Auxiliary decision-making method and device for wind power plant, electronic equipment and medium

Abstract

The invention discloses an auxiliary decision-making method and device for a wind power plant, electronic equipment and a medium, and relates to the technical field of artificial intelligence and wind power operation and maintaining.According to the scheme, multi-source heterogeneous information of the wind power plant is obtained in real time and converted into a structured Markdown format document, so that corresponding knowledge fragments are extracted from a knowledge base and a case base in the wind power field; and finally, splicing the knowledge fragments into a prompt template and inputting the prompt template into the pre-trained large language model to obtain the auxiliary decision content output by the pre-trained large language model, which is more accurate compared with the traditional manual experience, and is more accurate by depending on the wind power field knowledge base and case base and the pre-trained large language model. The finally output auxiliary decision content is more suitable for the actual problem of the wind power plant, the error rate is lower, and the efficiency is higher.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and wind power operation and maintenance technology, and particularly relates to a wind farm auxiliary decision-making method and device, an electronic equipment and a medium. BACKGROUND

[0002] At present, the debugging and operation and maintenance of the wind farm highly depend on artificial experience. The operation and maintenance personnel need to face complex and multi-source equipment data such as voice, image, sensor data and text description, which are difficult to process uniformly. In addition, the operation and maintenance personnel often need to refer to unstructured historical case texts, which not only is low in efficiency, but also is generally high in error rate. The traditional knowledge base question-answering system has problems such as poor relevance of search results, response content out of context and lack of reasoning ability. SUMMARY

[0003] The present application aims to provide a wind farm auxiliary decision-making method and device, an electronic equipment and a medium. The present application converts multi-source heterogeneous information of a wind farm into a structured Markdown format document, extracts corresponding knowledge fragments from a wind power field knowledge base and a case base, splices the knowledge fragments into a prompt template and inputs the prompt template into a pre-trained large language model, so as to obtain auxiliary decision-making content output by the pre-trained large language model. Compared with the traditional artificial experience, the present application is more accurate. In addition, the present application relies on the wind power field knowledge base and the case base and the pre-trained large language model, so that the auxiliary decision-making content output ultimately is more in line with the actual problems of the wind farm, has a lower error rate and is more efficient.

[0004] To solve the above technical problems, the present application provides a wind farm auxiliary decision-making method, which comprises the following steps:

[0005] acquiring multi-source heterogeneous information of a wind farm, wherein the multi-source heterogeneous information comprises voice information of an operation and maintenance personnel of the wind farm, visual information of a wind farm site, sensor data of the wind farm and inquiry text information of the operation and maintenance personnel;

[0006] converting the multi-source heterogeneous information into a Markdown format document;

[0007] extracting knowledge fragments corresponding to the Markdown format document from a wind power field knowledge base and a case base;

[0008] splicing the knowledge fragments into a prompt template and inputting the prompt template into a pre-trained large language model, so as to obtain auxiliary decision-making content output by the pre-trained large language model, wherein the auxiliary decision-making content comprises any one or more of fault diagnosis analysis, debugging path suggestion, safety precautions and historical case reference.

[0009] Optionally, the step of converting the multi-source heterogeneous information into a Markdown format document comprises the following steps:

[0010] convert the language information into readable text content by using an ASR technology;

[0011] identify character information in the visual information by using an OCR technology, extract key visual features in the visual information by using a convolutional neural network model, and determine image summary content according to the character information and the key visual features;

[0012] determine a real-time sensor stream composed of timestamp-value pairs according to the sensor data;

[0013] convert the inquiry text information into a corresponding Token sequence by using a preset semantic parsing technology;

[0014] align the semantic structures of the readable text content, the image summary content, the real-time sensor stream, and the Token sequence by using a hierarchical feature extractor, and map the readable text content, the image summary content, the real-time sensor stream, and the Token sequence after semantic structure alignment into the Markdown format document.

[0015] Optionally, the extracting the knowledge fragments corresponding to the Markdown format document from the wind power field knowledge base and case base comprises:

[0016] inputting the Markdown format document into a preset semantic embedding model to obtain N query vectors output by the preset semantic embedding model, N being a positive integer not less than 1;

[0017] performing fast similarity matching in the wind power field knowledge base and case base by using the N query vectors and an ANN search technology, to extract N candidate sets corresponding to the N query vectors from the wind power field knowledge base and case base, a recall rate corresponding to the candidate set being greater than a preset recall rate threshold;

[0018] calculating semantic matching scores of the N candidate sets and the N query vectors by using a cross-encoder model, sorting the semantic matching scores in descending order, and outputting the first M candidate sets, M being a positive integer not less than 1;

[0019] taking the M candidate sets as the knowledge fragments.

[0020] Optionally, after obtaining the auxiliary decision content output by the pre-trained large language model, the method further comprises:

[0021] obtaining an evaluation result of the auxiliary decision content by the operation and maintenance personnel;

[0022] optimizing the wind power field knowledge base and case base and the pre-training large language model according to the evaluation result.

[0023] Optionally, optimizing the wind power field knowledge base and case base according to the evaluation result comprises:

[0024] determining a weight adjustment coefficient corresponding to the evaluation result;

[0025] adjusting a similarity weighted recall order of quick similarity matching in the wind power field knowledge base and case base according to the weight adjustment coefficient, so as to change an M candidate set extracted from the wind power field knowledge base and case base.

[0026] Optionally, optimizing the pre-training large language model according to the evaluation result comprises:

[0027] constructing a fine-tuning training set for optimizing the pre-training large language model according to the evaluation result;

[0028] re-training the pre-training large language model by using the fine-tuning training set, so as to optimize the pre-training large language model.

[0029] Optionally, the method further comprises:

[0030] transmitting the auxiliary decision content output by the pre-training large language model to a host computer.

[0031] To solve the above technical problems, the application further provides an auxiliary decision device for a wind farm, comprising:

[0032] an acquisition module configured to acquire multi-source heterogeneous information of the wind farm, wherein the multi-source heterogeneous information comprises voice information of an operation and maintenance personnel of the wind farm, visual information of a site of the wind farm, sensor data of the wind farm, and inquiry text information of the operation and maintenance personnel;

[0033] a conversion module configured to convert the multi-source heterogeneous information into a Markdown format document;

[0034] an extraction module configured to extract a knowledge fragment corresponding to the Markdown format document from a wind power field knowledge base and case base;

[0035] an auxiliary decision module configured to splice the knowledge fragment into a prompt template and input the prompt template into a pre-training large language model, so as to obtain auxiliary decision content output by the pre-training large language model, wherein the auxiliary decision content comprises any one or more of fault diagnosis analysis, debugging path suggestion, safety precautions, and historical case reference.

[0036] To solve the above technical problems, the application further provides an electronic device, comprising:

[0037] a memory for storing a computer program;

[0038] a processor for implementing the steps of the method for assisting decision of a wind farm when the computer program is executed.

[0039] To solve the above technical problems, the application further provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program implements the steps of the method for assisting decision of a wind farm when executed by a processor.

[0040] The application aims to provide a method and device for assisting decision of a wind farm, an electronic device and a medium, which obtains multi-source heterogeneous information of a wind farm in real time, converts the information into a structured Markdown format document, extracts corresponding knowledge fragments from a wind power field knowledge base and a case base, splices the knowledge fragments into a prompt template and inputs the prompt template into a pre-trained large language model, so as to obtain the assisting decision content output by the pre-trained large language model, which is more accurate than the traditional experience relying on manual work, and the assisting decision content output finally is more suitable for the actual problems of the wind farm, has a lower error rate and is more efficient. BRIEF DESCRIPTION OF DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of the provided drawings.

[0042] Figure 1 A process flow chart of a method for assisting decision of a wind farm provided by the present application;

[0043] Figure 2 A process flow chart of another method for assisting decision of a wind farm provided by the present application;

[0044] Figure 3 A structural schematic diagram of a device for assisting decision of a wind farm provided by the present application;

[0045] Figure 4 A structural schematic diagram of an electronic device provided by the present application. DETAILED DESCRIPTION

[0046] The core of the present application is to provide an auxiliary decision-making method, device, electronic equipment and medium for a wind farm. The present application converts multi-source heterogeneous information of a wind farm into a structured Markdown format document, extracts corresponding knowledge fragments from a wind power field knowledge base and case library, splices the knowledge fragments into a prompt template and inputs them into a pre-trained large language model, and finally obtains the auxiliary decision-making content output by the pre-trained large language model. Compared with the traditional experience relying on manual work, the present application is more accurate, and the auxiliary decision-making content output by the present application is more suitable for the actual problems of the wind farm, has a lower error rate and is more efficient.

[0047] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0048] Please refer to Figure 1 , Figure 1 A process flow chart of an auxiliary decision-making method for a wind farm is provided in the present application. The auxiliary decision-making method for a wind farm comprises the following steps.

[0049] S11: Obtain multi-source heterogeneous information of a wind farm, wherein the multi-source heterogeneous information comprises voice information of an operation and maintenance personnel of the wind farm, visual information of a wind farm site, sensor data of the wind farm, and inquiry text information of the operation and maintenance personnel.

[0050] S12: Convert the multi-source heterogeneous information into a Markdown format document.

[0051] S13: Extract knowledge fragments corresponding to the Markdown format document from a wind power field knowledge base and case library.

[0052] S14: Splice the knowledge fragments into a prompt template and input them into a pre-trained large language model to obtain auxiliary decision-making content output by the pre-trained large language model, wherein the auxiliary decision-making content comprises any one or more of fault diagnosis analysis, debugging path suggestion, safety precautions, and historical case reference.

[0053] In the present application, considering that the current wind farm debugging and operation and maintenance work highly depends on manual experience, the operation and maintenance personnel need to face complex and multi-source equipment data such as voice, image, sensor data and text description, which is difficult to process uniformly, and often needs to refer to unstructured historical case text, which is not only low in efficiency, but also generally has a high error rate, therefore, the present scheme selects to automatically obtain multi-source heterogeneous information of the wind farm, and converts it into a Markdown format document, then extracts the knowledge fragments corresponding to the Markdown format document from the wind farm knowledge base and case base, the extraction process of the knowledge fragments is automatic and does not need to rely on manual experience, in addition, after the knowledge fragments are extracted, the auxiliary decision content of the wind farm also needs to be determined based on the knowledge fragments, and the present scheme selects to splice the knowledge fragments into a prompt template and input a pre-trained large language model to obtain the auxiliary decision content output by the pre-trained large language model, wherein the auxiliary decision content includes any one or more of fault diagnosis analysis, debugging path suggestion, safety precautions and historical case reference, therefore, the present scheme can automatically obtain multi-source heterogeneous information of the wind farm, and then automatically obtain the auxiliary decision content, and the finally output auxiliary decision content is more in line with the actual problems of the wind farm, has a lower error rate and is more efficient.

[0054] It should be noted that the pre-trained large language model is a large language model trained by wind power field data.

[0055] It should be noted that the multi-source heterogeneous information is: ; wherein, represents that the operation and maintenance personnel describe the current fault, consultation or operation intention through voice, and the system uses an automatic speech recognition module (ASR) to transcribe to generate readable text content as part of the natural language input; represents visual information such as device on-site photos, alarm interface screenshots, etc., the system first identifies character information (such as alarm code, device model) in the image through an OCR module, then uses an image classification and detection model to extract key visual features (such as component state, alarm icon, hot spot area, etc.), and generates image summary content; represents an implementation sensor stream composed of timestamp-value pairs; is a natural language inquiry directly input by the operation and maintenance personnel through the terminal, the text directly enters the semantic analysis module for subsequent retrieval and generation, and is represented as a Token sequence: .

[0056] It should also be noted that the present scheme selects to use a convolutional neural network (CNN) based image classification model, specifically using a ResNet-50 architecture to extract high-level semantic features of the image; for the input image wherein , , These represent the image's height, width, and number of channels, respectively. The feature vectors undergo a linear transformation through a fully connected layer, and then are converted into a class probability distribution using a softmax function. This determines the image's class. probability distribution: ;in For the fully connected layer corresponding to the first Class weight vector, For bias terms; Image feature vectors extracted by CNN; The preset number of categories for wind power operation and maintenance scene images is used, such as "equipment normal", "loose bolts", "blade damage", "instrument panel alarm", etc. Ultimately, the model classifies the image into the category with the highest probability, i.e.: The image target detection model uses YOLOv10 to identify and locate key objects in the image. The model training set was optimized for wind farm scenarios, including 12 core target categories such as blade damage (cracks, corrosion), bolt status (loosening, missing), equipment components (gearboxes, control cabinet instruments), and environmental elements (anemometers, wind vanes). The YOLOv10 detection output is structured data containing the target category (e.g., "blade crack"), bounding box coordinates (x1, y1, x2, y2), and confidence score (e.g., 0.91), as shown in Table 1.

[0057] Table 1

[0058]

[0059] It should also be noted that after perception and initial processing, all types of input data enter the semantic alignment and encoding module. This module uses a hierarchical feature extractor to perform semantic structure alignment on the inputs from different modalities. The features of each modality are uniformly encoded and mapped to a standard format document, represented using structured Markdown, to drive the semantic retrieval and large language model generation modules, such as:

[0060] ## Question via voice

[0061] Fan A2 displays abnormal voltage.

[0062] ## Image Attachment

[0063] - 

[0064] - Image recognition identified "Alarm Code E107"

[0065] ## Real-time parameters

[0066] - Wind speed: 12.3 m / s

[0067] - Motor temperature: 60°C

[0068] - Voltage: 420V

[0069] ## Maintenance staff notes

[0070] May be related to voltage fluctuations.

[0071] It should also be noted that the structured Markdown document in this application, as a unified semantic carrier, is input to the subsequent semantic retrieval and generation module. The precise recall and rearrangement of semantic knowledge retrieval includes two stages, namely the initial recall stage and the rearrangement and rearrangement stage, which are in a serial structure, and the input of the rearrangement and rearrangement stage is the output of the recall stage. This design guarantees the recall speed while greatly improving the semantic matching accuracy, and is suitable for the actual situation of redundant knowledge, information-intensive, and diversified queries in the wind power field. Among them, the initial recall stage uses a semantic embedding model to generate a high-dimensional dense query vector as follows: ; Through an approximate nearest neighbor search (ANN) in the knowledge base , a candidate set closest in semantics is recalled. In actual deployment, this application uses high-performance ANN engines such as FAISS (Facebook AI Similarity Search) to build indexes, supporting millisecond-level recall under large-scale knowledge bases (millions of paragraphs). At the same time, vector compression (such as PQ) technology is used to balance accuracy and performance: ; Where is the number of candidates to be returned. The goal of this stage is to maximize the coverage of the truly relevant document set in terms of query semantics. One of its evaluation indicators is recall (Recall), which is an important indicator in the field of information retrieval (Information Retrieval) to measure the coverage ability of the system, defined as follows: ; The wind farm full-life-cycle intelligent auxiliary decision-making system and method based on multi-modal perception and semantic retrieval guarantee the recall rate . The rearrangement and rearrangement stage uses a cross-encoder model to calculate the semantic matching score of each candidate paragraph and the query as follows: ; Where represents the element-wise multiplication of vectors and , emphasizing their similarity in each dimension, is a learnable weight matrix used for linear transformation to map the concatenated vector to a scalar score space. The system sorts all candidate paragraphs according to the matching score and outputs the top N documents as the final retrieval result: ;Utilizing large language models for fault diagnosis and debugging path recommendation includes a retrieval-augmented generation (RAG) module and decision output, which concatenates the input query and retrieval results as context: ;Where the prompt template is constructed as follows:

[0072] [System instruction] You are a wind power operation and maintenance expert, please analyze the fault based on the following context:

[0073] [Retrieval results]

[0074] [Input question]

[0075] [Requirements] Output in the following format:

[0076] 1. Generate fault cause analysis: P(cause 1) > P(cause 2) >...

[0077] 2. Suggest operation (debugging process)

[0078] 3. Safety tips

[0079] The RAG structure combines document retrieval capabilities and language generation capabilities, and compared to traditional knowledge graph or rule-based question and answer solutions, it has stronger semantic understanding, cross-modal fusion, and transferability. The results are input into a large language model (such as DeepSeek or a wind power-specific fine-tuned model) to generate fault cause analysis (including confidence), recommended operation steps (process debugging), and safety operation precautions. The example output structure is as follows:

[0080] ## Fault diagnosis

[0081] - Reason speculation: false alarm of grid-side voltage sensor or inverter bus overvoltage

[0082] - Confidence ranking:

[0083] 1. Bus overvoltage (0.82)

[0084] 2. Sensor drift (0.63)

[0085] ## Suggested operation

[0086] 1. Check if the inverter bus voltage is stable;

[0087] 2. Adjust the upper computer protection threshold;

[0088] 3. If false alarm, recalibrate the sensor.

[0089] ## Safety tips

[0090] - Need to power off before repair;

[0091] - There is high voltage inside the inverter, and the operator needs to have a certificate.

[0092] The system introduces a user feedback driven closed-loop optimization mechanism to realize the continuous evolution and performance improvement of large language models and semantic retrieval systems in wind power debugging and operation and maintenance scenarios. After the model completes fault diagnosis, debugging suggestions, and safety prompts, users can explicitly label the output results to form feedback signals for subsequent optimization processing. User feedback is collected through an interactive interface in the form of scoring or labeling, including: "effective", "partially effective", and "ineffective". The resulting feedback is used to adjust the weight of the knowledge base documents, supplement the prompt templates and context case library, or build a fine-tuning training set to optimize the semantic retrieval model.

[0093] Among them, the feedback label set is: ;

[0094] According to user feedback, the weight coefficient of the document segment R appearing in the retrieval candidate paragraph is adjusted, which is used for the next round of similarity weighted recall sorting: ; Where is an indicator function, and effective feedback improves the priority of document retrieval, is an adjustment factor, preset to 0.2. The online update rule of the knowledge base is: ; Similarly, when , the output is included in the positive sample prompt template set and the "high-quality historical case library" for prompt construction or few-shot examples in similar scenarios, improving the context learning effect of the model.

[0095] The embodiment provides an auxiliary decision-making method for a wind farm. The method converts multi-source heterogeneous information of the wind farm into structured Markdown format documents in real time, extracts corresponding knowledge segments from a wind power field knowledge base and a case library, and finally splices the knowledge segments into prompt templates and inputs them into a pre-trained large language model to obtain auxiliary decision-making content output by the pre-trained large language model. Compared with traditional manual experience, the method is more accurate, and the auxiliary decision-making content output by the method is more consistent with actual problems of the wind farm, has a lower error rate, and is more efficient.

[0096] On the basis of the above embodiment:

[0097] As an optional embodiment, converting multi-source heterogeneous information into a Markdown format document includes:

[0098] convert language information into readable text content using ASR technology;

[0099] identify character information within visual information using OCR technology, extract key visual features in visual information using a convolutional neural network model, and determine image summary content according to the character information and the key visual features;

[0100] determine a real-time sensor stream composed of timestamp-value pairs according to sensor data;

[0101] convert the inquiry text information into the corresponding Token sequence using the preset semantic parsing technology;

[0102] align the semantic structures of the readable text content, the image summary content, the real-time sensor stream, and the Token sequence using a hierarchical feature extractor, and map the readable text content, the image summary content, the real-time sensor stream, and the Token sequence after semantic structure alignment to a Markdown format document.

[0103] In the present application, considering that the Markdown format document belongs to pure text format, in order to convert multi-source heterogeneous information into a Markdown format document, it is necessary to convert language information into readable text content using ASR technology, similarly, identify character information within visual information using OCR technology, extract key visual features in visual information using a convolutional neural network model, and determine image summary content according to the character information and the key visual features, determine a real-time sensor stream composed of timestamp-value pairs according to sensor data, convert the inquiry text information into the corresponding Token sequence using the preset semantic parsing technology, and finally align the semantic structures of the readable text content, the image summary content, the real-time sensor stream, and the Token sequence using a hierarchical feature extractor, i.e. align and arrange the above data, and map the readable text content, the image summary content, the real-time sensor stream, and the Token sequence after semantic structure alignment to a Markdown format document, i.e. convert the above data to a pure text format with semantic structure alignment, ensuring the integrity of the scheme.

[0104] As an optional embodiment, knowledge fragments corresponding to the Markdown format document are extracted from a wind power field knowledge base and a case base, including:

[0105] input the Markdown format document into a preset semantic embedding model to obtain N query vectors output by the preset semantic embedding model, N being a positive integer not less than 1;

[0106] The N query vectors and the ANN search technology are used for fast similarity matching in the wind power field knowledge base and case base, so as to extract N candidate sets corresponding to the N query vectors from the wind power field knowledge base and case base, and the recall rate of the candidate set is greater than a preset recall rate threshold;

[0107] The semantic matching scores of the N candidate sets and the N query vectors are calculated by using a cross-encoder model, the semantic matching scores are sorted in descending order, and the first M candidate sets are output, M being a positive integer greater than or equal to 1;

[0108] The M candidate sets are used as knowledge fragments.

[0109] In the present application, considering that there are a large number of knowledge fragments in the wind power field knowledge base and case base, in order to extract knowledge fragments corresponding to the Markdown format document, the Markdown format document needs to be input into a preset semantic embedding model first to obtain N query vectors output by the preset semantic embedding model, and then the query vectors and the ANN search technology are used for fast similarity matching in the wind power field knowledge base and case base, so as to extract N candidate sets corresponding to the N query vectors from the wind power field knowledge base and case base, and the N candidate sets are a set with a higher similarity matching degree with the Markdown format document, and the recall rates of the N candidate sets are all greater than a preset recall rate threshold, so as to ensure the accuracy of the knowledge fragments extracted from the wind power field knowledge base and case base, finally, the semantic matching scores of the N candidate sets and the N query vectors are calculated by using a cross-encoder model, the semantic matching scores are sorted in descending order, and the first M candidate sets are output, that is, the M candidate sets with the highest semantic matching scores with the Markdown format document are selected as knowledge fragments, thereby improving the reliability of the scheme.

[0110] As an optional embodiment, after obtaining the auxiliary decision content output by the pre-trained large language model, the method further comprises:

[0111] obtaining an evaluation result of the auxiliary decision content by the operation and maintenance personnel;

[0112] optimizing the wind power field knowledge base and case base and the pre-trained large language model according to the evaluation result.

[0113] In the present application, as shown in Figure 2 considering that the operation and maintenance personnel can evaluate the practicability of the auxiliary decision content after the pre-trained large language model outputs the auxiliary decision content, the present scheme selects to obtain an evaluation result of the auxiliary decision content by the operation and maintenance personnel, and optimizes the wind power field knowledge base and case base and the pre-trained large language model according to the evaluation result, thereby ensuring the accuracy and practicability of the auxiliary decision content output by the pre-trained large language model.

[0114] As an optional embodiment, the wind power field knowledge base and the case base are optimized according to the evaluation result, including:

[0115] A weight adjustment coefficient corresponding to the evaluation result is determined.

[0116] The similarity weighted recall order of the fast similarity matching in the wind power field knowledge base and the case base is adjusted according to the weight adjustment coefficient, so as to change the M candidate sets extracted from the wind power field knowledge base and the case base.

[0117] In the present application, considering that the M candidate sets are extracted from the wind power field knowledge base and the case base, the wind power field knowledge base and the case base are optimized according to the evaluation result, which actually changes the M candidate sets output by the wind power field knowledge base and the case base. Because the M candidate sets are selected from the N candidate sets corresponding to the N query vectors, the similarity weighted recall order of the fast similarity matching in the wind power field knowledge base and the case base needs to be adjusted according to the weight adjustment coefficient, so as to change the N candidate sets corresponding to the N query vectors extracted from the wind power field knowledge base and the case base, and further change the M candidate sets output by the wind power field knowledge base and the case base, so as to ensure the accuracy of the optimization process.

[0118] As an optional embodiment, the pre-training large language model is optimized according to the evaluation result, including:

[0119] A fine-tuning training set for optimizing the pre-training large language model is constructed according to the evaluation result;

[0120] The pre-training large language model is retrained using the fine-tuning training set to optimize the pre-training large language model.

[0121] In the present application, because the pre-set large language model is trained in advance, the pre-set large language model is actually optimized according to the evaluation result. In fact, the fine-tuning training set for optimizing the pre-training large language model is constructed according to the evaluation result, and then the pre-training large language model is retrained using the fine-tuning training set to optimize the pre-training large language model, which ensures the integrity of the model optimization process and improves the field understanding and decision accuracy of the model.

[0122] As an optional embodiment, it further includes:

[0123] The auxiliary decision content output by the pre-training large language model is transmitted to the upper computer.

[0124] In the present application, considering that each auxiliary decision content corresponds to different operation and maintenance operations, in order to subsequently analyze the situation of the wind farm in detail, it is necessary to record each auxiliary decision content, therefore the present scheme selects the auxiliary decision content output by the pre-trained large language model to be transmitted to the upper computer, so as to record each auxiliary decision content by using the upper computer, and the situation of the wind farm can be analyzed in detail according to the auxiliary decision content by using the upper computer, which is more convenient for actual operation.

[0125] Please refer to Figure 3 , Figure 3 A structural schematic diagram of an auxiliary decision device for a wind farm is provided in the present application. The auxiliary decision device for the wind farm comprises:

[0126] The acquisition module 11 is configured to acquire multi-source heterogeneous information of the wind farm, wherein the multi-source heterogeneous information comprises voice information of an operation and maintenance personnel of the wind farm, visual information of a wind farm site, sensor data of the wind farm, and inquiry text information of the operation and maintenance personnel.

[0127] The conversion module 12 is configured to convert the multi-source heterogeneous information into a Markdown format document.

[0128] The extraction module 13 is configured to extract a knowledge segment corresponding to the Markdown format document from a wind farm domain knowledge base and a case base.

[0129] The auxiliary decision module 14 is configured to splice the knowledge segment into a prompt template and input the prompt template into a pre-trained large language model, so as to obtain auxiliary decision content output by the pre-trained large language model, wherein the auxiliary decision content comprises any one or more of fault diagnosis analysis, debugging path suggestion, safety precautions, and historical case reference.

[0130] The auxiliary decision device for the wind farm provided in the present embodiment corresponds to the above method, and has the same beneficial effects as the above method, therefore the embodiments of the auxiliary decision device for the wind farm are described in the description of the embodiments of the method, which will not be described here.

[0131] Please refer to Figure 4 , Figure 4 A structural schematic diagram of an electronic device is provided in the present application. The electronic device comprises:

[0132] The memory 20 is configured to store a computer program.

[0133] The processor 21 is configured to execute the computer program to implement the steps of the auxiliary decision method for the wind farm as described above.

[0134] The electronic device provided in the present embodiment can include but is not limited to a smart phone, a tablet computer, a notebook computer, or a desktop computer, etc.

[0135] The processor 21 can include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor 21 can be implemented in at least one of a hardware form of a digital signal processor (DSP), a field-programmable gate array (FPGA), a programmable logic array (PLA), etc. The processor 21 can also include a main processor and a coprocessor. The main processor is a processor for processing data in a wake-up state, also known as a central processing unit (CPU). The coprocessor is a low-power processor for processing data in a standby state. In some embodiments, the processor 21 can be integrated with a graphics processing unit (GPU) for rendering and drawing content required to be displayed by the display screen. In some embodiments, the processor 21 can further include an artificial intelligence (AI) processor for processing computing operations related to machine learning.

[0136] The memory 20 can include one or more computer-readable storage media that can be non-transitory. The memory 20 can further include a high-speed random access memory, and a nonvolatile memory such as one or more disk storage devices, flash storage devices. In the present embodiment, the memory 20 is at least used to store the following computer program 201, wherein the computer program is loaded and executed by the processor 21, and can implement the related steps of the auxiliary decision-making method of the wind farm disclosed in any of the preceding embodiments. In addition, the resources stored by the memory 20 can further include an operating system 202 and data 203, etc., and the storage manner can be temporary storage or permanent storage. The operating system 202 can include Windows, Unix, Linux, etc. The data 203 can include but is not limited to the auxiliary decision-making method of the wind farm, etc.

[0137] In some embodiments, the electronic device can further include a display screen 22, an input / output interface 23, a communication interface 24, a power supply 25, and a communication bus 26.

[0138] Those skilled in the art can understand that, Figure 4 The structure shown in the figure does not constitute a limitation on the electronic device, and can include more or fewer components than those shown.

[0139] The embodiment is aimed to provide an electronic device, wherein the memory 20 is used to store a computer program, and the processor 21 is used to execute the computer program to realize the steps of the auxiliary decision-making method of the wind farm as described above, so that the process of auxiliary decision-making is more efficient and accurate.

[0140] The application further provides an embodiment of a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program is executed by a processor to realize the steps of the auxiliary decision-making method of the wind farm as described above.

[0141] It can be understood that if the method in the above embodiment is realized in the form of a software function unit and sold or used as an independent product, the method can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the application essentially or the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and executes all or part of the steps of the method of each embodiment of the application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0142] The computer readable storage medium provided by the embodiment has the same beneficial effects as the above method, and therefore the embodiments of the computer readable storage medium are described in the description of the embodiments of the method, and will not be described here.

[0143] It should be noted that in the specification, the relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to the process, method, article or device. Without more limitations, the element defined by the statement "including a" does not exclude the presence of another identical element in the process, method, article or device including the element.

[0144] The foregoing description of the disclosed embodiments enables a person skilled in the art to make or use the application. Modifications of these embodiments will occur to persons of skill in the art, and that the appended claims are intended to cover all such modifications that do not depart from the true spirit and scope of the application. Therefore, the application is not limited to the embodiments shown but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method of assisted decision making for a wind farm, characterized in that, The method comprises the following steps: acquiring multi-source heterogeneous information of a wind farm, wherein the multi-source heterogeneous information comprises voice information of an operator of the wind farm, visual information of a site of the wind farm, sensor data of the wind farm, and inquiry text information of the operator; converting the multi-source heterogeneous information into a Markdown format document; extracting knowledge fragments corresponding to the Markdown format document from a wind power field knowledge base and a case base; splicing the knowledge fragments into a prompt template and inputting the prompt template into a pre-trained large language model to obtain auxiliary decision content output by the pre-trained large language model, wherein the auxiliary decision content comprises any one or more of fault diagnosis analysis, debugging path suggestion, safety precautions, and historical case reference.

2. The method of assisting decision making for a wind farm according to claim 1, characterized in that, The step of converting the multi-source heterogeneous information into a Markdown format document comprises the following steps: converting language information into readable text content by using an ASR technology; identifying character information in the visual information by using an OCR technology, extracting key visual features in the visual information by using a convolutional neural network model, and determining image abstract content according to the character information and the key visual features; determining a real-time sensor stream composed of timestamp-value pairs according to the sensor data; converting the inquiry text information into a corresponding Token sequence by using a preset semantic parsing technology; performing semantic structure alignment on the readable text content, the image abstract content, the real-time sensor stream, and the Token sequence by using a hierarchical feature extractor, and mapping the readable text content, the image abstract content, the real-time sensor stream, and the Token sequence after semantic structure alignment into the Markdown format document.

3. The method of aiding decision making for a wind farm of claim 1, wherein, The step of extracting knowledge fragments corresponding to the Markdown format document from a wind power field knowledge base and a case base comprises the following steps: inputting the Markdown format document into a preset semantic embedding model to obtain N query vectors output by the preset semantic embedding model, wherein N is a positive integer not less than 1; performing fast similarity matching in the wind power field knowledge base and the case base by using N query vectors and an ANN search technology to extract N candidate sets corresponding to the N query vectors from the wind power field knowledge base and the case base, wherein a recall rate corresponding to each candidate set is greater than a preset recall rate threshold; calculating semantic matching scores of N candidate sets and N query vectors by using a cross-encoder model, sorting the semantic matching scores in descending order, and outputting the first M candidate sets, wherein M is a positive integer not less than 1; taking the M candidate sets as the knowledge fragments.

4. The method of assisting decision making for a wind farm of claim 3, wherein, After obtaining the auxiliary decision content output by the pre-trained large language model, the method further comprises the following steps: acquiring an evaluation result of the auxiliary decision content by the operator; optimizing the wind power field knowledge base and the case base and the pre-trained large language model according to the evaluation result.

5. The method of aiding decision making for a wind farm of claim 4, wherein, The step of optimizing the wind power field knowledge base and the case base according to the evaluation result comprises the following steps: determining a weight adjustment coefficient corresponding to the evaluation result; According to the weight adjustment coefficient, the similarity weighted recall sequence of the quick similarity matching in the wind power field knowledge base and case base is adjusted to change the M candidate sets extracted from the wind power field knowledge base and case base.

6. The method of aiding decision making for a wind farm of claim 4, wherein, According to the evaluation result, the pre-training large language model is optimized, including: According to the evaluation result, a fine-tuning training set for optimizing the pre-training large language model is constructed; The pre-training large language model is retrained by using the fine-tuning training set to optimize the pre-training large language model.

7. The wind farm assistance decision method according to any of claims 1 to 6, characterized in that, Also includes: The auxiliary decision content output by the pre-training large language model is transmitted to the upper computer.

8. An aid to decision device for a wind farm, characterized in that Including: An acquisition module is configured to acquire multi-source heterogeneous information of a wind farm, the multi-source heterogeneous information including voice information of an operation and maintenance personnel of the wind farm, visual information of a site of the wind farm, sensor data of the wind farm, and inquiry text information of the operation and maintenance personnel; A conversion module is configured to convert the multi-source heterogeneous information into a Markdown format document; An extraction module is configured to extract a knowledge fragment corresponding to the Markdown format document from a wind power field knowledge base and case base; An auxiliary decision module is configured to splice the knowledge fragment into a prompt template and input the prompt template into a pre-training large language model to obtain auxiliary decision content output by the pre-training large language model, the auxiliary decision content including any one or more of fault diagnosis analysis, debugging path suggestion, safety precautions, and historical case reference.

9. An electronic device, comprising: Including: A memory is configured to store a computer program; A processor is configured to implement the steps of the auxiliary decision method of the wind farm according to any one of claims 1 to 7 when the computer program is executed.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the steps of the auxiliary decision method of the wind farm according to any one of claims 1 to 7.

Images