Wind power prediction method, system and equipment based on multi-site dual-space-time network and medium

Abstract

The invention relates to a wind power prediction method, system and device based on a multi-site dual-space-time network and a medium. The method comprises the following steps: acquiring historical data of a multi-site wind power plant, and preprocessing the historical data of the multi-site wind power plant to obtain preprocessed data; performing feature selection processing on the preprocessed data to obtain an optimized feature set; inputting the optimized feature set into a preset multi-site dual-time-space network to obtain a first future predicted value of each site and a second future predicted value of each site; the multi-site dual-time-space network comprises a long short-term memory network branch and a time convolution network branch; and carrying out weighted fusion on the first future predicted value and the second future predicted value to obtain a multi-station wind power predicted value. By adopting the method, the precision and robustness of multi-station wind power prediction can be improved, the clustering development requirement of the wind power plant can be better adapted, and powerful support is provided for power grid dispatching strategy optimization and safe and stable operation.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of wind power prediction, and particularly relates to a wind power prediction method, system, device and medium based on a multi-site double spatio-temporal network. BACKGROUND

[0002] As a clean and renewable energy, the development and utilization scale of wind energy continues to expand, and wind power prediction technology has become a key to ensuring stable power grid scheduling and improving wind power consumption efficiency. Current wind power prediction mainly relies on physical methods, statistical methods and deep learning methods. The physical method has high calculation cost and is not suitable for ultra-short-term prediction. The statistical method is difficult to capture the nonlinear relationship of data. Although the deep learning method has improved, it is mainly aimed at a single wind farm site, which is difficult to adapt to the development needs of wind farm clusters.

[0003] In the traditional technology, the geographical correlation of multi-site wind power data preprocessing is often ignored, and only abnormal values are removed by a fixed threshold and missing values are completed by simple interpolation. Feature selection relies on artificial experience or a single indicator, which cannot accurately select features strongly related to power. The prediction model mainly uses a single network branch, which is difficult to capture short-term and long-term time series features at the same time, and does not effectively mine the spatial correlation of multi-site data.

[0004] However, the current method has the following core problems: first, the multi-site data preprocessing does not consider geographical similarity, and the physical rationality of the data after repair is insufficient; second, the feature selection lacks quantitative evaluation of the nonlinear relationship between features and power, and redundant features affect the efficiency of the model; third, the single network branch cannot simultaneously model short-term and long-term time series, and the overall prediction accuracy and robustness of multi-site data cannot meet the needs of power grid for regional wind power output perception. SUMMARY

[0005] Therefore, it is necessary to provide a wind power prediction method, system, device and medium based on a multi-site double spatio-temporal network to improve the prediction accuracy and robustness of multi-site wind power.

[0006] In a first aspect, the application provides a wind power prediction method based on a multi-site double spatio-temporal network, comprising:

[0007] Obtaining multi-site wind farm historical data and preprocessing the multi-site wind farm historical data to obtain preprocessed data; the multi-site wind farm historical data includes the cabin direction, blade pitch angle, wind speed, relative wind direction, wind power, environmental temperature and cabin internal temperature of each site;

[0008] Performing feature selection processing on the preprocessed data to obtain an optimized feature set; the optimized feature set is a set of key data features related to wind power in the multi-site wind farm;

[0009] input the optimized feature set into a preset multi-site double space-time network to obtain first future prediction values of each site and second future prediction values of each site; the multi-site double space-time network includes a long short-term memory network branch and a time convolution network branch;

[0010] The first future prediction value and the second future prediction value are weighted and fused to obtain a multi-site wind power prediction value.

[0011] In one of the embodiments, the multi-site wind farm historical data is preprocessed to obtain preprocessed data, including:

[0012] The cabin direction and the blade pitch angle in the obtained multi-site wind farm historical data are subjected to outlier identification to obtain initial abnormal data; the initial abnormal data includes abnormal value data of the cabin direction exceeding a preset cabin direction range and abnormal value data of the blade pitch angle exceeding a preset blade pitch angle range;

[0013] The following formula is used to correct the abnormal values in the initial abnormal data to obtain data after correction of the abnormal values:

[0014]

[0015] wherein, represents the corrected cabin direction or blade pitch angle value, represents the minimum reasonable value of the cabin direction or blade pitch angle, represents the maximum reasonable value of the cabin direction or blade pitch angle, represents the original value of the cabin direction or blade pitch angle in the initial abnormal data;

[0016] The missing data of the cabin direction and the blade pitch angle in the data after correction of the abnormal values is completed to obtain data after feature completion;

[0017] The wind speed and the relative wind direction in the data after feature completion are subjected to abnormal value marking, and the abnormal values of the wind speed exceeding a preset wind speed range and the relative wind direction exceeding a preset relative wind direction range are set to null values to obtain the preprocessed data.

[0018] In one of the embodiments, obtaining the preprocessed data further includes:

[0019] The following formula is used to perform angle normalization processing on the cabin direction to obtain cabin direction normalized data:

[0020]

[0021] wherein, represents the normalized cabin direction, represents the original value of the cabin direction in the preprocessed data, represents a modulo operation;

[0022] The absolute wind direction is obtained by comparing the relative wind direction in the cabin orientation normalized data with the preprocessed data;

[0023] Normalize the absolute wind direction to obtain normalized wind direction data;

[0024] Based on the geographical coordinates of each station, missing values ​​in the wind direction normalization data are repaired to obtain the repaired data.

[0025] Outlier handling is performed on the repaired data to obtain the final preprocessed data; the final preprocessed data includes the complete repaired data.

[0026] In one embodiment, feature selection processing is performed on the preprocessed data to obtain an optimized feature set, including:

[0027] The preprocessed data is binned and discretized to obtain discretized feature data;

[0028] For the discretized feature data, a two-dimensional joint histogram of each feature and wind power is constructed, and the joint probability distribution, the marginal probability distribution of each feature, and the marginal probability distribution of wind power are calculated to obtain the probability distribution data.

[0029] Based on the probability distribution data, the entropy value is calculated using the following formula:

[0030]

[0031] in, Representing a single feature entropy, Representation of features The set of all discrete values, Representation of features With wind power joint entropy, Represents the set of all discrete values ​​of wind power. Representation of features Marginal probability distribution Represents the joint probability distribution;

[0032] Based on the entropy calculation results, the univariate mutual information data is obtained using the following formula:

[0033]

[0034] in, Representation of features With wind power mutual information, Representing a single feature entropy, Entropy represents wind power output. Representation of features With wind power The joint entropy;

[0035] Based on univariate mutual information data, the joint mutual information between different feature combinations and wind power is calculated to obtain an optimized feature set.

[0036] In one embodiment, the first future prediction value for each site is obtained using the following method:

[0037] The optimized feature set is normalized to obtain normalized feature data;

[0038] The normalized feature data is input into a two-layer long short-term memory network to obtain the first predicted value; the first predicted value is the output of the long short-term memory network at the last time step.

[0039] The first predicted value is input into the dual-path attention mechanism network to obtain the site feature association features;

[0040] Based on the correlation of site characteristics, the first future prediction value of each site is obtained.

[0041] In one embodiment, the second future prediction value is obtained using the following method:

[0042] The optimized feature set is normalized to obtain normalized feature data;

[0043] Normalized feature data is input into a multi-layer temporal convolutional network to obtain a second predicted value; the second predicted value is the output of the last time step of the temporal convolutional network.

[0044] The second predicted value is input into the dual-path attention mechanism network to obtain the second site feature association feature;

[0045] Based on the correlation of site characteristics, the second future prediction value of each site is obtained.

[0046] In one embodiment, the first future prediction value and the second future prediction value are weighted and fused to obtain the multi-site wind power prediction value, including:

[0047] Based on the first and second future forecasts, the multi-site wind power forecast is obtained using the following formula:

[0048]

[0049] in, For the first The predicted fusion value for each site, These are the weighting coefficients. For the first The first future forecast value for each site, For the first The second future forecast value for each site.

[0050] Secondly, this application also provides a wind power forecasting system based on a multi-site dual-temporal network, comprising:

[0051] The data acquisition and preprocessing module is used to acquire historical data from multiple wind farm sites and preprocess the historical data to obtain preprocessed data. The historical data from multiple wind farm sites includes the nacelle orientation, blade pitch angle, wind speed, relative wind direction, wind power, ambient temperature, and nacelle internal temperature for each site.

[0052] The feature selection module is used to perform feature selection processing on the preprocessed data to obtain an optimized feature set; the optimized feature set is a set of key data features related to wind power in a multi-site wind farm.

[0053] The preliminary prediction module is used to input the optimized feature set into a preset multi-site dual-temporal network to obtain the first future prediction value and the second future prediction value of each site; the multi-site dual-temporal network includes a long short-term memory network branch and a temporal convolutional network branch.

[0054] The wind power prediction module is used to weight and fuse the first future prediction value and the second future prediction value to obtain the wind power prediction value for multiple sites.

[0055] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect.

[0056] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect.

[0057] The aforementioned wind power forecasting method, system, equipment, and medium based on a multi-site dual-temporal network, through data preprocessing that corrects outliers and fills in missing values, and combining the geographical correlation of wind farm sites to repair data anomalies and missing data, ensures the integrity and physical rationality of the input data. Utilizing feature selection based on mutual information, it accurately selects key features for wind power prediction, reducing model redundancy and training complexity. The dual-branch architecture of the multi-site dual-temporal network efficiently captures the short-term dependencies and long-term trends of wind power data, and, in conjunction with a dual-path attention mechanism, mines multi-site spatial correlations and multi-feature dynamic correlations. Adaptive weighted fusion integrates the advantages of both branches, effectively improving the accuracy and robustness of multi-site wind power prediction, providing a reliable regional wind power output reference for grid dispatch, and contributing to wind power consumption and stable grid operation. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 This is a schematic diagram of a wind power prediction method based on a multi-site dual-temporal network in one embodiment;

[0060] Figure 2 This is a flowchart illustrating a multi-site wind power prediction method in one embodiment;

[0061] Figure 3 This is a schematic diagram of the structure of a wind power forecasting system based on a multi-site dual-temporal network in one embodiment;

[0062] Figure 4 This is a schematic diagram of a computer device in one embodiment. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0064] In one embodiment, such as Figure 1 As shown, a wind power forecasting method based on a multi-site dual-spatiotemporal network is provided. This embodiment illustrates the application of this method to a wind power forecasting terminal (hereinafter referred to as the terminal). It can be understood that this method can also be applied to a server, and can also be applied to a system including a terminal and a server, and implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0065] S1: Obtain historical data of wind farms at multiple sites and preprocess the historical data of wind farms at multiple sites to obtain preprocessed data.

[0066] For example, the wind power forecasting terminal first establishes a communication connection with the data acquisition system of each wind farm to obtain historical data from multiple wind farm sites. This historical data covers the core operational and environmental characteristics of all wind farm sites, including the nacelle orientation, blade pitch angle, wind speed, relative wind direction, wind power, ambient temperature, and nacelle internal temperature for each site. After acquiring the data, the wind power forecasting terminal preprocesses the data, eliminating data noise and incompleteness through operations such as outlier identification, correction, and missing value completion, ultimately obtaining preprocessed data. Among these, the nacelle orientation reflects the orientation angle of the wind turbine nacelle; the blade pitch angle affects the wind turbine's wind energy capture efficiency and is constrained by the unit's mechanical structure; wind speed is a key indicator for measuring wind energy resource intensity; relative wind direction is the wind direction angle relative to the nacelle and needs to be converted to wind direction in a geographic coordinate system based on the nacelle orientation; and wind power reflects the actual power generation output of the unit.

[0067] S2 performs feature selection processing on the preprocessed data to obtain an optimized feature set.

[0068] For example, after acquiring preprocessed data, the wind power forecasting terminal performs feature selection on the preprocessed data to obtain an optimized feature set. Feature selection is used to filter out key features strongly correlated with wind power from multiple features contained in the preprocessed data, eliminating redundant and irrelevant features to reduce the training complexity of subsequent models and minimize the risk of overfitting. The optimized feature set, as the filtering result, is essentially a set of key data features related to wind power in multi-site wind farms. These features can reflect the core factors affecting wind power changes to the greatest extent, such as angle-normalized absolute wind direction and corrected wind speed, providing high-information-density input features for the multi-site dual-temporal-spatial network.

[0069] S3, input the optimized feature set into the preset multi-site dual spatiotemporal network to obtain the first future prediction value and the second future prediction value of each site.

[0070] Specifically, the wind power forecasting terminal inputs the optimized feature set into a pre-defined multi-site dual-temporal network. This network is a deep learning architecture specifically designed for multi-site wind power forecasting, comprising a Long Short-Term Memory (LSTM) branch and a Temporal Convolutional Network (TCN) branch. The LSTM branch, through a forget gate, input gate, and output gate gating mechanism, selectively retains or discards information from time-series data, effectively capturing short-term dependencies and adapting to the rapid fluctuations in wind power data over short periods. The TCN branch, through a dilated causal convolution structure, expands the receptive field of the convolution without increasing computation, effectively extracting long-term dependencies and adapting to the long-term variation patterns of wind power data. The terminal computes in parallel through the two branches, outputting a first future prediction value and a second future prediction value for each site. The first future prediction value reflects the power change trend over a short period, while the second future prediction value reflects the power change trend over a longer period. Both the first and second future prediction values ​​can be short-term wind power predictions.

[0071] S4, weighted and fused the first future forecast value and the second future forecast value to obtain the wind power forecast value for multiple sites.

[0072] For example, the wind power forecasting terminal weights and fuses the first and second future forecasts for each site, integrating the forecasting advantages of both branches to improve the final forecast accuracy. The first future forecast is better at capturing short-term data fluctuations, while the second future forecast is better at grasping long-term data trends. Weighted fusion enables the complementary advantages of both. The terminal determines the weighting coefficients through a preset weighting calculation strategy, such as minimizing the forecast error based on validation set data, and then calculates the final forecast value for each site using the fusion formula. Finally, it integrates the final forecast values ​​of all sites to obtain multi-site wind power forecast values, providing a regional-level wind power output forecasting reference for grid dispatch.

[0073] The aforementioned wind power prediction method based on a multi-site dual-spatiotemporal network effectively improves the quality and physical rationality of input data through multi-dimensional data preprocessing, combining unit operation boundary constraints to correct outliers, and utilizing geographical similarity to fill in missing values. Mutual information feature selection is used to accurately filter key features, reduce redundant information interference, and optimize model input efficiency. The LSTM and TCN branches of the dual-spatiotemporal network efficiently capture short-term and long-term temporal dependencies, respectively, and are combined with a dual-path attention mechanism to mine multi-site spatial correlations and dynamic feature correlations, achieving comprehensive multi-scale spatiotemporal feature modeling. Finally, adaptive weighted fusion integrates the advantages of both branches to improve the accuracy and robustness of multi-site wind power prediction.

[0074] In an optional embodiment, historical data from multiple wind farm sites is preprocessed to obtain preprocessed data, including the following steps:

[0075] S11, outlier identification is performed on the nacelle orientation and blade pitch angle in the acquired historical data of wind farms from multiple sites to obtain initial outlier data.

[0076] The initial abnormal data includes outliers in nacelle orientation and blade pitch angle that exceed the preset range. For example, the wind power forecasting terminal prioritizes identifying outliers in two types of features—nacelle orientation and blade pitch angle—based on acquired historical data from multiple wind farm sites. The preset nacelle orientation range is a reasonable interval determined based on the mechanical rotation limits of the wind turbine; nacelle orientation data exceeding this range cannot be achieved through normal turbine operation and is therefore identified as anomalies. Similarly, the preset blade pitch angle range is a safe operating range set based on the structural strength of the blades and wind energy capture requirements; blade pitch angle data exceeding this range can lead to turbine operational risks or a sharp drop in power generation efficiency and is also identified as anomalies. The wind power forecasting terminal compares the nacelle orientation and blade pitch angle data for each site with the corresponding preset ranges, filters out data exceeding the ranges, and integrates them to form the initial abnormal data.

[0077] S12, use the following formula to correct outliers in the initial outlier data to obtain the outlier-corrected data:

[0078]

[0079] in, This indicates the corrected nacelle orientation or blade pitch angle value. This indicates the minimum reasonable value for the nacelle direction or blade pitch angle. This indicates the maximum reasonable value for the nacelle direction or blade pitch angle. This indicates the original values ​​of the nacelle orientation or blade pitch angle in the initial abnormal data.

[0080] In the above boundary constraint correction formula, This indicates the corrected nacelle orientation or blade pitch angle value, used to ensure that the corrected data is within a reasonable operating range. This indicates the minimum reasonable value for the nacelle direction or blade pitch angle. This indicates the maximum reasonable value of the nacelle direction or blade pitch angle, both of which are pre-set based on the design parameters and operating specifications of the wind turbine. This represents the initial value of the nacelle orientation or the initial value of the blade pitch angle in the initial anomaly data. The wind power forecasting terminal uses this formula to calculate values ​​less than... Replace the outlier with greater than Replace the outlier with Normal data remains unchanged, and the data after anomaly correction is obtained.

[0081] S13, complete the missing data of nacelle direction and blade pitch angle in the outlier-corrected data to obtain the feature-completed data.

[0082] Specifically, after receiving the outlier-corrected data, the wind power forecasting terminal uses a data integrity verification module to identify missing data regarding nacelle orientation and blade pitch angle. Missing data typically originates from temporary power outages in data acquisition equipment, sensor malfunctions, or communication link interruptions. Failure to complete this data will lead to data gaps in subsequent time-series modeling. The terminal uses linear interpolation for data completion. The technical principle is to construct a linear function to fit the missing data based on normal data from adjacent time points. For example, assuming the nacelle orientation data at a certain site at time t-1 and t+1 are a and b respectively, the missing data at time t is obtained through... Calculations show that this method can maximize the correlation between computational efficiency and the temporal variation trend of the data. Through completion operations, the terminal completes the missing data for nacelle orientation and blade pitch angle, forming data with complete features.

[0083] S14, mark outliers in the wind speed and relative wind direction in the feature-completed data, set outliers where the wind speed exceeds the preset wind speed range and the relative wind direction exceeds the preset relative wind direction range to null values, and obtain preprocessed data.

[0084] For example, the wind power forecasting terminal further marks outliers for wind speed and relative wind direction based on the data after feature completion. The preset wind speed range is determined based on the operating characteristics of the wind turbine, with the lower limit being the turbine's start-up wind speed and the upper limit being the turbine's cut-out wind speed. The wind speed range is... Wind speed data outside this range is considered abnormal; the preset relative wind direction range is... This range covers all possible wind direction angles. Relative wind direction data outside this range are considered abnormal due to acquisition errors or numerical overflow during data transmission. The wind power forecasting terminal sets these abnormal values ​​to null, completes all preprocessing steps, and finally obtains the preprocessed data. Specifically, wind speeds below the turbine's startup speed will prevent the turbine from rotating, while wind speeds above the turbine's cut-off speed will trigger a shutdown for safety. Null values ​​are marked as missing in the data storage.

[0085] In an optional embodiment, obtaining the preprocessed data further includes the following steps:

[0086] S21, Using the following formula, the cabin orientation is normalized to obtain normalized cabin orientation data:

[0087]

[0088] in, This indicates the normalized cabin orientation. This represents the original cabin orientation value in the preprocessed data. This indicates a modulo operation.

[0089] Specifically, the wind power forecasting terminal performs angle normalization processing on the nacelle orientation in the preprocessed data to eliminate modeling interference caused by angle periodicity. Angle periodicity refers to the fact that angular data such as the nacelle orientation can lead to model misjudgments due to numerical cycles. In the above normalization formula, This represents the normalized cabin orientation, with its value range strictly constrained within... This ensures the physical continuity of the angle. This indicates the original value of the nacelle orientation in the preprocessed data, which may change due to continuous rotation of the unit. Accumulation of values ​​in multiples of integers. This represents a modulo operation used to map any raw angle value to a specified interval. The terminal uses this formula to normalize the cabin orientation, obtaining normalized cabin orientation data.

[0090] S22, based on the relative wind direction in the cabin orientation normalized data and preprocessed data, the absolute wind direction is obtained.

[0091] Specifically, the wind power forecasting terminal uses the nacelle orientation normalized data and the relative wind direction in the preprocessed data to calculate the absolute wind direction in the geographic coordinate system. The relative wind direction only reflects the angular relationship between the wind direction and the nacelle; for example, the nacelle orientation... At that time, relative wind direction This represents the actual wind direction. However, the relative wind direction cannot be directly used for spatial correlation analysis between multiple sites; while the absolute wind direction can reflect the true geographical wind direction and is a key feature for mining spatial correlation in multi-site wind power forecasting. During the calculation process, the terminal directly adds the relative wind direction of each site to the corresponding nacelle direction normalized data to obtain the absolute wind direction of that site, realizing the wind direction conversion from the relative coordinate system to the geographical coordinate system.

[0092] S23, normalize the absolute wind direction to obtain normalized wind direction data.

[0093] After obtaining the absolute wind direction, the wind power forecasting terminal needs to perform normalization processing to ensure the physical consistency of the wind direction data. Since the absolute wind direction is the sum of the normalized data of the relative wind direction and the nacelle direction, it may show discrepancies. The extent of the situation, such as relative wind direction Normalized data with cabin orientation Adding them together gives The wind direction needs to be normalized using the same modulo operation formula as the relative wind direction. By normalizing the absolute wind direction, the terminal constrains the absolute wind direction of all stations within the range, avoiding model training bias caused by angle overflow, and obtains wind direction normalized data.

[0094] S24. Based on the geographical coordinates of each station, the missing values ​​in the wind direction normalization data are repaired to obtain the repaired data.

[0095] Specifically, the wind power forecasting terminal repairs missing values ​​in the wind direction normalization data based on the geographical coordinates of each station. First, the terminal acquires the set of geographical coordinates for all stations. By calculating the Euclidean distance between any two stations to reflect their spatial proximity, stations with close geographical distances are grouped into the same meteorological similarity cluster. This means that geographically close stations are influenced by the same atmospheric circulation system, and their meteorological characteristics such as wind speed and direction have strong spatial correlation. Then, the terminal uses data from stations with normal data within the same cluster to calculate the mean and complete the missing values ​​for the missing stations. For example, if the wind direction data for station A is missing, the mean wind direction value of other normal stations in its cluster is the repaired value for station A. This method ensures that the repaired data conforms to the consistency of regional meteorological characteristics. The repaired data is obtained by repairing the missing values ​​in the wind direction normalization data through the terminal.

[0096] S25, perform outlier processing on the repaired data to obtain the final preprocessed data.

[0097] Specifically, the wind power forecasting terminal performs secondary outlier processing on the repaired data to improve data quality. The processing covers all features in the repaired data, including wind speed, absolute wind direction, nacelle direction, blade pitch angle, wind power, relative wind direction, ambient temperature, and nacelle internal temperature. After processing outliers in nacelle direction and blade pitch angle, based on the repaired data, missing values ​​are repaired for wind speed, absolute wind direction, ambient temperature, and nacelle internal temperature (significantly affected by the environment). Specifically, outliers in ambient temperature, nacelle internal temperature, and relative wind direction are set to null. A hierarchical clustering imputation method based on geographical location is used to fill in missing data from geographically proximate stations by leveraging meteorological similarities between them, improving the physical plausibility and spatial consistency of the imputation results. Finally, the terminal verifies and repairs the wind power feature after processing all other input features, resulting in the final preprocessed data. For example, the terminal sets a reasonable fluctuation range for each feature and filters out extreme outliers that exceed the fluctuation range. These outliers usually originate from sudden sensor failures or extreme weather interference. If they are retained, the model will learn incorrect patterns. The terminal sets these extreme outliers to null and uses the intra-cluster mean completion method to repair them, finally obtaining preprocessed data without outliers or missing values. The complete repaired data consists of all the feature data obtained after outlier processing and feature completion of all the repaired data.

[0098] In an optional embodiment, feature selection processing is performed on the preprocessed data to obtain an optimized feature set, including the following steps:

[0099] S31, the preprocessed data is binned and discretized to obtain discretized feature data.

[0100] For example, the wind power forecasting terminal performs binning discretization on continuous features in the preprocessed data. Continuous features include wind speed, absolute wind direction, and wind power. These features have a wide range of values ​​and exhibit numerous subtle differences; directly using them for mutual information calculation would lead to a surge in computational load and susceptibility to noise interference. Binning discretization, by dividing the range of continuous features into several non-overlapping intervals and mapping different values ​​within each interval to the same discrete value, effectively reduces data complexity and highlights data distribution patterns. The terminal employs a binning strategy based on data distribution (such as equal-frequency binning) to ensure a relatively balanced amount of data within each interval, avoiding information loss due to improper interval division, ultimately obtaining discretized feature data. The binning strategy can be equal-frequency binning.

[0101] S32. For the discretized feature data, construct a two-dimensional joint histogram of each feature and wind power, calculate the joint probability distribution and the marginal probability distribution of each feature and the marginal probability distribution of wind power, and obtain the probability distribution data.

[0102] For example, the wind power forecasting terminal constructs a two-dimensional joint histogram for each feature and wind power for the discrete feature data. The two-dimensional joint histogram uses the discrete values ​​of the feature on the horizontal axis and the discrete values ​​of wind power on the vertical axis. It counts the frequency of each feature value and power value combination in historical data, and then divides the frequency by the total number of samples to convert it into a joint probability distribution. , This represents the probability that feature X takes the value x and wind power Y takes the value y. Simultaneously, the terminal counts the frequency of occurrence of all discrete values ​​for each feature, converting this into a marginal probability distribution for the feature. The frequency of occurrence of all discrete values ​​of wind power is statistically analyzed and converted into the marginal probability distribution of wind power. The terminal combines the probability distribution data to form the probability distribution data, which is the core input for subsequent entropy and mutual information calculations.

[0103] S33, Based on probability distribution data, the entropy value is calculated using the following formula:

[0104]

[0105] in, Representing a single feature entropy, Representation of features The set of all discrete values, Representation of features With wind power joint entropy, Represents the set of all discrete values ​​of wind power. Representation of features Marginal probability distribution This represents the joint probability distribution.

[0106] Specifically, the terminal calculates the entropy values ​​of all features using a preset formula, obtaining the entropy calculation result. In the aforementioned entropy calculation formula, Representing a single feature Entropy is used to measure the uncertainty of feature values. The smaller the entropy value, the more concentrated the feature values ​​are and the higher the certainty of the information. Representation of features The set of all discrete values, Representation of features With wind power The joint entropy is used to measure the uncertainty of the value that both factors take together. Represents the set of all discrete values ​​of wind power. Representation of features Marginal probability distribution This represents the joint probability distribution.

[0107] S34. Based on the entropy calculation results, the univariate mutual information data is obtained using the following formula:

[0108]

[0109] in, Representation of features With wind power mutual information, Representing a single feature entropy, Entropy represents wind power output. Representation of features With wind power The joint entropy.

[0110] Specifically, the wind power forecasting terminal calculates the univariate mutual information between each feature and wind power based on the entropy calculation results, using a mutual information formula to obtain univariate mutual information data. Mutual information is used to quantify the strength of the nonlinear correlation between two random variables. In the above formula for calculating mutual information, Representation of features With wind power The greater the mutual information value, the greater the contribution of the feature to the prediction of wind power. Representing a single feature entropy, Entropy represents wind power output. Representation of features With wind power The joint entropy.

[0111] S35. Based on univariate mutual information data, the joint mutual information between different feature combinations and wind power is calculated to obtain an optimized feature set.

[0112] For example, the wind power forecasting terminal further calculates the joint mutual information between different feature combinations and wind power based on univariate mutual information data. A single feature may not fully reflect the complex factors affecting wind power, while feature combinations may have synergistic effects. Joint mutual information is used to assess the correlation strength between feature combinations and wind power under such synergistic effects. Its calculation principle is to expand the calculation dimension of mutual information by constructing a multidimensional joint probability distribution of multiple features and wind power. The terminal prioritizes selecting features with high univariate mutual information for combination, calculates the joint mutual information of each combination, and finally selects feature combinations with a small number of features and high joint mutual information. Such combinations can reduce model complexity with fewer features while fully covering the key information affecting wind power. The selected feature combinations are the optimized feature set.

[0113] In an optional embodiment, the first future prediction value for each site is obtained using the following method:

[0114] S41, normalize the optimized feature set to obtain normalized feature data.

[0115] For example, the wind power forecasting terminal performs normalization processing on the optimized feature set to eliminate the interference of differences in the physical dimensions of different features on model training. Differences in the dimensions of different features (e.g., wind speed in m / s, angle in °) can cause the model to be overly biased towards features with large vectors during training, affecting the training effect. The terminal uses a max-min normalization method, which maps all values ​​of the features to the [0,1] interval, as shown in the formula. (in Let be the normalized feature values, and x be the original feature values ​​in the optimized feature set. , These are the minimum and maximum values ​​of the feature, respectively. The terminal calculates its value by traversing each feature in the optimization feature set. and Then, substitute the values ​​into the formula to complete the normalization transformation of each feature value, and finally obtain the normalized feature data, ensuring that each feature has equal weight and influence in model training.

[0116] S42, input the normalized feature data into the two-layer long short-term memory network to obtain the first predicted value.

[0117] Specifically, the wind power forecasting terminal inputs normalized feature data into a pre-set two-layer LSTM network. The two-layer LSTM network consists of two stacked LSTM units. The first LSTM unit performs preliminary temporal feature extraction on the normalized feature data, learning basic short-term variation patterns (such as the impact of minute-level wind speed fluctuations on power). The second LSTM unit, based on the output of the first layer, further mines more complex short-term dependency information (such as power change patterns under the synergistic effect of multiple features). During network training, the terminal continuously adjusts the network parameters using the Backpropagation Through Time (BPTT) algorithm to minimize the prediction error. The network parameters include the weight matrix and bias vector, and the prediction error includes the mean squared error. After training, the terminal extracts the output of the last time step of the two-layer LSTM network as the first predicted value.

[0118] S43, input the first predicted value into the dual-path attention mechanism network to obtain the site feature association features.

[0119] Specifically, the wind power prediction terminal inputs the first predicted value into a dual-path attention mechanism network, which includes a site attention path and a feature attention path. The core principle of the attention mechanism is to focus on information more important to the prediction task by calculating attention weights. In the site attention path, the terminal treats each wind farm site as an independent processing unit, i.e., a token. By calculating the attention weights between different sites, it captures the spatial dependencies between multiple sites; the higher the weight, the stronger the spatial influence between the corresponding sites. In the feature attention path, the terminal transposes the first predicted value, making the feature dimension the main processing dimension. Similarly, it calculates the attention weights between different features through the attention mechanism to explore the potential correlations between features. The terminal concatenates and integrates the outputs of the two paths according to the feature dimensions to obtain site feature association features that simultaneously contain multi-site spatial correlation information and multi-feature correlation information.

[0120] S44. Based on the site characteristics and associated features, the first future prediction value of each site is obtained.

[0121] Specifically, the wind power forecasting terminal generates the first future forecast value for each site based on site-specific features and related features. First, the terminal decomposes the site-specific features and related features according to the site dimension, separating the related features corresponding to each site to obtain an independent feature representation for each site. This independent feature representation contains the short-term time-series information of the corresponding site, spatial correlation information with other sites, and correlation information of various features, comprehensively reflecting the factors affecting the short-term power output of that site. Then, the terminal configures an independent output header for each site. The output header is essentially a linear transformation layer that maps the independent feature representation of the site to the power prediction space, that is, converts the high-dimensional features into power values ​​through linear calculation, obtaining the first future forecast value for that site. The terminal performs the above operations on all sites, ultimately obtaining the first future forecast values ​​for all sites.

[0122] In an optional embodiment, the second future prediction value is obtained using the following method:

[0123] S51, normalize the optimized feature set to obtain normalized feature data.

[0124] For example, the wind power forecasting terminal performs normalization processing on the optimized feature set, using the max-min normalization method to map each feature value to the [0,1] interval. The purpose of this step is to ensure that the data input to the multi-layer TCN network and the data input to the LSTM network have the same dimensional range, avoiding model training bias caused by differences in data scale and ensuring consistency in training conditions between the two branches. The terminal calculates the minimum and maximum values ​​of each feature in the optimized feature set, substitutes them into the max-min normalization formula, and completes the normalization transformation of all feature values ​​to obtain normalized feature data, providing standardized data for the subsequent input to the TCN network.

[0125] S52, normalized feature data is input into a multi-layer temporal convolutional network to obtain the second predicted value.

[0126] For example, the wind power forecasting terminal inputs normalized feature data into a pre-defined multi-layer TCN network. The multi-layer TCN network consists of multiple stacked TCN unit layers. Each TCN unit employs a dilated causal convolution structure. Dilated convolution, by introducing a gap (dilation factor) into the convolution kernel, expands the receptive field of the convolution without increasing the number of parameters, enabling the network to capture dependencies over longer time series. Causal convolution ensures that the network only uses historical data for prediction, conforming to the logical constraint that time series prediction does not use future data. During network training, the terminal optimizes the network's ability to extract long-term time series features by adjusting the dilation factor and convolution kernel parameters of each layer. After training, the terminal extracts the output of the last time step of the multi-layer TCN network as the second predicted value.

[0127] S53, input the second predicted value into the dual-path attention mechanism network to obtain the second site feature association feature.

[0128] For example, the wind power forecasting terminal inputs the second predicted value into a dual-path attention mechanism network. In the site attention path, the terminal uses each site as a token and calculates the spatial correlation weights between sites through the attention mechanism to capture the spatial dependencies of multiple sites over long-term time scales, such as the synchronous power changes of multiple sites under the influence of regional meteorological systems. In the feature attention path, the terminal transposes the second predicted value, uses features as tokens, calculates the correlation weights between features, and mines the potential correlations between features over long-term time scales, such as the indirect influence of temperature on wind speed over long time scales. The terminal integrates the outputs of the two paths to obtain the second site feature association features, which contain long-term spatial correlation information and feature association information of multiple sites.

[0129] S54. Based on the site characteristics and associated features, the second future prediction value of each site is obtained.

[0130] For example, the wind power forecasting terminal generates a second future forecast value for each site based on the second site feature association features. First, the terminal decomposes the second site feature association features by site dimension, obtaining an independent long-term association feature representation for each site. This representation includes the corresponding site's long-term time-series information, long-term spatial association information, and long-term feature association information, comprehensively reflecting the factors affecting the site's long-term power output. Then, the terminal inputs the independent long-term association feature representation for each site into its configured independent output header. Through linear transformation of the output header, the association features are mapped to power forecast values; that is, high-dimensional features are converted into power values ​​through linear calculation. The terminal performs the above operation on all sites, ultimately obtaining the second future forecast value for all sites.

[0131] In an optional embodiment, the first future prediction value and the second future prediction value are weighted and fused to obtain the multi-site wind power prediction value, including:

[0132] Based on the first and second future forecasts, the multi-site wind power forecast is obtained using the following formula:

[0133]

[0134] in, For the first The predicted fusion value for each site, These are the weighting coefficients. For the first The first future forecast value for each site, For the first The second future forecast value for each site.

[0135] For example, the wind power forecasting terminal calculates the multi-site wind power forecast value based on the first and second future forecast values ​​of each site using the above formula. In the formula for calculating the fused forecast value, ... For the first The combined forecast value of each site, i.e. the final wind power forecast value. These are weighting coefficients, obtained by the terminal through optimization based on validation set data to minimize prediction error (such as mean absolute error). They are used to balance the contributions of short-term and long-term predictions. If the prediction error of the short-term branch at a certain site is small, then... The value tends to be too large, and vice versa. For the first The first future forecast value for each site, For the first The second future forecast value for each site. The terminal integrates all sites' data. Ultimately, the predicted wind power values ​​for multiple sites were obtained.

[0136] The aforementioned wind power prediction method based on a multi-site dual-spatiotemporal network improves the completeness and physical rationality of multi-site wind power data by preprocessing data, correcting outliers by combining unit operation boundary constraints, and repairing missing values ​​by utilizing geographical similarity. It also employs feature selection based on mutual information to filter key features strongly correlated with wind power, reducing model redundancy and improving training efficiency. The dual-branch architecture of the multi-site dual-spatiotemporal network captures short-term temporal dependencies through an LSTM branch and a TCN branch, which is better suited for capturing long-term dependencies. A dual-path attention mechanism is combined to mine spatial correlations and dynamic feature correlations, achieving comprehensive modeling of multi-scale spatiotemporal information. Adaptive weighted fusion integrates the advantages of dual-branch prediction, improving the accuracy, robustness, and generalization ability of multi-site wind power prediction, providing a more reliable regional wind power output reference for grid dispatch.

[0137] To further illustrate the solutions of the embodiments of this application, a specific example is provided below. For example... Figure 2 As shown, the multi-site wind power forecasting method includes the following steps:

[0138] S1: Acquire historical data of wind farms, preprocess key features such as nacelle orientation and blade pitch angle, and handle outliers in all data.

[0139] S1.1: For the nacelle direction and blade pitch angle data, since this feature is strictly constrained by the operating boundary of the wind turbine, these two data should be within the range of nacelle direction [-720, 720] and blade pitch angle [0, 89]. Data outside this range are considered outliers and are corrected using the boundary constraint correction method.

[0140]

[0141] S1.2: The missing data in the nacelle orientation and blade pitch angle features are filled in using linear interpolation.

[0142] S1.3: The ambient temperature and the cabin interior temperature should be within the ranges of [-21℃, 60℃] and [-21℃, 70℃] respectively, the relative wind direction should be within the range of [-180°, 180°], and the wind speed should be within the range of... The unit is m / s; when these characteristics do not meet the requirements, they are considered abnormal data and marked as empty, as missing values.

[0143] S2: Perform angle normalization and absolute wind direction calculation on cabin direction and wind direction data to eliminate periodic interference and ensure physical consistency of direction.

[0144] S1.2: To eliminate modeling interference caused by the periodicity of the cabin's orientation angle, the original values ​​are normalized to the interval [-180°, 180°] using modular arithmetic. The specific formula is as follows:

[0145]

[0146] S2.2: To obtain the true wind direction information in the geographic coordinate system, the relative wind direction is added to the normalized cabin direction to calculate the absolute wind direction; then, the calculated absolute wind direction is also normalized to the interval [-180°, 180°] to maintain its directional continuity and physical consistency. The specific formula is:

[0147]

[0148] S3: Based on the repaired data, missing values ​​of wind speed, absolute wind direction, ambient temperature, cabin interior temperature, and power characteristics are repaired to improve the integrity and physical rationality of the data.

[0149] S3.1: Spatial clustering is performed based on the geographical coordinates of each station, dividing them into several regions with similar meteorological characteristics.

[0150] The specific groupings are as follows: the set of geographic coordinates for all stations is... ,in Indicates the first The coordinates of each station, then any two stations and The Euclidean distance between them is defined as:

[0151]

[0152] Construct a distance matrix between all sites ,in Subsequently, hierarchical clustering was performed using the Ward connectivity criterion, whose merging strategy minimizes the increase in the sum of squares within groups:

[0153]

[0154] in, , Clusters to be merged and The number of sites, , Its centroid coordinates, This represents the increase in the squared error within each cluster after merging. By cutting the clustering book, a specified number K= clusters are obtained, ultimately dividing all sites into mutually exclusive groups. ,Right now:

[0155] S3.2: For missing data on wind speed, wind direction, ambient temperature, and cabin interior temperature, a location-based hierarchical clustering imputation method is used for repair.

[0156]

[0157] in, Indicates at time With the A collection of data from other sites that belong to the same cluster and whose data is normal.

[0158] S3.3: For the data after the above processing, check that the wind power is greater than or equal to 0, and when the wind speed is greater than 2.5m / s, the wind power must be greater than 0. Any abnormal values ​​that do not meet the requirements are regarded as missing values, and then the same hierarchical clustering imputation method based on geographical location is used for repair.

[0159] S4: After completing the data repair, the input features are filtered to improve the model's predictive performance and computational efficiency.

[0160] S4.1: First, calculate the mutual information between each feature and the target variable using an entropy-based estimation method. First, discretize the continuous variables by binning; the number of bins is determined according to the Freedman-Diaconis rule, i.e.

[0161]

[0162] in, For target variable interquartile range, The number of samples is [number], and the number of bins is [number]. And restricted to the range [5,50]; then a two-dimensional joint histogram is constructed, the joint probability distribution p(x,y) and marginal distributions p(x) and p(y) are calculated, and the entropy is calculated based on this:

[0163]

[0164]

[0165] Univariate mutual information is calculated by the following formula:

[0166]

[0167] S4.2: Calculate the joint mutual information between each feature combination and the target variable. By constructing a multivariate joint probability distribution, evaluate the nonlinear correlation strength between the feature combinations and the target variable. Then, based on the mutual information calculation results of univariate and multivariate combinations, while ensuring information gain, prioritize feature combinations with fewer features and complete information coverage as the output features for the final modeling.

[0168] S5: After completing data repair and feature selection, normalize the selected features and construct samples.

[0169] S5.1: The maximum-minimum normalization method is used to map each feature value to the interval [0,1].

[0170] S5.2: Construct the normalized data into a sample structure with a fixed time window, arranged chronologically. The time window length is 60, and each sample is [N, T, F], where N is the number of stations, T is the time step, and F is the number of features. The corresponding target variable value is [N, Tp], where Tp is the step size to be predicted. Finally, divide the constructed sample set into training, validation, and test sets in a 7:2:1 ratio.

[0171] S6: Construct a bi-branch prediction model

[0172] S6.1: Constructing the LSTM-Transformer Structure

[0173] S6.1.1: Input the constructed data into a two-layer LSTM network, with each layer having a hidden unit dimension of 32.

[0174] S6.1.2: The output of the last time step of the LSTM network is fed into a dual-path Transformer structure. Each Transformer structure has a hidden layer dimension of 64 and two attention heads. The Station Transformer treats each station as a token; the Feature Transformer takes the transpose of the output of the last time step of the LSTM as input and treats each feature as a token.

[0175] S6.1.3: Transpose the output of the feature Transformer back to its original format, and then concatenate it with the output of the site Transformer along the feature dimension.

[0176] S6.1.4: The concatenated output is split according to the site dimension to obtain the hidden representation of each site. The hidden representation of each site is input into its corresponding independent output header, and finally the power prediction value of each site is obtained.

[0177] S6.2: Constructing the TCN-Transformer structure

[0178] S6.2.1: The constructed data is input into a multi-layer TCN network, and temporal features are extracted through multi-layer dilated causal convolution. Each layer has 64 output channels. The expansion factor of the layer is .

[0179] S6.2.2: The output of the last time step of the multi-layer TCN network is fed into a dual-path Transformer structure. Each Transformer structure has a hidden layer dimension of 64 and two attention heads. The Station Transformer treats each station as a token; the Feature Transformer takes the transpose of the output of the last time step of the LSTM as input and treats each feature as a token.

[0180] S6.2.3: Transpose the output of the feature Transformer back to its original format, and then concatenate it with the output of the site Transformer along the feature dimension.

[0181] S6.2.4: The concatenated output is split according to the site dimension to obtain the hidden representation of each site. The hidden representation of each site is input into its corresponding independent output header, and finally the power prediction value of each site is obtained.

[0182] S7: Input the outputs of the LSTM-Transformer model and the TCN-Transformer model into the adaptive ensemble module. This module assigns an independent learnable weight parameter to each site and generates the final power prediction value by weighted fusion of the two prediction results.

[0183] Multi-site wind power forecasting methods demonstrate excellent performance in wind power forecasting tasks. As shown in Table 1:

[0184] Table 1

[0185] Experiments across multiple prediction durations demonstrate that our proposed method outperforms existing mainstream prediction models, such as CNN_LSTM, GCN_LSTM, GAT_LSTM, LSTM, and Transformer, in key metrics including mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²). Experimental results show that the multi-site wind power prediction method can more accurately capture the temporal variations and spatial correlations of wind power data. Through innovative fusion dual-branch structure and data processing, it improves the accuracy and robustness of wind power prediction.

[0186] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0187] Based on the same inventive concept, this application also provides a wind power forecasting system based on a multi-site dual-spatiotemporal network for implementing the wind power forecasting method based on a multi-site dual-spatiotemporal network described above. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the wind power forecasting system based on a multi-site dual-spatiotemporal network provided below can be found in the limitations of the wind power forecasting method based on a multi-site dual-spatiotemporal network described above, and will not be repeated here.

[0188] In one exemplary embodiment, such as Figure 3 As shown, a wind power forecasting system 300 based on a multi-site dual-temporal network is provided, including:

[0189] The data acquisition and preprocessing module 301 is used to acquire historical data of wind farms at multiple sites and preprocess the historical data of wind farms at multiple sites to obtain preprocessed data. The historical data of wind farms at multiple sites includes the nacelle direction, blade pitch angle, wind speed, relative wind direction, wind power, ambient temperature and nacelle internal temperature of each site.

[0190] The feature selection module 302 is used to perform feature selection processing on the preprocessed data to obtain an optimized feature set; the optimized feature set is a set of key data features related to wind power in a multi-site wind farm.

[0191] The preliminary prediction module 303 is used to input the optimized feature set into a preset multi-site dual-temporal network to obtain the first future prediction value and the second future prediction value of each site; the multi-site dual-temporal network includes a long short-term memory network branch and a temporal convolutional network branch.

[0192] The wind power prediction module 304 is used to weight and fuse the first future prediction value and the second future prediction value to obtain the wind power prediction value of multiple sites.

[0193] Furthermore, the data acquisition and preprocessing module 301 is also used for:

[0194] Anomaly identification is performed on the nacelle orientation and blade pitch angle in the acquired historical data of wind farms from multiple sites to obtain initial anomaly data. The initial anomaly data includes anomaly data of nacelle orientation exceeding the preset nacelle orientation range and anomaly data of blade pitch angle exceeding the preset blade pitch angle range.

[0195] The following formula is used to correct outliers in the initial outlier data, resulting in outlier-corrected data:

[0196]

[0197] in, This indicates the corrected nacelle orientation or blade pitch angle value. This indicates the minimum reasonable value for the nacelle direction or blade pitch angle. This indicates the maximum reasonable value for the nacelle direction or blade pitch angle. This indicates the original values ​​of the nacelle orientation or blade pitch angle in the initial abnormal data;

[0198] The missing data for nacelle orientation and blade pitch angle in the outlier-corrected data are filled in to obtain the feature-completed data;

[0199] Outlier values ​​are marked for wind speed and relative wind direction in the feature-completed data. Outliers with wind speed exceeding the preset wind speed range and relative wind direction exceeding the preset relative wind direction range are set to null values ​​to obtain preprocessed data.

[0200] Furthermore, the data acquisition and preprocessing module 301 is also used for:

[0201] The cabin orientation is normalized using the following formula to obtain normalized cabin orientation data:

[0202]

[0203] in, This indicates the normalized cabin orientation. This represents the original cabin orientation value in the preprocessed data. Represents modulo operation;

[0204] The absolute wind direction is obtained by comparing the relative wind direction in the cabin orientation normalized data with the preprocessed data;

[0205] Normalize the absolute wind direction to obtain normalized wind direction data;

[0206] Based on the geographical coordinates of each station, missing values ​​in the wind direction normalization data are repaired to obtain the repaired data.

[0207] Outlier handling is performed on the repaired data to obtain the final preprocessed data; the final preprocessed data includes the complete repaired data.

[0208] Furthermore, the feature selection module 302 is also used for:

[0209] The preprocessed data is binned and discretized to obtain discretized feature data;

[0210] For the discretized feature data, a two-dimensional joint histogram of each feature and wind power is constructed, and the joint probability distribution, the marginal probability distribution of each feature, and the marginal probability distribution of wind power are calculated to obtain the probability distribution data.

[0211] Based on the probability distribution data, the entropy value is calculated using the following formula:

[0212]

[0213] in, Representing a single feature entropy, Representation of features The set of all discrete values, Representation of features With wind power joint entropy, Represents the set of all discrete values ​​of wind power. Representation of features Marginal probability distribution Represents the joint probability distribution;

[0214] Based on the entropy calculation results, the univariate mutual information data is obtained using the following formula:

[0215]

[0216] in, Representation of features With wind power mutual information, Representing a single feature entropy, Entropy represents wind power output. Representation of features With wind power The joint entropy;

[0217] Based on univariate mutual information data, the joint mutual information between different feature combinations and wind power is calculated to obtain an optimized feature set.

[0218] Furthermore, the preliminary prediction module 303 is also used for:

[0219] The optimized feature set is normalized to obtain normalized feature data;

[0220] The normalized feature data is input into a two-layer long short-term memory network to obtain the first predicted value; the first predicted value is the output of the long short-term memory network at the last time step.

[0221] The first predicted value is input into the dual-path attention mechanism network to obtain the site feature association features;

[0222] Based on the correlation of site characteristics, the first future prediction value of each site is obtained.

[0223] Furthermore, the preliminary prediction module 303 is also used for:

[0224] The optimized feature set is normalized to obtain normalized feature data;

[0225] Normalized feature data is input into a multi-layer temporal convolutional network to obtain a second predicted value; the second predicted value is the output of the last time step of the temporal convolutional network.

[0226] The second predicted value is input into the dual-path attention mechanism network to obtain the second site feature association feature;

[0227] Based on the correlation of site characteristics, the second future prediction value of each site is obtained.

[0228] Furthermore, the wind power prediction module 304 is also used for:

[0229] Based on the first and second future forecasts, the multi-site wind power forecast is obtained using the following formula:

[0230]

[0231] in, For the first The predicted fusion value for each site, These are the weighting coefficients. For the first The first future forecast value for each site, For the first The second future forecast value for each site.

[0232] In one embodiment, such as Figure 4 A computer device 400 is provided, comprising:

[0233] At least one processor 401, and at least one memory 402 communicatively connected to said processor 401; said memory stores application code executable by said processor, said application code being executed by said processor to enable said processor to perform the steps of the wind power forecasting method based on a multi-site dual-temporal network as described above;

[0234] Computer equipment may also include: sensor 403;

[0235] The processor 401, memory 401, and sensor 403 can be connected via bus 404 or other means. The figure shows an example of connection via bus 404. Figure 4 The character is represented by a single thick line, but this does not mean that there is only one bus or a type of bus.

[0236] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0237] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0238] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A wind power prediction method based on a multi-site dual space-time network, characterized in that, The method comprises: acquiring multi-site wind farm historical data, and preprocessing the multi-site wind farm historical data to obtain preprocessed data; the multi-site wind farm historical data comprises nacelle direction, blade pitch angle, wind speed, relative wind direction, wind power, environmental temperature and nacelle internal temperature of each site; performing feature selection processing on the preprocessed data to obtain an optimized feature set; the optimized feature set is a set of key data features related to wind power in a multi-site wind farm; inputting the optimized feature set into a preset multi-site double spatiotemporal network to obtain a first future prediction value of each site and a second future prediction value of each site; the multi-site double spatiotemporal network comprises a long short-term memory network branch and a time convolution network branch; performing weighted fusion on the first future prediction value and the second future prediction value to obtain a multi-site wind power prediction value.

2. The method of claim 1, wherein, The preprocessing of the multi-site wind farm historical data to obtain preprocessed data comprises: performing outlier identification on the nacelle direction and the blade pitch angle in the acquired multi-site wind farm historical data to obtain initial abnormal data; the initial abnormal data comprises abnormal value data of the nacelle direction out of a preset nacelle direction range and abnormal value data of the blade pitch angle out of a preset blade pitch angle range; using the following formula to correct the abnormal values in the initial abnormal data to obtain data after correction of abnormal values: wherein, represents a corrected nacelle direction or blade pitch angle value, represents a minimum reasonable value for the nacelle direction or the blade pitch angle, represents a maximum reasonable value for the nacelle direction or the blade pitch angle, represents a nacelle direction raw value or a blade pitch angle raw value in the initial abnormal data; completing missing data of the nacelle direction and the blade pitch angle in the data after correction of abnormal values to obtain data after feature completion; performing abnormal value marking on the wind speed and the relative wind direction in the data after feature completion, setting abnormal values of the wind speed out of a preset wind speed range and the relative wind direction out of a preset relative wind direction range as null values to obtain the preprocessed data.

3. The method of claim 1, wherein, The obtaining of the preprocessed data further comprises: using the following formula to perform angle normalization processing on the nacelle direction to obtain nacelle direction normalized data: wherein, represents the normalized cabin direction, represents the cabin direction raw value in the pre-processed data, represents a modulo operation; obtaining absolute wind direction according to the nacelle direction normalized data and the relative wind direction in the preprocessed data; performing normalization on the absolute wind direction to obtain wind direction normalized data; based on geographical coordinates of each site, repairing missing values in the wind direction normalized data to obtain repaired data; performing abnormal value processing on the repaired data to obtain final preprocessed data; the final preprocessed data comprises complete repaired data.

4. The method of claim 3, wherein, The feature selection processing on the preprocessed data to obtain an optimized feature set comprises: performing binning discretization on the preprocessed data to obtain discretized feature data; constructing a two-dimensional joint histogram of each feature and wind power based on the discretized feature data, calculating joint probability distribution and edge probability distribution of each feature and edge probability distribution of wind power to obtain probability distribution data; based on the probability distribution data, using the following formula to obtain an entropy value calculation result: wherein, denotes the entropy of a single feature , denotes the set of all discrete values of a feature , denotes the joint entropy of a feature and the wind power , denotes the set of all discrete values of the wind power , denotes the marginal probability distribution of a feature denotes the joint probability distribution; based on the entropy value calculation result, using the following formula to obtain univariate mutual information data: wherein, represents the mutual information of the feature and the wind power, represents the entropy of the individual feature , represents the entropy of the wind power, represents the joint entropy of the feature and the wind power, ;​ based on the univariate mutual information data, calculating joint mutual information of different feature combinations and the wind power to obtain the optimized feature set.

5. The method of claim 1, wherein, The first future prediction value of each station is obtained using the following method: The optimization feature set is normalized to obtain normalized feature data; The normalized feature data is input into a double-layer long short-term memory network to obtain a first prediction value; the first prediction value is the output of the last time step of the long short-term memory network; The first prediction value is input into a double-path attention mechanism network to obtain station feature correlation features; The first future prediction value of each station is obtained according to the station feature correlation features.

6. The method of claim 1, wherein, The second future prediction value is obtained using the following method: The optimization feature set is normalized to obtain normalized feature data; The normalized feature data is input into a multi-layer time convolution network to obtain a second prediction value; the second prediction value is the output of the last time step of the time convolution network; The second prediction value is input into a double-path attention mechanism network to obtain second station feature correlation features; The second future prediction value of each station is obtained according to the station feature correlation features.

7. The method of claim 1, wherein, The first future prediction value and the second future prediction value are weighted and fused to obtain a multi-station wind power prediction value, including: The first future prediction value and the second future prediction value are used to obtain the multi-station wind power prediction value using the following formula: wherein, is a fusion prediction value for the th site, is a weight coefficient, is a first future prediction value for the th site, is a second future prediction value for the th site.

8. A wind power prediction system based on multi-site dual space-time network, characterized in that, The system comprises: A data acquisition and preprocessing module for acquiring multi-station wind farm historical data and preprocessing the multi-station wind farm historical data to obtain preprocessed data; the multi-station wind farm historical data includes the cabin direction, blade pitch angle, wind speed, relative wind direction, wind power, environmental temperature, and cabin internal temperature of each station; A feature selection module for performing feature selection processing on the preprocessed data to obtain an optimization feature set; the optimization feature set is a set of key data features related to wind power in a multi-station wind farm; A preliminary prediction module for inputting the optimization feature set into a preset multi-station double space-time network to obtain a first future prediction value of each station and a second future prediction value of each station; the multi-station double space-time network includes a long short-term memory network branch and a time convolution network branch; A wind power prediction module for weighting and fusing the first future prediction value and the second future prediction value to obtain a multi-station wind power prediction value. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor executes the computer program to implement the method of any one of claims 1-7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the method of any one of claims 1-7.

Images