Typhoon weather wind power prediction method based on high-order coding and cluster division

Abstract

The invention discloses a typhoon weather wind power prediction method based on high-order coding and cluster division, and relates to the technical field of wind power prediction, and the method comprises the steps: obtaining weather prediction data in a future set time, and carrying out the dynamic extraction and cluster division of the weather prediction data; acquiring historical power data and meteorological data, and constructing a data set as training data based on the historical power data and the meteorological data; extracting training data in each sub-cluster according to a cluster division result, and performing modeling training according to the training data in the sub-clusters; and each sub-cluster loads a test set corresponding to the sub-cluster, prediction is carried out by using a model after modeling training is completed, and all sub-cluster results are summarized to form overall wind power prediction output in a future set time. The method can better adapt to uncertainty caused by wind speed fluctuation in extreme weather, improves the robustness and accuracy of wind power prediction, and provides more reliable data support for power grid dispatching optimization and wind power plant operation.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power prediction, and in particular to a method for predicting wind power in typhoon weather based on high-order coding and cluster division. Background Art

[0002] As one of the important new energy sources, wind power is highly sensitive to changes in weather conditions and exhibits strong volatility and randomness. Especially in the current situation of complex climate change and typhoon extreme weather events, large wind shedding events will lead to a sharp increase or decrease in wind power output, and even damage to wind turbines, affecting the supply and demand balance of the power grid and national energy security. Today, the distribution scale of wind farms has gradually changed from the decentralized, small-scale distribution in the early stages of development to clustered, large-scale distribution. Large-scale access to wind power clusters will have a certain impact, so the importance of conducting power forecasts for wind power clusters under typhoon extreme weather is becoming increasingly prominent.

[0003] Under extreme weather conditions such as typhoons, wind speed and direction change dramatically, resulting in a significant increase in the uncertainty of wind power forecasting. Traditional methods that rely on numerical weather forecasts (NWP) and historical power data are susceptible to interference from measurement noise, data missing and sudden changes in wind speed, and the prediction accuracy is significantly reduced. Summary of the invention

[0004] The purpose of the present invention is to overcome the defect in the prior art that wind power prediction is easily affected by measurement noise, data missing and sudden changes in wind speed, resulting in a significant decrease in prediction accuracy, and to provide a typhoon weather wind power prediction method based on high-order coding and cluster division.

[0005] The purpose of the present invention is achieved through the following technical solutions: The typhoon weather wind power prediction method based on high-order coding and cluster division includes: Obtain weather forecast data (NWP) within a set time in the future, and dynamically extract and cluster the weather forecast data; Obtain historical power data and meteorological data, and construct a data set based on the historical power data and meteorological data as training data; Extract training data in each sub-cluster according to the cluster division results, and perform modeling training based on the training data in the sub-cluster; Each subcluster loads its own corresponding test set and uses the model that has completed modeling training to make predictions separately. The results of all subclusters are aggregated to form the overall wind power prediction output within the future set time.

[0006] Preferably, the key meteorological factors in the meteorological data are determined by the Pearson correlation coefficient, the correlation between the meteorological factors in the meteorological data and the power is analyzed, and the first several meteorological factors with the largest absolute value of the correlation coefficient are taken as the key meteorological factors.

[0007] Preferably, the obtaining of the meteorological prediction data within a future set time includes the wind speed, wind direction, air pressure, temperature and humidity data at 100 meters, 30 meters and 10 meters above the ground.

[0008] Preferably, if the sample number of the meteorological prediction data is less than the minimum required sample number for model prediction, the meteorological prediction data is expanded by a small sample expansion method.

[0009] Preferably, the historical power data is obtained through the SCADA system. If there are short-term missing, jump or drift errors in the historical power data obtained through the SCADA system, they are identified and corrected through a mechanism based on sliding window statistics, change point detection and anomaly classification, and the quality of the historical power data is improved through an error confidence and weighting strategy.

[0010] Preferably, the dynamic extraction and cluster division of the meteorological prediction data are specifically as follows: The meteorological prediction data is divided into several groups of data sets with different time periods, and the duration range within each data set is the same; the dynamic clustering algorithm is used to divide the cluster structure within each data set to obtain the sub-cluster division result for this time period.

[0011] Preferably, the dynamic clustering algorithm is an improved dynamic clustering algorithm that integrates multi-feature weighting and time period comparison mechanisms.

[0012] Preferably, the process of the improved dynamic clustering algorithm is as follows: Within each sliding window, the original sample features are extracted, a multi-feature weighted distance metric function is constructed, the weighting coefficients are set according to the influence degree of meteorological variables on wind power, and initial clustering is performed; The BIC evaluation criterion is used to select the optimal number of clusters and dynamically adjust the number of sub-clusters; A time series comparison constraint optimization mechanism is introduced to construct positive sample pairs within the same time period and negative sample pairs in adjacent time periods, and a contrast loss is introduced to strengthen the stability and discriminability of the cluster boundary; The dynamic sub-cluster division result is output for subsequent separate training and prediction of the model.

[0013] Preferably, the modeling training is based on the modeling training of a high-order shrinkage autoencoder and a Transformer model.

[0014] Preferably, a high-order derivative regularization term of the input with respect to the hidden representation is introduced into the loss function of the high-order contractive autoencoder to limit the sensitivity of the encoder to small perturbations in the input space.

[0015] The beneficial effects of the present invention are as follows: The present invention proposes a typhoon weather wind power prediction method based on high-order encoding and cluster division, introducing a clustering optimization strategy to improve the prediction accuracy and stability of wind power under extreme weather conditions such as typhoons. First, the present invention uses a high-order contractive autoencoder (HO-CAE) to preprocess the SCADA (the source of historical power data and meteorological data) system and NWP data, removing measurement noise and outliers and improving data quality. By introducing a second-order gradient constraint on the input data, the high-order contractive autoencoder can effectively reduce small perturbations in the data and make the model more robust to extreme meteorological conditions. Subsequently, future NWP prediction data is divided into multiple time periods, and clustering analysis is performed based on key features such as wind speed and wind direction. By introducing a time series contrast constraint mechanism, the wind farm can be divided into multiple sub-clusters with similar behaviors, and then a Transformer model is trained for each sub-cluster respectively to further improve the accuracy of time series modeling. Through the self-attention mechanism, the Transformer model can effectively capture the complex non-linear relationship between wind speed and power output, and learn stable time series patterns within different sub-clusters, thereby reducing the impact of wind speed mutations on the overall prediction accuracy. Finally, the prediction results of each sub-cluster are aggregated to obtain the global wind power prediction. Compared with traditional methods, the present invention can better adapt to the uncertainty brought by wind speed fluctuations under extreme weather, improve the robustness and accuracy of wind power prediction, and provide more reliable data support for grid dispatching optimization and wind farm operation. Description of the Drawings

[0016] Figure 1 is a flowchart of the present invention; Figure 2 is a comparison chart of wind speed changes at SF1-SF3 stations during typhoon passage; Figure 3 is a flowchart of the WCC-TCC clustering method; Figure 4 is a schematic diagram of HO-CAE; Figure 5 is a structure diagram of HO-CAE; Figure 6 is a structure diagram of the attention mechanism; Figure 7 is a structure diagram of the Transformer model; Figure 8 is the overall modeling process; Figure 9 is the wind farm cluster division result; Figure 10 It is a schematic diagram of the wind power prediction result under normal conditions (without generator tripping) during typhoon weather; Figure 11 It is a schematic diagram of the wind power prediction result under the generator tripping scenario during typhoon weather. Specific implementation manners

[0017] Now, example embodiments will be described more fully with reference to the accompanying drawings. However, the example embodiments can be implemented in various forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this application will be more complete and comprehensive, and will fully convey the concept of the example embodiments to those skilled in the art.

[0018] In addition, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to give a thorough understanding of the embodiments of this application. However, those skilled in the art will realize that the technical solutions of this application can be practiced without one or more of the specific details, or other methods, components, devices, steps, etc. can be adopted. In other cases, well-known methods, devices, implementations, or operations are not shown or described in detail to avoid obscuring aspects of this application.

[0019] The block diagrams shown in the drawings are only functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0020] The flowcharts shown in the drawings are only illustrative and not necessarily include all the contents and operations / steps, nor are they necessarily executed in the described order. For example, some operations / steps can be decomposed, and some operations / steps can be combined or partially combined, so the actual execution order may change according to the actual situation.

[0021] Embodiment 1: A wind power prediction method for typhoon weather based on high-order coding and cluster partitioning, as Figure 1 shown, includes: Obtain meteorological prediction data within a future set time, and perform dynamic extraction and cluster partitioning on the meteorological prediction data; Obtain historical power data and meteorological data, and construct a data set based on the historical power data and meteorological data as training data; Extract the training data within each sub-cluster according to the cluster partitioning result, and perform modeling training according to the training data within the sub-cluster; Each sub-cluster loads its corresponding test set, uses the models that have completed modeling training to make predictions respectively, and summarizes the results of all sub-clusters to form the overall wind power prediction output within the set future time.

[0022] Specifically, this embodiment proposes a typhoon weather wind power prediction method based on high-order coding and cluster division. This method is divided into three stages, namely: data preprocessing and feature extraction; cluster dynamic division method during typhoon; sub-cluster power prediction based on HO-CAE-Transformer.

[0023] Stage 1: Data preprocessing and feature extraction During the typhoon passing through, there are obvious differences in the wind speeds of different wind circles such as the typhoon eye, typhoon edge, or typhoon center. Using the predicted wind speed at the anemometer tower as the feature input, it is difficult for existing prediction methods to reflect the spatial differences in the meteorological conditions of the wind farm during the typhoon passing through, which is one of the reasons for the increase in typhoon weather prediction errors. To illustrate the wind speed differences at adjacent spatial positions during the typhoon passing through, this embodiment analyzes the wind speed changes of three adjacent wind farms at the same moment when the typhoon "Talim" passed through in 2023. The true wind speed measurement data of SF1 - SF3 stations are as Figure 2 shown.

[0024] As Figure 2 can be seen, overall, the rapid ramp change trends of SF1 and SF2 are consistent, but the occurrence times are different, indicating that the wind speed change has a certain lag. For SF3 passing through the typhoon eye, its wind speed curve shows an M shape because the wind speed in the area near the typhoon eye is relatively low. Therefore, it can be seen that the wind speed during the typhoon passing through has strong spatio-temporal differences.

[0025] Furthermore, the Pearson correlation coefficient is further used to analyze the correlation between key meteorological factors and wind power under typhoon weather. The greater the absolute value of the correlation coefficient, the stronger the correlation between the two time series. The correlation coefficients of key meteorological factors are shown in Table 1.

[0026] Table 1 Correlation coefficients of key meteorological factors ; As can be seen from Table 1, there is a significant correlation between wind speed and power, followed by wind direction, air pressure, humidity, and documents. The correlation between other meteorological elements and power is relatively weak. Therefore, under the typhoon scenario, the average values of wind speed, wind direction, humidity, pressure, and temperature are used as feature inputs to characterize the spatial differences.

[0027] Stage 2: Cluster dynamic division method during typhoon During the passage of a typhoon, there are significant differences in wind speeds in different wind circles such as the typhoon eye, the typhoon edge, or the typhoon center. Therefore, at different moments of the same typhoon process, the similarity between the wind resource matrices of each station also fluctuates, and the static clustering method cannot capture this fluctuation. To more accurately identify the spatial structure and temporal characteristics of the power output of a wind farm under extreme typhoon weather conditions, the present invention proposes an improved dynamic clustering method (WCC-TCC) that combines multi-feature weighting and time-period comparison mechanisms. This method combines an adaptive measure of feature importance, an automatic evaluation of the clustering structure, and a comparison constraint optimization mechanism based on temporal consistency, significantly enhancing the physical rationality, robustness, and distinguishability of sub-cluster division.

[0028] (1) Construction of multi-feature weighted distance First, based on the correlation analysis results between wind power and each meteorological feature, a feature weight vector is assigned , and the distance between samples is weighted. The weighted Euclidean distance is defined as follows: ; where represents the weighted Euclidean distance between sample i and sample j, which is used to measure the similarity between two samples in the feature space; d is the total number of feature dimensions, that is, the number of features each sample has (such as wind speed, wind direction, air pressure, humidity, temperature, power, etc.); k is the index of the feature dimension; represents the embedding value of sample i on the k-th feature; represents the embedding value of sample j on the k-th feature; represents the corresponding weight. This weighting strategy ensures the dominant modeling ability for key variables such as wind speed and wind direction.

[0029] The weight can be determined by two mechanisms: one is to initialize it through linear discriminant analysis (LDA) or maximum information gain; the other is to use this weight as a learnable parameter and optimize it through backpropagation during the clustering training stage to ensure better adaptation to the specific characteristics of the samples.

[0030] (2) Adaptive selection mechanism for the number of clusters To avoid the problems of underfitting or overfitting that may be caused by manually setting the number of clusters, the present invention introduces the Bayesian information criterion (BIC) as an evaluation standard to select the optimal number of clusters in each sliding window . That is, within each 4-hour time period, the division quality index under multiple numbers of clusters is evaluated, and the optimal structure is automatically determined to improve the stability and generalization ability of the cluster model.

[0031] (3) Discriminant enhancement mechanism introducing time-period comparison constraints Considering that the wind field state during typhoon passage has significant temporal evolution, that is, the wind field behavior is relatively consistent within the same time period, while the difference between different time periods is significant. Therefore, the present invention constructs time period-level "positive sample pairs" and "negative sample pairs": Positive sample pairs: Different site samples within the same time period (4h); Negative sample pairs: Sample pairs from adjacent different sliding time periods.

[0032] In the embedding space, the contrastive loss function (Contrastive Loss) is adopted: ; Among them, P is the set of positive sample pairs, representing two sample pairs from the same time period (such as the same 4-hour window); N is the set of negative sample pairs, representing sample pairs from different time periods (such as the front and back two sliding windows); , , respectively represent the vector representations of samples , , in the embedding space (i.e., the clustering representation space); m is the set interval boundary; represents the Euclidean distance of the vector, which is used to measure the similarity between samples. This mechanism enhances the discriminant boundary stability of sub-cluster division by maximizing the cross-time discriminability and minimizing the within-time consistency.

[0033] (4)Overall process and output In summary, the process of the proposed WCC-TCC clustering method is as Figure 3 shown, including: Within each 4-hour sliding window, extract the original sample features, including hub wind speed, wind direction, air pressure, humidity, temperature, historical power, etc.; Construct a multi-feature weighted distance metric function, set the weighted coefficients according to the influence degree of meteorological variables on wind power, and perform initial clustering; Adopt evaluation criteria such as BIC to select the optimal number of clusters and dynamically adjust the number of sub-clusters; Introduce a time series contrast constraint optimization mechanism, construct positive sample pairs within the same time period and negative sample pairs in adjacent time periods, and introduce contrastive loss to strengthen the stability and discriminability of the clustering boundary; Output the dynamic sub-cluster division result for subsequent separate training and prediction of the model.

[0034] This clustering method combines the feature evaluation guided by physical importance, the adaptive structure determination, and the representation learning mechanism of time contrast enhancement, and has strong robustness, time series perception ability, and noise suppression ability, providing a structurally reasonable and temporally consistent basic support for wind power cluster prediction.

[0035] Stage 3: Sub-cluster Power Prediction Method Based on Higher-order Contractive Autoencoder and Transformer Under typhoon extreme weather conditions, meteorological variables such as wind speed and wind direction often exhibit sudden and drastic changes, resulting in strong non-linear fluctuations in the power output time series. This characteristic not only introduces high-noise interference but also poses significant challenges to the stability and generalization ability of traditional neural network models. Therefore, the present invention proposes to combine the Higher-order Contractive Autoencoder (HO-CAE) with the Transformer structure for wind power prediction modeling of each dynamic sub-cluster.

[0036] (1) SCADA Error Modeling and Compensation Mechanism In the process of wind power prediction modeling, key operating parameters such as wind speed, wind direction, temperature, humidity, and historical power collected by the SCADA system are important sources of model input, and their data quality directly determines the effectiveness of the prediction model. However, under extreme weather conditions such as typhoons, the remote communication link of the SCADA system is easily affected by factors such as wind and rain interference and electromagnetic signal attenuation, often resulting in problems such as information delay, short-term interruption, data drift, or mutation, further exacerbating the prediction error. Therefore, to improve the adaptability and robustness of the model to SCADA transmission errors, the present invention introduces a SCADA error modeling and compensation mechanism before the HO-CAE feature extraction module to dynamically detect, label, and correct abnormal data.

[0037] This mechanism mainly includes the following three steps: Firstly, multi-scale adaptive anomaly detection: Use the historical SCADA power sequence to construct local statistical features within the sliding window, such as mean, variance, skewness, kurtosis, etc., and combine the change point detection method to identify mutation values, drift segments, and communication interruption signals. In the detection process, a multi-scale sliding window combination strategy is adopted to adapt to error types with different time granularities, ensuring the comprehensiveness and accuracy of abnormal point identification.

[0038] Secondly, error type classification and compensation scheme matching: According to the anomaly detection results, the abnormal data is divided into three main types: "short-term missing", "instantaneous jump", and "continuous drift". Among them, short-term missing data is filled by time series linear interpolation and similar feature backfilling strategy of adjacent wind farms; instantaneous jump data is suppressed transiently by historical trend reconstruction; while for continuous drift segments, the main trend signal is extracted based on empirical mode decomposition, a steady-state reference curve is constructed and offset correction is performed through least squares registration.

[0039] Finally, the error weight constraint at the embedding level is implemented: the error annotation result is passed to the HO-CAE input layer, a position weight mask mechanism is introduced, and a dynamic weight attenuation factor is set for the error region to reduce its interference with the encoder training process. Specifically, the following error correction term is added to the encoder’s embedding representation: (3); in, is the original embedding feature, is the reference feature after smooth interpolation, is the error confidence weight at time step t, which is adaptively adjusted based on the error intensity and type. This mechanism can dynamically weaken the impact of error information during model training and improve the quality of feature representation.

[0040] Through the coordinated cooperation of the above-mentioned error detection, type classification and embedded correction mechanism, the present invention realizes intelligent fault tolerance and correction of SCADA signal transmission disturbances while ensuring the quality of model input data, further improving the model stability and prediction accuracy under extreme meteorological conditions.

[0041] (2) High-order contraction autoencoder feature extraction Based on the traditional denoising autoencoder, the high-order contraction autoencoder effectively limits the sensitivity of the encoder to small perturbations in the input space by introducing a high-order derivative regularization term of the input to the hidden representation in the loss function, thereby obtaining a more stable, smooth and robust feature representation. This structure is particularly suitable for modeling the drastic fluctuations of wind speed data and uncertainty of power response during typhoons. Its structure and principle are as follows Figure 4 and Figure 5 shown.

[0042] Assume that the time series of wind power and meteorological characteristics is: ; In the formula, is the power and meteorological characteristics at time t, is a d-dimensional data space. Due to measurement noise, data missing and other issues, the actual observed data is: ; In the formula, is the actual observed value of power and meteorological characteristics at time t; is the noise term, which is usually assumed to obey a Gaussian distribution, i.e. .

[0043] The high-order contraction autoencoder consists of an encoder and a decoder, where the encoder is: ; Among them, is the encoder function that maps the input features to a low-dimensional embedding space; is the weight matrix of the encoder; is the bias vector of the encoder; is the non-linear activation function; is the embedded representation (latent space features) output by the encoder.

[0044] The decoder is: (7); Among them, is the output at the t-th time step of the model reconstruction, which is the noise reduction and restoration of the original features; is the decoder function that restores the embedded features to the original feature space; is the weight matrix of the decoder; is the bias term of the decoder; is the non-linear activation function.

[0045] The model training objective is to minimize the following high-order regularization loss function: ; Among them, is the total loss function of HO-CAE. The first term is the reconstruction error, and the second term is the high-order shrinkage regularization term. is the regularization weight hyperparameter. represents the second-order derivative of the hidden layer with respect to the input (Hessian matrix). represents the Frobenius norm. This regularization term can effectively limit the response of the model to input perturbations, thereby suppressing the impact of wind speed mutations on prediction stability.

[0046] (3) Transformer Modeling and Power Prediction Take the embedded features processed by the high-order shrinkage autoencoder as the input of the Transformer model for time series modeling. The Transformer structure includes input embedding, positional encoding, multi-head self-attention mechanism, and feed-forward network. Its core computational mechanism is as follows: Embedding layer: ; In the formula, represents the input embedding vector at time step t, which is used to transform the noise-reduced input data into a high-dimensional feature space that the Transformer can process; is the weight matrix of the embedding layer; is the bias term.

[0047] Position encoding (2i-th and 2i + 1-th dimensions): (10); Multi-head attention mechanism: ; In the formula, are the query, key, and value matrices respectively; is the dimension of the key vector for scaling; is the normalization function that converts a numerical vector into a probability distribution such that the sum of all its elements is 1. The schematic diagram of the attention mechanism is as shown in Figure 6 shown.

[0048] The Transformer structure is as shown in Figure 7 shown. Its output sequence is y1, y2, …, yT. The power prediction value at each time step is calculated through a fully connected layer: (12); where, is the wind power output value at the (t + 1)-th moment predicted by the model; is the hidden state output vector of the Transformer model at time step t, which contains the comprehensive feature representation of historical time series information; is the weight matrix of the output layer (fully connected layer) for mapping features to the power space; is the bias term of the output layer for adjusting the offset of the output value.

[0049] The prediction training objective function is the mean squared error: ; where, is the loss function of the prediction task for measuring the error between the model output power and the true power; T is the total number of time steps, that is, the length of the prediction sequence; is the true wind power value at the (t + 1)-th moment; is the wind power value predicted by the model corresponding to the (t + 1)-th moment.

[0050] (4) Model training and integrated prediction output This method takes each sub-cluster divided by the improved clustering method as an independent modeling unit, and trains a set of HO-CAE + Transformer network structure models for each sub-cluster respectively. Each model has anti-noise stability in the input space and time-dependent learning ability in the modeling process, jointly achieving high-precision prediction at the sub-cluster level. Finally, the prediction results of each sub-cluster are summarized to form the cluster power prediction curve of the entire wind farm for the next 24 hours.

[0051] The overall modeling process of this stage is as Figure 8 shown, including three sub-modules: embedded space construction, Transformer time series learning, and power output mapping. The three cooperate to form the main framework of wind power modeling.

[0052] After the above three stages, more accurate wind power prediction results under typhoon weather can be obtained.

[0053] Example 2: This example provides a case analysis to verify the effectiveness of the method proposed in the present invention. The original dataset includes the daily forecast values of 8 meteorological factors and the measured wind power values of numerical weather forecasts in Guangdong from 2021 to 2022, with a time resolution of 15 minutes, including 5 typhoon events. The data during the typhoon passage in Guangdong Province under the influence of Typhoon "Talim" in July 2023 is used for model verification.

[0054] Cluster the cluster containing 20 stations into multiple sub-clusters. If the number of sub-clusters is too large, it may occur that a single station forms a separate class, thus losing the integrity of the cluster; if the number of sub-clusters is too small, it may lead to a huge difference in the number of stations in different sub-clusters, thus losing balance. Based on the principle of minimizing the distance within the cluster and maximizing the distance between clusters, the number of clustering clusters in this example is determined to be 3, and the cluster containing 10 wind farms is divided into 3 sub-clusters. The result of a dynamic division of the cluster during a certain operation is as Figure 9 shown. The 6 sub-graphs represent 6 groups of 4-hour time periods in sequence, and the five-pointed star, triangle, and circle correspond to the 3 sub-clusters after division respectively. It can be seen that the stations in the same sub-cluster are basically in similar geographical locations.

[0055] Figure 10 and Figure 11 give the comparison diagrams of the prediction results under normal scenarios and generator tripping scenarios. In addition to the method proposed in the present invention, two control models are set up. From Figure 10 it can be seen that in the normal situation without strong wind generator tripping events, the prediction effects of the three methods are not much different. Among them, the method proposed in the present invention has the best effect, the Transformer method ranks second, and the LSTM method has the worst effect. Figure 11Among them, the problem of inaccurate prediction during the wind farm restoration operation period has been effectively improved, and the proposed solution can effectively improve the prediction accuracy.

[0056] Table 2 shows the comparison of prediction errors. It can be seen that the solution proposed by the present invention has higher accuracy.

[0057] Table 2 Comparison of Prediction Errors ; After considering the specification and practicing the embodiments disclosed herein, those skilled in the art will readily conceive of other embodiments of the present application. The present application is intended to cover any variations, uses, or adaptations of the present application, which follow the general principles of the present application and include well-known common knowledge or conventional technical means in the technical field not disclosed in the present application.

[0058] It should be understood that the present application is not limited to the exact structures described above and shown in the drawings, and various modifications and changes can be made without departing from its scope. The scope of the present application is only limited by the appended claims.

Claims

1. A typhoon weather wind power prediction method based on high-order coding and cluster division, characterized in that include: Obtain weather forecast data within a set time in the future, and dynamically extract and cluster the weather forecast data; Obtain historical power data and meteorological data, and construct a data set based on the historical power data and meteorological data as training data; Extract training data in each sub-cluster according to the cluster division results, and perform modeling training based on the training data in the sub-cluster; Each subcluster loads its own corresponding test set and uses the model that has completed modeling training to make predictions separately. The results of all subclusters are aggregated to form the overall wind power prediction output within the future set time.

2. The typhoon weather wind power prediction method based on high-order coding and cluster division according to claim 1, wherein The key meteorological factors in the meteorological data are determined by the Pearson correlation coefficient, the correlation between the meteorological factors in the meteorological data and power is analyzed, and the first several meteorological factors with the largest absolute value of the correlation coefficient are taken as the key meteorological factors.

3. The typhoon weather wind power prediction method based on high-order coding and cluster division according to claim 2, wherein The weather forecast data obtained within a set time in the future includes wind speed, wind direction, air pressure, temperature and humidity data at 100 meters, 30 meters and 10 meters above the ground.

4. The typhoon weather wind power prediction method based on high-order coding and cluster division according to any one of claims 1-3, characterized in that, If the sample size of the meteorological forecast data is less than the minimum required sample size predicted by the model, the small sample expansion method is used to expand the meteorological forecast data.

5. The typhoon weather wind power prediction method based on high-order coding and cluster division according to claim 1, characterized in that, The historical power data is obtained through the SCADA system. If the historical power data obtained through the SCADA system has short-term missing, jump or drift errors, it is identified and corrected through sliding window statistics, change point detection and anomaly classification mechanisms, and the quality of the historical power data is improved through error confidence and weighting strategies.

6. The typhoon weather wind power prediction method based on high-order coding and cluster division according to claim 1, characterized in that The dynamic extraction and clustering of weather forecast data is specifically as follows: The meteorological forecast data is divided into several groups of data sets for different time periods, and the time range in each data set is consistent; a dynamic clustering algorithm is used to divide the cluster structure in each data set to obtain the sub-cluster division results of the time period.

7. The typhoon weather wind power prediction method based on high-order coding and cluster division according to claim 6, characterized in that The dynamic clustering algorithm is an improved dynamic clustering algorithm that integrates multi-feature weighting and time period comparison mechanism.

8. The typhoon weather wind power prediction method based on high-order coding and cluster division according to claim 7, characterized in that, The process of the improved dynamic clustering algorithm is as follows: In each sliding window, the original sample features are extracted, a multi-feature weighted distance measurement function is constructed, and the weighting coefficient is set according to the impact of meteorological variables on wind power to perform initial clustering; The BIC evaluation criteria is used to select the optimal number of clusters and dynamically adjust the number of subclusters; A time series contrast constraint optimization mechanism is introduced to construct positive sample pairs in the same time period and negative sample pairs in adjacent time periods, and contrast loss is introduced to enhance the stability and discriminability of cluster boundaries. Output dynamic sub-clustering results for subsequent model training and prediction.

9. The typhoon weather wind power prediction method based on high-order coding and cluster division according to claim 1, characterized in that The modeling training is based on high-order contraction autoencoder and Transformer model.

10. The typhoon weather wind power prediction method based on high-order coding and cluster division according to claim 9, characterized in that, The high-order contractive autoencoder introduces a high-order derivative regularization term of the input to the hidden representation in the loss function to limit the sensitivity of the encoder to small perturbations in the input space.

Images