

[0186] Then, the recommended results are displayed in a designated location on the configuration tool software interface, such as in the drop-down menu of the configuration box, and recommended to the engineering configuration personnel, who can then decide whether to select the recommended mapping point or manually find the correct mapping point.

[0187] To better understand the content of this invention, the technical solution of this invention will be further described in detail below with reference to embodiments and accompanying drawings:

[0188] Example 1: Data mapping application of a wind farm;

[0189] Taking a regional new energy company operating wind farms as an example, the region includes wind turbines from three manufacturers: Goldwind (6.7MW), Mingyang (8.3MW), and Envision (6MW). The company's control center's operation and maintenance system deploys intelligent agent applications, involving measurement points for a total of 200 physical quantities, including power, wind speed, temperature, and vibration. The naming rules for measurement points are completely different for each manufacturer, and they need to be uniformly mapped to the new energy company's standard measurement point library.

[0190] Combination such as Figure 1 The overall flowchart shown below illustrates the implementation process as follows:

[0191] (1) Data preparation; collect historical data and construct a training set: 5000 triplet samples. Each triplet contains the anchor point measurement name (e.g., "P_Gen"), the positive sample measurement name (e.g., "ActivePower"), and the negative sample measurement name (e.g., "WindSpeed"). According to the triplet construction rules of the present invention, the samples cover various naming patterns.

[0192] Construct a validation set: 1000 test point pairs (each source test point corresponds to the correct target), used for semantic model hyperparameter tuning and subsequent fusion weight calculation. The validation set does not overlap with the training set.

[0193] Construct a corpus: all measurement point name texts, used for entropy weight method to statistically analyze term distribution.

[0194] (2) Semantic model training: The pre-trained BERT-base-chinese model is used as the base encoder to encode the measurement point name text into a 768-dimensional semantic vector. A trainable linear mapping layer is added after the BERT output layer to project the semantic vector from 768 dimensions to 128 dimensions.

[0195] Adopting such Figure 2 The principle is to train the linear mapping parameters using the following triplet loss function:

[0196] ;

[0197] Set margin=0.5 and train for 30 epochs. After training, you will get a semantic similarity calculation model.

[0198] (3) Statistical model construction: Based on the corpus, the distribution of each word item is statistically analyzed, and the information entropy of each word item is calculated.

[0199] Feature vectors are constructed based on entropy weights. The vector dimension is the size of the corpus vocabulary, and the value of each dimension is the number of times a word appears multiplied by its entropy weight, as shown in the appendix. Figure 3 As shown in the flowchart, the corresponding statistical list is obtained. Note: the lower the information entropy (the more concentrated the term distribution), the higher the weight. It can be seen that "gearbox" has a low information entropy (0.35) and a high weight (0.23); "device" has a high information entropy (0.85) and a low weight (0.05).

[0200] For any data name, construct a feature vector V based on entropy weights, and obtain the formula for calculating the statistical similarity between two measurement point names A and B:

[0201] ;

[0202] (4) Calculate the fusion weights; evaluate the accuracy and recall of the two sub-models on the validation set respectively:

[0203] Semantic similarity calculation model: Acc_sem=82.5%, Recall_sem=79.8%;

[0204] Statistical similarity calculation model: Acc_sta=71.2%, Recall_sta=68.7%;

[0205] The weight adjustment factor λ is set to 0.6 (determined through cross-validation, taking into account both the correctness and completeness of the mapping).

[0206] Calculate the overall score:

[0207] Score_sem=0.6×82.5%+0.4×79.8%=0.495+0.319=0.814;

[0208] Score_sta=0.6×71.2%+0.4×68.7%=0.427+0.275=0.702;

[0209] Calculate the fusion weights:

[0210] α=0.814 / (0.814+0.702)=0.537;

[0211] β=0.702 / (0.814+0.702)=0.463;

[0212] (5) Testing the fusion model; calculating the fusion similarity on the test set:

[0213] Sim_fusion=0.537×Sim_sem+0.463×Sim_sta;

[0214] For each source measurement point, the candidate target with the highest similarity is selected as the recommended mapping result.

[0215] The fusion model test results show an accuracy of 86.3% and a recall of 84.7%. The results comparing the similarity judgment accuracy with other methods are as follows: Figure 4 As shown.

[0216] (6) Interface encapsulation and auxiliary configuration: The trained model is encapsulated as a standardized service plugin and deployed to the wind farm configuration tool. In this example, the number of candidate results returned is 2.

[0217] Engineers select the source measurement points to be mapped in the configuration tool interface. The system automatically calculates the fusion similarity with each measurement point in the candidate target library and displays two recommended results in a drop-down list. Engineers can directly click on the recommended results or manually search to complete the mapping configuration.

[0218] Example 2: Verification of adaptive performance at different sites;

[0219] To verify the adaptive capability of this invention ("one policy per wind farm"), tests were conducted at three wind farms with different data characteristics. A validation set was independently constructed for each wind farm, and the performance of the semantic model and statistical model were evaluated separately. Specific fusion weights were calculated, and then the mapping accuracy of the fusion model was tested. The test results are shown in Table 1.

[0220] Table 1 Test Results

[0221] Results analysis:

[0222] Wind field A (primarily described in English): The semantic model performed well, while the statistical model performed poorly. The fusion weight α = 0.591 (the semantic model has a higher weight), and the fusion accuracy reached 88.3%.

[0223] Wind field B (mainly Chinese): The two models performed relatively well, with a fusion weight of α=0.526 and a fusion accuracy of 85.2%.

[0224] C wind field (hybrid naming): The semantic model performed well, with a fusion weight α=0.539 and a fusion accuracy of 86.7%.

[0225] Further calculations were performed on the stability indices of the fusion model across the three sites:

[0226] S=1−(88.3%−85.2%) / 86.73%=0.964;

[0227] Based on the above stability indicators, it can be shown that the model performs stably on wind farms with different naming styles and has good generalization ability.

[0228] Conclusion: This invention can automatically adjust the fusion weights according to the characteristics of the site data, achieving better results than any single model at each site, thus verifying the effectiveness of the "one site, one policy" adaptive mechanism.

[0229] Example 3: Demonstration of an engineering application scenario (data mapping during system upgrade);

[0230] Scenario Description: A wind farm is upgrading an advanced application system, requiring the mapping of 4,000 measurement points from the old system (using Manufacturer A's naming convention) to the new system. Traditionally, this would require one engineer working for three days (approximately 24 person-hours) to complete the configuration, a task that is labor-intensive and prone to errors.

[0231] Implementation process:

[0232] (1) Model building and encapsulation: Using historical wind farm operation data, a training set, a validation set, and a corpus are built. The semantic model and statistical model are trained according to the steps in Example 1, and the fusion weights are calculated. Finally, the trained model is encapsulated as a service plugin, and its interface design and definition are shown in Tables 2 and 3 below.

[0233] Table 2 Input Interface

[0234] Table 3 Output Interfaces

[0235] (2) Plugin deployment; such as Figure 5 The schematic diagram shows that the intelligent mapping algorithm of this invention is encapsulated as a standardized service plugin and deployed in the wind farm server. The plugin provides services to the configuration tool through a service registry, achieving communication decoupling, and the configuration tool can easily call the intelligent data mapping function.

[0236] (3) Auxiliary Configuration: Engineers open the system upgrade configuration tool and import the old system's list of 4000 measurement points and the target standard measurement point library. The configuration tool automatically calculates the recommended mapping target and similarity score for each source measurement point. During the configuration process, the configuration tool interface displays the recommendation results. In this example: approximately 3400 measurement points are recommended with high confidence, and engineers can directly click to complete the mapping; approximately 200 measurement points are recommended with medium confidence, and engineers quickly review and confirm the correct mapping points; approximately 400 measurement points are recommended with low confidence, and engineers manually find the correct mapping points. The work takes 4-5 hours for one engineer to complete, and the configuration accuracy is high.

[0237] Engineering benefits: Supports hot-swapping, facilitates easy access to mapped services, and is plug-and-play. Reduces manual configuration and lowers the risk of human error. Significantly shortens system upgrade cycles.

[0238] According to a second embodiment of the present invention, a model fusion-based intelligent mapping system for multi-source heterogeneous data of wind farms is provided, the system comprising:

[0239] The data preprocessing module is used to collect historical monitoring data of wind farms and construct a training set for parameter learning, a validation set for hyperparameter tuning and fusion weight calculation, and a corpus for term distribution statistics.

[0240] The first model building module is used to perform text encoding processing on the training set using the pre-trained BERT model to obtain high-dimensional semantic vectors. It then combines the cosine similarity function and preset linear mapping parameters to construct a semantic similarity function. Finally, it uses the triplet loss function, which combines the triplet construction rules of wind power scenarios, to train the preset linear mapping parameters in the semantic similarity function, thereby obtaining a semantic similarity calculation model.

[0241] The second model building module is used to construct statistical feature vectors based on the corpus using the entropy weight method, and to construct a similarity calculation model based on statistical features based on the statistical feature vectors.

[0242] The weight calculation module is used to evaluate the performance of the semantic-based similarity calculation model and the statistical feature-based similarity calculation model using the validation set, and calculate the fusion weight of the semantic-based similarity calculation model and the statistical feature-based similarity calculation model according to the performance evaluation results.

[0243] The model fusion module is used to perform weighted fusion of the semantic-based similarity calculation model and the statistical feature-based similarity calculation model according to the fusion weight, so as to obtain a fusion similarity calculation model for calculating fusion similarity.

[0244] The system interface module includes input and output interfaces encapsulated as standardized service plugins. The input interface is used to obtain the source data to be mapped and input it into the fusion similarity calculation model. The output interface is used to output the source data to be mapped and the data with the highest similarity values ​​among the various mappable data, so as to realize the interaction with the information interface of the preset configuration tool.

[0245] According to a third embodiment of the present invention, an electronic device is provided, the electronic device comprising: one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to perform the steps in any of the above method embodiments.

[0246] According to a fourth embodiment of the present invention, a computer-readable storage medium is provided, wherein a computer program is stored in the computer-readable storage medium, wherein the computer program, when running, controls the device where the computer-readable storage medium is located to perform the steps in any of the above method embodiments.

[0247] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard disk, magnetic disk, or optical disk.

[0248] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, optical storage, etc.) containing computer-usable program code.

[0249] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent mapping of multi-source heterogeneous data from wind farms based on model fusion, characterized in that, Includes the following steps: S1. Collect historical monitoring data of wind farms, and construct a training set for parameter learning, a validation set for hyperparameter tuning and fusion weight calculation, and a corpus for term distribution statistics. S2. Use the pre-trained BERT model to perform text encoding processing on the training set to obtain high-dimensional semantic vectors. Combine the cosine similarity function and the preset linear mapping parameters to construct a semantic similarity function. Use the triplet loss function that combines the triplet construction rules of wind power scenario to train the preset linear mapping parameters in the semantic similarity function to obtain a semantic similarity calculation model. S3. Based on the corpus, construct statistical feature vectors using the entropy weight method, and construct a similarity calculation model based on statistical features based on the statistical feature vectors. S4. Using the validation set, evaluate the performance of the semantic-based similarity calculation model and the statistical feature-based similarity calculation model respectively, and calculate the fusion weight of the semantic-based similarity calculation model and the statistical feature-based similarity calculation model according to the performance evaluation results. S5. Based on the fusion weight, the semantic-based similarity calculation model and the statistical feature-based similarity calculation model are weighted and fused to obtain a fusion similarity calculation model for calculating fusion similarity. S6. Encapsulate the fusion similarity calculation model into a standardized service plugin, and configure the input and output interfaces for the service plugin; S7. Obtain the source data to be mapped and input it into the fusion similarity calculation model through the input interface. Calculate the similarity value between the source data to be mapped and each mappable data through the fusion similarity calculation model. Select the data with the highest similarity value and output it to the preset configuration tool interface through the output interface.

2. The intelligent mapping method for multi-source heterogeneous data of wind farms based on model fusion according to claim 1, characterized in that, The process of using a pre-trained BERT model to perform text encoding on the training set to obtain high-dimensional semantic vectors, and then constructing a semantic similarity function by combining a cosine similarity function and preset linear mapping parameters, and training the preset linear mapping parameters in the semantic similarity function using a triplet loss function that incorporates the triplet construction rules of wind power scenarios, to obtain a semantic-based similarity calculation model includes the following steps: S21. Use a word segmenter to segment and encode the training set to obtain the input tensor; S22. Use the pre-trained BERT model as the base encoder, and input the input tensor into the base encoder. Through text encoding processing, obtain the high-dimensional semantic vector. S23. Add linear mapping parameters that can be obtained through training to the high-dimensional semantic vector, and use the linear mapping parameters to optimize the cosine similarity function to obtain the semantic similarity function. S24. Based on the engineering characteristics of wind farms, construct the triplet construction rules for wind power scenarios, and combine them with preset training parameters. Through the triplet loss function, train the preset linear mapping parameters to obtain a semantic similarity calculation model.

3. The intelligent mapping method for multi-source heterogeneous data of wind farms based on model fusion according to claim 2, characterized in that, The semantic-based similarity calculation model is used to calculate the semantic similarity value; the formula for calculating the semantic similarity value is: ； In the formula, h_A and h_B are high-dimensional semantic vectors output by the BERT model, W and b are linear mapping parameters obtained during training, and Sim_sem(A,B) is the semantic similarity value.

4. The intelligent mapping method for multi-source heterogeneous data of wind farms based on model fusion according to claim 1, characterized in that, The process of constructing statistical feature vectors using the entropy weight method based on the corpus, and then constructing a similarity calculation model based on statistical features using these vectors, includes the following steps: S31. Perform statistics on each term in the corpus and construct a term-document matrix based on the statistical results; S32. Based on the term-document matrix, calculate the information entropy of each term and the weight of each term based on the information entropy; S33. Construct a statistical feature vector based on entropy weight according to the weight of each term; S34. Combine the entropy-weighted statistical feature vector with the cosine similarity formula to form a similarity calculation model based on statistical features.

5. The intelligent mapping method for multi-source heterogeneous data of wind farms based on model fusion according to claim 4, characterized in that, The statistical feature-based similarity calculation model is used to calculate the statistical feature similarity value, and the formula for calculating the statistical feature similarity value is as follows: ； In the formula, Sim_sta(A,B) represents the statistical feature similarity value, and V A V is the statistical characteristic vector of the measurement point name A. B Let B be the statistical characteristic vector of the measurement point.

6. The intelligent mapping method for multi-source heterogeneous data of wind farms based on model fusion according to claim 1, characterized in that, The process of evaluating the performance of the semantic-based similarity calculation model and the statistical feature-based similarity calculation model using a validation set, and then calculating the fusion weights of the two models based on the evaluation results, includes the following steps: S41. Based on preset complementary indicators, calculate the performance data of the semantic similarity calculation model and the statistical feature similarity calculation model on the validation set respectively. S42. Based on the preset weight adjustment factor, calculate the comprehensive evaluation performance score of the semantic similarity calculation model and the statistical feature-based similarity calculation model respectively. S43. Normalize the comprehensive evaluation performance scores of the semantic-based similarity calculation model and the statistical feature-based similarity calculation model to obtain the fusion weight of the semantic-based similarity calculation model and the fusion weight of the statistical feature-based similarity calculation model.

7. The intelligent mapping method for multi-source heterogeneous data of wind farms based on model fusion according to claim 1, characterized in that, The expression for the fusion similarity calculation model is: ； In the formula, α is the fusion weight value of the semantic similarity calculation model, β is the fusion weight value of the statistical feature similarity calculation model, Sim_sem(x,y) is the semantic similarity calculation value, Sim_sta(x,y) is the statistical feature similarity calculation value, and Sim_fusion(x,y) is the fusion similarity calculation value.

8. The intelligent mapping method for multi-source heterogeneous data of wind farms based on model fusion according to claim 6, characterized in that, The preset complementary metrics include accuracy and recall. The accuracy rate is the ratio of the number of source measurement points that are correctly predicted to the total number of source measurement points in the validation set, where the candidate target with the highest similarity is selected as the prediction mapping point for each source measurement point data in the validation set. The recall rate is calculated as the proportion of all source measurement data that should be mapped to the target and correctly predicted for each real target measurement point data in the validation set, and the average value of all targets is taken.

9. The intelligent mapping method for multi-source heterogeneous data of wind farms based on model fusion according to claim 1, characterized in that, The rules for constructing triplets for wind power scenarios include: anchor point selection rules, positive sample construction rules, and negative sample construction rules.

10. A model fusion-based intelligent mapping system for multi-source heterogeneous data of wind farms, used to implement the model fusion-based intelligent mapping method for multi-source heterogeneous data of wind farms as described in any one of claims 1-9, characterized in that, The system includes: The data preprocessing module is used to collect historical monitoring data of wind farms and construct a training set for parameter learning, a validation set for hyperparameter tuning and fusion weight calculation, and a corpus for term distribution statistics. The first model building module is used to perform text encoding processing on the training set using the pre-trained BERT model to obtain high-dimensional semantic vectors. It then combines the cosine similarity function and preset linear mapping parameters to construct a semantic similarity function. Finally, it uses the triplet loss function, which combines the triplet construction rules of wind power scenarios, to train the preset linear mapping parameters in the semantic similarity function, thereby obtaining a semantic similarity calculation model. The second model building module is used to construct statistical feature vectors based on the corpus using the entropy weight method, and to construct a similarity calculation model based on statistical features based on the statistical feature vectors. The weight calculation module is used to evaluate the performance of the semantic-based similarity calculation model and the statistical feature-based similarity calculation model using the validation set, and calculate the fusion weight of the semantic-based similarity calculation model and the statistical feature-based similarity calculation model according to the performance evaluation results. The model fusion module is used to perform weighted fusion of the semantic-based similarity calculation model and the statistical feature-based similarity calculation model according to the fusion weight, so as to obtain a fusion similarity calculation model for calculating fusion similarity. The system interface module includes input and output interfaces encapsulated as standardized service plugins. The input interface is used to obtain the source data to be mapped and input it into the fusion similarity calculation model. The output interface is used to output the source data to be mapped and the data with the highest similarity values ​​among the various mappable data, so as to realize the interaction with the information interface of the preset configuration tool.

Images