Wind power condition-related descriptive character accurate fast-growing technology based on three-dimensional spatio-temporal data

Abstract

The invention relates to the technical field of meteorology, and discloses a wind power condition-related descriptive text accurate fast-growing technology based on three-dimensional spatiotemporal data, which comprises the following steps of: acquiring a three-dimensional spatiotemporal data set comprising time, geographic space and wind power parameters, extracting wind power characteristics by using a three-dimensional convolutional neural network (3D-CNN), and obtaining a three-dimensional data set comprising time, geographic space and wind power parameters; and the features are converted into descriptive characters by adopting a sequence-to-sequence (Seq2Seq) model based on an attention mechanism. The 3D-CNN effectively captures spatial-temporal characteristics, and the Seq2Seq model ensures the accuracy and vividness of character generation. According to the technology, accurate and rapid generation of wind power condition description is realized, and powerful support is provided for the fields of weather forecast, climate analysis, wind energy evaluation and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorology, and in particular to a technology for quickly and accurately generating descriptive text about wind conditions based on three-dimensional spatiotemporal data. Background Art

[0002] In the field of meteorology, accurate description of wind conditions is crucial for weather forecasting, climate analysis, wind energy assessment, and many wind-related industries and applications. Traditional descriptions of wind conditions mainly rely on data records from meteorological observation sites and the empirical judgment of professionals. Although this method can provide a certain degree of accuracy, it has many limitations. First, the distribution of observation sites is often uneven, resulting in missing wind data in some areas, affecting the comprehensiveness of the description. Second, the empirical judgment of professionals is affected by their personal knowledge and experience level, and is somewhat subjective, making it difficult to achieve standardization and automation.

[0003] With the development of information technology and big data, the application of three-dimensional spatiotemporal data in the field of meteorology is becoming more and more extensive. Three-dimensional spatiotemporal data not only includes the traditional time dimension and geographic space dimension, but also introduces the wind parameter dimension, such as wind speed, wind direction, gust, etc., which provides a rich data basis for the detailed description of wind conditions. However, how to extract useful wind characteristics from massive three-dimensional spatiotemporal data and convert them into accurate and easy-to-understand descriptive text is the current technical challenge.

[0004] Most existing technical methods rely on statistical models or simple machine learning algorithms, which have many shortcomings when dealing with complex three-dimensional spatiotemporal data. For example, statistical models often make predictions based on the average value or trend of historical data, and it is difficult to capture the instantaneous changes and local characteristics of wind conditions. Simple machine learning algorithms may make model training difficult due to high data dimensions and complex features, and their generalization capabilities are limited. The application of natural language generation technology in the description of wind conditions is also still in its infancy. Existing natural language generation models are often optimized for specific text types or fields, and the generation effect is often unsatisfactory for data such as wind conditions that have professional and spatiotemporal characteristics. In particular, when describing the fine features and spatiotemporal changes of wind conditions, existing models are often difficult to be accurate, vivid and fast. Summary of the invention

[0005] The purpose of the present invention is to provide a technology for quickly and accurately generating descriptive text about wind conditions based on three-dimensional spatiotemporal data to solve the problems raised in the above-mentioned background technology.

[0006] To achieve the above-mentioned purpose, the present invention provides the following technical solution: a technology for accurately and quickly generating descriptive text about wind conditions based on three-dimensional spatiotemporal data, the technology comprising: Acquire a three-dimensional spatiotemporal dataset including a time dimension, a geographic space dimension, and a wind parameter dimension, wherein the wind parameter includes wind speed, wind direction, and gust; Preprocess three-dimensional spatiotemporal data, including data cleaning, missing value processing, outlier detection and processing; Construct a wind condition feature extraction model to extract wind features from preprocessed three-dimensional spatiotemporal data; Using natural language generation models, the extracted wind features are converted into descriptive text; Output accurate and quickly generated descriptive text of wind conditions.

[0007] Preferably, the wind condition feature extraction model adopts a three-dimensional convolutional neural network 3D-CNN, whose structure includes an input layer, multiple 3D convolutional layers, a pooling layer, and a fully connected layer, which is used to capture the spatiotemporal features in three-dimensional spatiotemporal data, wherein the calculation formula of the 3D convolutional layer is: in, For input data, is the 3D convolution kernel, is the bias term, is the output feature.

[0008] Preferably, during the training of the 3D convolutional neural network, a cross entropy loss function is used to measure the difference between the model prediction and the true label.

[0009] Preferably, the natural language generation model adopts a sequence-to-sequence model Seq2Seq based on an attention mechanism, wherein the encoder part adopts an LSTM network to encode the wind force characteristics, and the decoder part adopts an LSTM network with an attention mechanism to generate descriptive text, and the calculation formula of the attention mechanism is: in, is the attention weight, The score calculated for the alignment model, is the hidden state of the decoder at the previous time step, is the hidden state of the encoder, a is the alignment function, and it is implemented using a fully connected layer plus an activation function.

[0010] Preferably, during the training process of the Seq2Seq model, a teacher forcing strategy is adopted for training, and the input of each time step is the true label of the previous time step rather than the output predicted by the model.

[0011] Preferably, the method further comprises post-processing the generated descriptive text, including grammar checking, spelling checking and deletion of redundant words.

[0012] Preferably, the three-dimensional spatiotemporal data set also includes other meteorological parameters related to wind conditions, including at least temperature, humidity and air pressure, so as to enhance the comprehensiveness and accuracy of the description of wind conditions.

[0013] Preferably, when constructing the wind condition feature extraction model, other meteorological parameters are fused with the wind parameters by feature splicing and feature crossing.

[0014] Preferably, the method further comprises personalizing the output descriptive text, adjusting the description style, level of detail and use of professional terminology according to user needs.

[0015] Preferably, the implementation methods of personalized customization include: Design a variety of description templates and select or combine templates according to user needs; Adopt conditional generation technology to generate descriptive text that meets the requirements according to the description style and professional level set by the user; Provide a user feedback mechanism to iteratively optimize the generated descriptive text based on user feedback, gradually approaching the description style and content expected by users.

[0016] Compared with the prior art, the present invention has the following beneficial effects: By acquiring a three-dimensional spatiotemporal data set containing time dimensions, geographic space dimensions, and wind parameter dimensions, the present invention can make full use of the rich wind information and effectively make up for the data missing problem caused by the uneven distribution of observation sites. At the same time, the three-dimensional convolutional neural network (3D-CNN) is used to extract wind condition features, which can accurately capture the instantaneous changes and local characteristics of wind conditions, significantly improving the comprehensiveness and accuracy of the description.

[0017] Traditional methods rely on the experience and judgment of professionals and are subjective. The present invention uses automated feature extraction and natural language generation technology to reduce human intervention, achieve objectivity and standardization of wind condition descriptions, and improve the consistency and credibility of descriptions. The 3D-CNN model used in the present invention can efficiently process high-dimensional and complex three-dimensional spatiotemporal data, while the sequence-to-sequence (Seq2Seq) model based on the attention mechanism can quickly convert the extracted wind features into descriptive text. This combination enables the present invention to greatly improve the efficiency and speed of description generation while maintaining the accuracy of the description, meeting the needs of real-time or quasi-real-time applications.

[0018] By optimizing the natural language generation model, the present invention can generate professional and vivid descriptive text about wind conditions. This not only helps professionals to conduct accurate meteorological analysis and decision-making, but also enables non-professionals to more easily understand and perceive changes in wind conditions, thereby improving the readability and dissemination of information. The 3D-CNN and Seq2Seq models used in the present invention both have strong generalization capabilities and can adapt to the needs of describing wind conditions in different regions and time periods. At the same time, model training through the cross-entropy loss function can further optimize model performance and improve the model's ability to recognize and describe complex wind conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a working principle diagram of a technology for accurately and quickly generating descriptive text about wind conditions based on three-dimensional spatiotemporal data according to the present invention; Figure 2 Flowchart for training the model for wind condition feature extraction; Figure 3 Generate a flow chart for descriptive text. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] See also Figure 1-3 The present invention provides a technical solution: a technology for accurately and quickly generating descriptive text about wind conditions based on three-dimensional spatiotemporal data, the technology comprising: Step 1: Obtain a three-dimensional spatiotemporal dataset: Data source: Three-dimensional spatiotemporal data containing timestamps, geographic locations (longitude, latitude, altitude) and wind parameters (wind speed, wind direction, gusts) are collected from meteorological observation stations, remote sensing satellites, drones and other channels.

[0022] Data format: Make sure the data is stored in a structured format such as nc, CSV, JSON or a database table to facilitate subsequent processing.

[0023] Step 2: Data preprocessing: Data cleaning: remove duplicate records and invalid values ​​(such as negative wind speed, unreasonable wind direction, etc.).

[0024] Missing value processing: Missing values ​​can be processed by using interpolation, mean filling, or inference based on adjacent data points.

[0025] Outlier detection and processing: Use statistical methods (such as the 3σ principle) or machine learning algorithms (such as isolation forests) to detect outliers and correct or eliminate them.

[0026] Step 3: Construct a wind condition feature extraction model: A three-dimensional convolutional neural network (3D-CNN) is used to capture the spatiotemporal features in three-dimensional spatiotemporal data.

[0027] Step 4: Generate a model using natural language: Model selection: A sequence-to-sequence model (Seq2Seq) based on the attention mechanism is adopted, in which the encoder part uses the long short-term memory network (LSTM) to encode the wind characteristics, and the decoder part uses the LSTM with attention mechanism to generate descriptive text.

[0028] Model training: Encoder: Input the wind feature sequence into LSTM to learn its temporal dependencies.

[0029] Decoder: Based on the encoder output and the attention mechanism, it gradually generates descriptive text.

[0030] Loss function: Use cross entropy loss or BLEU score to evaluate the difference between the generated text and the reference text and optimize the model parameters.

[0031] Step 5: Output descriptive text: Text generation: The preprocessed 3D spatiotemporal data is input into the feature extraction model to extract the wind characteristics, and then input into the natural language generation model to generate accurate and fast descriptive text of the wind conditions.

[0032] Post-processing: Perform grammar checking and rhetoric optimization on the generated text to ensure the accuracy and readability of the text.

[0033] Output method: Descriptive text can be output through various channels such as web pages, mobile applications, weather reports, etc., providing convenient wind condition information for different users.

[0034] The present invention will be further described below in conjunction with Examples 1 to 3: Embodiment 1: In the present invention, the steps of constructing a wind condition feature extraction model include: Data preprocessing: First, the original three-dimensional spatiotemporal data is preprocessed. This includes checking and processing missing values, for example, filling missing data through interpolation; detecting and processing outliers, such as removing data points that exceed the range by setting a reasonable range. Then, the data is standardized to ensure that the data in each dimension has similar dimensions and distribution, which is conducive to model training.

[0035] Build 3D-CNN model: Design and build a three-dimensional convolutional neural network (3D-CNN) model to extract wind characteristics. The model structure is as follows: Input layer: Receives preprocessed three-dimensional spatiotemporal data with a shape of (T,H,W,C), where T is the time step, H and W are the height and width of the geographic space, respectively, and C is the number of wind parameter channels (3 in this case, namely wind speed, wind direction, and gust).

[0036] 3D convolution layer: multiple 3D convolution layers are set, each of which uses convolution kernels of different sizes and numbers to capture spatiotemporal features of different scales. The convolution operation is as described in claim 2, and the feature map is calculated on the three-dimensional data by sliding the convolution kernel. The calculation formula of the 3D convolution layer is: in, For input data, is the 3D convolution kernel, is the bias term, is the output feature.

[0037] Pooling layer: Add a pooling layer after each 3D convolution layer to reduce the feature dimension and reduce the amount of computation while retaining important information. In this example, the Max Pooling operation is used.

[0038] Fully connected layer: After the last pooling layer, a fully connected layer is added to map high-dimensional features to a low-dimensional space to facilitate subsequent natural language generation.

[0039] Output layer: Depending on the task requirements, the output layer can be a classification label (such as wind force level) or a feature vector (for subsequent natural language generation).

[0040] Model training: The 3D-CNN model is trained using the preprocessed 3D spatiotemporal dataset. During the training process, the cross entropy loss function (as described in claim 3) is used to measure the difference between the model prediction and the true label. The gradient is calculated by the back-propagation algorithm, and the model parameters are updated using an optimizer (such as Adam or SGD), and the training is iterated until the loss function converges or the preset training rounds are reached.

[0041] Suppose there is a dataset containing one year of wind data, with a time step of 24 hours (i.e. one data point per day), a geospatial dimension of 100x100 grid, and each grid point contains three parameters: wind speed, wind direction, and gust. First, preprocess the data, and then build and train the 3D-CNN model. During the training process, the cross entropy loss function is used to calculate the difference between the model prediction and the true wind level label, and the model parameters are updated through back propagation and optimizer. After training, the model is able to extract wind features, providing a basis for subsequent natural language generation.

[0042] During the training process of the 3D-CNN model, the cross entropy loss function is used to measure the difference between the model prediction and the true label. The specific calculation steps are as follows: Calculate the predicted probability: At the output layer of the model, a softmax function is used to convert the raw output into a probability distribution.

[0043] Calculate cross entropy loss: Calculate cross entropy loss based on true label and predicted probability. For the entire dataset, the loss function is the average of all sample losses.

[0044] Optimize model parameters: Calculate the gradient of the loss function to the model parameters through the back-propagation algorithm, and use an optimizer (such as Adam) to update the parameters to minimize the loss function. This process is repeated until the loss function converges or the preset training rounds are reached.

[0045] When training the 3D-CNN model, assume that the dataset contains 1,000 samples, each of which corresponds to a wind condition within 24 hours. The true label is the wind force level (such as 0-12). During the training process, the original value output by the model is converted into a predicted probability distribution through the softmax function. Then, the cross entropy loss function is used to calculate the difference between the model prediction and the true wind force level label. Through backpropagation and optimizers (such as Adam), the model parameters are iteratively updated until the loss function converges. In this way, the trained 3D-CNN model can accurately extract wind features and provide a reliable basis for subsequent natural language generation.

[0046] Embodiment 2: When building a natural language generation model, a sequence-to-sequence (Seq2Seq) model based on an attention mechanism is specifically adopted, and its core lies in the design of the encoder and decoder. In the encoder part, the long short-term memory network (LSTM) is used to encode wind features. The LSTM network can capture long-term dependencies in sequence data and is suitable for processing changes in wind features in the time dimension. The LSTM network of the encoder receives the wind feature sequence extracted by 3D-CNN, processes and updates its hidden state time by time, and finally obtains a context vector containing all time step information.

[0047] The decoder also uses the LSTM network, but introduces the attention mechanism on this basis. The attention mechanism allows the decoder to dynamically focus on different parts of the encoder output when generating each word, that is, to assign different weights to the hidden state of each time step. In specific implementation, for each time step of the decoder, the correlation score between the current hidden state and all the hidden states of the encoder is calculated (i.e. ), the score is converted into attention weight through the softmax function (i.e. ). Then, the attention weights are used to weight the encoder’s hidden states to get the contextual representation of the current time step. This contextual representation is used together with the decoder’s current hidden state to generate the next word.

[0048] The calculation formula of the attention mechanism is as follows: in, is the attention weight, The score calculated for the alignment model, is the hidden state of the decoder at the previous time step, is the hidden state of the encoder, a is the alignment function, and it is implemented using a fully connected layer plus an activation function.

[0049] In the training process of the Seq2Seq model, in order to improve training efficiency and model performance, the Teacher Forcing strategy is adopted. Specifically, during the training phase, the input of the decoder is not the output generated by the model itself (i.e., the predicted word at the previous time step), but the real label sequence (i.e., the real word at the previous time step). The advantage of this is that the model can receive more accurate and stable input during the training process, so as to learn the correct output distribution faster.

[0050] When implementing the teacher forcing strategy, the first time step of the decoder is usually fed with a special start symbol (such as " <start>”). For subsequent time steps, the input is the word of the corresponding time step in the true label sequence. During the forward propagation of the model, the loss is calculated for each time step, that is, the cross entropy loss between the probability distribution predicted by the model and the true label. These losses are then accumulated and averaged to obtain the loss value of the entire sequence for backpropagation and parameter update.

[0051] By adopting the teacher forcing strategy, the model can more directly learn the mapping relationship between the input sequence and the output sequence during the training process, avoiding the error accumulation caused by the model's own prediction errors. This helps to accelerate the convergence of the model and improve the stability and efficiency of training. At the same time, after the model training is completed, when reasoning (i.e. generating descriptive text), the input of the decoder is switched to the output generated by the model itself to realize the autoregressive text generation process.

[0052] Example 3 After generating the descriptive text, in order to further improve its accuracy and readability, the present invention implements a series of post-processing steps, including grammar checking, spelling checking and redundant word removal. These steps are implemented by natural language processing (NLP) technology.

[0053] Grammar checking: Use open source grammar checking tools (such as Grammarly API or LanguageTool) to perform grammatical analysis on the generated text, identify and correct grammatical errors, and ensure that the sentence structure is correct and the grammar conforms to the standards.

[0054] Spell Checking: Use a spell checking library (such as Hunspell or Aspell) to verify the spelling of each word in the text, correct spelling errors, and ensure that all words are correct.

[0055] Redundant word removal: A set of redundant word recognition rules are designed to analyze sentence structure through NLP technology, identify and delete redundant words or phrases, such as repeated adjectives, unnecessary adverbs, etc., to make the description more concise and clear.

[0056] For example, the generated original description "Today's wind speed is very fast, reaching 30 kilometers per hour, which is a very high speed" may be modified to "Today's wind speed reached 30 kilometers per hour, extremely high" after post-processing.

[0057] The three-dimensional spatiotemporal data set in the present invention includes not only wind parameters such as wind speed, wind direction, and gusts, but also other meteorological parameters such as temperature, humidity, and air pressure to provide a more comprehensive description of wind conditions.

[0058] Dataset construction: Data are collected through various means such as meteorological stations, satellite remote sensing, and drone observations to ensure the temporal and spatial coverage and accuracy of the data. The dataset is organized by time dimension (such as every hour), geographic space dimension (such as longitude and latitude grid), and wind and meteorological parameter dimension.

[0059] Feature fusion: When constructing the wind condition feature extraction model, other meteorological parameters are fused with wind parameters by feature splicing and feature crossover. Feature splicing is to directly connect the feature vectors of different parameters into a longer feature vector; feature crossover is to capture the correlation between parameters by calculating the interactive features between parameters (such as the product of wind speed and temperature, the combination of wind direction and humidity, etc.).

[0060] For example, for a certain time point t and geographic location (x, y), the wind feature vector may be [wind speed, wind direction, gust], and the meteorological feature vector may be [temperature, humidity, air pressure]. Through feature concatenation, the fused feature vector is [wind speed, wind direction, gust, temperature, humidity, air pressure]; through feature crossover, new features such as wind speed × temperature or wind direction × humidity may be generated.

[0061] In order to meet the needs of different users, the present invention provides a personalized descriptive text customization service, which is specifically implemented as follows: Design description templates: Design a variety of description templates based on common wind conditions, such as concise version, detailed version, professional version, etc. Users can select or combine templates as needed to generate descriptive text that suits their own style.

[0062] Conditional generation technology: Using conditional natural language generation technology, such as conditional random fields (CRF) or conditional generative adversarial networks (CGAN), to generate descriptive text that meets the requirements according to the description style and professional level set by the user. For example, the user can choose the "easy to understand" style, and the system will generate an easy-to-understand description; if the "professional and detailed" style is selected, a description containing more professional terms and detailed data will be generated.

[0063] User feedback mechanism: Provides a user feedback interface to allow users to evaluate and modify the generated descriptive text. The system iteratively optimizes based on user feedback and gradually adjusts the generation strategy to make the generated descriptive text closer to the user's expected description style and content. For example, if the user feedback hopes to add a description of the wind force change trend, the system will add the corresponding content in the subsequent generation.

[0064] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0065] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.< / start>

Claims

1. A technology for accurately and quickly generating descriptive text about wind conditions based on three-dimensional spatiotemporal data, characterized in that: The techniques include: Acquire a three-dimensional spatiotemporal dataset including a time dimension, a geographic space dimension, and a wind parameter dimension, wherein the wind parameter includes wind speed, wind direction, and gust; Preprocess three-dimensional spatiotemporal data, including data cleaning, missing value processing, outlier detection and processing; Construct a wind condition feature extraction model to extract wind features from preprocessed three-dimensional spatiotemporal data; Using natural language generation models, the extracted wind features are converted into descriptive text; Output accurate and quickly generated descriptive text of wind conditions.

2. According to claim 1, a technology for accurately and quickly generating descriptive text about wind conditions based on three-dimensional spatiotemporal data, characterized in that: The wind condition feature extraction model adopts a three-dimensional convolutional neural network 3D-CNN, whose structure includes an input layer, multiple 3D convolutional layers, a pooling layer, and a fully connected layer, which is used to capture the spatiotemporal features in three-dimensional spatiotemporal data. The calculation formula of the 3D convolutional layer is: in, For input data, is the 3D convolution kernel, is the bias term, is the output feature.

3. The accurate and rapid generation technology of descriptive text about wind conditions based on three-dimensional spatiotemporal data according to claim 2 is characterized by: During the training process of 3D convolutional neural network, the cross entropy loss function is used to measure the difference between the model prediction and the true label.

4. The accurate and rapid generation technology of descriptive text about wind conditions based on three-dimensional spatiotemporal data according to claim 1 is characterized in that: The natural language generation model adopts a sequence-to-sequence model Seq2Seq based on an attention mechanism, in which the encoder part adopts an LSTM network to encode wind characteristics, and the decoder part adopts an LSTM network with an attention mechanism to generate descriptive text. The calculation formula of the attention mechanism is: in, is the attention weight, The score calculated for the alignment model, is the hidden state of the decoder at the previous time step, is the hidden state of the encoder, a is the alignment function, and it is implemented using a fully connected layer plus an activation function.

5. The technology for accurately and quickly generating descriptive text about wind conditions based on three-dimensional spatiotemporal data according to claim 4 is characterized by: During the training process of the Seq2Seq model, a teacher forcing strategy is adopted for training, and the input of each time step is the true label of the previous time step rather than the output predicted by the model.

6. The technology for accurately and quickly generating descriptive text about wind conditions based on three-dimensional spatiotemporal data according to claim 1 is characterized by: The method also includes post-processing the generated descriptive text, including grammar checking, spelling checking and redundant word deletion.

7. The technology for accurately and quickly generating descriptive text about wind conditions based on three-dimensional spatiotemporal data according to claim 1 is characterized by: The three-dimensional spatiotemporal data set also includes other meteorological parameters related to wind conditions, including at least temperature, humidity and air pressure, so as to enhance the comprehensiveness and accuracy of the description of wind conditions.

8. The accurate and rapid generation technology of descriptive text about wind conditions based on three-dimensional spatiotemporal data according to claim 7 is characterized in that: When constructing the wind condition feature extraction model, other meteorological parameters are integrated with wind parameters through feature splicing and feature crossing.

9. The technology for accurately and quickly generating descriptive text about wind conditions based on three-dimensional spatiotemporal data according to claim 1 is characterized in that: The method also includes personalizing the output descriptive text and adjusting the description style, detail level and use of professional terms according to user needs.

10. The accurate and rapid generation technology of descriptive text about wind conditions based on three-dimensional spatiotemporal data according to claim 9 is characterized in that: Ways to achieve personalized customization include: Design a variety of description templates and select or combine templates according to user needs; Adopt conditional generation technology to generate descriptive text that meets the requirements according to the description style and professional level set by the user; Provide a user feedback mechanism to iteratively optimize the generated descriptive text based on user feedback, gradually approaching the description style and content expected by users.

Images