Wind resource assessment report generation method based on large model technology

Abstract

The invention discloses a wind resource assessment report generation method based on a large model technology. The method comprises the following steps: processing multi-source data, and constructing a domain knowledge base and a fine tuning database; field adaptive RAG enhancement is carried out, professional literatures, historical cases and industry specifications in a field knowledge base are retrieved by adopting a retrieval enhancement generation technology, and wind resource assessment contents meeting technical standards are generated through a multi-modal fusion word vector technology; performing large model fine tuning, and optimizing the generative large model through a gradient-free optimization algorithm Wind-DFO on the basis of a fine tuning database by adopting a field self-adaptive fine tuning strategy; and automatically generating a report, arranging and matching multiple templates through an AI workflow, embedding a data visualization chart, executing verification of a grammar layer, a logic layer and a compliance layer, and outputting a standardized wind resource assessment report. The accuracy, specialty and efficiency of report generation are remarkably improved, and the problems that a traditional method depends on artificial experience, the data utilization efficiency is low, and standardization is insufficient are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind power generation, and specifically relates to a method for generating a wind resource assessment report based on large model technology. Background Art

[0002] Wind resource assessment reports have strategic guiding significance for the entire life cycle of wind power projects. As the core basis for project development, they provide scientific support for site optimization and unit selection through multi-dimensional data quantitative analysis. However, traditional wind resource assessment methods have bottlenecks: first, they rely on manual full-process operations, making it difficult to achieve spatiotemporal alignment and dynamic updates of cross-modal data (such as mesoscale simulation data and ground observations); second, traditional assessment methods have a cycle of up to 7-10 days, with many processes, which are time-consuming and labor-intensive, making it difficult to meet the needs of rapid assessment and decision-making; third, the assessment process relies on the experience of wind resource engineers. Wind resource engineers need a lot of practical experience to increase the accuracy of assessments of complex projects. Without the in-depth participation of experienced engineers, it is difficult to ensure that the assessment accuracy of the wind resource assessment report meets the requirements.

[0003] As the global energy transition accelerates, the large-scale development of wind power, as a core pillar of clean energy, places higher demands on accurate and intelligent wind resource assessment. The preparation of traditional wind resource assessment reports relies heavily on manual experience and has the following defects:

[0004] (1) Data fusion limitations: Existing methods lack intelligent processing capabilities for multi-source heterogeneous data, making it difficult to achieve spatiotemporal feature alignment and dynamic knowledge updates. For example, forest cover area modeling still uses a roughness simplification method that ignores the spatial heterogeneity of leaf area density distribution and vegetation drag coefficient or a single numerical simulation method, which is unable to adapt to the meteorological characteristics of different regions. Although existing automated tools can improve computing speed, they lack multimodal data fusion capabilities (such as semantic association between topographic maps and standard texts), resulting in a lack of interpretability of key indicators (turbulence intensity, 50-year maximum wind speed).

[0005] (2) Data processing is inefficient and time-consuming: Manually collecting, cleaning, and processing large amounts of meteorological and geographic data is very time-consuming, especially when dealing with multi-source data. It is prone to delays and errors: manual operations can easily introduce human errors, such as data entry errors or calculation errors, which affect the accuracy of the final results.

[0006] (3) Poor analytical consistency: Different analysts may interpret and process the same set of data in different ways, resulting in a lack of consistency and comparability in the evaluation results.

[0007] (4) Insufficient standardization: Due to the lack of a unified standard process, the evaluation methods and results between different projects may be difficult to directly compare, which reduces the credibility of the evaluation.

[0008] (5) Inefficient report generation: Manually writing reports requires a lot of time for typesetting and formatting, especially when they contain a large number of charts and data. There is a lot of repetitive work: for multiple similar projects, reports need to be written from scratch each time, and existing templates and content cannot be effectively reused, which increases the workload.

[0009] (6) Knowledge updating is lagging behind and professional knowledge updating is slow: Manual writing relies on the analyst’s knowledge level and experience accumulation, and may not be able to keep up with the latest industry standards and technological developments in a timely manner.

[0010] (7) Inconvenience in collaboration and sharing, and complexity in teamwork: In projects involving multiple people, it is more complicated to coordinate and synchronize the work of various parts, which is prone to miscommunication and information asymmetry. Personal experience and knowledge are difficult to record and pass on systematically, and the cost of training new employees is high, which is not conducive to long-term development. Summary of the Invention

[0011] In view of this, the purpose of the present invention is to provide a wind resource assessment report generation method based on large model technology. Through the collaboration of RAG technology and dynamic optimization strategy, the accuracy, professionalism and efficiency of report generation are significantly improved, and the pain points of traditional methods such as reliance on manual experience, low data utilization, and insufficient standardization are solved. It provides efficient and intelligent decision-making support for wind power project development and promotes the large-scale development and digital transformation of clean energy.

[0012] In order to achieve the above object, the present invention provides the following technical solutions:

[0013] A method for generating a wind resource assessment report based on large model technology includes the following steps:

[0014] Multi-source data processing: Integrate wind resource knowledge and wind resource assessment reports to build a domain knowledge base covering wind power terminology, specifications, and cases. Intelligently extract wind resource characteristics, including average wind speed, turbulence intensity, and wind shear index, through machine learning algorithms, and dynamically update the knowledge base. Construct a fine-tuning dataset based on a triple structure and expand and enhance the fine-tuning dataset.

[0015] Domain-adaptive RAG enhancement: This technology uses retrieval-enhanced generation technology to dynamically retrieve professional literature, historical cases, and industry specifications from the domain knowledge base based on a generative large model. Multimodal word embedding technology is then used to generate wind resource assessment content that meets technical standards. This multimodal word embedding technology includes a dual-stream encoding network and a cross-attention mechanism to achieve joint encoding and bidirectional interaction of text and image features.

[0016] Large model fine-tuning: Using a domain-adaptive fine-tuning strategy, we optimize the generative large model using the Wind-DFO gradient-free optimization algorithm, which includes dynamic Latin hypercube sampling, Bayesian surrogate model construction, and adaptive learning rate updates.

[0017] Automated report generation: Through AI workflow orchestration, multiple templates are matched, data visualization charts are embedded, and syntax, logic, and compliance layer verification are performed to output standardized wind resource assessment reports.

[0018] Furthermore, the method steps for multi-source data processing are as follows:

[0019] 11) Knowledge base construction

[0020] 111) Data integration: Integrate wind resource assessment reports and wind resource knowledge including IEC standard library, national standard library, industry standard library, industry design specifications, industry technical regulations and equipment white papers, and build a domain knowledge base covering wind power terminology, specifications and cases;

[0021] 112) Constructing a professional terminology network graph: Based on the GraphSAGE algorithm, a wind power professional terminology graph network is constructed to achieve joint coding of associated parameters;

[0022] 113) Dynamic knowledge update: Develop a dynamic knowledge update engine that automatically identifies standard document version changes through a rule engine and triggers the knowledge base reconstruction process to ensure that the knowledge base content is synchronized with the latest industry standards;

[0023] 12) Fine-tuning dataset construction

[0024] 121) Structured dataset definition: Using Alpaca-52k enhanced format, construct a triple structured dataset;

[0025] 122) Data format standardization: Use dataset_info.json to define dataset metadata, convert raw data into standard JSON format through scripts, and align field naming;

[0026] 123) Text processing and enhancement:

[0027] Word segmentation and length control: Use sliding windows to segment long texts to avoid information loss;

[0028] Data augmentation: Back-translation technology is used to generate multilingual variants, and the SimBERT model is used for semantic replacement to generate semantically consistent synonymous expressions to improve the diversity of the dataset.

[0029] Furthermore, the method steps for domain adaptive RAG enhancement are as follows:

[0030] 21) Hybrid Search: This includes sparse search, dense search, and compliance constraint channels, efficiently retrieving multimodal information from the knowledge base to ensure the professionalism and compliance of generated content. The sparse search is based on keyword matching to quickly screen relevant paragraphs in technical documents. The dense search encodes text and numerical features into a unified semantic space and calculates semantic similarity. The compliance constraint channel accesses the industry standard database in real time to perform mandatory logical verification of key indicators.

[0031] 22) Modal vector fusion: Encode text and image features into a unified semantic space to improve cross-modal information fusion capabilities;

[0032] 221) Design a two-stream encoding network: use the improved attention_word2vec model to encode the text stream, and use convolutional network layers and linear regression layers to encode the image stream;

[0033] 222) Interpretable Causal Mechanism: A causal gating mechanism is introduced to construct an interpretable weight distribution module, and the interaction weights of text and image features are calculated through MLP;

[0034] 223) Bidirectional feature enhancement mechanism: Use the cross-attention mechanism to generate image-driven text enhancement features and text-driven image enhancement features, and fuse the text enhancement features and image enhancement features through residual connections to generate bidirectional outputs;

[0035] 23) Semantic Contrast Alignment: Through contrastive learning mechanism, the consistency of image features and text features in the semantic space is enforced.

[0036] Furthermore, in step 221), the method for encoding the text stream is:

[0037] attention_word2vec[text]+cosine rotation position encoding cosRoPE[pos]→text_feature

[0038] The improved attention_word2vec model introduces a self-attention mechanism based on the CBOW architecture to perform hierarchical feature extraction, including:

[0039] Word embedding layer: uses dynamic word2vec to capture local semantics through sliding windows;

[0040] Attention enhancement: Multi-head attention is introduced to calculate the global word relationship weight, and the output and input of each attention layer are weighted and fused;

[0041] Cosine Rotational Position Encoding (cosRoPE) combines Rotational Position Encoding (RoPE) with cosine modulation. The principle is as follows:

[0042]

[0043] Where: R(θ pos ) is a learnable rotation matrix; pos is the position index; i is the dimension index; θ pos is a learnable rotation angle parameter; is the matrix multiplication operator.

[0044] Furthermore, in step 221), the method for encoding the image stream is:

[0045] convolution[image]→linear feature projection→image_feature

[0046] Among them: convolution is convolutional network; linear is linear regression;

[0047] The principle of the convolutional network layer is:

[0048] O conv =f(I*K+b)

[0049] Where: K is the convolution kernel; I*K represents the convolution operation applied to the input image I; b is the bias term; f is the activation function; O conv It is the output feature map obtained after the convolution operation;

[0050] The linear regression layer flattens the feature map output by the convolutional network layer into a one-dimensional vector, and aligns it with the text feature dimension with LayerNorm normalization. The principle of the linear regression layer is:

[0051] y = LayerNorm(Wx+b′)

[0052] Where: W is the weight matrix; b′ is the bias term; LayerNorm represents the layer normalization operation, which is used to standardize the output so that the final output is aligned with the text feature dimension.

[0053] Furthermore, in step 222), the principle of the causal gating mechanism is:

[0054] causal_gate=MLP(Text_Feature⊙Image_Feature)

[0055] Among them: causal_gate is the causal gating mechanism; MLP is the multi-layer perceptron; Text_Feature is the text feature; Image_Feature is the image feature;

[0056] The interaction weights of text and image features are:

[0057] gate=σ(Conv1D(text_feature)*causal_gate)

[0058] Where: gate is the overall gating mechanism; Conv1D is a one-dimensional convolutional network; σ is the sigma function.

[0059] Furthermore, in step 223), the bidirectional output obtained by fusing the text enhancement features and the image enhancement features through residual connection is:

[0060] F final =MCA(V→T)+α·MCA(T→V)

[0061] Among them: F final To fuse text enhancement features and image enhancement features through residual connection, MCA(V→T) is the text enhancement feature; MCA(T→V) is the image enhancement feature; α is the weight coefficient; V is the visual image; T is the text.

[0062] Furthermore, in step 23), the loss function of semantic comparison alignment is:

[0063]

[0064] in: Represents the similarity score indexation result of the positive sample pair; sim(v,t + ) is the visual feature v and the corresponding positive sample text feature t + The cosine similarity of is used to quantify the degree of semantic alignment between the two; τ is the temperature coefficient, which controls the steepness of the probability distribution; Contains the sum of similarities between the positive sample and all negative samples; t ― is the negative sample text feature.

[0065] Furthermore, the steps for optimizing the generative large model using the non-gradient optimization algorithm Wind-DFO are as follows:

[0066] 31) Dynamic parameter space partitioning and initialization: Generate candidate parameter points through dynamic Latin hypercube sampling:

[0067]

[0068] Where: θ (i) is the i-th candidate point; is a uniform distribution, indicating that the parameter range in each dimension is from θ min to θ max ; N is the number of candidate points generated;

[0069] Calculate the initial loss and set the current optimal solution;

[0070] 32) Bayesian surrogate model construction: Based on the Gaussian process surrogate model, the mapping relationship between the parameter space and the loss function is established:

[0071]

[0072] Where: f(θ) is the objective function value under parameter θ; μ(θ) is the mean function; k(θ,θ′) is the kernel function, which defines the similarity between different parameter points;

[0073] Adopting improved kernel function to adapt to high-dimensional non-stationary optimization:

[0074]

[0075] Where: r is the Euclidean distance between two parameter points, that is, r = ‖θ―θ′‖2; ‖·‖2 is the L2 norm;

[0076] 33) Multi-objective acquisition function optimization: Designing improved expected improvement functions:

[0077]

[0078] Where: α EI (θ) is the expected improvement value at parameter θ; f min is the currently known optimal loss value; f(θ) is the loss value predicted by the proxy model; σ 2 (θ) is the variance of the proxy model prediction; ∈1 is the numerical stability coefficient to prevent the denominator from being zero;

[0079] Generate a candidate direction set through the CMA-ES algorithm;

[0080] 34) Gradient-independent direction search: Evaluate the actual loss of candidate directions and select the Pareto frontier solution to determine the optimal descent direction;

[0081] 35) Determine whether the rate of change of the actual loss for m consecutive iterations is lower than the set threshold: if so, terminate the iteration, obtain the optimal parameters, and use the Bayesian surrogate model to fit the error; if not: execute step 36)

[0082] 36) Dynamic learning rate update: Learning rate is calculated based on momentum adaptation mechanism:

[0083]

[0084] Where: η t is the learning rate at step t; η base is the basic learning rate; g t is the pseudo gradient estimate, which represents the gradient estimate at step t; ‖g t‖ is the L2 norm of the pseudo gradient estimate; ∈2 is the numerical stability coefficient to prevent the denominator from being zero;

[0085] Update parameters:

[0086] θ best ←θ best +η t ·g t

[0087] Where: θ best is the current optimal parameter; η t is the learning rate of the current step; g t is the pseudo gradient estimate;

[0088] 37) Determine whether the current iteration step t is equal to the maximum allowed iteration step T: If so, terminate the iteration, obtain the optimal parameters, and use the Bayesian surrogate model to fit the error; if not, set t = t + 1 and loop through step 31).

[0089] Furthermore, the method steps for automatic report generation are:

[0090] 41) Dynamic Orchestration Engine: Uses a finite state machine to control the generation process and uses Python's transitions library to define state transitions at each stage, including initial state, data verification, project overview, wind energy resource analysis, power generation calculation, and conclusion.

[0091] 42) Multi-dimensional output: Use the Jinja2 template engine to generate technical reports in Markdown format, and automatically insert visual charts drawn by Matplotlib to generate reports;

[0092] During the report generation process, multiple levels of validation are employed, including:

[0093] Syntax layer: Use LangChain to perform syntax layer verification to ensure that the text content complies with grammatical rules;

[0094] Logic layer: Use CLIPS rule engine to verify logical consistency and ensure that all calculation results are reasonable and correct;

[0095] Compliance layer: Automatically mark relevant standard clause indexes in reports to ensure compliance.

[0096] The beneficial effects of the present invention are:

[0097] The wind resource assessment report generation method based on large-scale model technology in this invention significantly improves the accuracy, professionalism, and efficiency of report generation through the collaboration of RAG technology and dynamic optimization strategy. It solves the pain points of traditional methods such as reliance on manual experience, low data utilization, and insufficient standardization. It provides efficient and intelligent decision-making support for wind power project development, promotes the large-scale development and digital transformation of clean energy, and has the following technical effects:

[0098] (1) Improve the intelligence level of data processing and feature extraction

[0099] During the multi-source data processing step, by integrating various data sources and building a domain knowledge base, the system achieves intelligent extraction and dynamic updating of wind resource characteristics. This not only improves data quality and consistency, but also automatically identifies and extracts key features, reducing the need for manual intervention and ensuring efficient and accurate data processing. In particular, as new data is continuously input, the system dynamically updates the knowledge base, maintaining sensitivity to the latest wind resource characteristics, thereby improving the reliability of long-term forecasts.

[0100] (2) In the domain-adaptive RAG enhancement step, a generative large model and retrieval enhancement technology RAG are used to dynamically retrieve professional literature, historical cases, and industry specifications from the knowledge base to generate wind resource assessment content that meets technical standards. Through multimodal fusion word vector technology, using a dual-stream encoding method and an interpretable causal mechanism, interpretable weights are constructed, enabling the model to trace the source of feature contributions and improve the model's interpretability. Furthermore, a two-way interactive channel is introduced through a cross-attention mechanism, which enhances the model's interpretability and integration, making the generated report not only accurate but also more transparent, making it easier for users to understand the basis behind each conclusion.

[0101] (3) In the large model fine-tuning step, domain-adaptive fine-tuning of the large model is adopted, and the QLoRA optimization strategy is introduced. In particular, the gradient-free optimization algorithm Wind-DFO integrates multi-objective optimization theory, efficient sampling strategy, and adaptive learning rate technology, providing a new optimization paradigm for fine-tuning in the LLM domain. The optimization strategy of the present invention not only improves the performance of the model in specific scenarios, but also significantly reduces computing resource consumption, improves training efficiency, and enables the model to more flexibly respond to wind resource assessment needs in different regions and conditions.

[0102] (4) In the automated report generation step, AI workflow orchestration is adopted to support multi-template matching, data visualization and compliance verification, and a wind resource assessment report can be output with one click, which greatly reduces the time and complexity of manual operations and improves the efficiency and standardization of report generation. The present invention can select appropriate report templates for data visualization processing according to user needs, and intuitively display key information such as wind resource distribution and power generation potential. The built-in compliance verification function ensures that the generated report complies with industry standards and regulatory requirements, further enhancing the professionalism and credibility of the report.

[0103] (5) Solve the pain points of traditional methods and promote industry development

[0104] ① Reduce manual dependence: Compared with traditional methods, this invention significantly reduces dependence on manual experience, reduces errors caused by human operational errors or deviations, and improves the accuracy of reports.

[0105] ② Improve data utilization: By integrating multi-source data and advanced machine learning technology, data utilization is greatly improved, ensuring that every data point can provide effective support for the final result.

[0106] ③ Standardization and normalization: Through unified templates and logical rules, the report generation process is standardized and normalized, making reports between different projects more comparable and consistent.

[0107] In summary, the present invention significantly improves the accuracy, professionalism and efficiency of wind resource assessment report generation through the organic combination of four steps: multi-source data processing, domain-adaptive RAG enhancement, large model fine-tuning and automated report generation. The present invention not only solves the pain points of traditional methods such as reliance on manual experience, low data utilization, and insufficient standardization, but also provides efficient and intelligent decision-making support for wind power project development, promoting the large-scale development and digital transformation of clean energy. The method of the present invention not only helps to improve the economic and social benefits of wind power projects, but also opens up new directions for future research and applications, and promotes the continuous progress and development of the industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0108] In order to make the purpose, technical solutions and beneficial effects of the present invention more clear, the present invention provides the following drawings for illustration:

[0109] Figure 1 This is a main flow chart of an embodiment of a method for generating a wind resource assessment report based on a large model technology according to the present invention;

[0110] Figure 2 A detailed flow chart of an embodiment of a method for generating a wind resource assessment report based on a large model technology according to the present invention;

[0111] Figure 3Flowchart generated for large RAG-based models;

[0112] Figure 4 Flowchart for optimizing a large generative model using the gradient-free optimization algorithm Wind-DFO. DETAILED DESCRIPTION

[0113] The present invention will be further described below with reference to the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it. However, the embodiments are not intended to limit the present invention.

[0114] Big model technology refers to machine learning models with a large number of parameters built using deep learning algorithms, particularly models developed based on the Transformer architecture, which has made significant progress in fields such as natural language processing (NLP) and computer vision. These models typically contain hundreds of millions or even more parameters, enabling them to learn and capture more complex and detailed patterns and features in the data. By training on massive amounts of text data, big models can understand and generate high-quality natural language text, supporting a variety of application scenarios from translation, question-answering, to text creation. Advances in big models have driven the development of the field of artificial intelligence, providing more powerful and flexible tools to solve practical problems and offering researchers new avenues for exploring deeper levels of machine intelligence. A method for generating wind resource assessment reports based on big model technology, relying on the Transformer architecture and self-attention mechanism, achieves complex tasks such as quantifying wind energy potential and optimizing turbine layout by fusing massive parameters and multimodal data. This technology not only reconstructs the entire process from data cleaning to compliance verification (traditional manual work takes 15-30 days, and the standardization rate is less than 60%), but also directly empowers the "dual carbon" strategy through the carbon sink certification function, and promotes wind power development from experience-driven to a new paradigm of AI precise decision-making.

[0115] Llama is a large-scale language model open-sourced by Meta (formerly Facebook) designed to advance research and application development in the field of natural language processing (NLP). Based on the Transformer architecture and trained on large amounts of text data, the Llama model possesses powerful text generation and comprehension capabilities. A notable feature is its open source nature, which allows researchers and developers to freely access, modify, and deploy the model, accelerating the pace of innovation and reducing development costs. By providing a series of pre-trained models, Llama supports a variety of application scenarios, including but not limited to text generation, dialogue systems, and translation services. Furthermore, Llama's openness fosters community collaboration and knowledge sharing, enabling more practitioners to participate in the development of large-scale model technology and jointly advance the advancement of artificial intelligence. Whether for academic research or commercial applications, Llama provides users with a flexible and powerful toolset to meet diverse task requirements.

[0116] The large model technology based on Retrieval-Augmented Generation (RAG) combines the advantages of traditional generative models and information retrieval systems, aiming to improve the accuracy and relevance of generation tasks. The RAG model consists of a retrieval component and a generation component: first, the retrieval component retrieves the most relevant context or document fragments from a large-scale document library based on the input query; then, the generation component uses this retrieved information to generate the final answer or content. This architecture not only enables the model to access and utilize the latest information in external knowledge bases, but also overcomes the limitations of traditional methods that rely solely on internal parameters of the model to store knowledge, such as outdated knowledge or insufficient coverage. In this way, large models based on RAG demonstrate higher accuracy and flexibility in question-answering systems, text summarization, and other natural language generation tasks, opening up new avenues for achieving more intelligent and data-driven applications.

[0117] Big model technology, based on AI workflow orchestration, automates and intelligently processes complex tasks by integrating and optimizing multiple AI components or modules. This technology leverages advanced workflow orchestration tools to organically combine steps such as data collection and preprocessing, feature engineering, model training and tuning, and inference and post-processing into an efficient and flexible workflow. In big model applications, this means dynamically adjusting the resources and technology stack used in each step based on specific task requirements, maximizing performance and efficiency. For example, in natural language processing (NLP) tasks, workflow orchestration can implement the following process: first, raw text is processed using specially designed data cleansing and preprocessing modules; then, the most appropriate feature extraction method, such as word embedding or sentence vector representation, is selected based on the task requirements; then, the most suitable large-scale pre-trained model for the task is invoked for further analysis or generation; finally, customized post-processing steps ensure the quality and usability of the output results. The advantage of this architecture is that it not only supports highly automated end-to-end solution development but also allows for rapid iteration and experimentation. Developers can easily plug in new algorithms or adapt existing components to adapt to changing requirements. Furthermore, workflow orchestration facilitates efficient resource management, making load balancing and fault recovery simpler and more reliable in large-scale distributed computing environments. Therefore, large-scale model technology based on AI workflow orchestration provides strong support for building more intelligent, flexible, and scalable AI systems, significantly improving efficiency across the entire process from research to production.

[0118] Specifically, such as Figure 1-2 As shown, the wind resource assessment report generation method based on the large model technology in this embodiment includes multi-source data processing, domain adaptive RAG enhancement, large model fine-tuning and automatic report generation.

[0119] 1. Multi-source data processing:

[0120] Integrate wind resource knowledge and wind resource assessment reports to build a domain knowledge base covering wind power terminology, specifications, and case studies. Intelligently extract wind resource characteristics, including average wind speed, turbulence intensity, and wind shear index, through machine learning algorithms, and dynamically update the knowledge base. Construct a fine-tuning dataset based on a triple structure and expand and enhance the fine-tuning dataset.

[0121] By building a domain knowledge base, this embodiment achieves intelligent extraction of wind resource characteristics and automatic updating of the dynamic knowledge base, which not only improves the efficiency of data processing, but also enhances the quality and reliability of data, solving the problem of relying on manual experience and low data utilization in traditional methods. Specifically, in this embodiment, the method steps for multi-source data processing are as follows:

[0122] 11) Knowledge base construction

[0123] 111) Data integration: Integrate wind resource assessment reports and wind resource knowledge including IEC standard library, national standard library, industry standard library, industry design specifications, industry technical regulations and equipment white papers, and build a domain knowledge base covering wind power terminology, specifications and cases.

[0124] 112) Construct a professional terminology network graph: Based on the GraphSAGE algorithm, a wind power professional terminology graph network is constructed to achieve joint coding of related parameters such as "wind speed-power generation".

[0125] 113) Dynamic knowledge update: Develop a dynamic knowledge update engine that automatically identifies standard document version changes (such as NB / T 31146-2023) through a rule engine (Drools) and triggers the knowledge base reconstruction process to ensure that the knowledge base content is synchronized with the latest industry standards.

[0126] 12) Fine-tuning dataset construction: An example of constructing a fine-tuning dataset is as follows.

[0127] 121) Structured dataset definition: Using Alpaca-52k enhanced format, construct a triple structured dataset:

[0128] {

[0129] "context":"The annual average wind speed at a height of 120m in a certain wind farm is 6.8m / s",

[0130] "question":"What are the recommended models and corresponding hours?",

[0131] "answer":"Based on a single unit capacity of 6.7MW, the estimated operating hours are 2900h..."

[0132] }

[0133] 122) Data format standardization: Use dataset_info.json to define dataset metadata (such as name, path, processor type), convert raw data (such as CSV / text) to standard JSON format through scripts, and align field naming (such as instruction vs prompt).

[0134] 123) Text processing and enhancement:

[0135] Word segmentation and length control: Use sliding window segmentation for long texts. For example, segment long texts by 512 tokens, retaining context overlap (overlap rate 15%) to avoid information loss.

[0136] Data augmentation: Back-translation technology (such as Chinese-English translation) is used to generate multilingual variants, and the SimBERT model is used for semantic replacement to generate semantically consistent synonymous expressions to improve the diversity of the dataset.

[0137] 2. Domain-adaptive RAG enhancement:

[0138] Retrieval-enhanced generation technology is adopted to dynamically retrieve professional literature, historical cases and industry specifications in the domain knowledge base based on a generative large model, and multimodal fusion word vector technology is used to generate wind resource assessment content that meets technical standards; the multimodal fusion word vector technology includes a dual-stream encoding network and a cross-attention mechanism to achieve joint encoding and two-way interaction of text and image features.

[0139] This embodiment uses a generative large model, llama, combined with retrieval enhancement technology (RAG), which can dynamically retrieve professional literature, historical cases, and industry specifications from the knowledge base to generate wind resource assessment content that meets technical standards. In particular, this embodiment innovatively proposes a multimodal fusion word vector technology, using a dual-stream encoding method and an interpretable causal mechanism to construct weights. This allows the model to not only trace the source of feature contributions and improve the model's interpretability, but also build a two-way interactive channel through a cross-attention mechanism, further enhancing the model's integration and interpretability.

[0140] Specifically, such as Figure 3 As shown, in this embodiment, the method steps of domain adaptive RAG enhancement are as follows.

[0141] 21) Hybrid retrieval: including sparse retrieval, dense retrieval and compliance constraint channels, efficiently retrieves multimodal information from the knowledge base to ensure the professionalism and compliance of the generated content.

[0142] The hybrid search in this example uses the WeKnow-RAG framework, integrating:

[0143] Sparse search: Use the BM25 algorithm to match keywords and quickly filter relevant paragraphs in technical documents.

[0144] Dense retrieval: Use the Contriever model to encode text and numerical features into a unified semantic space and calculate semantic similarity.

[0145] Compliance constraint channel: Real-time access to the industry standard database to perform mandatory logical verification of key indicators such as "wind farm spacing design".

[0146] 22) Modal vector fusion: Encode text and image features into a unified semantic space to improve cross-modal information fusion capabilities.

[0147] 221) Design a two-stream encoding network: Use the improved attention_word2vec model to encode the text stream. This example constructs a 768-dimensional unified semantic space to construct new word vectors. This example uses convolutional network layers and linear regression layers to encode the image stream.

[0148] In this embodiment, the method for encoding the text stream by the dual-stream encoding network is:

[0149] attention_word2vec[text]+cosine rotation position encoding cosRoPE[pos]→text_feature

[0150] The improved attention_word2vec model introduces a self-attention mechanism based on the CBOW architecture to perform hierarchical feature extraction, including:

[0151] Word embedding layer: Dynamic word2vec is used to capture local semantics through a sliding window. The window size in this embodiment is 5.

[0152] Attention enhancement: Introducing multi-head attention to calculate global word relationship weights:

[0153]

[0154] Among them: Q / K / V are word vector projection residual connections from different subspaces; d k Represents the dimension of the word vector.

[0155] The number of attention heads in the multi-head attention layer of this embodiment is 4. The output and input of each attention layer are weighted fused.

[0156] In this embodiment, cosine rotational position encoding cosRoPE is a combination of rotational position encoding RoPE and cosine modulation. The principle is:

[0157]

[0158] Where: R(θ pos ) is a learnable rotation matrix; pos is the position index; i is the dimension index; θ pos is a learnable rotation angle parameter; is the matrix multiplication operator.

[0159] Add the results of attention_word2vec and cosRoPE to get the final text_feature.

[0160] In this embodiment, the method for encoding the image stream is:

[0161] convolution[image]→linear feature projection→image_feature

[0162] Among them: convolution is convolutional network; linear is linear regression;

[0163] Assume that the input image is I, and its size is W×H×c (width×height×number of channels). The principle of the convolutional network layer is:

[0164] O conv =f(I*K+b)

[0165] Where: K is the convolution kernel, also known as the filter; I*K represents the convolution operation applied to the input image I; b is the bias term; f is the activation function, such as ReLU; conv It is the output feature map obtained after the convolution operation.

[0166] The linear regression layer flattens the feature map output by the convolutional network layer into a one-dimensional vector, and uses LayerNorm to normalize it so that it aligns with the text feature dimension. The principle of the linear regression layer is:

[0167] y=LaterNorm(Wx+b′)

[0168] Wherein: W is the weight matrix. In this embodiment, the matrix shape of the weight matrix W is 512×2048, which is used to map the input feature map from 2048 dimensions to 512 dimensions; b′ is the bias term; LayerNorm represents the layer normalization operation, which is used to standardize the output so that the final output is aligned with the text feature dimension.

[0169] 222) Interpretable causal mechanism: A causal gating mechanism is introduced to construct an interpretable weight distribution module, and the interaction weights of text and image features are calculated through MLP.

[0170] Specifically, the principle of the causal gating mechanism is:

[0171] causal_gate=MLP(Text_Feature⊙Image_Feature)

[0172] Among them: causal_gate is the causal gating mechanism; MLP is the multi-layer perceptron; Text_Feature is the text feature; Image_Feature is the image feature.

[0173] The interaction weights of text and image features are:

[0174] gate=σ(Conv1D(text_feature)*causal_gate)

[0175] Where: gate is the overall gating mechanism; Conv1D is a one-dimensional convolutional network; σ is the sigma function.

[0176] This method can trace the source of feature contributions (such as identifying the driving effect of the "turbulence intensity" area in the image on the text description) and improve the interpretability of the model.

[0177] 223) Bidirectional feature enhancement mechanism: Use the cross attention mechanism (MCA) to generate image-driven text enhancement features and text-driven image enhancement features, and fuse the text enhancement features and image enhancement features through residual connections for bidirectional output.

[0178] Image → Text: Use image features as query and text as key / value

[0179] Text → Image: Take text features as query and image as key / value and finally fuse the two-way output through residual connection. Specifically, the two-way output obtained by fusing text enhancement features and image enhancement features through residual connection is:

[0180] F final =MCA(V→T)+α·MCA(T→V)

[0181] Among them: F final To fuse text enhancement features and image enhancement features through residual connection, MCA(V→T) is the text enhancement feature; MCA(T→V) is the image enhancement feature; α is the weight coefficient; V is the visual image; T is the text.

[0182] 23) Semantic Contrast Alignment: In cross-modal feature fusion, the role of the semantic contrast alignment constraint is to enforce the consistency of image features and text features in the semantic space. Specifically, in this embodiment, the loss function of semantic contrast alignment is:

[0183]

[0184] in:

[0185] Molecular part: Represents the similarity score indexation result of the positive sample pair; sim(v,t + ) is the visual feature v and the corresponding positive sample text feature t + The cosine similarity is used to quantify the degree of semantic alignment between the two. τ is the temperature coefficient, usually ranging from 0.05 to 0.2, which is used to control the steepness of the probability distribution. A smaller τ will amplify the discrimination of difficult negative samples.

[0186] Denominator: Contains the sum of similarities between the positive sample and all negative samples; t― It is the negative sample text feature, that is, the text that does not match the current image, obtained by random sampling or hard negative mining.

[0187] Overall loss function: By maximizing the similarity ratio (numerator / denominator) of positive sample pairs, the model is forced to learn discriminative features for cross-modal alignment. The smaller the loss value, the more similar the positive sample pairs are than the negative sample pairs.

[0188] This constraint uses contrastive learning to map image and text features with the same semantics to nearby regions in a unified semantic space, while pushing irrelevant features away. For example, a "turbulence intensity map" in a wind farm image should be highly similar to its corresponding text description in vector space.

[0189] 3. Fine-tuning of large models:

[0190] A domain-adaptive fine-tuning strategy is adopted to optimize the generative large model through the gradient-free optimization algorithm Wind-DFO, which includes dynamic Latin hypercube sampling, Bayesian surrogate model construction and adaptive learning rate update.

[0191] During large-model fine-tuning, this example proposes the Wind-DFO optimization algorithm, which integrates multi-objective optimization theory, efficient sampling strategies, and adaptive learning rate technology. Compared with traditional optimization methods, Wind-DFO provides a new optimization paradigm for fine-tuning in the LLM (Large Language Model) domain, significantly improving the effectiveness and efficiency of model fine-tuning, especially for high-dimensional non-stationary optimization problems.

[0192] The domain adaptive fine-tuning strategy is:

[0193] Using QLoRA optimization strategy, key parameter configuration:

[0194] lora_rank=64;

[0195] lora_alpha=32;

[0196] lora_dropout=0.1.

[0197] Introduction to course learning mechanism:

[0198] Initial training: 20% basic questions and answers + 80% standard clause analysis;

[0199] Mid-term addition: 50% complex scenario decisions (such as extreme climate adaptability design);

[0200] Late-stage enhancement: 30% economic optimization tasks (LCOE minimization).

[0201] In this embodiment, the non-gradient optimization algorithm Wind-DFO implements dynamic Latin hypercube sampling in the parameter space and selects the steepest descent direction based on Bayesian optimization. Figure 4 As shown, the method steps for optimizing the generative large model by using the non-gradient optimization algorithm Wind-DFO in this embodiment are as follows.

[0202] 31) Dynamic parameter space partitioning and initialization:

[0203] Input: Model parameter initial point θ0 (QLoRA adaptation layer parameters, dimension).

[0204] Operation: Generate candidate parameter points through dynamic Latin hypercube sampling:

[0205]

[0206] Where: θ (i) is the i-th candidate point; is a uniform distribution, indicating that the parameter range in each dimension is from θ min to θ max ; N is the number of candidate points generated.

[0207] Calculate the initial loss and set the current optimal solution.

[0208] 32) Bayesian surrogate model construction:

[0209] Goal: Establish a mapping relationship from parameter space to loss function

[0210] Operation: Establish a mapping relationship from parameter space to loss function based on the Gaussian process surrogate model:

[0211]

[0212] Where: f(θ) is the objective function value under parameter θ; μ(θ) is the mean function; k(θ,θ′) is the kernel function, which defines the similarity between different parameter points.

[0213] An improved kernel function is used to adapt to high-dimensional non-stationary optimization. In this embodiment, the kernel function is Matérn 5 / 2:

[0214]

[0215] Where: r is the Euclidean distance between two parameter points, that is, r = ‖θ―θ′‖2; ‖·‖2 is the L2 norm, that is, the Euclidean distance.

[0216] 33) Multi-objective acquisition function optimization:

[0217] Goal: Balance exploration (unknown areas) and utilization (current optimal areas);

[0218] Operation: Expected improvement function for design improvement:

[0219]

[0220] Where: α EI (θ) is the expected improvement value at parameter θ; f min is the currently known optimal loss value; f(θ) is the loss value predicted by the proxy model; σ 2 (θ) is the variance of the surrogate model prediction; ∈1 is the numerical stability coefficient to prevent the denominator from being zero.

[0221] The candidate direction set is generated by the CMA-ES algorithm.

[0222] 34) Gradient-independent direction search:

[0223] Objective: To determine the optimal descent direction;

[0224] operate:

[0225] (1) The candidate direction sets are evaluated in parallel and the actual loss is calculated.

[0226] (2) Select the Pareto frontier solution: After comprehensive sorting based on loss reduction and directional stability, the optimal direction is determined.

[0227] 35) Determine whether the rate of change of the actual loss for m consecutive iterations is lower than a set threshold. If so, terminate the iteration, obtain the optimal parameters, and use the Bayesian surrogate model fitting error for subsequent model diagnosis. If not, execute step 36. In this embodiment, the value of m is 5.

[0228] 36) Dynamic learning rate update:

[0229] Goal: Adaptively adjust the parameter update step size.

[0230] Operation: Calculate the learning rate based on the momentum adaptation mechanism:

[0231]

[0232] Where: η t is the learning rate at step t; η base is the basic learning rate; g t is the pseudo gradient estimate, which represents the gradient estimate at step t; ‖g t ‖ is the L2 norm of the pseudo gradient estimate; ∈2 is the numerical stability coefficient to prevent the denominator from being zero.

[0233] Update parameters:

[0234] θ best ←θ best +η t ·g t

[0235] Where: θ best is the current optimal parameter; η t is the learning rate of the current step; g t is the pseudo gradient estimate.

[0236] The Wind-DFO algorithm is essentially a black-box optimizer based on Bayesian optimization. It replaces explicit gradient calculations with surrogate models and is suitable for:

[0237] (1) High-dimensional non-convex parameter space (such as LLM fine-tuning);

[0238] (2) Scenarios with limited hardware resources (low video memory requirements);

[0239] (3) Multi-objective trade-off optimization (such as accuracy-efficiency balance).

[0240] The algorithm combines multi-objective optimization theory, efficient sampling strategy, and adaptive learning rate technology, providing a new optimization paradigm for LLM field fine-tuning.

[0241] 37) Determine whether the current iteration step t is equal to the maximum allowed iteration step T: If so, terminate the iteration, obtain the optimal parameters, and use the Bayesian surrogate model to fit the error; if not, set t = t + 1 and loop through step 31).

[0242] 4. Automated report generation:

[0243] Through AI workflow orchestration and matching of multiple templates, data visualization charts are embedded, and syntax, logic, and compliance layer verification are performed to output a standardized wind resource assessment report.

[0244] The automated report generation module utilizes AI workflow orchestration, supporting multi-template matching, data visualization, and compliance verification, providing a one-stop service from data analysis to report generation. This implementation significantly simplifies the report generation process, reduces human intervention, ensures professional and standardized reporting, and meets the diverse needs of wind power project development.

[0245] In this embodiment, the method steps for automatic report generation are as follows.

[0246] 41) Dynamic Orchestration Engine: Uses a finite state machine to control the generation process:

[0247] Initial state → Data verification → Project overview → Wind energy resource analysis → Power generation calculation → Conclusion

[0248] That is, this embodiment defines the state transitions of each stage through Python's transitions library, including initial state, data verification, project overview, wind energy resource analysis, power generation calculation and conclusion.

[0249] 42) Multi-dimensional output: Use the Jinja2 template engine to generate technical reports in Markdown format, and automatically insert visual charts drawn by Matplotlib to generate reports.

[0250] During the report generation process, multiple levels of validation are employed, including:

[0251] Syntax layer: Use LangChain to perform syntax layer verification to ensure that the text content complies with grammatical rules;

[0252] Logic layer: Use CLIPS rule engine to verify logical consistency and ensure that all calculation results are reasonable and correct;

[0253] Compliance layer: Automatically mark relevant standard clause indexes in reports to ensure compliance.

[0254] In this embodiment, in order to realize AI intelligent report generation, the finite state machine (FSM) is first used to control the generation process, and the state transitions of each stage are defined through Python's transitions library, including the initial state, data verification, project overview, wind energy resource analysis, power generation calculation and conclusion. Then, the Jinja2 template engine is used to generate a technical report in Markdown format, and a visual chart drawn by Matplotlib is automatically inserted. In the process of generating the report, a multi-level verification mechanism is implemented: first, LangChain is used for syntax verification to ensure that the text content complies with the grammatical rules; then the CLIPS rule engine is used to verify the logical consistency to ensure that the calculation results are reasonable and correct; finally, the relevant standard clause index is automatically marked in the report to ensure compliance. The specific steps are as follows:

[0255] Dynamic orchestration engine: Defines a finite state machine (FSM) that starts from the initial state and sequentially performs data verification, generates a project overview, analyzes wind energy resources, calculates power generation, and finally generates a conclusion.

[0256] Multi-dimensional output system:

[0257] Use the Jinja2 template engine to generate reports in Markdown format and embed charts drawn by Matplotlib.

[0258] Syntax verification: LangChain calls the language model to check and correct grammatical errors in the text.

[0259] Logical layer verification: Load and run the CLIPS rule engine to verify the logical consistency of the calculation results according to predefined rules.

[0260] Compliance layer verification: Automatically mark relevant standard clause indexes in the report to ensure that the content complies with industry regulations.

[0261] The entire process not only achieves a high degree of automation in report generation, but also ensures report quality and compliance through a multi-level verification mechanism. This solution can be flexibly applied to various vertical fields, quickly generating high-quality professional reports, as shown in Table 1.

[0262] Table 1 Comparison of technical advantages

[0263] index Traditional methods The present invention Improvement Report generation cycle 2-3 weeks ≤16 hours 7.5 times Compliance verification coverage 70％-80％ 85％ 5％-15％

[0264] This embodiment has been verified through actual projects, and the report passing rate has been improved, while the workload of expert review has been reduced.

[0265] The above embodiments are merely preferred embodiments for the purpose of fully illustrating the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are within the scope of protection of the present invention. The scope of protection of the present invention shall be subject to the claims.

Claims

1. A method for generating a wind resource assessment report based on large model technology, characterized by: The following steps are involved: Multi-source data processing: Integrate wind resource knowledge and wind resource assessment reports to build a domain knowledge base covering wind power terminology, specifications, and cases. Intelligently extract wind resource characteristics, including average wind speed, turbulence intensity, and wind shear index, through machine learning algorithms, and dynamically update the knowledge base. Construct a fine-tuning dataset based on the triple structure, and expand and enhance the fine-tuning dataset; Domain-adaptive RAG enhancement: This technology uses retrieval-enhanced generation technology to dynamically retrieve professional literature, historical cases, and industry specifications from the domain knowledge base based on a generative large model. Multimodal word embedding technology is then used to generate wind resource assessment content that meets technical standards. This multimodal word embedding technology includes a dual-stream encoding network and a cross-attention mechanism to achieve joint encoding and bidirectional interaction of text and image features. Large model fine-tuning: Using a domain-adaptive fine-tuning strategy, we optimize the generative large model based on the fine-tuning dataset using the Wind-DFO gradient-free optimization algorithm. The Wind-DFO algorithm includes dynamic Latin hypercube sampling, Bayesian surrogate model construction, and adaptive learning rate updates. Automated report generation: Through AI workflow orchestration, multiple templates are matched, data visualization charts are embedded, and syntax, logic, and compliance layer verification are performed to output standardized wind resource assessment reports.

2. The method for generating a wind resource assessment report based on large model technology according to claim 1 is characterized in that: The steps for multi-source data processing are as follows: 11) Knowledge base construction 111) Data integration: Integrate wind resource assessment reports and wind resource knowledge including IEC standard library, national standard library, industry standard library, industry design specifications, industry technical regulations and equipment white papers, and build a domain knowledge base covering wind power terminology, specifications and cases; 112) Constructing a professional terminology network graph: Based on the GraphSAGE algorithm, a wind power professional terminology graph network is constructed to achieve joint coding of associated parameters; 113) Dynamic knowledge update: Develop a dynamic knowledge update engine that automatically identifies standard document version changes through a rule engine and triggers the knowledge base reconstruction process to ensure that the knowledge base content is synchronized with the latest industry standards; 12) Fine-tuning dataset construction 121) Structured dataset definition: Using Alpaca-52k enhanced format, construct a triple structured dataset; 122) Data format standardization: Use dataset_info.json to define dataset metadata, convert raw data into standard JSON format through scripts, and align field naming; 123) Text processing and enhancement: Word segmentation and length control: Use sliding windows to segment long texts to avoid information loss; Data augmentation: Back-translation technology is used to generate multilingual variants, and the SimBERT model is used for semantic replacement to generate semantically consistent synonymous expressions to improve the diversity of the dataset.

3. The method for generating a wind resource assessment report based on large model technology according to claim 1 is characterized in that: The method steps for domain adaptive RAG enhancement are as follows: 21) Hybrid Search: This includes sparse search, dense search, and compliance constraint channels, efficiently retrieving multimodal information from the knowledge base to ensure the professionalism and compliance of generated content. The sparse search is based on keyword matching to quickly screen relevant paragraphs in technical documents. The dense search encodes text and numerical features into a unified semantic space and calculates semantic similarity. The compliance constraint channel accesses the industry standard database in real time to perform mandatory logical verification of key indicators. 22) Modal vector fusion: Encode text and image features into a unified semantic space to improve cross-modal information fusion capabilities; 221) Design a two-stream encoding network: use the improved attention_word2vec model to encode the text stream, and use convolutional network layers and linear regression layers to encode the image stream; 222) Interpretable Causal Mechanism: A causal gating mechanism is introduced to construct an interpretable weight distribution module, and the interaction weights of text and image features are calculated through MLP; 223) Bidirectional feature enhancement mechanism: Use the cross-attention mechanism to generate image-driven text enhancement features and text-driven image enhancement features, and fuse the text enhancement features and image enhancement features through residual connections to generate bidirectional outputs; 23) Semantic Contrast Alignment: Through contrastive learning mechanism, the consistency of image features and text features in the semantic space is enforced.

4. The method for generating a wind resource assessment report based on large model technology according to claim 3 is characterized in that: In step 221), the method for encoding the text stream is: attention_word2vec[text]+cosine rotation position encoding cosRoPE[pos]→text_feature The improved attention_word2vec model introduces a self-attention mechanism based on the CBOW architecture to perform hierarchical feature extraction, including: Word embedding layer: uses dynamic word2vec to capture local semantics through sliding windows; Attention enhancement: Multi-head attention is introduced to calculate the global word relationship weight, and the output and input of each attention layer are weighted and fused; Cosine Rotational Position Encoding (cosRoPE) combines Rotational Position Encoding (RoPE) with cosine modulation. The principle is as follows: Where: R(θ pos ) is a learnable rotation matrix; pos is the position index; i is the dimension index; θ pos is a learnable rotation angle parameter; is the matrix multiplication operator.

5. The method for generating a wind resource assessment report based on large model technology according to claim 3 is characterized in that: In step 221), the method for encoding the image stream is: convolution[image]→linear feature projection→image_feature Among them: convolution is convolutional network; linear is linear regression; The principle of the convolutional network layer is: O conv =f(I*K+b) Where: K is the convolution kernel; I*K represents the convolution operation applied to the input image I; b is the bias term; f is the activation function; O conv It is the output feature map obtained after the convolution operation; The linear regression layer flattens the feature map output by the convolutional network layer into a one-dimensional vector, and aligns it with the text feature dimension with LayerNorm normalization. The principle of the linear regression layer is: y = LayerNorm(Wx+b′) Where: W is the weight matrix; b′ is the bias term; LayerNorm represents the layer normalization operation, which is used to standardize the output so that the final output is aligned with the text feature dimension.

6. The method for generating a wind resource assessment report based on large model technology according to claim 3 is characterized in that: In step 222), the principle of the causal gating mechanism is: causal_gate=MLP(Text_Feature⊙Image_Feature) Among them: causal_gate is the causal gating mechanism; MLP is the multi-layer perceptron; Text_Feature is the text feature; Image_Feature is the image feature; The interaction weights of text and image features are: gate=σ(Conv1D(text_feature)*causal_gate) Where: gate is the overall gating mechanism; Conv1D is a one-dimensional convolutional network; σ is the sigma function.

7. The method for generating a wind resource assessment report based on large model technology according to claim 3 is characterized in that: In step 223), the bidirectional output obtained by fusing the text enhancement features and the image enhancement features through residual connection is: F final =MCA(V→T)+α·MCA(T→V) Among them: F final To fuse text enhancement features and image enhancement features through residual connection, MCA(V→T) is the text enhancement feature; MCA(T→V) is the image enhancement feature; α is the weight coefficient; V is the visual image; T is the text.

8. The method for generating a wind resource assessment report based on large model technology according to claim 3 is characterized in that: In step 23), the loss function of semantic comparison alignment is: Among them: e sim(v,t+) / τ Represents the similarity score indexation result of the positive sample pair; sim(v,t + ) is the visual feature v and the corresponding positive sample text feature t + The cosine similarity of is used to quantify the degree of semantic alignment between the two; τ is the temperature coefficient, which controls the steepness of the probability distribution; Contains the sum of similarities between the positive sample and all negative samples; t ― is the negative sample text feature.

9. The method for generating a wind resource assessment report based on large model technology according to claim 1, characterized in that: The steps for optimizing the generative large model using the non-gradient optimization algorithm Wind-DFO are as follows: 31) Dynamic parameter space partitioning and initialization: Generate candidate parameter points through dynamic Latin hypercube sampling: Where: θ (i) is the i-th candidate point; is a uniform distribution, indicating that the parameter range in each dimension is from θ min to θ max ; N is the number of candidate points generated; Calculate the initial loss and set the current optimal solution; 32) Bayesian surrogate model construction: Based on the Gaussian process surrogate model, the mapping relationship between the parameter space and the loss function is established: Where: f(θ) is the objective function value under parameter θ; μ(θ) is the mean function; k(θ,θ′) is the kernel function, which defines the similarity between different parameter points; Adopting improved kernel function to adapt to high-dimensional non-stationary optimization: Where: r is the Euclidean distance between two parameter points, that is, r = ‖θ―θ′‖2; ‖·‖2 is the L2 norm; 33) Multi-objective acquisition function optimization: Designing improved expected improvement functions: Where: α EI (θ) is the expected improvement value at parameter θ; f min is the currently known optimal loss value; f(θ) is the loss value predicted by the proxy model; σ 2 (θ) is the variance of the proxy model prediction; ∈1 is the numerical stability coefficient to prevent the denominator from being zero; Generate a candidate direction set through the CMA-ES algorithm; 34) Gradient-independent direction search: Evaluate the actual loss of candidate directions and select the Pareto frontier solution to determine the optimal descent direction; 35) Determine whether the rate of change of the actual loss for m consecutive iteration steps is lower than the set threshold: if so, terminate the iteration, obtain the optimal parameters, and use the Bayesian surrogate model to fit the error; if not: execute step 36); 36) Dynamic learning rate update: Learning rate is calculated based on momentum adaptation mechanism: Where: η t is the learning rate at step t; η base is the basic learning rate; g t is the pseudo gradient estimate, which represents the gradient estimate at step t; ‖g t ‖ is the L2 norm of the pseudo gradient estimate; ∈2 is the numerical stability coefficient to prevent the denominator from being zero; Update parameters: i best ←θ best +n t ·g t Where: θ best is the current optimal parameter; η t is the learning rate of the current step; g t is the pseudo gradient estimate; 37) Determine whether the current iteration step t is equal to the maximum allowed iteration step T: If so, terminate the iteration, obtain the optimal parameters, and use the Bayesian surrogate model to fit the error; if not, set t = t + 1 and loop through step 31).

10. The method for generating a wind resource assessment report based on large model technology according to claim 1, characterized in that: The steps for automated report generation are: 41) Dynamic Orchestration Engine: Uses a finite state machine to control the generation process and uses Python's transitions library to define state transitions at each stage, including initial state, data verification, project overview, wind energy resource analysis, power generation calculation, and conclusion. 42) Multi-dimensional output: Use the Jinja2 template engine to generate technical reports in Markdown format, and automatically insert visual charts drawn by Matplotlib to generate reports; During the report generation process, multiple levels of validation are employed, including: Syntax layer: Use LangChain to perform syntax layer verification to ensure that the text content complies with grammatical rules; Logic layer: Use CLIPS rule engine to verify logical consistency and ensure that all calculation results are reasonable and correct; Compliance layer: Automatically mark relevant standard clause indexes in reports to ensure compliance.

Images