Prediction method of bridge wind field change pattern based on LSTM deep embedded clustering model

Abstract

This invention discloses a method for predicting wind field variation patterns in bridges based on an LSTM deep embedded clustering model, belonging to the technical field of bridge wind field prediction. The method obtains wind speed and direction information at a mountain canyon bridge site, performs preprocessing, and generates a wind speed sequence dataset. A fused LSTM deep embedded clustering model is constructed and the wind speed sequence dataset is input into the LSTM deep embedded clustering model for training until the loss function stabilizes. The trained LSTM deep embedded clustering model is then used to reconstruct cluster centers through a decoder, forming a library of diurnal wind speed variation patterns and partitioning the wind speed sequences through soft allocation. Finally, the wind speed data to be predicted is matched with the library of diurnal wind speed variation patterns, with the most similar diurnal wind speed variation pattern being used as the prediction result. This solution effectively extracts time series features and innovatively proposes using a softmax function to calculate soft allocation within the deep embedded clustering layer, achieving efficient end-to-end training of neural networks.

Description

Technical Field

[0001] The present invention belongs to the technical field of bridge wind field prediction, and specifically relates to a bridge wind field change pattern prediction method based on an LSTM deep embedded clustering model. Background Art

[0002] Because long-span bridges have relatively low stiffness and damping, they are susceptible to wind-induced vibration under wind loads. Therefore, the wind-resistant design of long-span bridge structures often plays a critical role in controlling wind loads. In mountainous canyon environments, wind fields are complex and variable, influenced by the topography. Long-term monitoring of wind fields at bridge sites and predicting wind field variations based on this data can aid in decision-making regarding the proper construction, operation, and maintenance of bridges.

[0003] Existing wind field prediction methods include numerical weather forecasting: by collecting meteorological data for numerical calculations, solving a set of fluid mechanics and thermodynamic differential equations, and predicting future wind speeds; statistical forecasting methods: establishing a predictive regression model, using historical data as model input, and training to obtain specific model parameters for wind speed prediction, specifically including time series method and machine learning method.

[0004] However, there are many problems with existing wind field prediction methods. Numerical weather forecasting methods have significant computational complexity and are highly sensitive to initial conditions. Numerical models cannot adapt well to the climatic characteristics of complex terrain areas such as mountainous canyons. A single statistical forecasting method is difficult to adapt to the prediction of complex and variable wind fields: the time series method cannot perfectly fit the nonlinear relationship between variables, and the performance of time series methods for medium-term or long-term predictions is poor; supervised machine learning algorithms require a large amount of data for training and prediction, which is difficult and time-consuming to train. Bridge wind field data is highly discrete, and existing methods have poor prediction effects on the high-frequency part of wind field data. There is a lack of a method for predicting the daily variation pattern of wind speed. Summary of the Invention

[0005] To predict diurnal wind speed patterns, clustering methods can be used to build a database of diurnal wind speed patterns, and then predict these patterns through feature matching. Clustering can automatically group samples in a dataset based on a certain similarity measure. It is widely used in fields such as data mining, pattern recognition, image processing, and market analysis. In particular, clustering can help discover underlying structures or patterns in data without pre-labeled data. However, traditional clustering algorithms have significant limitations, including sensitivity to initial values and noise, assumptions about simple cluster shapes, and poor ability to handle high-dimensional data. To overcome these shortcomings, many new clustering algorithms have emerged in recent years, such as deep learning-based clustering methods, which are better able to handle complex and high-dimensional data and offer greater flexibility and robustness. However, existing technologies do not fully extract data features and ignore the relationship between feature extraction and clustering, resulting in limited improvement in clustering performance using deep learning. To address these issues, this paper proposes a method for predicting diurnal wind speed patterns on bridges based on a LSTM deep embedded clustering model. This improved deep embedded clustering algorithm is then used to implement diurnal wind speed pattern prediction.

[0006] The present invention provides a bridge wind field change pattern prediction method based on an LSTM deep embedded clustering model, comprising the following steps:

[0007] S1. Data preparation: Obtain wind speed and direction information at the mountain canyon bridge site, perform preprocessing, and generate a wind speed sequence dataset;

[0008] S2. Model construction: Build a fused LSTM deep embedded clustering model, which includes an LSTM autoencoder layer and a deep embedded clustering layer;

[0009] S3, model training: inputting the wind speed sequence dataset into the LSTM deep embedded clustering model for training until the loss function is stable;

[0010] S4. Construction of a diurnal variation pattern library: Using the trained LSTM deep embedded clustering model, the decoder is used to reconstruct the cluster centers, forming a diurnal wind speed variation pattern library and dividing the wind speed sequence through soft allocation;

[0011] S5. Daily variation pattern prediction: Match the wind speed data to be predicted with the wind speed daily variation pattern library, and take the most similar wind speed daily variation pattern as the prediction result.

[0012] Preferably, the S1 specifically includes:

[0013] S11, dividing the wind speed and direction data at predetermined time intervals, and eliminating the measured data that does not meet the preset length and standard deviation;

[0014] S12. Decompose the wind speed and wind direction into wind speed components in specific directions, and calculate the average wind speed and wind direction angle;

[0015] S13. Divide the average wind speed into equal time intervals, use a filter to obtain the change trend, and perform standardization to form a wind speed sequence data set.

[0016] Preferably, the S2 specifically includes:

[0017] S21. Construct an LSTM autoencoder layer, including an encoder composed of an LSTM layer and a decoder composed of an LSTM layer and a fully connected layer;

[0018] S22. Build a deep embedded clustering layer to classify the latent variables, use k-means to initialize the cluster centers, calculate the soft allocation through the softmax function and construct the auxiliary allocation, and use the KL divergence of the two as the clustering loss;

[0019] S23, the reconstruction loss and clustering loss are integrated into the LSTM deep embedded clustering model loss function;

[0020] S24. Set the learning rate, initialize the clustering loss weight and the clustering loss weight growth rate.

[0021] Preferably, the soft allocation is specifically expressed as:

[0022] ;

[0023] Where, express Belong to j The probability of the class, express Features , is the temperature parameter, k is the number of clusters; m is the sequence number of the cluster center, from 1 to k change; c m It is m class probability; c j It is j Cluster centers of the classes;

[0024] The auxiliary allocation is specifically expressed as:

[0025] ;

[0026] Where, for Belong to j class probability; for Belong to m class probability; for Belong to m class probability; x is the sequence number of the wind speed sequence; n is the number of wind speed sequences.

[0027] This formula is given calculate , that is, Features , belongs to j The probability of a class, once given, cannot be changed. In the summation formula, we need to traverse All elements belong to j The probability of the class, where x is a variable. x = i situation, that is, z i It is certain. z x is changing, the two sets are the same, but the ranges they represent are different. z x Indicates the x The characteristics of a variable.

[0028] Preferably, the LSTM deep embedded clustering model loss function is specifically expressed as:

[0029] ;

[0030] Where, represents the weight of clustering loss; ; To initialize the weight of clustering loss, t is the iteration round, is the weight growth rate of clustering loss; is the reconstruction loss function; is the clustering loss.

[0031] Preferably, the S3 specifically includes:

[0032] The wind speed sequence dataset is input into the LSTM deep embedded clustering model, the number of clusters and the total number of training rounds are set, and training is performed; wherein, in each training session, the wind speed sequence dataset is passed through the LSTM deep embedded clustering model to obtain the LSTM deep embedded clustering model loss function, and the LSTM network parameters and the cluster centers of the deep embedded clustering layer are optimized by stochastic gradient descent with momentum until the set total number of rounds are completed, and the training is stopped after the total number of rounds is completed. The number of rounds when the total loss is reduced to a stable state is less than the specified total number of training rounds, so that the total loss is stable after all rounds of training are completed. The total loss here represents the LSTM deep embedded clustering model loss function.

[0033] The judgment basis for the stability of total loss is as follows:

[0034] ;

[0035] Where, is the total loss in round t, is the total loss in the t+1th round. When the change in loss is less than 0.0002, the loss can be considered stable.

[0036] Preferably, the S4 specifically includes:

[0037] S41, using the optimized LSTM decoder to reconstruct the optimized cluster centers to form a wind speed diurnal variation pattern library;

[0038] S42. Calculate the probability of the wind speed sequence in the latent variable belonging to each cluster center through the soft allocation, and divide the wind speed sequence into the cluster center with the calculated maximum probability.

[0039] Preferably, the S5 specifically includes:

[0040] S51. For any segment of wind speed data divided by the same time interval as in S1, obtain an average wind speed sequence according to step S1 and perform maximum value normalization to obtain a standardized average wind speed;

[0041] S52. Standardize the diurnal wind speed variation pattern in the diurnal wind speed variation pattern library, extract the wind speed within the same time interval as the standardized average wind speed sequence to be predicted from the standardized diurnal wind speed variation pattern to obtain a standardized wind field pattern fragment, and find the wind speed variation pattern with the smallest Euclidean distance to the standardized average wind speed as the closest wind speed variation pattern within the same time interval from the standardized wind field pattern fragment and perform denormalization processing to obtain the final diurnal wind speed variation pattern prediction data.

[0042] Preferably, the normalization of the diurnal variation pattern of wind speed is specifically expressed as follows:

[0043] ;

[0044] Where, To obtain the variable Maximum value function within the period; For the j A standardized diurnal variation pattern of wind speed; For the j Diurnal variation pattern of wind speed.

[0045] Preferably, the final wind speed diurnal variation pattern prediction data is specifically expressed as:

[0046] ;

[0047] Where, is the predicted diurnal variation pattern of normalized wind speed; is the average wind speed series; is the average wind speed series The maximum value of .

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] 1. This solution aims to predict diurnal wind farm patterns by innovatively integrating LSTM into a deep embedded clustering model. This method effectively extracts time series features and innovatively proposes using the softmax function to calculate soft assignment within the deep embedded clustering layer, enabling efficient end-to-end training of neural networks. Clustering loss weights are introduced into the model loss function to dynamically adjust the weights of reconstruction and clustering tasks. The examples show that the bridge wind farm diurnal pattern prediction method based on the LSTM deep embedded clustering model is highly effective in predicting diurnal wind farm patterns.

[0050] 2. This solution predicts the daily variation pattern of wind speed at the site of a long-span bridge, obtaining daily wind speed trends to support bridge construction and operation and maintenance decisions. The proposed LSTM deep embedded clustering algorithm, which incorporates an LSTM autoencoder into the deep embedded clustering, can better extract time series features and achieve end-to-end time series clustering, simplifying the process and reducing human intervention. Furthermore, this method is highly versatile and can be widely applied to various wind farm monitoring tasks, adapting to complex wind farm environments. It is easy to program, more efficient to train, and has a low time cost.

[0051] 3. This application realizes the dynamic fusion of autoencoder and clustering algorithm:

[0052] The LSTM deep embedded clustering model in this paper integrates feature extraction and clustering tasks to achieve end-to-end training, and one model completes both tasks at the same time. However, the two tasks are sequential, and the features extracted by the autoencoder are clustered. Therefore, feature extraction should be implemented before clustering. Therefore, a parameter is set , achieving different emphases on the two tasks at the beginning and end of the task. Compared to a static fusion loss, we focus on different tasks at different stages. We propose a dynamically changing fusion loss, optimize computing resource allocation, accelerate model convergence, and improve model training efficiency. Furthermore, focusing on different tasks at different stages improves the model's performance on different tasks.

[0053] 4. This application uses the softmax function to calculate auxiliary allocation:

[0054] Conventional clustering algorithms deterministically assign a set of data to a particular cluster, lacking smooth probability distribution properties. This makes constructing a loss function difficult or impossible. A soft assignment strategy can be used to train clustering tasks. However, conventional soft assignment methods, such as the Student's t-distribution, are unstable, computationally complex, and converge slowly. Therefore, we propose incorporating the soft assignment method using the softmax function into the model. This soft assignment method constructs a loss function for clustering training, enabling end-to-end integrated model training. Compared to the traditional Student's t-distribution, soft assignment using the softmax function is computationally simpler and more efficient. It is highly compatible with common loss functions (such as the cross-entropy loss) and relatively easy to compute gradients. Softmax is highly efficient for large datasets and high-dimensional data. Furthermore, it avoids vanishing and exploding gradients, making it particularly suitable for training deep neural networks. In the softmax method, the membership of each data point to each cluster center is expressed probabilistically, making the model results more interpretable and intuitive.

[0055] 5. Improved Deep Embedded Clustering - Using LSTM Autoencoders instead of general autoencoders:

[0056] Time series clustering tasks require attention to temporal features, but standard autoencoders are not well suited to extracting these features. By constructing an LSTM autoencoder layer to replace the standard autoencoder layer, the deep embedded clustering method is improved. In deep embedded clustering, using an LSTM autoencoder instead of a standard autoencoder significantly improves the model's ability to process time series data. LSTMs can capture long-term dependencies and temporal relationships in data, generating latent representations with more temporal information. They are suitable for variable-length inputs and are better at filtering out noise. In contrast, standard autoencoders primarily focus on static features and struggle to effectively model the dynamic changes in time series data. Therefore, when processing data with temporal dependencies, LSTM autoencoders excel in feature extraction, noise robustness, and clustering accuracy.

[0057] 6. Prediction method - holistic matching prediction

[0058] Bridge wind farm data is highly discrete, and existing methods have poor prediction performance for the high-frequency components of wind farm data. There is also a lack of methods specifically tailored to predict diurnal variations in wind speed. The results obtained by integrating the LSTM deep embedded clustering method demonstrate excellent prediction of diurnal variations. Common sliding prediction methods (such as ARIMA and LSTM) require stationary assumptions on the data, which may not be suitable for certain complex time series data. Furthermore, conventional prediction methods rely on the previous prediction, resulting in large cumulative errors and increased computational complexity. Time series matching prediction, which seeks historically similar patterns to make predictions, effectively captures long-term dependencies and complex dynamics, avoiding the error accumulation issues inherent in sliding prediction methods. It can adaptively process series data at different time scales, enhancing sensitivity to unexpected events and trend changes while reducing noise interference on prediction results. Furthermore, the matching prediction method is independent of the previous prediction, avoiding cumulative errors and exhibiting strong adaptability, making it particularly suitable for time series data with complex nonlinear and time-varying characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 A schematic diagram of a flow chart of an embodiment of the present invention;

[0060] Figure 2 This is a schematic diagram of the arrangement of wind farm data acquisition equipment in an embodiment of the present invention;

[0061] Figure 3 This is a schematic diagram of the structure of the LSTM deep embedded clustering model network integrated in an embodiment of the present invention;

[0062] Figure 4 This is a schematic diagram of the results after clustering and reconstruction in an embodiment of the present invention;

[0063] Figure 5A comparison chart of the wind speed diurnal variation model prediction results and the measured results in an embodiment of the present invention;

[0064] Figure 6 Schematic diagram of the prediction results of the diurnal wind speed variation model for a certain day in an embodiment of the present invention. DETAILED DESCRIPTION

[0065] Example 1: This application solves the problem that existing methods have poor prediction effects on the high-frequency portion of wind field data. In addition, the present invention has the following objectives:

[0066] First, engineering needs: Wind speed trends provide a preliminary reference for bridge construction, operation, and maintenance technicians, allowing them to make appropriate decisions based on daily wind speed readings. Furthermore, different wind speed trends, such as sudden increases and decreases or gradual increases and decreases, can lead to significantly different bridge responses.

[0067] Secondly, the research found that in complex mountainous and canyon areas, when analyzing the observed wind field data, it was found that the changing trend of wind speed data has a certain periodicity. One of the obvious regular fragments was found:

[0068] Over a daily period, wind speeds are low in the early morning, then increase, reaching their maximum in the afternoon, and then decrease again to lower values. We found many similar patterns. Therefore, we imagine being able to summarize this daily variation in wind speed, analyzing its specificities and related patterns, on a daily basis.

[0069] The present invention obtains wind speed and direction information at a bridge site in a mountainous canyon, performs preprocessing, and generates a wind speed sequence dataset. A fusion LSTM deep embedded clustering model is constructed, and the wind speed sequence dataset is input into the LSTM deep embedded clustering model for training until the loss function stabilizes. The trained LSTM deep embedded clustering model is then used to reconstruct cluster centers through a decoder to form a wind speed diurnal variation pattern library and divide the wind speed sequence through soft allocation. Finally, the wind speed data to be predicted is matched with the wind speed diurnal variation pattern library, and the most similar wind speed diurnal variation pattern is used as the prediction result.

[0070] Based on the aforementioned inventive concept, the implementation ideas of this application are:

[0071] Average wind speed extraction is a mature data preprocessing method for wind field research. 10 minutes is often used as the basic time interval in engineering. That is to say, the originally collected data may be one per second, but it is too discrete. In engineering, it is often divided into 10 minutes to obtain an average wind speed within each 10 minutes. Then, many 10-minute data sets are combined to form an average wind speed sequence for research.

[0072] During the research, we found that even the 10-minute average wind speed still has strong fluctuations. However, despite the fluctuations, the data also has a clear trend. The research direction of this application is to focus on and extract this trend. Here, we use Gaussian filtering to extract the wind speed trend, remove high-frequency signals, and focus only on the trend of wind speed changes.

[0073] To summarize the daily variation of this time series, the first thing that comes to mind is the clustering algorithm, which is also a commonly used and highly efficient method. The more common data for clustering is two-dimensional, which is displayed on the graph as a group of scattered points. Each point can be represented by the coordinates ( , ) indicates that each point is described by two numerical values. The time series is ( , , ,…, ), each point is represented by There are usually hundreds of n, so this point can be called a "bar". A piece of data contains many values. dimensional data. Clustering high-dimensional data like time series presents significant challenges: the sheer volume of data requires significant computer storage space. Furthermore, calculating the distance between two pieces of data when clustering high-dimensional data is complex, resulting in slow clustering convergence. Therefore, data dimensionality reduction is necessary. Numerous methods exist for dimensionality reduction, and autoencoders are particularly suitable, offering high efficiency, low error, and ease of use. Therefore, autoencoders are used to reduce the data dimensionality, and then the reduced data is clustered. This clustering yields libraries of types with distinct patterns of variation. This results in consistent wind speed trends within each library, making them highly representative. This is because clustering utilizes extensive historical data, encompassing virtually all types of variation. Therefore, future daily wind speed trends are largely encompassed by these libraries, allowing them to be used to predict wind speeds.

[0074] After you have a preliminary idea, you need to continue to refine it.

[0075] First, the wind speed database is summarized by combining an autoencoder with clustering. The autoencoder can reduce the length of a 144-day wind speed sequence, facilitating clustering. However, during research, it was discovered that this step always requires training the autoencoder, extracting features, and then using these features for clustering. Therefore, even with new data, this process still needs to be performed step by step. This operation (processing) is very inconvenient and involves excessive human intervention, increasing the potential risk of error, increasing operator workload and time, and reducing efficiency. Therefore, this application combines the autoencoder and clustering into a single model that directly outputs clustering results based on input. This significantly improves efficiency and facilitates program development, allowing for easy replication of experiments and use by other technical personnel. Therefore, the idea was to use a deep embedded clustering approach, integrating clustering into training. Clustering is trained alongside the autoencoder and outputs clustering results at the end of training.

[0076] Among them, how to achieve fusion is a very important issue, because the result of traditional clustering is a deterministic value, so the error can only be expressed as "yes" or "no". Specifically, if a piece of data is judged to be the first category, if it is indeed the first category, the error is 0, otherwise the error is 100%, there are only these two possibilities. In simulation training, taking the autoencoder as an example, its error is expressed as the difference between the reconstructed data and the original data. This difference can be calculated by a formula, such as MSE, which calculates the square root of the difference between the original data and the reconstructed data. The calculated value can be any value, so this loss function can be continuous and smooth, and the derivative can be calculated. Calculating the derivative is a necessary step to update the aforementioned model during the training process. Therefore, the autoencoder can be trained using MSE, while traditional clustering cannot differentiate the discrete results of judging "yes" or "no", so the model cannot be updated. This application introduces the softmax function for classification tasks in machine learning, and uses a probabilistic method to define whether a piece of data belongs to a certain class, quantifying the clustering results. That is, it outputs the probability that a piece of data belongs to each class, rather than assuming that the probability of belonging to one class is 1 and the rest is 0. Since the clustering results are quantized, the computational error is also quantized. The change in the computational error is no longer a transformation between two values, but rather a process that can be described by a specific function. Using gradient descent and derivatives, we can change the parameters of the clustering model (in this application's method, the parameters are the cluster centers) to continuously optimize the results. This application completely changes the existing clustering determination model and results.

[0077] However, this application is to cluster the features extracted by the autoencoder. In the prior art, such feature extraction and clustering tasks cannot be performed simultaneously. They must be focused on different stages. This application proposes a dynamic loss function for the embedded clustering algorithm. In the first stage, the clustering loss is basically 0, so the loss will be more of the reconstruction error of the autoencoder, mainly to optimize the autoencoder. In the second stage, the reconstruction error of the autoencoder is already very small. At this time, the autoencoder already has good performance, and the clustering loss of the overall model becomes the dominant loss. At this time, the cluster center will be continuously optimized to eventually obtain the best clustering result.

[0078] Since this application is for feature extraction of time series, it utilizes a long short-term memory (LSTM) neural network to effectively extract time-related information from the data, improving the accuracy of the autoencoder's reconstruction of the original data. Good data reconstruction indicates that the autoencoder is able to extract key information. This led to the development of an LSTM deep embedded clustering method, which yielded a library of diurnal wind speed variation patterns and summarized the patterns of diurnal wind speed variation.

[0079] After obtaining the wind speed diurnal variation pattern library, we can see that wind speed variations within each class are similar, while classes are not similar. Furthermore, time series have strong autocorrelation: data from one period influences data from the next, and changes in the next period are somewhat dependent on the previous period. Therefore, once the previous period is determined, the trend for the entire day can be determined. Therefore, for a wind speed series that is less than a day (for example, the previous 14 hours), it is effective to match the wind speed data from the previous period (the previous 14 hours) in the wind speed diurnal variation pattern library and use the complete wind speed diurnal variation pattern to represent the wind speed variation for that day. Because it integrates a large amount of historical data, the wind speed diurnal variation pattern library can more comprehensively summarize the patterns of wind speed variation, making it effective to use matching methods to predict wind speed variation patterns.

[0080] In the prediction method of the diurnal variation pattern of wind speed mentioned in this method, the diurnal variation pattern of wind speed refers to the trend of wind speed, such as Figure 6 The red curve in Figure 1 (blue represents raw data, green represents fluctuations, and red represents trends) indicates that this method can serve as a basis for preliminary project decisions and provide technical personnel with a reference for determining wind speed trends. This method achieves a high level of accuracy. However, this method does not include the prediction of fluctuations, making it impossible to predict short-term sudden changes in wind speed.

[0081] like Figure 1-6 As shown in the figure, the bridge wind field change pattern prediction method based on the LSTM deep embedded clustering model includes:

[0082] S1. Data preparation: Obtain wind speed and direction information at the mountain canyon bridge site, perform preprocessing, and generate a wind speed sequence dataset;

[0083] S2. Model construction: Build a fused LSTM deep embedded clustering model, which includes an LSTM autoencoder layer and a deep embedded clustering layer;

[0084] S3, model training: inputting the wind speed sequence dataset into the LSTM deep embedded clustering model for training until the loss function is stable;

[0085] S4. Construction of a diurnal variation pattern library: Using the trained LSTM deep embedded clustering model, the decoder is used to reconstruct the cluster centers, forming a diurnal wind speed variation pattern library and dividing the wind speed sequence through soft allocation;

[0086] S5. Daily variation pattern prediction: Match the wind speed data to be predicted with the wind speed daily variation pattern library, and take the most similar wind speed daily variation pattern as the prediction result.

[0087] Furthermore, the specific method of data preparation in step S1 is to convert wind speed Corresponding wind direction Split the data into 10-minute intervals. For any data, its length should be:

[0088]

[0089] Where, For any data length, The minimum sampling number requirement for sampling equipment in ten minutes.

[0090] The measured data should meet the following requirements:

[0091]

[0092]

[0093] In the formula Indicates the average measured wind speed. is the standard deviation of the measured wind speed, Indicates the average value of the measured wind direction. is the standard deviation of the measured wind direction.

[0094] The measured data that do not meet the above inequality conditions are eliminated. and wind direction Decompose into Direction and Direction, that is:

[0095]

[0096]

[0097] In the formula and Respectively indicate along Direction and The above two directions are the specific directions mentioned in this application. The specific directions here are commonly used in the vector decomposition method of wind speed. In the specific research process, in order to facilitate the analysis of the data obtained by the Wind3D 6000 wind laser radar, the instrument is set to the north as the wind direction angle. 0° direction, when rotating clockwise, the wind direction angle Increase. The general method of wind field research, vector decomposition method (Xiang Haifan, Ge Yaojun, Zhu Ledong. Modern Bridge Wind Resistance Theory and Practice [M]. Beijing: People's Transportation Press, 2005) is used to decompose the wind speed into two mutually orthogonal directions; after this treatment, Along =0° direction, Along =90° direction. After decomposing the instantaneous wind speed into these two directions, take 10 minutes as the basic time interval, and calculate two average values of the wind speed in the two directions within each interval. , then square and sum them to synthesize the average wind speed and wind direction angle. The average wind speed obtained in this way is the main wind speed, and the average wind direction is the main wind direction. This is the vector decomposition method of wind speed.

[0098] Then calculate the ten-minute average wind speed and wind direction ,Right now:

[0099]

[0100]

[0101] In the formula and Respectively indicate along Direction and The average wind speed component in the direction of the wind is then calculated. The data is segmented in 24-hour units. The starting time of each data item is 00:00:00. For any data item, the following conditions must be met:

[0102]

[0103] In the formula Indicates the length of each data in 24 hours. Data that does not meet the above formula is directly eliminated. Then a Gaussian filter is used to obtain the change trend of each data. The maximum normalization method is used to standardize the data, that is:

[0104]

[0105] In the formula represents the normalized wind speed, The ten-minute average wind speed. is the maximum value function.

[0106] Integrate all Forming a wind speed series data set , .

[0107] Contains n sequences, each of which is a time series , represents the normalized wind speed data of one day, each sequence Contains 144 normalized values of the ten-minute average wind speed maximum, arranged in chronological order from the beginning to the end.

[0108] Furthermore, in the model building process in step S2, the fused LSTM deep embedded clustering model constructed includes an LSTM autoencoder layer and a deep embedded clustering layer, wherein the specific structure of the LSTM autoencoder layer includes:

[0109] Encoder: It consists of LSTM layers and is used to extract input features. The latent variables it outputs represent the key features of the input data and are expressed as:

[0110]

[0111] In the formula represents the encoder, Represents the input of the model, here is the wind speed series data set, represents the encoder model parameters of the LSTM autoencoder, Is the output of the model, representing the wind speed series data set The feature set extracted from There are n sequences, each of which are all one-dimensional vectors with length less than 144, express characteristics;

[0112] Decoder: It consists of LSTM layers and fully connected layers, used to reconstruct the original data from the latent variables, expressed as:

[0113]

[0114] In the formula Represents the set of wind speed sequences reconstructed by the decoder by decoding the features extracted by the encoder. , Describes the decoder, is the input of the decoder model, representing the wind speed sequence data set The feature set extracted from Represents the decoder model parameters of the LSTM autoencoder;

[0115] Latent space: The latent variables output by the encoder serve as a low-dimensional feature representation of the input data.

[0116] Furthermore, the model building process in step S2 constructs a fused LSTM deep embedded clustering model, which includes an LSTM autoencoder layer and a deep embedded clustering layer, wherein the specific behaviors of the deep embedded clustering layer include:

[0117] will be collected n sequences in Divided into k classes, the class center is Indicates that , use k-means to initialize the cluster center ,calculate Soft assignment to cluster centers and auxiliary distribution , and the KL divergence composed of soft assignment and auxiliary assignment is used as the clustering loss.

[0118] Furthermore, in the model building process in step S2, the fused LSTM deep embedded clustering model is constructed, which includes an LSTM autoencoder layer and a deep embedded clustering layer, and uses the fusion of reconstruction loss and clustering loss as the loss function of the model. The reconstruction loss is expressed as:

[0119]

[0120] Where, represents the reconstruction loss function, , . Use the softmax function to calculate Soft assignment to cluster centers ,Right now:

[0121]

[0122] In the formula express The probability of belonging to class j, , Is the temperature parameter that controls the hardness or softness of clustering. When the temperature is high, the samples are evenly distributed among the clusters, while when the temperature is low, the samples are more concentrated in a certain cluster. , which can also be called target allocation, that is:

[0123]

[0124] Where, express The probability of belonging to the jth class is given by Calculated, with higher credibility, requires Getting closer ,when equal The clustering loss is zero when , so the clustering loss can be defined as the KL divergence between the soft assignment and the auxiliary assignment, that is:

[0125]

[0126] In the formula Indicates the calculation of the KL divergence between P and Q. The LSTM autoencoder reconstruction loss With Deep Embedded Clustering Loss Fusion, forming the loss function of the fused LSTM deep embedded clustering model, namely:

[0127]

[0128] In the formula represents the weight of clustering loss, It has dynamic changing characteristics, namely:

[0129]

[0130] In the formula is the weight for initializing clustering loss, t is the number of iterations, is the weight growth rate of clustering loss.

[0131] Furthermore, in the model building process in step S2, the fused LSTM deep embedded clustering model is constructed, and the LSTM network parameters are optimized by the stochastic gradient descent method with momentum. 、 and the cluster centers of the deep embedded clustering layer LSTM network parameters 、 The optimization calculation is as follows:

[0132]

[0133]

[0134] In the formula is the encoder model parameter in the t-th round LSTM autoencoder, is the encoder model parameter in the LSTM autoencoder of the t+1th round, are the decoder model parameters in the t-th round LSTM autoencoder, are the decoder model parameters in the LSTM autoencoder of the t+1th round, represents the learning rate, Express about The gradient operator, Express about The gradient operator of . Cluster centers of deep embedded clustering layer The optimization calculation is as follows:

[0135]

[0136] In the formula is the deep embedded cluster center of the tth round, is the deep embedded clustering center of the t+1th round, represents the learning rate, Express about The gradient operator.

[0137] Furthermore, the model construction process in step S2 includes an LSTM autoencoder layer and a deep embedded clustering layer. Each component is built using code, including all the computational processes described in step S2. Each component is connected together through code to form a trainer for the entire model. The overall model requires inputs such as the wind speed series dataset, the selected number of clusters, and the total number of training rounds. The output is the encoder, decoder, and cluster centers.

[0138] Furthermore, in the model training process in step S3, the data prepared in step S1 is input into the model constructed in step S2, the selected number of clusters and the total number of training rounds are input, and the code is run to start training. In each round, after the data set passes through the model, the total loss formula (16) described in S2 is calculated, and the model and cluster centers are updated using formulas (18)-(20) to make the model and cluster centers more accurate. The round ends and the next round is carried out. After the specified total number of rounds is completed, the training stops. The total loss represents the error. As the number of rounds increases, the total loss decreases. It should be ensured that the number of rounds when the total loss is reduced to a stable state and no longer changes significantly is less than the specified total number of training rounds.

[0139] Furthermore, the process of constructing the daily variation pattern library in step S4 uses the optimized LSTM decoder The optimized cluster centers Reconstruction is performed to form a wind speed diurnal variation pattern library, namely:

[0140]

[0141] In the formula , represents the wind speed daily variation pattern library, which contains k wind speed sequences, each sequence is a one-dimensional vector of length 144, represents the network parameters of the optimal LSTM decoder. All sequences in , each sequence are all one-dimensional vectors with length less than 144, express The characteristics of each cluster center are calculated by formula (13) Probability , assign the class label with the maximum probability to the sequence: for example The largest, then Belongs to category j.

[0142] Furthermore, the daily variation pattern prediction process in step S5 is performed on the basis of steps S1-S4. Each sequence in the data set obtained in step S1 All of them start from 00:00:00, so the wind speed data needed for model prediction should also start from 00:00:00. For any section of wind speed data starting from 00:00:00 and lasting no longer than 24 hours, the start and end times are 、 ,use Time period display to First, calculate the ten-minute average wind speed sequence of period T according to equations (1)-(6) in step S1 , is a one-dimensional vector with a length of N and N is less than 144, and the formula (9) is used to Perform maximum value normalization to obtain the standardized average wind speed . Then , and standardize, namely:

[0143]

[0144] Where, To obtain the variable Function of the maximum value within the period; is the jth normalized diurnal variation pattern of wind speed; is the j-th diurnal variation pattern of wind speed; from the standardized diurnal variation pattern of wind speed Intercept Wind speed during the period , The length of is also N, and One-to-one correspondence, and then according to the Euclidean distance from Find and The closest Wind speed change mode during the period, if , which corresponds to , and Perform inverse normalization to obtain the final wind speed diurnal variation pattern forecast data ,Right now:

[0145]

[0146] Where, is the predicted diurnal variation pattern of normalized wind speed; is the average wind speed series; is the average wind speed series The maximum value of .

[0147] Example 2: A method for predicting bridge wind field change patterns based on an LSTM deep embedded clustering model, comprising:

[0148] S1 Data preparation: Obtain the original information of wind speed and direction at the mountain canyon bridge site, perform data preprocessing, and generate a data set.

[0149] This embodiment uses a laser radar anemometer to measure wind field data, such as Figure 2 As shown. The laser radar anemometer is 80m away from one side of the bridge tower in the direction of the bridge axis and 40m away from the bridge tower in the direction perpendicular to the bridge axis. The vertical elevation difference with the bridge deck is 4m. The laser radar anemometer can measure wind speed at various heights simultaneously. In this embodiment, the wind speed at a height of 56m is selected as the measured wind speed. ,wind direction The bridge axis is 0° and clockwise is positive.

[0150] First, the wind speed Corresponding wind direction Split the data into 10-minute intervals. For any data, its length should be:

[0151]

[0152] Where, For any data length, The minimum sampling number requirement for the sampling device in ten minutes. Take 100, and remove data with a length less than 100. The measured data should meet the following requirements:

[0153]

[0154]

[0155] In the formula Indicates the average measured wind speed. is the standard deviation of the measured wind speed, Indicates the average value of the measured wind direction. is the standard deviation of the measured wind direction. The measured data that do not meet the above inequality conditions are eliminated. and wind direction Decompose into The bridge axis direction and The direction perpendicular to the bridge axis, that is

[0156]

[0157]

[0158] In the formula and Respectively indicate along Direction and The wind speed component of the direction. Then calculate the ten-minute average wind speed and wind direction ,Right now:

[0159]

[0160]

[0161] In the formula and Respectively indicate along Direction and The average value of the wind speed components in the direction.

[0162] Then the ten-minute average wind speed The data is segmented in 24-hour units. The starting time of each data item is 00:00:00. For any data item, the following conditions must be met:

[0163]

[0164] In the formula Indicates the length of each data in 24-hour units. Data that does not meet the above formula is directly eliminated, and finally 130 data are obtained. One data represents a wind speed data sequence. A sequence contains 144 values, and one value represents the average wind speed for ten minutes. Here, a data from July 1, 2023 is used as an example, as shown in Table 1:

[0165] Table 1 Average wind speed data on July 1, 2023

[0166]

[0167] Then use Gaussian filter to obtain the change trend of each data and set the parameters in the Gaussian filter =8.

[0168] The above data changes to the data shown in Table 2:

[0169] Table 2 Average wind speed data after Gaussian filter

[0170]

[0171] The maximum normalization method is used to standardize the data, that is:

[0172]

[0173] In the formula represents the normalized wind speed, The ten-minute average wind speed. is the maximum value function.

[0174] The data in Table 2 above changes to the data in Table 3 below :

[0175] Table 3 Standardized average wind speed data

[0176]

[0177] Integrate all Forming a wind speed series data set , . Contains 130 sequences, each of which is a time series , represents the normalized wind speed data of one day, each sequence Contains 144 normalized values of the ten-minute average wind speed maximum value, arranged in chronological order from front to back. Wind speed series data set , where each row represents a ,therefore There are 130 rows and 144 columns.

[0178] S2 model construction: This model is a fusion LSTM deep embedded clustering model, consisting of two parts: LSTM autoencoder and deep embedded clustering layer.

[0179] After obtaining the processed data set, we then use Python code to build a fused LSTM deep embedded clustering model, which includes an LSTM autoencoder layer and a deep embedded clustering layer. The specific structure of the LSTM autoencoder layer includes: Encoder: It consists of an LSTM layer and is used to extract input features. The latent variables it outputs represent the key features of the input data and are expressed as:

[0180]

[0181] In the formula represents the encoder, Represents the input of the model, here is the wind speed series data set, represents the encoder model parameters of the LSTM autoencoder, Is the output of the model, representing the wind speed series data set The feature set extracted from There are 130 sequences in are all one-dimensional vectors with a length of 32. express Features, and One-to-one correspondence; decoder: consists of LSTM layers and fully connected layers, used to reconstruct the original data from the latent variables, expressed as:

[0182]

[0183] In the formula Represents the set of wind speed sequences reconstructed by the decoder by decoding the features extracted by the encoder. , Describes the decoder, is the input of the decoder model, representing the wind speed sequence data set The feature set extracted from Represents the decoder model parameters of the LSTM autoencoder; Latent space: the latent variables output by the encoder, which serve as the low-dimensional feature representation of the input data. The structure of the entire LSTM autoencoder is as follows Figure 3 shown.

[0184] The reconstruction loss and clustering loss are fused as the loss function of the model. The reconstruction loss is expressed as:

[0185]

[0186] Where, represents the reconstruction loss function, , The code is as follows:

[0187] reconstruction_loss =tf.reduce_mean(tf.square(X_train -reconstructed))

[0188] After constructing the LSTM autoencoder structure, a deep embedded clustering layer is constructed. The specific data processing process of the clustering layer is as follows: The 130 sequences in the dataset are divided into 8 clusters. The silhouette coefficient method is used to select the number of clusters. The silhouette coefficient method can measure the compactness and separation of each data point after clustering, and then evaluate the quality of the entire cluster. The silhouette coefficient s(i) is calculated as follows:

[0189]

[0190] Where, represents the average distance from the data point 𝑖 to the nearest other clusters (i.e., the degree of separation between clusters), Represents the average distance from the data point 𝑖 to its cluster (i.e., the closeness within the cluster). The silhouette coefficients of all data points are added together to obtain the total silhouette coefficient S. The larger S is, the better the clustering effect is. Different cluster numbers are used for clustering and S is calculated. The cluster number when S is the largest is selected as the cluster number of this embodiment. The cluster center is used Indicates that , use k-means to initialize the cluster center , the code is as follows:

[0191] # Initialize cluster centers using K-means

[0192] from sklearn.cluster import KMeans

[0193] kmeans =KMeans(n_clusters=num_clusters)

[0194] encoded_features = encoder.predict(X_train) # Use encoder to extract features

[0195] initial_centers =kmeans.fit(encoded_features).cluster_centers_

[0196] calculate Soft assignment to cluster centers and auxiliary distribution , and the KL divergence composed of soft allocation and auxiliary allocation is used as the clustering loss. Use the softmax function to calculate Soft assignment to cluster centers ,Right now:

[0197]

[0198] In the formula express The probability of belonging to class j, , Is the temperature parameter, which is used to control the hardness of clustering. When the temperature is high, the samples will be evenly distributed to each cluster, while when the temperature is low, the samples will be more concentrated in a certain cluster. Here the temperature parameter is 0.993. Construct a high-confidence allocation as an auxiliary allocation , which can also be called target allocation, that is:

[0199]

[0200] In the formula express The probability of belonging to the jth class is given by Calculated, with higher credibility, requires Getting closer ,when equal The clustering loss is zero when , so the clustering loss can be defined as the KL divergence between the soft assignment and the auxiliary assignment, that is:

[0201]

[0202] In the formula Indicates calculating the KL divergence between P and Q.

[0203] Reconstructing the LSTM Autoencoder Loss With Deep Embedded Clustering Loss Fusion, forming the loss function of the fused LSTM deep embedded clustering model, namely:

[0204]

[0205] In the formula represents the weight of clustering loss, It has dynamic changing characteristics, namely:

[0206]

[0207] In the formula To initialize the weight of clustering loss, take 0.01, t is the iteration round, is the weight growth rate of clustering loss, which is 1.005.

[0208] S3 model training: Input data into the model instance and train until the loss function stabilizes.

[0209] A total loss is calculated in each round, and the LSTM network parameters need to be optimized by stochastic gradient descent with momentum. 、 and the cluster centers of the deep embedded clustering layer , which reduces the total loss of the next round. LSTM network parameters 、 The optimization calculation is as follows:

[0210]

[0211]

[0212] In the formula is the encoder model parameter in the t-th round LSTM autoencoder, is the encoder model parameter in the LSTM autoencoder of the t+1th round, are the decoder model parameters in the t-th round LSTM autoencoder, are the decoder model parameters in the LSTM autoencoder of the t+1th round, Indicates the learning rate, which is 0.001. Express about The gradient operator, Express about The gradient operator.

[0213] Cluster centers of deep embedded clustering layer The optimization calculation is as follows:

[0214]

[0215] In the formula is the deep embedded cluster center of the tth round, is the deep embedded clustering center of the t+1th round, Indicates the learning rate, which is 0.001. Express about The gradient operator.

[0216] The constructed model includes an LSTM autoencoder layer and a deep embedded clustering layer. Each part is built through code, including all the calculation processes described in step S2. Each part is connected together through code to form a trainer for the entire model. The overall model requires the wind speed sequence dataset as input. , as well as the selected number of clusters and the total number of training rounds, the output is the encoder, decoder and cluster centers.

[0217] After building the overall model trainer, prepare the data , after adding one dimension, the shape becomes (130, 1, 144), making it suitable for the input of the model.

[0218] The dataset is input into the constructed overall model trainer, the number of clusters is defined as 8, the total number of training rounds is 600, and the output is LSTM autoencoder (autoencoder), encoder (encoder), decoder (decoder) and cluster centers (cluster_centers).

[0219] When running the trainer of the above code, the total loss formula (16) described in S2 is calculated after each round of the data set passes through the model, and the model and cluster centers are updated using formulas (18)-(20) to make the model and cluster centers more accurate. The round ends and the next round is carried out. Training stops after the specified total number of rounds. The total loss represents the error. As the number of rounds increases, the total loss decreases. The total loss of each round is printed. When the total loss reaches 600 rounds, the change in the loss is less than 0.0002, so it can be judged that the loss function value has stabilized at this time. The condition for stopping training here is that the change in the loss is less than 0.0002 (customized according to the specific project) or the set training rounds are completed. In either case, it is necessary to ensure that the number of rounds when the total loss is reduced to a stable state and no longer changes significantly is less than the specified total number of training rounds.

[0220] After the training is completed, the LSTM autoencoder, encoder, decoder and cluster center can be obtained. The encoder is used to train the dataset with 130 rows and 144 columns. Encode to get , There are 130 rows and 32 columns. Each line is from The features extracted from each row of .

[0221] The length of each cluster center is 32, 8 cluster centers .

[0222] S4 Daily Variation Pattern Library Construction: The cluster centers are reconstructed using the LSMT decoder to form a wind speed daily variation pattern library.

[0223] For the above 8 cluster centers Using optimized LSTM decoder Reconstruction is performed to form a wind speed diurnal variation pattern library, namely:

[0224]

[0225] In the formula , represents the wind speed daily variation pattern library, which contains 8 wind speed sequences, each sequence is a one-dimensional vector of length 144, Represents the network parameters of the optimal LSTM decoder. There are 8 rows and 144 columns.

[0226] Will All sequences in , each sequence are all one-dimensional vectors with length less than 144, express The characteristics of each cluster center are calculated by formula (13) Probability , assign the class label with the maximum probability to the sequence: for example The largest, then Belongs to the jth category, thus dividing All sequences in The class to which it belongs.

[0227] Data belonging to the same class Corresponding Draw it in the same picture with the center of this class, such as Figure 4 As shown, there are eight subgraphs, representing eight classes respectively. is a set of wind speed series data The length of the sequence is 144. is a collection A sequence of length 32, express Features, and One to one correspondence. Figure 4 In each subgraph, the horizontal axis represents time, the vertical axis represents the maximum normalized wind speed data, and the thin dotted line in each subgraph is the wind speed series data in this category. The thick solid line represents this type of diurnal wind speed variation pattern. Therefore, there are eight types of diurnal wind speed variation patterns. The wind speed series data in each sub-graph are similar, and each type of diurnal wind speed variation pattern can represent the diurnal variation pattern of this type of wind speed.

[0228] S5 Daily Variation Pattern Prediction: Match the known wind speed data with the corresponding area of the wind speed daily variation pattern, and find the most similar wind speed daily variation pattern as the prediction result.

[0229] Finally, the daily variation pattern prediction is carried out. The daily variation pattern prediction is based on the steps S1-S4. Each sequence in the data set obtained in S1 All of them start from 00:00:00 (the same time), so the wind speed data that needs to be predicted by the model should also start from 00:00:00. For any wind speed data starting from 00:00:00 and lasting no longer than 24 hours (24 hours is for diurnal variation prediction), the start and end times are 、 ,use Time period display to Time, here , First, calculate the ten-minute average wind speed sequence for period T according to equations (1)-(6) in step S1 , is a one-dimensional vector with a length of 84, and is expressed using formula (9) Perform maximum value normalization to obtain the standardized average wind speed . Then , and standardize, namely:

[0230]

[0231] In the formula To obtain the variable The function of the maximum value within the period. From the normalized diurnal variation pattern of wind speed Intercept Wind speed during the period , The length of is also 84, and One-to-one correspondence, and then according to the Euclidean distance from Find and The closest Wind speed change mode during the period, if , which corresponds to , and Perform inverse normalization to obtain the final wind speed diurnal variation pattern forecast data ,Right now:

[0232]

[0233] You only need to input the data (new_data_cut) for which you want to predict the pattern, and you can complete the prediction of the daily variation pattern of wind speed and draw the prediction results.

[0234] In order to observe whether the prediction results are accurate, this embodiment uses the complete wind speed daily variation pattern data of a certain day, intercepts the data of the 00:00:00-14:00:00 segment, which is 84 in length, and uses the above method to make predictions. The prediction results are plotted on Figure 5 The intercepted data to be predicted and the predicted data are as follows:

[0235] Example 1: To be predicted ; Prediction results .

[0236] In order to verify the prediction effect, The data from 00:00:00 to 14:00:00 are intercepted from the complete wind speed sequence of a known day. By comparing the predicted value with the actual value, the effect of the prediction can be known. Figure 5 The solid line in the figure is the curve of the complete wind speed series data. Figure 5 The middle dashed line corresponds to the curve drawn based on the prediction results of the wind speed series. Figure 5 The horizontal axis represents time, and the vertical axis represents wind speed. Figure 5 It can be seen that the wind speed sequence to be predicted from 14:00:00 to 24:00:00 is very consistent with the real sequence. The calculated MAPE is equal to 0.1138, and the prediction error is very small, indicating that this method is effective in predicting the diurnal variation pattern of wind speed.

[0237] The MAPE calculation formula is as follows:

[0238]

[0239] Where, is the number of data points, is the predicted data, The MAPE value is between 0 and 1. The smaller the MAPE value, the smaller the prediction error.

[0240] This solution aims to predict diurnal wind farm patterns by innovatively integrating LSTM into a deep embedded clustering model. This method effectively extracts time series features and innovatively proposes using a softmax function to calculate soft allocation within the deep embedded clustering layer, enabling efficient end-to-end training of neural networks. Clustering loss weights are introduced into the model loss function to dynamically adjust the weights of reconstruction and clustering tasks. The examples show that the bridge wind farm diurnal pattern prediction method based on the LSTM deep embedded clustering model is highly effective in predicting diurnal wind farm patterns.

Claims

1. A bridge wind field change pattern prediction method based on the LSTM deep embedded clustering model is characterized by: The following steps are involved: S1. Data preparation: Obtain wind speed and direction information at the mountain canyon bridge site, perform preprocessing, and generate a wind speed sequence dataset; S2. Model construction: Build a fused LSTM deep embedded clustering model, which includes an LSTM autoencoder layer and a deep embedded clustering layer; S21. Construct an LSTM autoencoder layer, including an encoder composed of an LSTM layer and a decoder composed of an LSTM layer and a fully connected layer; S22. Build a deep embedded clustering layer to classify the latent variables, use k-means to initialize the cluster centers, calculate the soft allocation through the softmax function and construct the auxiliary allocation, and use the KL divergence of the two as the clustering loss; S23, the reconstruction loss and clustering loss are integrated into the LSTM deep embedded clustering model loss function; S24, set the learning rate, initialize the clustering loss weight and the clustering loss weight growth rate; The loss function of the LSTM deep embedded clustering model is specifically expressed as: ； Where, represents the weight of clustering loss; ; is the weight for initializing clustering loss, t is the number of iterations, is the weight growth rate of clustering loss; is the reconstruction loss function; is the clustering loss; S3. Model training: inputting the wind speed sequence dataset into the LSTM deep embedded clustering model for training until the loss function is stable; S4. Construction of a diurnal variation pattern library: Using the trained LSTM deep embedded clustering model, the decoder is used to reconstruct the cluster centers, forming a diurnal wind speed variation pattern library and dividing the wind speed sequence through soft allocation; S5. Daily variation pattern prediction: Match the wind speed data to be predicted with the wind speed daily variation pattern library, and take the most similar wind speed daily variation pattern as the prediction result.

2. The method according to claim 1, wherein Said S1 specifically includes: S11, dividing the wind speed and direction data at predetermined time intervals, and eliminating the measured data that does not meet the preset length and standard deviation; S12. Decompose the wind speed and wind direction into wind speed components in specific directions, and calculate the average wind speed and wind direction angle; S13. Divide the average wind speed into equal time intervals, use a filter to obtain the change trend, and perform standardization to form a wind speed sequence data set.

3. The method according to claim 1, wherein The soft allocation is specifically expressed as: ； Where, express The probability of belonging to class j, express Features , is the temperature parameter, k is the number of clusters; m is the number of cluster centers, ranging from 1 to k; c m is the cluster center of the mth class; c j is the cluster center of the jth class; The auxiliary allocation is specifically expressed as: ； Where, for The probability of belonging to class j; for The probability of belonging to the mth class; for The probability of belonging to the mth category; x is the sequence number of the wind speed sequence; n is the number of wind speed sequences.

4. The method according to claim 1, wherein The S3 specifically includes: The wind speed sequence dataset is input into the LSTM deep embedded clustering model, the number of clusters and the total number of training rounds are set, and training is performed; wherein, in each training, the wind speed sequence dataset obtains the LSTM deep embedded clustering model loss function through the LSTM deep embedded clustering model, and the LSTM network parameters and the cluster center of the deep embedded clustering layer are optimized by the stochastic gradient descent method with momentum, and the training is stopped after the set total number of rounds are completed, ensuring that the number of rounds when the total loss is reduced to a stable state is less than the specified total number of training rounds, so that the total loss has stabilized after all rounds of training are completed.

5. The method according to claim 1, wherein The S4 specifically includes: S41, using the optimized LSTM decoder to reconstruct the optimized cluster centers to form a wind speed diurnal variation pattern library; S42. Calculate the probability of the wind speed sequence in the latent variable belonging to each cluster center through the soft allocation, and divide the wind speed sequence into the cluster center with the calculated maximum probability.

6. The method according to claim 1, wherein The S5 specifically includes: S51. For any segment of wind speed data divided by the same time interval as in S1, obtain an average wind speed sequence according to step S1 and perform maximum value normalization to obtain a standardized average wind speed; S52. Standardize the diurnal wind speed variation pattern in the diurnal wind speed variation pattern library, extract the wind speed within the same time interval as the standardized average wind speed sequence to be predicted from the standardized diurnal wind speed variation pattern to obtain a standardized wind field pattern fragment, and find the wind speed variation pattern with the smallest Euclidean distance to the standardized average wind speed as the closest wind speed variation pattern within the same time interval from the standardized wind field pattern fragment and perform denormalization processing to obtain the final diurnal wind speed variation pattern prediction data.

7. The method according to claim 6, wherein The normalized diurnal variation pattern of wind speed is specifically expressed as follows: ； Where, To obtain the variable Maximum value function within the period; is the jth normalized diurnal variation pattern of wind speed; is the j-th diurnal variation pattern of wind speed.

8. The method according to claim 6, wherein The final wind speed daily variation pattern prediction data is specifically expressed as: ； Where, is the predicted diurnal variation pattern of normalized wind speed; is the average wind speed series; is the average wind speed series The maximum value of .

Images