Wind resource assessment report generation method based on large model technology

Abstract

This invention discloses a method for generating wind resource assessment reports based on large-scale model technology, comprising: multi-source data processing to construct a domain knowledge base and a fine-tuning database; domain-adaptive RAG enhancement, employing retrieval enhancement generation technology to retrieve professional literature, historical cases, and industry standards from the domain knowledge base, and generating wind resource assessment content that conforms to technical standards through multimodal fusion word vector technology; large-scale model fine-tuning, adopting a domain-adaptive fine-tuning strategy, optimizing the generative large-scale model based on the fine-tuning database and using the gradient-free optimization algorithm Wind-DFO; and automated report generation, using AI workflow orchestration to match multiple templates, embedding data visualization charts, and performing syntax, logic, and compliance layer checks to output a standardized wind resource assessment report. This invention significantly improves the accuracy, professionalism, and efficiency of report generation, solving the pain points of traditional methods such as reliance on human experience, low data utilization efficiency, and insufficient standardization.

Description

Technical Field

[0001] This invention belongs to the field of wind power generation technology, specifically a method for generating wind resource assessment reports based on large model technology. Background Technology

[0002] Wind resource assessment reports have strategic guiding significance for wind power projects throughout their entire lifecycle. As the core basis for project development, they provide scientific support for site selection optimization and turbine selection through multi-dimensional data quantitative analysis. However, traditional wind resource assessment methods have bottlenecks: First, they rely on manual, end-to-end 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 7-10 days, involve numerous procedures, and are time-consuming and labor-intensive, making it difficult to meet the needs of rapid assessment and decision-making; third, the assessment process depends on the experience of wind resource engineers, who need extensive practical experience to increase the accuracy of assessments for complex projects. Without the deep involvement of experienced engineers, it is difficult to ensure that the accuracy of the wind resource assessment report meets the requirements.

[0003] With the accelerated global energy transition, wind power, as a core pillar of clean energy, faces increasingly higher demands for accurate and intelligent wind resource assessments as its large-scale development progresses. Traditional wind resource assessment reports rely heavily on human experience, resulting in the following shortcomings:

[0004] (1) Limitations of data fusion: 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 roughness simplification methods, ignoring the spatial heterogeneity of leaf area density distribution and vegetation drag coefficient, or using single numerical simulation methods, which cannot adapt to the meteorological characteristics of different regions. Although existing automated tools can improve computational speed, they lack multimodal data fusion capabilities (such as semantic association between topographic maps and normative texts), resulting in a lack of interpretability for key indicators (turbulence intensity, maximum wind speed returning once in 50 years).

[0005] (2) Low data processing efficiency and long processing time: Manually collecting, cleaning and processing large amounts of meteorological and geographical data is very time-consuming, especially when dealing with multi-source data, which is prone to delays and errors: Human errors are easily introduced during manual operation, 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 differently, leading to 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 of different projects may be difficult to compare directly, which reduces the credibility of the evaluation.

[0008] (5) Low report generation efficiency: Manually writing reports requires a lot of time for layout and formatting, especially when there are a lot of charts and data. There is a lot of repetitive work: For multiple similar projects, reports need to be written from scratch each time, which cannot effectively reuse existing templates and content, increasing the workload.

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

[0010] (7) Inconvenient collaboration and sharing, complex team collaboration: In projects involving multiple people, coordinating and synchronizing the work of each part is more complicated, which can easily lead to poor communication and information asymmetry. It is difficult to systematically record and pass on personal experience and knowledge, and the training cost for new employees is high, which is not conducive to long-term development. Summary of the Invention

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

[0012] To achieve the above objectives, the present invention provides the following technical solution:

[0013] A method for generating wind resource assessment reports 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, standards 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. Build a fine-tuning dataset based on the triplet structure and expand and enhance the fine-tuning dataset.

[0015] Domain-Adaptive RAG Enhancement: Employing retrieval enhancement generation technology, this method dynamically retrieves professional literature, historical cases, and industry standards from the domain knowledge base based on a generative large model. It then generates wind resource assessment content that conforms to technical standards through multimodal fusion word vector technology. This multimodal fusion word vector technology includes a dual-stream coding network and a cross-attention mechanism to achieve joint encoding and bidirectional interaction of text and image features.

[0016] Large model fine-tuning: A domain-adaptive fine-tuning strategy is adopted, and the generative large model is optimized by the gradient-free optimization algorithm Wind-DFO. The Wind-DFO algorithm includes dynamic Latin hypercube sampling, Bayesian surrogate model construction and adaptive learning rate update.

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

[0018] Furthermore, the 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, to build a domain knowledge base covering wind power terminology, specifications and cases;

[0021] 112) Constructing a technical terminology network: Based on the GraphSAGE algorithm, construct a wind power technical terminology network to achieve joint encoding of associated parameters;

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

[0023] 12) Fine-tuning dataset construction

[0024] 121) Definition of structured dataset: Construct a triplet structured dataset using the Alpaca-52k enhanced format;

[0025] 122) Data format standardization: Use dataset_info.json to define dataset metadata, and use a script to convert the raw data into standard JSON format, aligning field names;

[0026] 123) Text Processing and Enhancement:

[0027] Word segmentation and length control: Long texts are segmented using a sliding window to avoid information loss;

[0028] Data augmentation: Back-translation technology is used to generate multilingual variants, and the SimBERT model is used for semantic substitution to generate semantically consistent synonyms, thereby improving the diversity of the dataset.

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

[0030] 21) Hybrid retrieval: This includes sparse retrieval, dense retrieval, and a compliance constraint channel, which efficiently retrieves multimodal information from the knowledge base to ensure the professionalism and compliance of the generated content. The sparse retrieval is based on keyword matching to quickly filter relevant paragraphs in technical documents. The dense retrieval 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 on key indicators.

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

[0032] 221) Design a dual-stream encoding network: use an 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) Explainable causal mechanism: Introduce a causal gating mechanism to construct an explainable weight allocation module, and calculate the interaction weights of text and image features through MLP;

[0034] 223) Bidirectional feature enhancement mechanism: Use 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 connection for bidirectional output;

[0035] 23) Semantic contrast alignment: Through contrastive learning mechanism, the consistency between image features and text features in semantic space is forced to be unified.

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

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

[0038] The improved attention_word2vec model introduces a self-attention mechanism on top of the CBOW architecture for hierarchical feature extraction, including:

[0039] Word embedding layer: Employs dynamic word2vec to capture local semantics through a sliding window;

[0040] Attention enhancement: Multi-head attention is introduced to calculate global word relationship weights, and the outputs and inputs 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 It is a learnable rotation angle parameter; This is the matrix multiplication operator.

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

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

[0046] Where: convolution is a convolutional network; linear is linear regression;

[0047] The principle of convolutional network layers is:

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

[0049] Where: K is the convolution kernel; I*K represents applying a convolution operation 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 convolution operation;

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

[0051] y = LayerNorm(Wx + b′)

[0052] Where: W is the weight matrix; b′ is the bias term; LayerNorm represents the layer normalization operation, 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 as follows:

[0054] causal_gate=MLP(Text_Feature⊙Image_Feature)

[0055] Where: causal_gate is the causal gating mechanism; MLP is the multilayer perceptron; Text_Feature is the text feature; Image_Feature is the image feature;

[0056] The interaction weights between 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 image enhancement features through residual connections is:

[0060] F final =MCA(V→T)+α·MCA(T→V)

[0061] Wherein: F final To fuse text enhancement features and image enhancement features through residual connections for bidirectional output; MCA(V→T) represents text enhancement features; MCA(T→V) represents image enhancement features; α represents weight coefficients; V represents visual image; T represents text.

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

[0063]

[0064] in: This represents the indexed similarity score of positive sample pairs; sim(v,t) + () represents 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 a temperature coefficient that controls the steepness of the probability distribution. The sum of similarities between positive samples and all negative samples; t ― These are negative sample text features.

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

[0066] 31) Dynamic parameter space partitioning and initialization: Candidate parameter points are generated through dynamic Latin hypercube sampling:

[0067]

[0068] Where: θ (i) It is the i-th candidate point; It is a uniform distribution, meaning the range of parameters 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: Establishing a mapping relationship from the parameter space to the loss function based on the Gaussian process surrogate model:

[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 points with different parameters;

[0073] An improved kernel function is used to adapt to high-dimensional nonstationary optimization:

[0074]

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

[0076] 33) Multi-target acquisition function optimization: Design of the desired improved function:

[0077]

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

[0079] Candidate direction sets are generated using the CMA-ES algorithm;

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

[0081] 35) Determine if the rate of change of the actual loss over m consecutive iterations is lower than a set threshold: if yes, terminate the iteration, obtain the optimal parameters, and fit the error using a Bayesian surrogate model; if no, proceed to step 36).

[0082] 36) Dynamic learning rate update: Calculating the learning rate based on a momentum adaptive mechanism:

[0083]

[0084] Where: η t Let η be the learning rate at step t; base Based on the learning rate; g t For pseudo-gradient estimation, let g represent the gradient estimate at step t; ||g t|| represents 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 It is the current optimal parameter; η t g is the learning rate for the current step. t This is a pseudo-gradient estimate;

[0088] 37) Determine if the current iteration step t is equal to the maximum allowed iteration step T: if yes, terminate the iteration, obtain the optimal parameters, and fit the error using a Bayesian surrogate model; if no, let t = t + 1, and repeat step 31).

[0089] Furthermore, the steps for automated report generation are as follows:

[0090] 41) Dynamic orchestration engine: The generation process is controlled by a finite state machine. The state transitions of each stage are defined by Python's transitions library, including the initial state, data verification, project overview, wind energy resource analysis, power generation calculation and conclusion.

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

[0092] The report generation process employs multi-level validation, including:

[0093] Syntax layer: LangChain is used for syntax layer validation to ensure that the text content conforms to the syntax rules;

[0094] Logic layer: Uses CLIPS rule engine to verify logical consistency and ensure that all calculation results are reasonable and error-free;

[0095] Compliance layer: The report automatically includes indexes of relevant standard clauses to ensure compliance.

[0096] The beneficial effects of this invention are as follows:

[0097] This invention presents a wind resource assessment report generation method based on large model technology. Through the synergy of RAG technology and dynamic optimization strategies, it significantly improves the accuracy, professionalism, and efficiency of report generation. It addresses the pain points of traditional methods, such as reliance on human experience, low data utilization, and insufficient standardization. This provides efficient and intelligent decision 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] In the multi-source data processing steps, by integrating multiple data sources and constructing a domain knowledge base, intelligent extraction and dynamic updating of wind resource characteristics are achieved. This not only improves data quality and consistency but also automatically identifies and extracts key features, reducing the need for manual intervention and ensuring the efficiency and accuracy of data processing. In particular, with the continuous input of new data, the system can dynamically update 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 adopted. By dynamically retrieving professional literature, historical cases, and industry standards from the knowledge base, wind resource assessment content that conforms to technical standards is generated. Through multimodal fusion word vector technology, using dual-stream coding and an interpretable causal mechanism, interpretable weights are constructed, enabling the model to trace the source of feature contributions and improving the model's interpretability. Furthermore, a bidirectional interaction channel is introduced through a cross-attention mechanism, enhancing the model's interpretability and fusion, making the generated report not only accurate but also more transparent, facilitating users' understanding of the basis behind each conclusion.

[0101] (3) In the large model fine-tuning step, a domain-adaptive fine-tuning approach is adopted for the large model, and the QLoRA optimization strategy is introduced. In particular, by integrating the gradient-free optimization algorithm Wind-DFO with multi-objective optimization theory, efficient sampling strategies, and adaptive learning rate techniques, a new optimization paradigm for fine-tuning in the LLM field is provided. The optimization strategy of this invention not only improves the performance of the model in specific scenarios, but also significantly reduces the consumption of computational resources, improves training efficiency, and enables the model to more flexibly cope with the wind resource assessment needs under different regions and conditions.

[0102] (4) In the automated report generation process, an AI workflow orchestration method is adopted, supporting multi-template matching, data visualization, and compliance verification. A wind resource assessment report is output with one click, significantly reducing the time and complexity of manual operations and improving the efficiency and standardization of report generation. This invention can select appropriate report templates for data visualization processing according to user needs, intuitively displaying key information such as wind resource distribution and power generation potential. The built-in compliance verification function ensures that the generated report meets industry standards and regulatory requirements, further enhancing the report's professionalism and credibility.

[0103] (5) Addressing the pain points of traditional methods and promoting industry development

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

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

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

[0107] In summary, this invention significantly improves the accuracy, professionalism, and efficiency of wind resource assessment report generation by organically combining four steps: multi-source data processing, domain-adaptive RAG enhancement, large-scale model fine-tuning, and automated report generation. This invention not only addresses the pain points of traditional methods, such as reliance on human experience, low data utilization, and insufficient standardization, but also provides efficient and intelligent decision support for wind power project development, promoting the large-scale development and digital transformation of clean energy. This method not only helps improve the economic and social benefits of wind power projects but also opens up new directions for future research and application, promoting the continuous progress and development of the industry. Attached Figure Description

[0108] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration:

[0109] Figure 1 This is the main flowchart of an embodiment of the wind resource assessment report generation method based on large model technology of the present invention;

[0110] Figure 2 This is a detailed flowchart of an embodiment of the wind resource assessment report generation method based on large model technology of the present invention;

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

[0112] Figure 4 This is a flowchart illustrating the Wind-DFO algorithm, a gradient-free optimization algorithm, used to optimize large generative models. Detailed Implementation

[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 and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0114] Large-scale model technology refers to machine learning models with a large number of parameters built using deep learning algorithms, especially those 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 nuanced patterns and features in data. Large-scale models, trained on massive amounts of text data, can understand and generate high-quality natural language text, supporting various application scenarios from translation and question answering to text creation. Advances in large-scale models have driven the development 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. The wind resource assessment report generation method based on large-scale model technology relies on the Transformer architecture and self-attention mechanism, fusing massive parameters and multimodal data to achieve complex tasks such as wind energy potential quantification and turbine layout optimization. This technology not only restructures the entire process from data cleaning to compliance verification (traditional manual processing takes 15-30 days and the standardization rate is less than 60%), but also directly empowers the "dual carbon" strategy through carbon sink verification, promoting wind power development from experience-driven to a new paradigm of AI-based precise decision-making.

[0115] Llama is a large-scale language model open-sourced by Meta (formerly Facebook) to advance research and applications in Natural Language Processing (NLP). Based on the Transformer architecture and trained on vast amounts of text data, Llama boasts powerful text generation and understanding capabilities. A key feature is its open-source nature, allowing researchers and developers free access to, modification, and deployment of the model, accelerating innovation and reducing development costs. By providing a suite of pre-trained models, Llama supports diverse 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 collectively advance artificial intelligence. Whether for academic research or commercial applications, Llama provides users with a flexible and powerful toolset to meet diverse task requirements.

[0116] Retrieval-Augmented Generation (RAG) based large-scale modeling combines the advantages of traditional generative models and information retrieval systems, aiming to improve the accuracy and relevance of generation tasks. A 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 repository 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 up-to-date information from external knowledge bases but also overcomes the limitations of traditional methods that rely solely on storing knowledge through internal model parameters, such as outdated or insufficient knowledge coverage. In this way, RAG-based large-scale models demonstrate higher accuracy and flexibility in question-answering systems, text summarization, and other natural language generation tasks, opening new avenues for more intelligent and data-driven applications.

[0117] Large-scale model technology based on AI workflow orchestration automates and intelligently processes complex tasks by integrating and optimizing multiple AI components or modules. This technology utilizes advanced workflow orchestration tools to organically combine data collection and preprocessing, feature engineering, model training and tuning, inference, and post-processing steps into an efficient and flexible workflow. In large-scale model applications, this means that the resources and technology stack used in each step can be dynamically adjusted according to specific task requirements, thereby maximizing performance and efficiency. For example, in Natural Language Processing (NLP) tasks, workflow orchestration can achieve the following process: First, the raw text is processed using a specially designed data cleaning and preprocessing module; then, the most suitable scheme is selected from various feature extraction methods, such as word embedding or sentence vector representation, according to task requirements; next, the large-scale pre-trained model best suited for the current task is invoked for further analysis or generation operations; 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 insert new algorithms or adjust existing components to adapt to constantly changing needs. Furthermore, workflow orchestration facilitates effective 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 boosting efficiency across the entire process from research to production.

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

[0119] 1. Multi-source data processing:

[0120] This project integrates wind resource knowledge and wind resource assessment reports to construct a domain knowledge base covering terminology, standards, and cases in the wind power field. Wind resource characteristics, including average wind speed, turbulence intensity, and wind shear index, are intelligently extracted using machine learning algorithms, and the knowledge base is dynamically updated. A fine-tuning dataset is built based on a triplet structure, and this dataset is then expanded and enhanced.

[0121] By constructing a domain knowledge base, this embodiment achieves intelligent extraction of wind resource characteristics and automatic updating of the dynamic knowledge base. This not only improves data processing efficiency but also enhances data quality and reliability, solving the problems of reliance on manual experience and low data utilization in traditional methods. Specifically, in this embodiment, the 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, to build a domain knowledge base covering wind power terminology, specifications and cases.

[0124] 112) Construct a network graph of technical terms: Based on the GraphSAGE algorithm, construct a network graph of wind power technical terms to realize the joint encoding of related parameters such as "wind speed-power generation".

[0125] 113) Dynamic knowledge update: Develop a dynamic knowledge update engine, which automatically identifies standard document version changes (such as NB / T 31146-2023) through the 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 shown below.

[0127] 121) Structured Dataset Definition: Using the Alpaca-52k enhanced format, construct a triplet structured dataset:

[0128] {

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

[0130] "question":"Suggested machine model and corresponding operating 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), and use scripts to convert raw data (such as CSV / text) into standard JSON format, aligning field names (such as instruction vs prompt).

[0134] 123) Text Processing and Enhancement:

[0135] Word segmentation and length control: Long texts are segmented using a sliding window, such as segmenting long texts into 512 tokens, preserving the context overlap area (overlap rate 15%), and avoiding information loss;

[0136] Data augmentation: Back-translation techniques (such as Chinese-English translation) are used to generate multilingual variants, and the SimBERT model is used for semantic substitution to generate semantically consistent synonyms, thereby improving the diversity of the dataset.

[0137] 2. Domain-Adaptive RAG Enhancement:

[0138] The method employs retrieval-enhanced generation technology, which dynamically retrieves professional literature, historical cases, and industry standards from a domain knowledge base based on a generative large model. It then generates wind resource assessment content that meets technical standards through multimodal fusion word vector technology. The multimodal fusion word vector technology includes a two-stream coding network and a cross-attention mechanism to achieve joint encoding and bidirectional interaction of text and image features.

[0139] This embodiment employs a generative large-scale model, LLAM, combined with retrieval enhancement technology (RAG), to dynamically retrieve professional literature, historical cases, and industry standards from the knowledge base, generating wind resource assessment content that conforms to technical standards. In particular, this embodiment innovatively proposes a multimodal fusion word vector technique, 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, improving its interpretability, but also to construct a bidirectional interaction channel through a cross-attention mechanism, further enhancing the model's fusion and interpretability.

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

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

[0142] This embodiment uses the WeKnow-RAG framework for hybrid retrieval, integrating:

[0143] Sparse Search: Keyword matching based on the BM25 algorithm is used to quickly filter relevant paragraphs in technical documents.

[0144] Dense retrieval: The Contriever model is used 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 on key indicators such as "wind farm spacing design".

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

[0147] 221) Design of a dual-stream encoding network: An improved attention_word2vec model is used to encode the text stream. This embodiment constructs a 768-dimensional unified semantic space to build novel word vectors. This embodiment uses convolutional network layers and linear regression layers to encode the image stream.

[0148] In this embodiment, the method by which the dual-stream coding network encodes the text stream is as follows:

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

[0150] The improved attention_word2vec model introduces a self-attention mechanism on top of the CBOW architecture for 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] Where: Q / K / V come from word vector projection residual connections from different subspaces; d k This represents the dimension of the word vector.

[0155] In this embodiment, the multi-head attention mechanism uses four attention heads. The outputs and inputs of each attention layer are weighted and fused.

[0156] In this embodiment, the cosine rotational position encoding (cosRoPE) combines rotational position encoding (RoPE) with cosine modulation. The principle is as follows:

[0157]

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

[0159] The results obtained from attention_word2vec and cosRoPE are added together to obtain the final text_feature.

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

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

[0162] Where: convolution is a convolutional network; linear is linear regression;

[0163] Assume the input image is I, with dimensions W×H×c (width×height×number of channels). The principle of a convolutional network layer is as follows:

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

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

[0166] The linear regression layer flattens the feature map output from the convolutional network layer into a one-dimensional vector, and then uses LayerNorm for normalization to align it with the text feature dimension. The principle of the linear regression layer is as follows:

[0167] y = LaterNorm(Wx + b′)

[0168] Where: W is the weight matrix. In this embodiment, the weight matrix W has a shape of 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) Explainable causal mechanism: Introduce a causal gating mechanism to construct an explainable weight allocation module, and calculate the interaction weights of text and image features through MLP.

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

[0171] causal_gate=MLP(Text_Feature⊙Image_Feature)

[0172] Where: causal_gate is the causal gating mechanism; MLP is the multilayer perceptron; Text_Feature is the text feature; Image_Feature is the image feature.

[0173] The interaction weights between 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 role of the "turbulence intensity" region in an image on the text description), thereby improving the interpretability of the model.

[0177] 223) Bidirectional feature enhancement mechanism: Use cross attention machine (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 connection for bidirectional output.

[0178] Image to Text: Using image features as the query and text as the key / value pair.

[0179] Text → Image: Using text features as the query and images as the key / value pair, the bidirectional output is fused through residual connections. Specifically, the bidirectional output obtained by fusing text enhancement features and image enhancement features through residual connections is as follows:

[0180] F final =MCA(V→T)+α·MCA(T→V)

[0181] Wherein: F final To fuse text enhancement features and image enhancement features through residual connections for bidirectional output; MCA(V→T) represents text enhancement features; MCA(T→V) represents image enhancement features; α represents weight coefficients; V represents visual image; T represents text.

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

[0183]

[0184] in:

[0185] Molecular part: This represents the indexed similarity score of positive sample pairs; sim(v,t) + () represents the visual feature v and the corresponding positive sample text feature t. + The cosine similarity is used to quantify the semantic alignment between the two samples; τ is a temperature coefficient, usually ranging from 0.05 to 0.2, used to control the steepness of the probability distribution. A smaller τ will amplify the discriminative power of difficult negative samples.

[0186] Denominator: The sum of similarities between positive samples and all negative samples; t― These are negative sample text features, i.e., text that does not match the current image, obtained through random sampling or hard negative mining.

[0187] The overall loss function works by maximizing the similarity ratio (numerator / denominator) of positive sample pairs, forcing the model to learn discriminative features for cross-modal alignment. A smaller loss value indicates that the similarity of positive sample pairs is significantly higher than that of negative sample pairs.

[0188] This constraint, through contrastive learning, maps semantically identical image and text features to neighboring regions within a unified semantic space, while pushing away irrelevant features. For example, the "turbulence intensity map" in a wind farm image should be highly similar to its corresponding text description in the vector space.

[0189] 3. Fine-tuning of the large model:

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

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

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

[0193] The QLoRA optimization strategy is adopted, and key parameter configurations are as follows:

[0194] lora_rank = 64;

[0195] lora_alpha = 32;

[0196] lora_dropout = 0.1.

[0197] Introduction to course learning mechanism:

[0198] Initial training: 20% basic Q&A + 80% standard clause analysis;

[0199] Mid-term addition: 50% of decisions will be made in complex scenarios (such as extreme climate adaptability design);

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

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

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

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

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

[0205]

[0206] Where: θ (i) It is the i-th candidate point; It is a uniform distribution, meaning the range of parameters 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 Proxy Model Construction:

[0209] Objective: To establish a mapping relationship between the parameter space and the loss function.

[0210] Operation: Establish a mapping relationship between the parameter space and the 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 points with different parameters.

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

[0214]

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

[0216] 33) Optimization of multi-target acquisition function:

[0217] Objective: To balance exploration (of unknown areas) with utilization (of the current optimal area);

[0218] Operation: Design an improved expected function:

[0219]

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

[0221] Candidate direction sets are generated using the CMA-ES algorithm.

[0222] 34) Gradient-independent direction search:

[0223] Objective: Determine the optimal descent direction;

[0224] operate:

[0225] (1) Perform parallel evaluation of the candidate direction set and calculate the actual loss.

[0226] (2) Select Pareto front solution: After comprehensively ranking based on the magnitude of loss reduction and directional stability, determine the optimal direction.

[0227] 35) Determine whether the rate of change of the actual loss over m consecutive iterations is lower than a set threshold: if yes, terminate the iteration, obtain the optimal parameters, and use the Bayesian surrogate model to fit the error for subsequent model diagnosis; if no, proceed to step 36). In this embodiment, m is 5.

[0228] 36) Dynamic learning rate update:

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

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

[0231]

[0232] Where: η t Let η be the learning rate at step t; base Based on the learning rate; g t For pseudo-gradient estimation, let g represent the gradient estimate at step t; ||g t ‖ represents 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 It is the current optimal parameter; η t g is the learning rate for the current step. t This is a pseudo-gradient estimate.

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

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

[0238] (2) Hardware resource-constrained scenarios (low video memory requirements);

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

[0240] This algorithm integrates multi-objective optimization theory, efficient sampling strategies, and adaptive learning rate techniques, providing a new optimization paradigm for fine-tuning in the LLM field.

[0241] 37) Determine if the current iteration step t is equal to the maximum allowed iteration step T: if yes, terminate the iteration, obtain the optimal parameters, and fit the error using a Bayesian surrogate model; if no, let t = t + 1, and repeat step 31).

[0242] 4. Automated report generation:

[0243] By using AI workflow orchestration to match multiple templates, embed data visualization charts, and perform syntax, logic, and compliance layer checks, a standardized wind resource assessment report is output.

[0244] The automated report generation module employs AI workflow orchestration, supporting multi-template matching, data visualization, and compliance verification, achieving a one-stop service from data analysis to report generation. This embodiment significantly simplifies the report generation process, reduces human intervention, ensures the professionalism and standardization of reports, and meets various needs in the wind power project development process.

[0245] In this embodiment, the method steps for generating automated reports are as follows.

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

[0247] Initial State → Data Validation → Project Overview → Wind Energy Resource Analysis → Power Generation Calculation → Conclusion

[0248] In this embodiment, the state transitions for each stage are defined using Python's transitions library, including the 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 visualization charts drawn by Matplotlib to generate reports.

[0250] The report generation process employs multi-level validation, including:

[0251] Syntax layer: LangChain is used for syntax layer validation to ensure that the text content conforms to the syntax rules;

[0252] Logic layer: Uses CLIPS rule engine to verify logical consistency and ensure that all calculation results are reasonable and error-free;

[0253] Compliance layer: The report automatically includes indexes of relevant standard clauses to ensure compliance.

[0254] In this embodiment, to achieve AI-powered intelligent report generation, a Finite State Machine (FSM) is first used to control the generation process. The state transitions for each stage are defined using Python's transitions library, including the initial state, data validation, project overview, wind energy resource analysis, power generation calculation, and conclusions. Then, the Jinja2 template engine is used to generate a Markdown-formatted technical report, and Matplotlib-generated visualizations are automatically inserted. During report generation, a multi-level validation mechanism is implemented: first, LangChain is used for syntax validation to ensure the text content conforms to grammatical rules; then, the CLIPS rule engine is used to verify logical consistency, ensuring that all calculation results are reasonable and error-free; finally, relevant standard clause indexes are automatically added to the report to ensure compliance. The specific operation steps are as follows:

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

[0256] Multi-dimensional output system:

[0257] Generate a Markdown-formatted report using the Jinja2 template engine and embed charts drawn by Matplotlib.

[0258] Syntax layer validation: The language model is called via LangChain 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: The report automatically marks the index of relevant standard clauses to ensure that the content complies with industry standards.

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

[0262] Table 1 Comparison of Technical Advantages

[0263] 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, resulting in an improved report pass rate and a reduced workload for expert review.

[0265] The above-described embodiments are merely preferred embodiments provided to fully illustrate 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 all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A method for generating a wind resource assessment report based on large model technology, characterized in that: Includes the following steps: Multi-source data processing: Integrate wind resource knowledge and wind resource assessment reports to build a domain knowledge base covering wind power terminology, standards 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; A fine-tuning dataset is constructed based on the triplet structure, and the fine-tuning dataset is expanded and enhanced. Domain-Adaptive RAG Enhancement: Employing retrieval enhancement generation technology, this method dynamically retrieves professional literature, historical cases, and industry standards from the domain knowledge base based on a generative large model. It then generates wind resource assessment content that conforms to technical standards through multimodal fusion word vector technology. This multimodal fusion word vector technology includes a dual-stream coding network and a cross-attention mechanism to achieve joint encoding and bidirectional interaction of text and image features. Large model fine-tuning: A domain-adaptive fine-tuning strategy is adopted. Based on the fine-tuning dataset, the generative large model is optimized using the gradient-free optimization algorithm Wind-DFO. The Wind-DFO algorithm includes dynamic Latin hypercube sampling, Bayesian surrogate model construction, and adaptive learning rate update. Automated report generation: Multiple templates are matched through AI workflow orchestration, data visualization charts are embedded, and syntax, logic and compliance checks are performed to output standardized wind resource assessment reports; The steps for optimizing large generative models using the gradient-free optimization algorithm Wind-DFO are as follows: 31) Dynamic parameter space partitioning and initialization: Candidate parameter points are generated through dynamic Latin hypercube sampling: in: It is the first One candidate point; It is a uniform distribution, meaning the range of parameters in each dimension is from arrive ; It is the number of candidate points generated; Calculate the initial loss and set the current optimal solution; 32) Bayesian surrogate model construction: Establishing a mapping relationship between the parameter space and the loss function based on the Gaussian process surrogate model: where: is the objective function value under the parameters ; is the mean function; is the kernel function, which defines the similarity between different parameter points; An improved kernel function is used to adapt to high-dimensional nonstationary optimization: in: The Euclidean distance between the two parameter points is... ; It is an L2 norm; 33) Multi-target acquisition function optimization: Design of the desired improved function: in: For parameters Expected improvement value at the location; The current known optimal loss value; The loss value predicted by the surrogate model; The variance predicted by the surrogate model; This is a numerical stability coefficient to prevent the denominator from being zero; Candidate direction sets are generated using the CMA-ES algorithm; 34) Gradient-independent direction search: Evaluate the actual loss of candidate directions and select the Pareto front solution to determine the optimal descent direction; 35) Determine whether the rate of change of the actual loss over m consecutive iterations is lower than a set threshold: if yes, terminate the iteration, obtain the optimal parameters, and fit the error using a Bayesian surrogate model; if no, proceed to step 36). 36) Dynamic learning rate update: The learning rate is calculated based on a momentum adaptive mechanism. in: In the first The learning rate of each step; Base learning rate; For pseudo-gradient estimation, it means that in the first... Gradient estimation of the step; The L2 norm for pseudo-gradient estimation; This is a numerical stability coefficient to prevent the denominator from being zero; Update parameters: where: is the current optimal parameter; is the learning rate for the current step; is the pseudo-gradient estimate; 37) Determine if the current iteration step t is equal to the maximum allowed iteration step T: if yes, terminate the iteration, obtain the optimal parameters, and fit the error using a Bayesian surrogate model; if no, let t = t + 1, and repeat step 31). 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, to build a domain knowledge base covering wind power terminology, specifications and cases; 112) Constructing a technical terminology network: Based on the GraphSAGE algorithm, construct a wind power technical terminology network to achieve joint encoding of associated parameters; 113) Dynamic knowledge update: Develop a dynamic knowledge update engine, which automatically identifies standard document version changes and triggers the knowledge base reconstruction process through a rule engine to ensure that the knowledge base content is synchronized with the latest industry standards; 12) Fine-tuning dataset construction 121) Definition of structured dataset: Construct a triplet structured dataset using the Alpaca-52k enhanced format; 122) Data format standardization: Use dataset_info.json to define dataset metadata, and use a script to convert the raw data into standard JSON format, aligning field names; 123) Text Processing and Enhancement: Word segmentation and length control: Long texts are segmented using a sliding window to avoid information loss; Data augmentation: Back-translation technology is used to generate multilingual variants, and the SimBERT model is used for semantic substitution to generate semantically consistent synonyms, thereby improving the diversity of the dataset.

2. The method of claim 1, wherein the method further comprises: The steps of the domain-adaptive RAG enhancement method are as follows: 21) Hybrid retrieval: This includes sparse retrieval, dense retrieval, and a compliance constraint channel, which efficiently retrieves multimodal information from the knowledge base to ensure the professionalism and compliance of the generated content. The sparse retrieval is based on keyword matching to quickly filter relevant paragraphs in technical documents. The dense retrieval 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 on key indicators. 22) Modal vector fusion: Encodes text and image features into a unified semantic space to improve cross-modal information fusion capabilities; 221) Design a dual-stream encoding network: use an improved attention_word2vec model to encode the text stream, and use convolutional network layers and linear regression layers to encode the image stream; 222) Explainable causal mechanism: Introduce a causal gating mechanism to construct an explainable weight allocation module, and calculate the interaction weights of text and image features through MLP; 223) Bidirectional feature enhancement mechanism: Use 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 connection for bidirectional output; 23) Semantic contrast alignment: Through contrastive learning mechanism, the consistency between image features and text features in semantic space is forced to be unified.

3. The method of claim 2, wherein: In step 221), the method for encoding the text stream is as follows: attention_word2vec[text] + cosine rotation position encoding cosRoPE[pos] → text_feature The improved attention_word2vec model introduces a self-attention mechanism on top of the CBOW architecture for hierarchical feature extraction, including: Word embedding layer: Employs dynamic word2vec to capture local semantics through a sliding window; Attention enhancement: Multi-head attention is introduced to calculate global word relationship weights, and the outputs and inputs 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: wherein: is a learnable rotation matrix; is a position index; is a dimension index; is a learnable rotation angle parameter; is a matrix multiplication operator.

4. The method of claim 3, wherein the method further comprises: In step 221), the method for encoding the image stream is as follows: convolution[image] → linear feature projection → image_feature Where: convolution refers to a convolutional network; linear refers to linear regression; The principle of convolutional network layers is: in: It is a convolution kernel; Indicates the input image Apply convolution operations; It is a bias term; It is an activation function; It is the output feature map obtained after convolution operation; The linear regression layer flattens the feature map output by the convolutional network layer into a one-dimensional vector, and then... Standardization, aligned with text feature dimensions, the principle of the linear regression layer is as follows: where: is a weight matrix; is a bias term; denotes a layer normalization operation that normalizes the output so that the final output is aligned with the text feature dimension.

5. The method of claim 3, wherein the method further comprises: In step 222), the principle of the causal gating mechanism is as follows: causal_gate = MLP(Text_Feature ⊙ Image_Feature) Where: causal_gate is the causal gating mechanism; MLP is the multilayer perceptron; Text_Feature is the text feature; Image_Feature is the image feature; The interaction weights between 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.

6. The method of claim 3, wherein the method further comprises: In step 223), the bidirectional output obtained by fusing text enhancement features and image enhancement features through residual connection is: in: To fuse text enhancement features and image enhancement features through residual connections for bidirectional output; Enhance text features; Enhance image features; These are the weighting coefficients; For visual images; text.

7. The method of claim 3, wherein the method further comprises: In step 23), the loss function for semantic contrast alignment is: in: This represents the indexed similarity score of positive sample pairs. It is a visual feature Corresponding positive sample text features The cosine similarity is used to quantify the degree of semantic alignment between the two; τ is a temperature coefficient that controls the steepness of the probability distribution. This includes the sum of similarities between positive samples and all negative samples; These are negative sample text features.

8. The method for generating wind resource assessment reports based on large-scale model technology according to claim 1, characterized in that: The steps for generating automated reports are as follows: 41) Dynamic orchestration engine: The generation process is controlled by a finite state machine. The state transitions of each stage are defined by Python's transitions library, including the initial state, data verification, project overview, wind energy resource analysis, power generation calculation and conclusion. 42) Multi-dimensional output: Generate technical reports in Markdown format using the Jinja2 template engine, and automatically insert visualization charts drawn with Matplotlib to generate reports; The report generation process employs multi-level validation, including: Syntax layer: LangChain is used for syntax layer validation to ensure that the text content conforms to the syntax rules; Logic layer: Uses CLIPS rule engine to verify logical consistency and ensure that all calculation results are reasonable and error-free; Compliance layer: The report automatically includes indexes of relevant standard clauses to ensure compliance.

Images