Ultra-short-term wind power prediction method based on multi-source data fusion

Abstract

The invention discloses an ultra-short-term wind power prediction method based on multi-source data fusion, and relates to the technical field of wind power generation, and the method comprises the steps: respectively collecting data from numerical weather forecast, wind power plant meteorological monitoring equipment and a fan operation state monitoring system, carrying out the preprocessing, and forming a multi-source wind power monitoring data flow; then, performing feature extraction and weighted fusion on the data stream to obtain a target wind power fusion feature set; thirdly, training and optimizing wind power multi-source historical data, and constructing a wind power prediction model; and finally, performing ultra-short-term wind power prediction on the target fusion feature set based on the model. The technical problems of low prediction precision and poor real-time performance caused by lack of multi-source data fusion in an existing wind power prediction method are solved, and the technical effect of improving the precision and the real-time performance of wind power prediction is achieved by fusing multi-source information such as numerical weather forecast, meteorological data and fan operation data.

Description

Technical Field

[0001] The present application relates to the technical field of wind power generation, and in particular to an ultra-short-term wind power prediction method based on multi-source data fusion. Background Art

[0002] Wind power forecasting is a key technology to improve the efficiency of power system dispatching and ensure the stability of the power grid. However, existing technologies have great deficiencies in forecasting accuracy and real-time performance, mainly due to the lack of multi-source data fusion. Traditional wind power forecasting methods usually rely on a single data source, such as using only numerical weather forecasts or local meteorological data of wind farms, resulting in low forecasting accuracy, especially in the prediction of ultra-short-term wind power fluctuations. The error is large. In addition, existing technologies fail to effectively integrate data from multiple sources, such as wind turbine operating status and environmental data, and fail to give full play to the advantages of multi-dimensional information, affecting the reliability and real-time performance of the forecast results. Therefore, how to improve the accuracy and real-time performance of wind power forecasting through multi-source data fusion has become a technical problem that needs to be solved urgently in the field of wind power.

[0003] At present, relevant technologies still face the technical problem of lack of multi-source data fusion in wind power prediction methods, which leads to low prediction accuracy and poor real-time performance. Summary of the invention

[0004] The present application solves the technical problems in existing wind power prediction methods that lack multi-source data fusion, resulting in low prediction accuracy and poor real-time performance, by providing an ultra-short-term wind power prediction method with multi-source data fusion.

[0005] The present application provides an ultra-short-term wind power prediction method based on multi-source data fusion, comprising:

[0006] Numerical weather forecast data, meteorological data and wind turbine operation data are collected and obtained from numerical weather forecast, local meteorological monitoring equipment of wind farm and wind turbine operation status monitoring system respectively; the numerical weather forecast data, meteorological data and wind turbine operation data are preprocessed to form a multi-source wind power monitoring data stream; feature extraction and weighted fusion are performed on the multi-source wind power monitoring data stream to obtain a target multi-source wind power fusion feature set; multi-source wind power historical data are collected and obtained, and training, optimization and updating of the multi-source wind power historical data are performed to construct a wind power prediction model; ultra-short-term wind power prediction is performed on the target multi-source wind power fusion feature set based on the wind power prediction model.

[0007] The ultra-short-term wind power prediction method based on multi-source data fusion proposed in this application first collects data from numerical weather forecasts, wind farm meteorological monitoring equipment, and wind turbine operation status monitoring systems, performs preprocessing, and forms a multi-source wind power monitoring data stream. Then, feature extraction and weighted fusion are performed on the data stream to obtain the target wind power fusion feature set. Next, training and optimization are performed using wind power multi-source historical data to construct a wind power prediction model. Finally, based on the model, an ultra-short-term wind power prediction is performed on the target fusion feature set. By integrating multi-source information such as numerical weather forecasts, meteorological data, and wind turbine operation data, the technical effect of improving the accuracy and real-time performance of wind power prediction is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to more clearly illustrate the technical solution of the embodiment of the present invention, the accompanying drawings of the embodiment of the present invention will be briefly introduced below. A flow chart is used in the present application to illustrate the operations performed by the system according to the embodiment of the present application. It should be understood that the preceding or following operations are not necessarily performed accurately in order. On the contrary, various steps can be processed in reverse order or simultaneously as needed. At the same time, other operations can also be added to these processes, or one or more operations can be removed from these processes.

[0009] Figure 1 A schematic diagram of a process flow of an ultra-short-term wind power prediction method based on multi-source data fusion provided in an embodiment of the present application;

[0010] Figure 2 A schematic diagram of a multi-source wind power monitoring data stream assembly process for an ultra-short-term wind power prediction method using multi-source data fusion provided in an embodiment of the present application. DETAILED DESCRIPTION

[0011] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below.

[0012] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings. The described embodiments should not be regarded as limiting the present application. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of this application.

[0013] In the following description, reference is made to "some embodiments", which describe a subset of all possible embodiments, but it is understood that "some embodiments" may be the same subset or different subsets of all possible embodiments, and may be combined with each other without conflict, and the terms "first\second" involved are merely to distinguish similar objects and do not represent a specific ordering of objects. The terms "including" and "having" and any variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or modules that are not clearly listed or inherent to these processes, methods, products, or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those generally understood by technicians in the technical field of this application. The terms used herein are for the purpose of describing the embodiments of the present application only.

[0014] The present application embodiment provides a multi-source data fusion ultra-short-term wind power prediction method, such as Figure 1 As shown, the method includes:

[0015] Step S100, numerical weather forecast data, meteorological data and wind turbine operation data are collected from numerical weather forecast, local meteorological monitoring equipment of wind farm and wind turbine operation status monitoring system respectively. Specifically, comprehensive and accurate wind power related data are collected from multiple data sources. From the aspect of numerical weather forecast, reliable data sources and interfaces should be determined, data elements such as wind speed and wind direction at different height layers and spatiotemporal resolution should be clarified, and programs or tools should be written according to interface specifications to obtain and classify and store data at intervals. For local meteorological monitoring equipment of wind farm, its functions and other information should be determined first and normal operation and calibration should be ensured, a higher acquisition frequency and transmission method should be determined, and integrated data should be received on the data processing platform. Starting from the wind turbine operation status monitoring system, the system should be familiar with the system and the permissions should be configured, key operating parameters and acquisition strategies such as wind turbine speed and blade angle should be determined, and programs should be written through interfaces to collect and properly store data, so as to lay a solid data foundation for subsequent wind power forecasting work, so that data from each data source can effectively serve the overall analysis process, improve the accuracy and reliability of the forecast, meet the actual needs of wind farm operation and power allocation, and help the efficient and stable development of the wind power industry.

[0016] Step S200, pre-process the numerical weather forecast data, meteorological data and wind turbine operation data to form a multi-source wind power monitoring data stream. Specifically, the collected numerical weather forecast data, meteorological data and wind turbine operation data are pre-processed to construct a multi-source wind power monitoring data stream. First, data cleaning is performed to check data integrity, and a small number of missing values ​​are filled by linear interpolation. If the missing values ​​are serious, they are marked or excluded; outliers are identified based on statistics and domain knowledge, such as unreasonable wind speed values ​​in numerical weather forecasts, abnormal data generated by equipment failures in local meteorological monitoring, and power anomalies in wind turbine operation, and the outliers are corrected or eliminated according to the situation. Unify the data format, convert the time data into a standard timestamp format, and unify the precision and encoding of numerical and character data respectively; and synchronize the time with the local meteorological monitoring equipment time as the reference, refine the time series of the numerical weather forecast data, and adjust the time stamp of the wind turbine operation data to align it with other data sources. The preprocessed data are integrated in chronological order, and the data from each data source at the same time are summarized with time as the main key, and arranged in sequence to form a multi-source wind power monitoring data stream, providing a high-quality data foundation for subsequent work, ensuring the accuracy and reliability of the wind power prediction system, and facilitating the efficient and stable operation and management of wind farms.

[0017] In one possible implementation, Figure 2As shown, the numerical weather forecast data, meteorological data and wind turbine operation data are preprocessed to form a multi-source wind power monitoring data stream, and step S200 further includes step S210, determining the multi-source wind turbine data application standard according to the numerical weather forecast, the local meteorological monitoring equipment of the wind farm and the wind turbine operation status monitoring system. Specifically, the numerical weather forecast system is deeply analyzed to master its data spatiotemporal resolution, prediction model accuracy and reliability, update frequency and meteorological element range, such as the wind speed prediction resolution, update interval and prediction accuracy under different meteorological conditions in a specific area; for the local meteorological monitoring equipment of the wind farm, it is necessary to know the model, measurement accuracy, installation location, acquisition frequency and stability and error range of each device under different environments, such as the measurement error of the anemometer in different wind speed sections, the accuracy of the thermometer and the drift under the influence of the environment; for the wind turbine operation status monitoring system, it is necessary to clarify the details of the wind turbine operation parameters provided by it, such as the measurement accuracy, adjustment range and normal operation relationship and change law between the parameters of the speed, power, blade angle, cabin temperature, etc. The characteristics of these three data sources and the actual operation needs of wind farms are combined to determine the application standards for multi-source wind turbine data, covering accuracy requirements, such as specifying the error range of wind speed data; integrity requirements, that is, the proportion of valid data that each data source should provide in each time period; consistency requirements, such as the difference limit of wind speed measurements from different sources at the same time; timeliness requirements, that is, the maximum delay time from data collection to application, etc. At the same time, data reliability assessment standards are formulated under different meteorological and wind turbine operating conditions to ensure that data quality and availability can be accurately judged during subsequent data processing.

[0018] Step S220, using the multi-source wind turbine data application standard to identify abnormal data on the numerical weather forecast data, meteorological data and wind turbine operation data, and obtain a multi-source wind turbine abnormal data set. Specifically, for numerical weather forecast data, based on the established data application standard, verify its accuracy. When the deviation between the predicted wind speed value of a certain area at a certain moment and the local historical average wind speed during the same period exceeds ±[X] m / s, and the difference between the wind speed measured by the local meteorological monitoring equipment at a similar time is greater than ±[X] m / s, mark the wind speed data as abnormal, and check its integrity at the same time. If there is a lack of data on some meteorological elements at a time point or in a region, it is also included in the abnormal data set. If the data update delay exceeds the specified time, it is also regarded as abnormal. For local meteorological monitoring data, anomalies are identified according to the equipment accuracy standard and the normal range of data changes. For example, if a certain anemometer has multiple consecutive measurement values ​​(no less than 5) unchanged, while other meteorological conditions change, or the wind speed exceeds the reasonable measurement range under the current environment, the wind speed data is judged to be abnormal. Temperature, humidity and other data are also judged accordingly. For example, if the temperature suddenly changes beyond ±[X]℃ and is inconsistent with the environmental thermal balance, or the humidity exceeds the normal range of the same period in history (such as relative humidity is higher than 95% or lower than 10%), they are marked as abnormal. In terms of wind turbine operation data, anomalies are detected according to the normal operating parameter range and mutual relationship of the wind turbine. If the power drops to zero but the speed is normal and the wind speed can generate electricity, or the speed increases abnormally beyond the rated speed ±[X]% and the power does not change reasonably, or the blade angle adjustment is illogical (such as the wind speed is stable but frequently adjusted significantly), and the cabin temperature exceeds the set alarm threshold (such as 80℃) without a reasonable reason, the relevant operation data will be listed as abnormal. After checking and comparing the data of the three data sources one by one, the data that meets the abnormal judgment conditions are summarized to form a multi-source wind turbine abnormal data set.

[0019] Step S230, according to the multi-source wind turbine abnormal data set, determine the abnormal data processing steps. Specifically, for the abnormal data in different data sources, take corresponding processing measures. In terms of numerical weather forecast data, if the abnormality is caused by the model prediction error, the individual data points are abnormal and the surrounding data is stable and reasonable, the wind speed data of adjacent time and space points can be corrected by linear interpolation or smoothing, such as replacing the abnormal wind speed value; if the data update is delayed, it can be downgraded or ignored in the subsequent analysis according to the timeliness requirements, and wait for new data; if a large area is missing or the quality is poor, it can be combined with backup sources, supplemented and corrected with historical data statistical models, or robustness can be enhanced in the prediction model to deal with uncertainty. For the abnormalities in local meteorological monitoring data, if a small amount of short-term abnormalities are caused by sensor failure, backup data can be used for backup data if there is a backup, and if there is no backup, it can refer to the surrounding normal sensors combined with wind direction and terrain to estimate and replace the faulty anemometer data; if the fault situation is complicated, the sensor needs to be repaired or replaced, the data is marked invalid and the fault information is recorded; if it is abnormal due to environmental factors, digital filtering is used for denoising under strong electromagnetic interference, and corrections and supplements are combined with relevant data under extreme weather. In the fan operation data, if individual points are abnormal due to short-term communication or small control fluctuations, such as abnormal power at a single moment but normal before and after, data smoothing or the before and after averaging method can be used for correction; if continuous abnormalities are caused by mechanical or electrical faults, such as abnormal increase in speed that cannot be restored, an alarm should be issued to notify the operation and maintenance personnel to shut down the machine for maintenance. Separately recording and analyzing data will help diagnose faults and evaluate the impact. In the prediction model, similar operating conditions or empirical models can be used to simulate and replace fault data based on historical data and fault types. After the fan is repaired, its operation data should be monitored to verify the repair effect and ensure normal operation.

[0020] Step S240, preprocessing the multi-source wind turbine abnormal data set based on the abnormal data processing step, and forming the multi-source wind power monitoring data stream according to the preprocessed data. Specifically, according to the determined abnormal data processing step, the multi-source wind turbine abnormal data set is processed one by one. For the abnormal data in the numerical weather forecast data, correction and optimization are performed by data interpolation, smoothing, ignoring or supplementing; for the abnormal data in the local meteorological monitoring data, data replacement, filtering, correction or marking invalid measures are adopted for processing; for the abnormal data in the wind turbine operation data, data correction, fault alarm, shutdown maintenance and data simulation replacement are performed according to the fault type and severity. During the processing, it is necessary to ensure that the processing of each abnormal data is recorded in detail, including the original value of the abnormal data, the processing method, the processing time and the processing reason, so as to facilitate the subsequent data tracing and quality assessment. After completing the abnormal data processing, the preprocessed numerical weather forecast data, meteorological data and wind turbine operation data are integrated and sorted in chronological order. Using a unified timestamp as an index, valid data from three data sources at the same time are combined to form a complete data record, including wind speed, wind direction, temperature, air pressure, humidity, fan speed, power output, blade angle, cabin temperature and other data fields. Then, these chronologically arranged data records are connected in sequence to form a multi-source wind power monitoring data stream. This data stream should be continuous, complete and accurate, and can truly reflect the operating status and meteorological conditions of the wind farm at different times, providing a high-quality and reliable data foundation for subsequent feature extraction, data fusion and the construction of wind power prediction models, thereby improving the performance and accuracy of the entire wind power prediction system and ensuring the safe and stable operation and efficient power generation of the wind farm.

[0021] Step S300, feature extraction and weighted fusion are performed on the multi-source wind power monitoring data stream to obtain a target multi-source wind power fusion feature set. Specifically, feature extraction and weighted fusion are performed on the multi-source wind power monitoring data stream to obtain a target multi-source wind power fusion feature set. In the feature extraction stage, the wind speed (including average wind speed, standard deviation, vertical shear), wind direction (duration distribution, change frequency amplitude), air pressure (gradient change, change rate) and other features at different altitudes are extracted from the numerical weather forecast data; the local meteorological monitoring data focuses on extracting features such as real-time fluctuations in wind speed and direction, temperature and humidity (daily change trend, correlation with wind speed and direction) and solar radiation intensity (if any); the wind turbine operation data extracts features such as power curve, speed (average speed, fluctuation range, matching relationship with wind speed), blade angle adjustment (frequency, amplitude, synergy with wind speed and direction), cabin temperature (change trend, peak frequency) and so on. Entering the weighted fusion stage, the weights are first determined based on the reliability, accuracy and impact of each data source on wind power. For example, the local meteorological monitoring data has high accuracy and strong correlation, so it can be set to 0.4, the numerical weather forecast data has an advantage in macro trends, so it can be set to 0.3, and the wind turbine operation data reflects the actual state, so it can be set to 0.3. Then, the weighted average method or machine learning fusion algorithm is used to calculate and fuse the feature dimensions of each data source according to the weight. For example, for the wind speed feature, the corresponding wind speed feature value of each data source is multiplied by the weight and added to obtain the fusion value. After such operation for all feature dimensions, a comprehensive, accurate and targeted multi-source wind power fusion feature set is obtained, which provides high-quality input for subsequent wind power prediction models, improves prediction accuracy and reliability, and strongly supports wind farm operation scheduling and stable operation of power systems.

[0022] In a possible implementation, feature extraction and weighted fusion are performed on the multi-source wind power monitoring data stream to obtain a target multi-source wind power fusion feature set. Step S300 further includes step S310, respectively performing correlation feature extraction and selection on the multi-source wind power monitoring data stream to obtain a target multi-source wind power key feature set. Specifically, in the correlation feature extraction link, analysis is performed on different data sources. For numerical weather forecast data, not only the average wind speed is focused on, but also statistical features such as standard deviation, skewness, and kurtosis of wind speed are extracted to show its fluctuation and distribution law. At the same time, the correlation and change trend of wind speed at different altitudes are calculated, such as using the vertical wind speed shear index to measure the change of wind speed with altitude. The correlation features between air pressure and wind speed and wind direction are also extracted, such as the relationship between air pressure gradient and wind speed and wind direction changes, and the key feature subset is obtained after screening; local meteorological monitoring data focuses on mining the real-time correlation features between wind speed and wind direction and other factors, such as analyzing the coordinated change relationship between temperature and wind speed, and extracting the interaction between humidity and wind speed and wind direction. Using features, and considering the relationship between solar radiation intensity (if any) and wind speed and direction, the key feature set is determined through optimization; the wind turbine operation data extracts the correlation features closely related to the wind turbine performance and power generation efficiency from the wind turbine operation status monitoring system, including building a power curve model and extracting its characteristic parameters, analyzing the coordinated law of wind turbine speed changes and wind speed, blade angle adjustment, and considering the correlation characteristics of cabin temperature and other operating parameters. After comprehensive screening, its key features are combined with the key feature subsets of the first two to form the target multi-source wind power key feature set, providing a basis for subsequent fusion analysis.

[0023] Step S320, the data source weights are assigned to the target multi-source wind power key feature set to obtain the wind power data source weight information. Specifically, the reliability and accuracy of each data source are evaluated. For numerical weather forecast data, its reliability is related to the accuracy of the prediction model, the accuracy of the observation data and the timeliness of the update. It is necessary to compare the historical forecast and the actual observation data, and analyze the prediction error distribution of each meteorological element under different meteorological conditions; the reliability of local meteorological monitoring data is affected by the equipment accuracy, installation location, maintenance and transmission stability. The equipment should be calibrated and maintained regularly, the data integrity and consistency should be checked, and the frequency and error range of abnormal values ​​should be counted; the reliability of wind turbine operation data is related to the accuracy of wind turbine sensors, the stability of the control system and the reliability of acquisition and transmission. It is necessary to analyze the noise level, data missing and consistency with the actual operating status. Then, the influence of each data source on wind power is determined through data analysis and actual experience. The local meteorological monitoring data can directly reflect the actual weather in the field and has a strong impact on the real-time changes of wind power. The wind turbine operation data is directly related to the power output, which helps to understand the power generation efficiency and potential problems. Although the numerical weather forecast data has errors, it can provide macro trends and information for early planning and disaster response. Finally, based on the reliability, accuracy assessment results and the degree of impact, the weight distribution scheme is determined using the hierarchical analysis method, expert scoring method or machine learning algorithm. For example, local meteorological monitoring data has high reliability and a large real-time impact, so it is assigned a weight of 0.4, wind turbine operation data is assigned a weight of 0.35, and numerical weather forecast data is assigned a weight of 0.25. These weight information are organized and recorded to form weight information, which provides a basis for subsequent weighted fusion.

[0024] Step S330, weighted fusion is performed on the target multi-source wind power key feature set according to the wind power data source weight information to obtain the target multi-source wind power fusion feature set. Specifically, when processing multi-source wind power data, it is necessary to first select a suitable weighted fusion method based on the data characteristics and actual needs. For example, the weighted average method is to linearly weight the sum of each feature in the target multi-source wind power key feature set according to the weight of the corresponding data source to obtain the fusion feature value; the distance-based weighted method determines the weight according to the distance or similarity between the data source data and the target feature, taking into account the local characteristic differences; the fusion method based on evidence theory models and fuses the data uncertainty to improve the reliability of the results; the fusion method implemented by the machine learning algorithm can automatically learn complex relationships and weight rules, but it requires a large amount of training data and computing resources. In practice, it is necessary to comprehensively consider factors such as data scale, quality and computing resources to select. Then, taking the weighted average method as an example, a weighted fusion operation is performed. For each feature dimension such as wind speed, wind direction, and power in the target multi-source wind power key feature set, the feature values ​​in the numerical weather forecast data, local meteorological monitoring data, and wind turbine operation data are multiplied by the corresponding weights and added together. All feature dimensions are calculated, and the resulting fused feature value set constitutes the target multi-source wind power fusion feature set. This set integrates multi-source key information, provides high-quality input for subsequent wind power prediction models, improves model accuracy and reliability, and assists in wind farm operation scheduling and stable operation of power systems.

[0025] In one possible implementation, the multi-source wind power monitoring data streams are respectively subjected to correlation feature extraction and selection to obtain a target multi-source wind power key feature set. Step S310 further includes step S311, respectively subjected to correlation feature extraction for the multi-source wind power monitoring data streams to obtain a multi-source wind power data correlation feature set. Specifically, in terms of correlation feature extraction, in-depth analysis is performed on different data sources. For numerical weather forecast data, the intrinsic connection between various meteorological elements and wind power is explored in depth. In addition to the average wind speed, the standard deviation of wind speed is calculated to measure the degree of fluctuation. The skewness and kurtosis are analyzed to understand the distribution pattern of wind speed. A vertical profile model of wind speed is constructed to grasp the law of vertical wind speed changes. The coordinated change characteristics of air pressure and wind speed and direction are extracted, and closely related features are screened out to form a subset of correlated features. Local meteorological monitoring data focuses on exploring the real-time dynamic correlation between wind speed and direction and other elements, and a regression or correlation model of temperature and wind speed is established to quantify the synergistic relationship. The mechanism of action of humidity and wind speed and direction is studied to extract combined features. Consider the correlation between solar radiation intensity (if any) and wind speed and direction, extract features, and integrate them to form a subset of correlated features; focus on extracting correlated features related to performance and power generation efficiency from wind turbine operation data, build a power curve model and extract key parameters to monitor faults and efficiency changes, analyze the coordinated laws of speed, wind speed and blade angle to extract relevant features to evaluate the control system, study the correlation between cabin temperature and other parameters to extract features to prevent faults, summarize features to form a subset of correlated features, and finally merge the subsets of the three data sources into a complete multi-source wind power data correlated feature set to provide a data basis for subsequent work.

[0026] Step S312, using the principal component analysis method to perform feature selection on the multi-source wind power data associated feature set, and obtain the target multi-source wind power key feature set. Specifically, when using the principal component analysis method, the multi-source wind power data associated feature set must first be subjected to data standardization processing, and the data of each feature is subtracted from its mean and divided by the standard deviation to make the dimensions and scales of different feature data consistent, and converted into standard normal distribution data with a mean of 0 and a standard deviation of 1, so as to create conditions for accurately extracting the principal component, because otherwise, the covariance matrix calculation will be affected by the dimension, thereby affecting the principal component extraction effect. Then the covariance matrix is ​​calculated for the standardized data, which can reflect the degree of linear correlation between the features. A positive and large covariance indicates that there may be redundancy in the strong positive correlation, a negative and large covariance indicates that there is a strong negative correlation, and a linear correlation close to 0 is weak. It is the key input of the principal component analysis, and the principal component representing the main information of the original data can be extracted through its eigenvalue decomposition. Then determine the principal components and sort them from large to small according to the eigenvalues. Usually, the first few principal components whose cumulative contribution rate reaches a specific threshold (such as more than 85%) are selected. The corresponding eigenvectors constitute the principal component transformation matrix. The original data can be multiplied with it to achieve dimensionality reduction to obtain the principal component data. The selection of the cumulative contribution rate threshold can reduce the dimension, avoid overfitting and improve the efficiency of subsequent analysis and model training while retaining most of the original information. Finally, based on the determined principal components, key features are selected from the original associated feature set to form the target multi-source wind power key feature set. The importance is determined by analyzing the size and sign of the original feature coefficients in the principal components, and the original features with large absolute values ​​of the coefficients in the principal components are selected. These key features remove redundancy and noise while retaining the main information. The information purity and representativeness are high, which can provide high-quality and concise input for the wind power prediction model, improve the model performance and accuracy, and more accurately predict the ultra-short-term wind power of wind farms, which strongly supports the operation and scheduling of wind farms and the stable operation of power systems.

[0027] In one possible implementation, the target multi-source wind power key feature set is weightedly fused according to the wind power data source weight information to obtain the target multi-source wind power fusion feature set, and step S330 further includes step S331, aligning the target multi-source wind power key feature set according to the time series to obtain the target multi-source wind power sequence feature set. Specifically, it is necessary to first clarify the time tag format and accuracy of the data from each data source. The numerical weather forecast data is often updated every 3 hours, and the time accuracy is the hour; the local meteorological monitoring data depends on the sampling time of the monitoring equipment, and may be recorded once a minute, with a higher resolution; the wind turbine operation data is generated according to the monitoring frequency of the wind turbine control system, and may be recorded every second. Then, using the time series of the local meteorological monitoring data as a unified reference standard, the data from different data sources in the target multi-source wind power key feature set are aligned in time series. If the update time of the numerical weather forecast data is inconsistent with the reference standard, such as updating at 9:00 and 12:00, and the data at 10:00 is needed, the linear interpolation method and other methods are used to supplement and adjust its characteristic value; if there is a transmission delay or inaccurate time mark in the wind turbine operation data, it is also necessary to calibrate and align. After completing the time series alignment, the key features of multi-source wind power aligned at each time point are combined to form a target multi-source wind power sequence feature set. This feature set can fully present the wind farm operation status information in chronological order, provide an orderly data basis for subsequent weighted fusion and analysis, and allow the data from various data sources to work together in the time dimension to more accurately reflect the changes in the operation status of the wind farm, which is of great significance to wind power prediction and analysis.

[0028] Step S332, weighted fusion is performed on the target multi-source wind power key feature set according to the wind power data source weight information to obtain the initial multi-source wind power fusion feature set. Specifically, before starting the weighted fusion operation, the wind power data source weight information needs to be clarified based on the reliability, accuracy and impact of each data source on wind power. For example, after preliminary data analysis and evaluation, given that local meteorological monitoring data has a large real-time impact on wind power and high reliability, it is given a weight of 0.4; wind turbine operation data is extremely critical to reflecting the actual operating status and performance of the wind turbine, and the weight is set to 0.35; although the numerical weather forecast data has prediction errors, it can provide macro-meteorological trend information, and the weight is 0.25. For each type of feature data at each time point in the target multi-source wind power sequence feature set, weighted fusion is performed according to the weight of the data source to which it belongs. For other key feature dimensions such as wind direction, temperature, air pressure, wind turbine power, speed, etc., similar weighted calculations are also carried out at each time point according to the weights of their respective data sources. By performing such a weighted fusion operation on all key features at each time point, a series of fused feature value sequences are obtained, which together constitute the initial multi-source wind power fusion feature set. This feature set integrates the key information of different data sources and performs reasonable weighting according to the importance of each data source, so that the fused features can more comprehensively and accurately reflect the operating status of the wind farm and the potential influencing factors of wind power. It builds a preliminary, weighted and integrated data foundation for subsequent verification and adjustment and the final determination of the target multi-source wind power fusion feature set, which helps to improve the input data quality of the wind power prediction model, thereby enhancing the accuracy and reliability of the prediction.

[0029] Step S333, verify and adjust the initial multi-source wind power fusion feature set to determine the target multi-source wind power fusion feature set. Specifically, multiple verification methods are used to evaluate the initial multi-source wind power fusion feature set. First, the historical data verification method is used to input the initial fusion feature set into the constructed wind power prediction model to predict the wind power in the past period of time. Then, the prediction results are compared and analyzed with the actual historical wind power data. By calculating the root mean square error (RMSE), mean absolute error (MAE) and other prediction error indicators, the accuracy of wind power prediction is quantitatively evaluated. If the prediction error is large, it means that there may be problems with the feature set and further analysis and adjustment are required. At the same time, cross-validation is carried out to divide the data set into multiple subsets, and one of the subsets is used as a verification set in turn, and the rest are used as training sets. This operation is repeated many times and the average error index of each verification is calculated. This method can more comprehensively evaluate the performance of the feature set on different data subsets, avoid overfitting, and ensure the stability and generalization ability of the feature set. If it is found that some features have large performance differences on different subsets or their contribution to the prediction results is unstable, it is necessary to consider adjusting or reselecting these features. According to the verification results, if it is found that the feature weight setting is unreasonable, resulting in large prediction errors, or that some features are highly correlated and there is information redundancy that affects the prediction effect, the initial multi-source wind power fusion feature set must be adjusted. The adjustment methods include redistributing feature weights through optimization algorithms such as genetic algorithms and particle swarm optimization algorithms to reduce prediction errors, or further screening and processing the features to remove features that contribute little to the prediction results or have high correlation to simplify the feature set and improve its information purity and effectiveness. After repeated verification and adjustment, the initial multi-source wind power fusion feature set is continuously optimized until it achieves satisfactory performance in various verification indicators. At this time, the fusion feature set determined is the target multi-source wind power fusion feature set. The final feature set can provide the wind power prediction model with the best quality and accurate input data, maximize the accuracy and reliability of wind power prediction, and provide strong support for wind farm operation and scheduling, stable operation of power systems, and reasonable allocation of energy, ensuring efficient and stable power generation of wind farms and improving the economic and social benefits of the entire wind power system.

[0030] Step S400, collect and obtain wind power multi-source historical data, train and optimize the wind power multi-source historical data, and build a wind power prediction model. Specifically, data collection is performed to determine data sources such as numerical weather forecast data, local meteorological monitoring data, and wind turbine operation data, and data is obtained and stored by corresponding methods, and then the original data is cleaned to handle missing and abnormal values, synchronize time and unified format. Then enter the training optimization and update stage, extract wind speed derivative features such as numerical weather forecast data, combined features of local meteorological monitoring data, and power curve features of wind turbine operation data from the pre-processed data, select appropriate models such as linear regression and neural network according to data characteristics, divide the training set and test set training model, use loss functions such as mean square error and optimization algorithms such as gradient descent to adjust parameters, use test sets to evaluate, calculate RMSE, MAE and other indicators, and if it is not ideal, optimize by adjusting hyperparameters, converting feature combinations, etc., and cross-validation and ensemble learning can also be used, and the model is updated with incremental learning as new data accumulates. Finally, a model with good performance is determined and deployed to the wind farm production environment. It is integrated with the monitoring and acquisition system, and wind power is predicted based on multi-source data collected in real time to provide a basis for operation and scheduling. At the same time, the model is continuously monitored and maintained to ensure its stable operation and prediction accuracy, thereby facilitating the efficient operation of the wind farm.

[0031] In one possible implementation, multi-source historical data of wind power is collected and acquired, and the multi-source historical data of wind power is trained, optimized and updated to construct a wind power prediction model. Step S400 further includes step S410, which pre-processes and extracts and fuses the multi-source historical data of wind power to obtain a historical multi-source wind power fusion feature set. Specifically, in the data pre-processing stage, data cleaning is first performed, and the numerical weather forecast, local meteorological monitoring and wind turbine operation data in the multi-source historical data of wind power are carefully checked. For the missing air pressure data at a certain moment in the numerical weather forecast, linear interpolation or meteorological model is used to fill it according to the previous and subsequent air pressure data and the air pressure trend in the surrounding area; the abnormal wind speed value of the local meteorological monitoring is corrected or marked as invalid in combination with the weather at that time and the data of the surrounding monitoring points; the power output of the wind turbine operation is estimated with reference to the power output under similar working conditions during the period when the power data is missing. Then, data normalization is carried out. For numerical features such as wind speed, temperature, and power, the minimum-maximum normalization method is used to map their values ​​to the interval [0,1] to eliminate the differences in the dimensions and value ranges of different feature data, so that the weights of the data from each data source are comparable in subsequent feature extraction and model training, avoiding model training deviations due to magnitude differences. Then, time synchronization is carried out, using the time series of local meteorological monitoring data as a reference. If the update interval of the numerical weather forecast data does not match the local data, it is updated every 3 hours and the local data is recorded every minute. In this case, the numerical weather forecast data is linearly interpolated within a 3-hour interval to generate approximate data every minute, ensuring that the data from each data source can accurately correspond to the operating status of the wind farm at the same time, providing time-aligned data for subsequent work. Entering the feature extraction and fusion stage, in feature extraction, the key features related to wind power are mined from the preprocessed data sources. In addition to the basic meteorological elements, the numerical weather forecast data also calculates the vertical shear characteristics of wind speed and the trend characteristics of air pressure change; the local meteorological monitoring data focuses on extracting the short-term fluctuation characteristics of wind speed and direction; the wind turbine operation data extracts the power curve characteristics and the coordinated change characteristics of wind turbine speed, wind speed, and blade angle. In terms of feature fusion, a weighted average-based method is used to determine the weight according to the degree of influence and reliability of each data source feature on wind power. For example, the weight of the wind speed and direction feature of the local meteorological monitoring data is set to 0.4, the weight of the power curve and speed of the wind turbine operation data is 0.35, and the weight of the numerical weather forecast data is 0.25. The fusion value is calculated by taking the fusion of wind speed features at a certain moment as an example. All feature dimensions are weighted and fused in this way. The obtained fusion feature sequence constitutes a historical multi-source wind power fusion feature set, which integrates multi-source key information and can provide high-quality input data for subsequent prediction model training.

[0032] Step S420, a recurrent neural network is used to perform prediction training on the historical multi-source wind power fusion feature set to obtain an initial power prediction model. Specifically, in terms of network structure design, a recurrent neural network structure including an input layer, a hidden layer and an output layer is constructed. The number of nodes in the input layer is determined according to the feature dimension of the historical multi-source wind power fusion feature set. For example, when the feature set contains 10 dimensions, the input layer is provided with 10 nodes to receive the processed wind power data features; the hidden layer uses LSTM units or GRU units to process the long-term dependencies of the time series, and 2-3 hidden layers are set according to data complexity and performance requirements, each layer contains 64-128 hidden units, and the output layer is provided with a node to output the wind power prediction value, and the gate control mechanism in the hidden layer uses a sigmoid function, and the hidden unit output uses a tanh function to enhance the nonlinear expression ability. In the model training stage, the historical multi-source wind power fusion feature set is first divided into a training set (70%-80%), a validation set (10%-15%) and a test set according to the time sequence to ensure that the model fully learns the features, adjusts the parameters to prevent overfitting and evaluates the generalization ability. Then, the network weights and bias parameters are randomly initialized, such as taking weights from a uniform distribution and setting the bias to 0 or a small constant, which helps the model converge quickly. Then, the training set is used for training, and the gradient update parameters are calculated based on the BPTT algorithm. The mean square error is used as the loss function. The time series data is input into the training batch to obtain the predicted value, and the loss is calculated based on the difference between the predicted value and the true value. Then, the parameters are updated using optimization algorithms such as Adam and Adagrad to make the predicted value close to the true value. During training, an appropriate learning rate (initial value 0.001 and decay according to the convergence situation) and the number of training rounds (determined according to the performance of the validation set to prevent overfitting) should be set. Although the initial power prediction model obtained in this way can preliminarily predict wind power, it still needs to be evaluated and optimized to improve performance reliability.

[0033] Step S430, evaluate, optimize and update the initial power prediction model to construct the wind power prediction model. Specifically, the model evaluation link includes two parts: indicator calculation and visual analysis. In terms of indicator calculation, the test set is used to evaluate the initial power prediction model. In addition to calculating the mean square error and root mean square error, the mean absolute error (whose formula can intuitively reflect the average absolute deviation between the predicted value and the true value) and the determination coefficient (the closer its value is to 1, the higher the goodness of fit and the better the prediction performance) are also calculated. This is used to quantitatively grasp the prediction error and fitting effect of the model on the test set, and provide a basis for subsequent optimization. In terms of visual analysis, in addition to indicator calculation, the model prediction results are analyzed by drawing time series graphs of predicted values ​​and true values, scatter plots and other visual methods. From the distribution of points in the graph, it is intuitively found whether the model has systematic deviations and insufficient prediction capabilities for special working conditions, so as to optimize it in a targeted manner. Model optimization covers hyperparameter adjustment, feature engineering optimization and model structure improvement. When adjusting hyperparameters, the number of hidden layers, number of hidden units, learning rate and other hyperparameters of the recurrent neural network are optimized based on the evaluation results. If the RMSE value is large and overfitting occurs, the number of layers or units is reduced, the regularization term is increased, and the learning rate is adjusted if the convergence is slow. The optimal value combination is determined through multiple experiments and cross-validation to improve performance and generalization ability. Feature engineering optimization requires further optimization of feature extraction and fusion methods, re-examination of features, and use of algorithms such as principal component analysis to remove features with small contributions and reduce dimensionality. At the same time, try to extract new features or dynamically adjust feature weights according to different weather conditions to enhance prediction capabilities under complex weather conditions. Model structure improvement can consider using a bidirectional recurrent neural network to use bidirectional information to enhance feature extraction, or introduce an attention mechanism to focus on important features to improve the efficiency of key information utilization and improve prediction accuracy. Model update includes incremental learning and model fusion. In incremental learning, this method is used to update the model as new data accumulates, and new data is gradually added to the training set for retraining. Replay-based or regularization methods are used to avoid forgetting old knowledge and achieve continuous optimization. Model fusion can combine multiple different initial power prediction models (such as models based on recurrent neural networks, support vector machines, etc.) using ensemble learning methods (such as random forests, gradient boosting trees, etc.), and obtain the final result through voting, weighted averaging, etc., giving full play to the advantages of each model to make up for the shortcomings and build a wind power prediction model with better performance. After multiple such evaluation, optimization and update processes, the final model with stable performance and accurate prediction can meet the requirements of wind farms and effectively support their production operations and management decisions.

[0034] In a possible implementation, the initial power prediction model is evaluated, optimized and updated to construct the wind power prediction model. Step S430 further includes step S431, using the mean square error function to perform loss evaluation calculation on the initial power prediction model to obtain model loss data. Specifically, data preparation is first performed to collect a data set for evaluating the initial power prediction model. The data set should have similar distribution characteristics to the data used in model training and should not be included in the training process to ensure the independence and accuracy of the evaluation. It can be divided into input features and corresponding real wind power values. The input features are multi-source wind power data after preprocessing and feature extraction fusion, such as wind speed, wind direction, air pressure, etc. in numerical weather forecast data, temperature and humidity in local meteorological monitoring data, and speed, blade angle and other information in wind turbine operation data. The input vector can reflect the operating status of the wind farm. The real wind power value is the wind farm power generation data actually measured at the corresponding time point, which is used for comparison and evaluation with the model output. Then the mean square error function is calculated. The input features in the data set are input into the initial power prediction model in turn to obtain the predicted wind power value of the model. The difference between the predicted value and the true value, i.e. the error value, is calculated for each sample data. The error value is then squared and the average of the squared error values ​​of all samples is calculated to obtain the mean square error (MSE). The calculation formula is: Where N represents the number of samples, y i is the actual wind power value of the i-th sample, It is the predicted wind power value of the model for the ith sample. The mean square error can quantify the overall deviation between the model prediction value and the true value. The smaller the value, the better the model performance. The larger the value, the lower the prediction accuracy and the greater the room for improvement. In actual calculations, programming tools such as Python's NumPy library can be used to efficiently implement it. The mean square error value is recorded as an important part of the model loss data, and its square root, namely the root mean square error (RMSE), can also be calculated. It can more intuitively reflect the average error amplitude and provide a more comprehensive reference basis for subsequent model analysis and optimization.

[0035] Step S432, based on the model loss data, the initial power prediction model is feedback optimized and updated to obtain the wind power prediction model. Specifically, the model loss data obtained by analysis and calculation includes mean square error, root mean square error, and other evaluation indicators such as mean absolute error and determination coefficient, so as to deeply analyze the performance characteristics and existing problems of the model in the prediction process. Through analysis, it can be checked whether there is a systematic deviation in the model. For example, the mean square error increases significantly in a specific time period or meteorological conditions. This may mean that the model has weak prediction ability under the corresponding circumstances, and the reasons need to be further analyzed. It may be that the data feature extraction is insufficient, or the model structure is difficult to adapt to complex conditions; at the same time, the trend of loss data changes is observed to judge the convergence of the model during training. If the loss value gradually decreases with training but slows down, it may be necessary to adjust the training parameters or methods to speed up convergence and improve efficiency. If the loss value fluctuates or does not converge, it is necessary to check the rationality of the model structure and whether the data is abnormal, and take optimization measures in time. Formulate feedback optimization strategies based on the analysis results of the model loss data, mainly to find out that the model has overfitting problems, that is, the performance is good in the training set, but the prediction accuracy is poor in the test set or actual application. Regularization techniques can be used, such as adding L1 or L2 regularization terms in the loss function, constraining weight parameters, and enhancing the generalization ability of the model to adapt to different data situations; if the loss value fluctuates or converges slowly due to unreasonable learning rate settings, the learning rate strategy can be adjusted, and the learning rate decay method can be used to gradually reduce the learning rate as the training progresses, so that the model can be fine-tuned in the later stage to avoid missing the optimal solution; in addition, if the model is too dependent on some features or there is redundant information between features, feature engineering can be re-performed to screen, transform or reorganize features, such as principal component analysis (PCA) and other methods to reduce dimensionality, extract the main feature components, reduce correlation, and improve training efficiency and prediction accuracy. Finally, the initial power prediction model is updated and retrained based on the formulated optimization strategy. If regularization technology is used, the loss function and training algorithm of the model should be adjusted accordingly, the loss value containing the regularization term should be accurately calculated, and the weight parameters should be constrained and updated as required during training. For learning rate adjustment, its size should be gradually changed during training according to the predetermined attenuation plan to ensure that the model converges stably to a better solution. After completing the feature engineering optimization, the training data needs to be re-prepared and input into the model training. During retraining, the mean square error function continues to be used as the loss evaluation indicator, and the change in loss value is monitored. The training parameters and strategies are adjusted according to the changes, such as increasing the number of training rounds, adjusting the batch size, etc. After multiple feedback optimization updates and training, when the performance indicators of the model on the validation set and the test set (such as mean square error, root mean square error, mean absolute error, etc.) meet the predetermined requirements, and the prediction results are stable and reliable in actual applications, the model is the final wind power prediction model, which can more accurately predict wind power and strongly support the operation and scheduling of wind farms and the stable operation of power systems.

[0036] In one possible implementation, the initial power prediction model is feedback optimized and updated based on the model loss data to obtain the wind power prediction model, and step S432 further includes step S4321, using the back propagation algorithm to optimize and update the parameters of the initial power prediction model based on the model loss data to obtain the model optimization parameters. Specifically, the back propagation algorithm is an efficient parameter update method based on the chain rule. The core is to guide the parameter update direction and step size by calculating the gradient of the loss function to the model parameters, so that the model prediction value is closer to the true value to reduce the loss function value. Before using this algorithm, it is necessary to clarify the model structure and parameter settings, such as determining the specific composition of the input layer, the hidden layer including the LSTM or GRU units used and their number, connection method, etc., and the output layer, clarifying the parameters such as the weight matrix and bias vector that need to be updated, and initializing them to appropriate random values ​​to lay the foundation for subsequent iterative updates. Perform the gradient calculation operation, input the prepared data set into the initial power prediction model, and after obtaining the predicted output, calculate the loss value between the predicted value and the true value based on the preset loss function (such as the mean square error function). Then, based on the chain rule, start from the loss function and gradually calculate the gradient of the loss value for each layer of parameters. For each neuron and connection weight in the neural network, the corresponding gradient value must be calculated. This involves complex partial derivative calculations, which require the clever use of differentiation rules and algorithm recursive characteristics to complete efficiently. For example, for the output of a simple fully connected neural network layer, the gradient of the loss function to the weight is calculated according to the chain rule, and the gradient values ​​of all parameters are calculated in this way. The parameter gradient value can reflect the degree and direction of the influence of the parameter on the loss function, providing key information for parameter updating. In the parameter update phase, after obtaining the gradient of the model parameters, the parameter update step size and direction are determined according to the optimization algorithm (such as stochastic gradient descent, Adagrad, Adadelta, Adam, etc.). Taking stochastic gradient descent as an example, according to its parameter update formula, the parameter size is adjusted by subtracting the learning rate from the current parameter value and multiplying it by the gradient value, so that the loss value can be reduced when the model predicts next time. In practical applications, adaptive optimization algorithms such as the Adam algorithm are often selected, which can automatically adjust the learning rate according to the historical information of the parameter gradient, making the parameter update more stable and efficient, and avoiding convergence problems caused by improper learning rate settings. After multiple iterations of calculating the gradient and updating the parameters, the model parameters are continuously adjusted to gradually improve the model prediction performance, and finally a set of optimization parameters that can enable the model to achieve a lower loss value on a given data set is obtained. They reflect the model's better fitting of data features and more accurate prediction of wind power, that is, the model optimization parameters, which lay the foundation for subsequent model configuration optimization.

[0037] Step S4322, based on the model optimization parameters, the configuration of the initial power prediction model is optimized to obtain the wind power prediction model. Specifically, parameter replacement and model reconstruction are performed to replace the original parameters of the initial power prediction model with the previously obtained model optimization parameters. This requires accurate positioning of the parameter storage location based on the specific implementation framework of the model (such as TensorFlow, PyTorch, etc.), accurately assigning the optimized parameter values ​​to the corresponding variables, ensuring that the model's subsequent predictions are calculated using new parameters, and carefully checking the integrity and consistency of the model structure to prevent parameter replacement errors or omissions, and ensure that the model runs normally and accurately. In the hyperparameter fine-tuning stage (this step is optional), even if the model already has optimized parameters, the hyperparameters can still be adjusted according to actual conditions to improve performance. The hyperparameters include the number of hidden layers, the number of hidden units, the learning rate decay strategy, the batch size, etc. These hyperparameters are changed within a certain range, and repeated tests and evaluations are performed on the validation set to observe the changes in performance indicators such as mean square error, root mean square error, and mean absolute error, and find the optimal hyperparameter combination. For example, although increasing the number of hidden layers may improve the fitting ability, it is easy to cause overfitting. At this time, it is necessary to adjust other hyperparameters (such as enhancing regularization, reducing learning rate, etc.) to balance the fitting and generalization capabilities, so that the model performs better in the test set and actual application. Model verification and finalization, use an independent test set to fully verify the optimized model, evaluate its prediction accuracy and reliability on unseen data, calculate various performance indicators and compare them with the previous version, confirm that the performance is significantly improved and meets the actual requirements, and test the stability and robustness of the model at the same time, such as predicting under different meteorological conditions and data noise levels, and observing whether the output is stable and accurate. If the model performs well in the verification, all indicators meet the standards and the actual application is stable and accurate, then the model is the final wind power prediction model, which can effectively integrate multi-source wind power data, accurately predict wind power, strongly support wind farm operation and scheduling, stable operation of the power system and reasonable allocation of energy, and ensure efficient and reliable operation and management of wind farms.

[0038] Step S500, based on the wind power prediction model, an ultra-short-term wind power prediction is performed on the target multi-source wind power fusion feature set. Specifically, data preparation and preprocessing are performed to ensure that the target multi-source wind power fusion feature set is complete and accurate, and possible data problems are checked and processed. Then, the data is processed according to the standardized method during model training to make its distribution characteristics consistent, and the dimension and order of the feature set are adapted. Then, the model is loaded and the parameters are set. The wind power prediction model is loaded from the storage to ensure that the model structure and parameters are correct, and the hyperparameters and internal state variables are checked and initialized. Then, an ultra-short-term wind power prediction is performed, and the processed feature set data is input into the model. The predicted value is obtained through a single-step or optional multi-step rolling prediction method. In the case of single-step prediction, the model outputs the predicted value at the current moment through internal calculation. In the case of multi-step prediction, attention should be paid to error accumulation and corrective measures should be taken. In the prediction, attention should be paid to the model calculation process, and problems should be promptly investigated and repaired. Finally, the prediction results are post-processed and output. First, the correction is verified and the predicted value is compared with the historical data or the results of a simple statistical model. If it is unreasonable, the cause is analyzed and adjusted, and the results are output in a suitable format. Visual display can also be selected to assist wind farm operation management and power dispatching decisions, thereby improving power generation efficiency and system stability.

[0039] The embodiment of the present application collects data from numerical weather forecasts, wind farm meteorological monitoring equipment, and wind turbine operation status monitoring systems, respectively, and performs preprocessing to form a multi-source wind power monitoring data stream. Then, the data stream is subjected to feature extraction and weighted fusion to obtain a target wind power fusion feature set. Next, multi-source historical data of wind power is used for training and optimization to construct a wind power prediction model. Finally, based on the model, an ultra-short-term wind power prediction is performed on the target fusion feature set, achieving the technical effect of improving the accuracy and real-time performance of wind power prediction by integrating multi-source information such as numerical weather forecasts, meteorological data, and wind turbine operation data.

[0040] The above specific implementation manner does not constitute a limitation to the protection scope of the present application. It should be understood by those skilled in the art that various modifications, combinations and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present application should be included in the protection scope of the present application. In some cases, the actions or steps recorded in the present application can be performed in an order different from that in the embodiment and can still achieve the desired results. In addition, the process depicted in the accompanying drawings does not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. A multi-source data fusion ultra-short-term wind power prediction method, characterized in that: The method comprises: Collect and obtain numerical weather forecast data, meteorological data and wind turbine operation data from numerical weather forecast, local meteorological monitoring equipment of wind farm and wind turbine operation status monitoring system respectively; Preprocessing the numerical weather forecast data, meteorological data and wind turbine operation data to form a multi-source wind power monitoring data stream; Performing feature extraction and weighted fusion on the multi-source wind power monitoring data stream to obtain a target multi-source wind power fusion feature set; Collect and obtain multi-source historical data of wind power, perform training, optimization and update on the multi-source historical data of wind power, and build a wind power prediction model; An ultra-short-term wind power prediction is performed on the target multi-source wind power fusion feature set based on the wind power prediction model.

2. The ultra-short-term wind power prediction method based on multi-source data fusion according to claim 1, characterized in that: The multi-source wind power monitoring data stream is composed of: Determine the application standard of multi-source wind turbine data based on the numerical weather forecast, the local meteorological monitoring equipment of the wind farm and the wind turbine operation status monitoring system; Using the multi-source wind turbine data application standard to identify abnormal data on the numerical weather forecast data, meteorological data and wind turbine operation data, to obtain a multi-source wind turbine abnormal data set; Determining abnormal data processing steps according to the multi-source wind turbine abnormal data set; The multi-source wind turbine abnormal data set is preprocessed based on the abnormal data processing step, and the multi-source wind power monitoring data stream is formed according to the preprocessed data.

3. The ultra-short-term wind power prediction method based on multi-source data fusion according to claim 1, characterized in that: The obtaining of the target multi-source wind power integration feature set comprises: Extract and select relevant features from the multi-source wind power monitoring data streams respectively to obtain a target multi-source wind power key feature set; Allocating data source weights for the target multi-source wind power key feature set to obtain wind power data source weight information; The target multi-source wind power key feature set is weightedly fused according to the wind power data source weight information to obtain the target multi-source wind power fusion feature set.

4. The ultra-short-term wind power prediction method based on multi-source data fusion according to claim 3, characterized in that: The step of obtaining a target multi-source wind power key feature set includes: Extracting correlation features from the multi-source wind power monitoring data streams respectively to obtain a multi-source wind power data correlation feature set; The principal component analysis method is used to perform feature selection on the multi-source wind power data associated feature set to obtain the target multi-source wind power key feature set.

5. The ultra-short-term wind power prediction method based on multi-source data fusion according to claim 4, characterized in that: The obtaining of the target multi-source wind power integration feature set comprises: Aligning the target multi-source wind power key feature set according to the time series to obtain a target multi-source wind power sequence feature set; The target multi-source wind power key feature set is weightedly integrated according to the wind power data source weight information to obtain an initial multi-source wind power integration feature set; The initial multi-source wind power fusion feature set is verified and adjusted to determine the target multi-source wind power fusion feature set.

6. The ultra-short-term wind power prediction method based on multi-source data fusion according to claim 1, characterized in that: The wind power prediction model is constructed, comprising: Preprocessing and feature extraction fusion of the wind power multi-source historical data to obtain a historical multi-source wind power fusion feature set; Using a recurrent neural network to perform prediction training on the historical multi-source wind power fusion feature set to obtain an initial power prediction model; The initial power prediction model is evaluated, optimized and updated to construct the wind power prediction model.

7. The ultra-short-term wind power prediction method based on multi-source data fusion according to claim 6, characterized in that: The step of constructing the wind power prediction model comprises: Using a mean square error function to perform loss assessment calculation on the initial power prediction model to obtain model loss data; The initial power prediction model is feedback optimized and updated based on the model loss data to obtain the wind power prediction model.

8. The ultra-short-term wind power prediction method based on multi-source data fusion according to claim 7, characterized in that: The obtaining of the wind power prediction model comprises: Using a back propagation algorithm to optimize and update the parameters of the initial power prediction model based on the model loss data to obtain model optimization parameters; The initial power prediction model is configured and optimized based on the model optimization parameters to obtain the wind power prediction model.

Images