Intelligent hybrid model-based ultra-short-term wind power prediction method and device

Abstract

This invention provides a method and apparatus for ultra-short-term wind power prediction based on an intelligent hybrid model, belonging to the field of wind power prediction technology. The method includes: extracting influencing factors from a historical power database of wind farms; designing an intelligent hybrid model architecture based on a set of wind power influencing factors; introducing an attention mechanism to perform multi-scale feature extraction and wind power prediction training on the historical power database based on the wind power influencing factor set and the intelligent hybrid model architecture; dynamically weighting multiple types of ultra-short-term wind power factor prediction models; and generating an ultra-short-term wind power prediction hybrid model for closed-loop wind power prediction. This invention solves the technical problem of low prediction accuracy and large fluctuations in wind power prediction when influenced by multiple factors due to the existence of single models or static prediction methods in existing technologies. It achieves the technical effect of improving the accuracy and stability of ultra-short-term wind power prediction through the fusion and dynamic weighting of multi-source factors in an intelligent hybrid model.

Description

Technical Field

[0001] This invention relates to the field of wind power forecasting technology, and specifically to a method and apparatus for ultra-short-term wind power forecasting based on an intelligent hybrid model. Background Technology

[0002] In the wind power industry, with the continuous increase in installed wind power capacity, the demand for accurate wind power forecasting by the power grid is becoming increasingly urgent. Especially in power grid dispatching and energy management scenarios, ultra-short-term wind power forecasting has become crucial to ensuring the safe and stable operation of the power grid. However, wind power is affected by various environmental factors such as wind speed, wind direction, temperature, and air density, and there are also differences in the characteristics of the wind turbines themselves, which makes wind power exhibit obvious nonlinear, non-stationary, and highly volatile characteristics.

[0003] Current wind power forecasting methods largely rely on single models or static methods based on historical data, such as time series analysis, regression, or single-layer neural networks. These methods assume stable statistical characteristics of the data or depend solely on the linear relationship between historical power and meteorological variables, making it difficult to address the nonlinear, non-stationary, and highly volatile characteristics of wind power. Factors such as sudden wind speed changes, wind direction deflections, and turbine dynamic responses cause power variations to be rapid and complex, making it difficult for single models to accurately capture instantaneous fluctuations. Static methods rely on fitting historical data, failing to reflect the inertial response of turbines to short-term wind speed changes and neglecting the spatial distribution of wind speed and the interaction of multiple meteorological variables. This results in low power forecasting accuracy and high volatility, making it difficult to meet the requirements of wind power grid-connected dispatch and stable grid operation.

[0004] Existing technologies rely on single models or static prediction methods, which leads to low prediction accuracy and large fluctuations when wind power is affected by multiple factors. Summary of the Invention

[0005] The purpose of this application is to provide a method and apparatus for ultra-short-term wind power prediction based on an intelligent hybrid model, in order to solve the technical problem that existing technologies use single models or static prediction methods, which result in low prediction accuracy and large fluctuations when wind power is affected by multiple factors.

[0006] In view of the above problems, this application provides a method and apparatus for ultra-short-term wind power prediction based on a smart hybrid model.

[0007] The first aspect of this application provides an ultra-short-term wind power prediction method based on an intelligent hybrid model. The method includes: constructing a historical power database of wind farms; extracting influencing factors from the historical power database to obtain a set of wind power influencing factors; designing an intelligent hybrid model architecture based on the set of wind power influencing factors, the intelligent hybrid model architecture including model architecture information for each type of influencing factor; introducing an attention mechanism to perform multi-scale feature extraction and wind power prediction training on the historical power database based on the set of wind power influencing factors and the intelligent hybrid model architecture, obtaining multiple types of ultra-short-term wind power factor prediction models; performing dynamic weighted decision-making on the multiple types of ultra-short-term wind power factor prediction models to generate an ultra-short-term wind power prediction hybrid model; and performing closed-loop wind power prediction using the ultra-short-term wind power prediction hybrid model.

[0008] Optionally, data cleaning and preprocessing are performed on the historical power database of the wind farm according to the data source application standards to obtain an available wind farm power dataset; the types of wind power influencing factors are obtained, including meteorological factors, equipment factors, time factors, grid dispatch factors, and terrain spatial factors; the available wind farm power dataset is spatiotemporally aligned and influencing factors are extracted according to the types of wind power influencing factors to obtain a preliminary wind power influencing factor set; principal component analysis and screening are performed on the preliminary wind power influencing factor set to obtain a complete wind power influencing factor set.

[0009] Optionally, the preliminary set of wind power influencing factors is standardized to obtain a standard set of wind power influencing factors features; the standard set of wind power influencing factors features is then decentralized and decomposed to obtain eigenvalues ​​and eigenvectors of the covariance matrix; principal component information is selected based on the eigenvalues, and the standard set of wind power influencing factors features is dimensionality-reduced and projected onto the eigenvectors based on the principal component information for merging to obtain a set of wind farm power dimensionality-reduced factors; the set of wind farm power dimensionality-reduced factors is then backtested and filtered to obtain a set of wind power influencing factors.

[0010] Optionally, the set of wind power influencing factors is classified to obtain wind farm influencing factor types, which include static deterministic factors, dynamic semi-deterministic factors, and meteorological uncertainties. Based on the wind farm influencing factor types, a set of wind farm factor type adaptation models is determined. The specific model architecture analysis of the wind farm influencing factor types is performed according to the set of wind farm factor type adaptation models to obtain multi-type influencing factor model architecture information. The multi-type influencing factor model architecture information is sequentially linked and combined to construct an intelligent hybrid model architecture.

[0011] Optionally, model fit analysis is performed on the static deterministic factors to obtain deterministic factor fit model types, which include physical mechanism models; based on the dynamic semi-deterministic factors, semi-deterministic factor fit model types are obtained, which include rule engine models and data-driven models; based on the meteorological uncertainties, uncertainty factor fit model types are obtained, which include deep learning models and statistical models; and based on the deterministic factor fit model types, semi-deterministic factor fit model types, and uncertainty factor fit model types, a wind farm factor type fit model set is determined.

[0012] Optionally, the set of wind power influencing factors is analyzed according to the intelligent hybrid model architecture to determine a set of multiple power prediction input and output factor variables; an attention mechanism is introduced to extract multi-scale features from the historical power database of the wind farm based on the set of multiple power prediction input and output factor variables to obtain a multi-scale feature dataset associated with multiple factors; the intelligent hybrid model architecture is used to train wind power prediction on the multi-scale feature dataset associated with multiple factors to obtain a multi-class ultra-short-term wind power factor prediction model.

[0013] Optionally, a preset ultra-short-term prediction time window is used to time-series label the multi-factor-related multi-scale feature dataset according to the ultra-short-term prediction time window, thereby obtaining a multi-factor-related multi-scale time-series dataset; input and output variable labels are applied to the multi-factor-related multi-scale time-series dataset to obtain a multi-factor-related multi-scale time-series sample set; and the intelligent hybrid model architecture is used to train the multi-factor-related multi-scale time-series sample set for wind power prediction, thereby obtaining a multi-type ultra-short-term wind power factor prediction model.

[0014] Optionally, the model prediction error parameters of the multi-type ultra-short-term wind power factor prediction model are calculated in real time within a preset adjacent time step threshold; the dynamic decision coefficients of the hybrid model are determined based on the model prediction error parameters; and the dynamic decision coefficients of the hybrid model are used to perform dynamic weighted decision-making on the multi-type ultra-short-term wind power factor prediction model to generate an ultra-short-term wind power prediction hybrid model.

[0015] Optionally, wind power prediction is performed using the ultra-short-term wind power prediction hybrid model to obtain wind power prediction deviation values, and feedback optimization and closed-loop prediction are performed on the ultra-short-term wind power prediction hybrid model based on the wind power prediction deviation values.

[0016] A second aspect of this application provides an ultra-short-term wind power prediction device based on an intelligent hybrid model. The device comprises: an influencing factor extraction module, used to construct a historical power database of wind farms and extract influencing factors from the historical power database to obtain a set of wind power influencing factors; a model architecture design module, used to design an intelligent hybrid model architecture based on the set of wind power influencing factors, the intelligent hybrid model architecture including model architecture information for each type of influencing factor; a prediction model acquisition module, used to introduce an attention mechanism to perform multi-scale feature extraction and wind power prediction training on the historical power database of wind farms based on the set of wind power influencing factors and the intelligent hybrid model architecture, obtaining multiple types of ultra-short-term wind power factor prediction models; and a wind power prediction module, used to perform dynamic weighted decision-making on the multiple types of ultra-short-term wind power factor prediction models to generate an ultra-short-term wind power prediction hybrid model, and to perform closed-loop wind power prediction through the ultra-short-term wind power prediction hybrid model.

[0017] One or more technical solutions provided in this application have at least the following technical effects or advantages: The method provided in this application constructs a historical power database of wind farms, extracts influencing factors from the historical power database to obtain a set of wind power influencing factors, designs an intelligent hybrid model architecture based on the set of wind power influencing factors, and includes model architecture information for each type of influencing factor, introduces an attention mechanism to perform multi-scale feature extraction and wind power prediction training on the historical power database based on the set of wind power influencing factors and the intelligent hybrid model architecture, and obtains multiple types of ultra-short-term wind power factor prediction models, performs dynamic weighted decision-making on the multiple types of ultra-short-term wind power factor prediction models, generates an ultra-short-term wind power prediction hybrid model, and performs closed-loop wind power prediction through the ultra-short-term wind power prediction hybrid model. This achieves the technical effect of improving the accuracy and stability of ultra-short-term wind power prediction through the fusion and dynamic weighting of multi-source factors in the intelligent hybrid model.

[0018] The above description is merely an overview of the technical solution of this application. To better understand the technical means of this application and to facilitate its implementation according to the description, and to make the above and other objects, features, and advantages of this application more apparent, specific embodiments of this application are described below. It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent through the following description. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0020] Figure 1 A flowchart illustrating the ultra-short-term wind power prediction method based on an intelligent hybrid model provided in this application.

[0021] Figure 2 A schematic diagram of the ultra-short-term wind power prediction device based on an intelligent hybrid model provided in this application.

[0022] Figure labeling: Module 11 for extracting influencing factors, Module 12 for model architecture design, Module 13 for obtaining the prediction model, and Module 14 for wind power prediction. Detailed Implementation

[0023] This application provides a method and apparatus for ultra-short-term wind power prediction based on an intelligent hybrid model. This addresses the technical problem of low prediction accuracy and large fluctuations in wind power prediction when influenced by multiple factors, which is often caused by existing single-model or static prediction methods. The method achieves the technical effect of improving the accuracy and stability of ultra-short-term wind power prediction through the fusion and dynamic weighting of multi-source factors using an intelligent hybrid model.

[0024] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be understood that the present invention is not limited to the exemplary embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. It should also be noted that, for ease of description, only the parts related to the present invention are shown in the accompanying drawings, not all of them.

[0025] Example 1, as Figure 1 As shown, this application provides an ultra-short-term wind power prediction method based on an intelligent hybrid model, which includes: A historical power database of wind farms is constructed, and influencing factors are extracted from the historical power database of wind farms to obtain a set of factors influencing wind power.

[0026] Furthermore, the wind power influencing factor set is obtained, including: performing data cleaning and preprocessing on the data from various sources in the historical power database of the wind farm according to the data source application standard to obtain an available wind farm power dataset; obtaining the types of wind power influencing factors, which include meteorological factors, equipment factors, time factors, grid dispatching factors, and terrain spatial factors; performing spatiotemporal alignment and influencing factor extraction on the available wind farm power dataset according to the types of wind power influencing factors to obtain a preliminary wind power influencing factor set; and performing principal component analysis and screening on the preliminary wind power influencing factor set to obtain a complete wind power influencing factor set.

[0027] Specifically, the historical power data, meteorological information, equipment status parameters and operating time data of the wind farm are continuously collected from multiple data sources such as the monitoring system of each wind turbine in the wind farm, meteorological station, and SCADA (Supervisory and Control System). At the same time, data such as dispatch instructions and load demand from the grid side are collected to build a historical power database of the wind farm.

[0028] Data cleaning and preprocessing are performed on data from various sources in the historical power database of wind farms according to data source application standards. These standards ensure data quality, including completeness, accuracy, and consistency. Data cleaning and preprocessing includes removing invalid data, filling in missing data, and outlier detection. Removing invalid data refers to removing abnormal data caused by equipment malfunctions or erroneous data recorded due to sensor failures. Filling in missing data involves using methods such as mean filling or median filling for power data with a small proportion of missing values. Outlier detection involves removing detected extreme values ​​or data that does not conform to physical laws. By sequentially performing the above data cleaning and preprocessing on data from various sources in the historical power database of wind farms, a usable wind farm power dataset is obtained. This usable wind farm power dataset refers to a dataset that meets predetermined data quality standards.

[0029] Based on the actual operation of wind farms, wind power output is affected by various factors. These factors include meteorological factors, equipment factors, time factors, grid dispatching factors, and terrain and spatial factors. Meteorological factors, such as wind speed, wind direction, temperature, and humidity, directly affect the power generation efficiency of wind turbines. Equipment factors, such as turbine speed, blade pitch angle, and unit status codes, affect the actual power output. Time factors include different times of day and seasons, with variations in wind speed, direction, and other meteorological conditions affecting wind power output. Grid dispatching factors involve the grid's capacity to accept wind power and its dispatching strategies, such as grid load demand and transmission line capacity limitations. Terrain and spatial factors include the terrain, altitude, and surrounding obstacles of the wind farm location, affecting the distribution of local wind fields and thus wind power output.

[0030] After obtaining the different types of wind power influencing factors, the available wind farm power dataset is spatiotemporally aligned according to the type of wind power influencing factor. Spatiotemporal alignment ensures the consistency of data from different sources in both time and space. Data sources are classified according to meteorological factors, equipment factors, time factors, grid dispatch factors, and terrain spatial factors, and the temporal resolution and spatial mapping relationship of each type of data are unified. For example, using the timestamp of power data recorded by the wind farm SCADA system as the reference time axis, meteorological data such as wind speed, wind direction, temperature, and air pressure from meteorological stations or numerical weather prediction (NWP) are aligned to the same time series through time resampling, such as 5-minute or 15-minute sampling intervals. For equipment factor data, such as turbine speed, blade pitch angle, and unit status code, equipment-level data matching is performed according to the turbine number and synchronized to the corresponding time node. For time factors, time feature variables such as hourly, daily, and seasonal cycles are extracted through timestamps. For grid dispatch factors, such as power curtailment orders and grid-connected power limits, event time matching is performed by matching dispatch record time with power time series. For terrain and spatial factors, such as wind turbine altitude and terrain slope, spatial mapping and association are performed through wind turbine GPS coordinates or wind farm layout maps to achieve spatiotemporal alignment of available wind farm power datasets.

[0031] After completing spatiotemporal alignment, influencing factors are extracted. Based on the wind power formation mechanism, key characteristic variables related to power changes are extracted from various data sources. For example, wind speed, wind direction, temperature, and humidity are extracted from meteorological data; turbine speed, blade angle, and operating status are extracted from equipment data; time, season, or periodic features are extracted from temporal data; and power curtailment orders or dispatch power indicators are extracted from grid data. These are combined with topographic spatial parameters, such as altitude, terrain, or turbine layout, to form a preliminary set of wind power influencing factors containing multi-source and multi-type characteristic variables. Principal component analysis and screening are then performed on this preliminary set of wind power influencing factors to obtain a final set of wind power influencing factors containing the most important influencing factors.

[0032] By comprehensively considering various factors such as meteorology, equipment, time, power grid dispatch, and terrain, it is possible to capture more comprehensive factors affecting wind power and their interactions, thereby improving the accuracy and reliability of ultra-short-term wind power forecasting.

[0033] Furthermore, principal component analysis and screening are performed on the preliminary set of wind power influencing factors to obtain a new set of wind power influencing factors. This includes: standardizing the preliminary set of wind power influencing factors to obtain a standard set of wind power influencing factor features; decentralizing and decomposing the standard set of wind power influencing factor features to obtain eigenvalues ​​and eigenvectors of the covariance matrix; selecting principal component information based on the eigenvalues; and merging the standard set of wind power influencing factor features onto the eigenvectors based on the principal component information to obtain a wind farm power dimensionality-reduced factor set; and backtesting and screening the wind farm power dimensionality-reduced factor set to obtain the new set of wind power influencing factors.

[0034] Specifically, the preliminary set of wind power influencing factors is standardized. For example, z-score standardization is used to calculate the difference between each feature data point and its mean in the preliminary set of wind power influencing factors. This difference is then divided by the standard deviation, adjusting the mean of each feature data point to 0 and the variance to 1. This allows feature data with different dimensions to be compared on the same scale, resulting in a standard set of wind power influencing factors. The standard set of wind power influencing factors is then decentralized. Decentralization involves subtracting the mean of each feature data point from all its sample values, making the mean of each feature data point zero. This eliminates the influence of the offset of different features on the covariance calculation, ensuring that feature changes only reflect the relative fluctuations of variables, rather than being affected by the absolute value. Finally, a covariance matrix is ​​constructed for the decentralized standard set of wind power influencing factors. The covariance matrix reflects the linear correlation and variation magnitude between features, and is obtained by calculating the covariance between feature i and feature j. The covariance matrix is ​​decomposed using eigenvalues. By solving the characteristic equation Cv = λv, the eigenvalues ​​λ and corresponding eigenvectors v of the covariance matrix are obtained. The eigenvalues ​​represent the contribution of each principal component to the total variance of the data, and the eigenvectors reflect the orientation of each principal component in the original feature space. Based on the eigenvalues, the k principal components with the largest variance contribution are selected. k can be determined by the cumulative variance contribution rate, which is the proportion of the sum of the variances of the k principal components to the sum of the variances of all principal components. Generally, when the cumulative variance contribution rate reaches 80%-90%, these k principal components are considered to contain most of the information in the original data. Based on the selected principal component information, the standard wind power influencing factor feature set is dimensionality-reduced and projected onto the eigenvectors for merging. This transforms the original data from a high-dimensional space to a low-dimensional space while preserving the main features of the data, resulting in the dimensionality-reduced factors for wind farm power.

[0035] Backtesting was conducted to validate the dimensionality-reduced factor set for wind farm power. The principal components obtained from the dimensionality reduction were used as inputs, and historical power data were combined. The contribution and prediction error of each principal component to ultra-short-term wind power prediction were evaluated through backtesting simulation or cross-validation methods, such as calculating the mean square error, correlation coefficient, or contribution rate. Principal components that had little effect on power prediction, low variance contribution, or significant noise in the time series were removed, retaining only those that were sensitive to power changes and had significant prediction performance. After this screening process, a set of wind power influencing factors was obtained, which retained the key features most sensitive to wind power changes while reducing redundancy and noise.

[0036] By standardizing and decomposing the preliminary set of wind power influencing factors, we can eliminate the dimensional differences and correlation interference between different influencing factors, making the data more standardized and easier to analyze. The dimensionality reduction operation of principal component analysis can greatly reduce the dimensionality of the data and reduce the computational complexity. Backtesting and screening can ensure that the final set of wind power influencing factors has high prediction accuracy and reliability, thereby effectively improving the reliability and effectiveness of the entire ultra-short-term wind power prediction.

[0037] Based on the set of factors affecting wind power, an intelligent hybrid model architecture is designed, which includes model architecture information for each type of influencing factor.

[0038] Furthermore, an intelligent hybrid model architecture is designed, including: classifying the set of wind power influencing factors to obtain wind farm influencing factor types, which include static deterministic factors, dynamic semi-deterministic factors, and meteorological uncertainties; determining a wind farm factor type adaptation model set based on the wind farm factor type; performing specific model architecture analysis on the wind farm influencing factor types according to the wind farm factor type adaptation model set to obtain multi-type influencing factor model architecture information; and sequentially combining the multi-type influencing factor model architecture information to construct an intelligent hybrid model architecture.

[0039] Furthermore, based on the types of wind farm influencing factors, a set of wind farm factor type adaptation models is determined, including: performing model adaptability analysis on the static deterministic factors to obtain deterministic factor adaptation model types, which include physical mechanism models; obtaining semi-deterministic factor adaptation model types based on the dynamic semi-deterministic factors, which include rule engine models and data-driven models; obtaining uncertainty factor adaptation model types based on the meteorological uncertainty factors, which include deep learning models and statistical models; and determining a set of wind farm factor type adaptation models based on the deterministic factor adaptation model types, the semi-deterministic factor adaptation model types, and the uncertainty factor adaptation model types.

[0040] Specifically, the wind power influencing factors are classified. First, the characteristics of each influencing factor in each set are analyzed. Based on these characteristics, the wind farm influencing factor types are obtained, including static deterministic factors, dynamic semi-deterministic factors, and meteorological uncertainties. Specifically, the characteristics of each influencing factor in each set are analyzed. If the physical mechanism of the influencing factor is clear, its changes are relatively stable, and it can be accurately predicted through formulas or simulations, such as wind speed, wind direction, air density, temperature, air pressure, terrain roughness, and wake effect, it is classified as a static deterministic factor. If the influencing factor changes over time and is partially controlled by human intervention or operational strategies, but can still be quantified through historical data or rules, such as turbine equipment status, blade pitch angle, power curtailment orders, and grid dispatch plans, it is classified as a dynamic semi-deterministic factor. If the influencing factor is highly random, greatly affected by the environment, and difficult to predict accurately, such as turbulence intensity, sudden wind speed changes, and cloud cover, it is classified as a meteorological uncertainty factor. By classifying, each influencing factor can be matched with the most suitable predictive model type, providing a clear factor hierarchy and processing logic for the architecture design of intelligent hybrid models.

[0041] For different types of wind farm influencing factors, model adaptability analysis was conducted to determine the corresponding wind farm factor type suitable model set. For static deterministic factors, due to their clear physical mechanisms, deterministic factor suitable model types were obtained. Among these, deterministic factor suitable model types include physical mechanism models, which are models constructed based on physical laws and mathematical equations, such as CFD or wake models for power calculation. For example, for wind power generation, according to aerodynamic principles, the wind force F on the wind turbine blades is related to factors such as wind speed v, air density ρ, and blade area A. The wind power calculation expression is P = 0.5ρv. 3 AC p η, where C p Let η be the wind energy utilization coefficient, and η be the generator efficiency. Through physical equations and calculation processes, the input static deterministic factors are transformed into wind power output.

[0042] For dynamic semi-deterministic factors that possess both regularity and data-driven characteristics, a semi-deterministic factor adaptation model type is obtained based on this data characteristic. This semi-deterministic factor adaptation model type includes a rule engine model and a data-driven model, mapping power changes through equipment status and scheduling instructions. Specifically, when analyzing dynamic semi-deterministic factors such as equipment status, blade pitch angle, gearbox temperature, power curtailment instructions, and grid scheduling plans, these factors exhibit partially quantifiable regularities, but are also subject to human intervention or strategy control. Therefore, based on the characteristics of this data type, a semi-deterministic factor adaptation model type is obtained, which includes a rule engine model and a data-driven model. The rule engine model consists of three parts: a rule base, a fact base, and an inference engine. The rule base stores wind farm operation experience rules and scheduling rules, such as reducing power output when the blade pitch angle is greater than a threshold, and reducing unit load when the gearbox temperature exceeds a safe value. The fact base stores the current equipment operating status and scheduling information in real time, such as pitch angle, gearbox temperature, unit speed, power curtailment instructions, and grid scheduling plans. The inference engine uses pattern matching algorithms, such as forward chaining inference, to match the state information in the fact base with the conditional rules in the rule base and trigger the corresponding decision rules, thereby generating power adjustment coefficients or control suggestions. Building upon the rule-based engine model, to further capture the complex relationship between equipment status and power, a data-driven model is introduced for prediction bias learning. Taking a multilayer perceptron (MLP) neural network as an example, its network structure includes an input layer, two hidden layers, and an output layer. The number of nodes in the input layer corresponds to the number of dynamic semi-deterministic factors. The first hidden layer has 64 neurons, and the second hidden layer has 32 neurons. The ReLU activation function is used in the hidden layers to enhance nonlinear expression. The output layer has one node and uses a linear activation function to output the wind power prediction bias value. The data-driven model is expressed as P=f(X dyn ;θ), where X dyn The input matrix is ​​a dynamic semi-deterministic factor matrix, where θ represents the network weights and bias parameters.

[0043] Meteorological uncertainties are characterized by strong randomness, nonlinearity, and dependence on historical data. Therefore, suitable models for these uncertainties include deep learning models and statistical models. Deep learning models, such as LSTM and GRU time-series networks, can capture the time dependence and nonlinear characteristics of wind speed changes. Statistical models, such as ARIMA combined with grey relational analysis for optimization, can predict short-term fluctuations and correct for extreme events. For example, to address the problem of meteorological uncertainties exhibiting significant time correlation and nonlinear fluctuation characteristics, a deep learning prediction model based on Long Short-Term Memory (LSTM) networks can be constructed. LSTM, through cell states and gating mechanisms—input gate, forget gate, and output gate—achieves selective memorization and updating of historical information, effectively capturing the long-term dependence of meteorological variables such as sudden wind speed changes and turbulence variations on wind power. The Long Short-Term Memory (LSTM) network model first constructs an input layer to receive time-series data on meteorological uncertainties. Then, two LSTM hidden layers are set up: the first layer contains 64 LSTM units, and the second layer contains 32 LSTM units, used to extract deep dynamic features from the time series layer by layer. A fully connected layer with 16 nodes is added after the hidden layers, and the ReLU activation function is used to enhance nonlinear expression. Finally, the output layer outputs the predicted wind power value sequence for future time periods. To address the short-term random fluctuations and extreme weather impacts in meteorological factors, an ARIMA model optimized by grey relational analysis is constructed in the statistical prediction layer. First, the stationarity of wind power and related meteorological sequences is tested, such as by the ADF test. If the sequence is non-stationary, a stationary sequence is obtained through differencing. The order parameters (p, d, q) of the ARIMA model are determined based on the autocorrelation function (ACF) and partial autocorrelation function (PACF), where p is the autoregressive order, d is the differencing order, and q is the moving average order. Subsequently, grey relational analysis was introduced to assess the correlation between meteorological factors. By calculating the grey relational degree between each meteorological variable and the wind power series, key variables with high correlation, such as wind speed mutation rate and turbulence intensity, were selected and used as exogenous correction factors for the ARIMA model to compensate for prediction errors under extreme weather or sudden fluctuations. In this way, the statistical model can improve prediction stability under extreme weather conditions while maintaining the linear prediction capability of time series data, and form a complementary prediction structure with the deep learning model to correct prediction biases under extreme weather conditions.

[0044] Then, based on the model types for deterministic factors, semi-deterministic factors, and uncertain factors, a set of model types for wind farm factor adaptation is determined. The set of model types for wind farm factor adaptation includes physical mechanism models, rule engine models, data-driven models, deep learning models, and statistical models.

[0045] After completing model type matching, the model architectures of various factors are analyzed in detail, extracting their input and output variables, feature processing methods, and prediction logic to form multi-type influencing factor model architecture information. Then, these model information are sequentially linked and combined according to time series or functional modules to construct a complete intelligent hybrid model architecture. This intelligent hybrid model architecture can comprehensively utilize physical laws, equipment operation laws, and meteorological randomness information within the same prediction framework, achieving multi-source data fusion and multi-level feature processing, thus improving the accuracy and reliability of wind power prediction. By matching different types of influencing factors to the most suitable prediction model and integrating the various model architectures, the intelligent hybrid model can fully utilize physical knowledge, rule logic, and data-driven prediction capabilities, ensuring accurate prediction of deterministic factors while flexibly responding to highly random meteorological factors, thereby improving the overall prediction accuracy and reliability. For example, wind power influencing factor data for a wind farm over a period of time includes static deterministic factors: average wind speed 10 m / s, wind direction 120 degrees, and air density 1.2 kg / m³. 3 The impact of factors such as temperature (20℃), air pressure (1010hPa), and terrain roughness (0.1m) on wind power can be accurately calculated using CFD and wake models. For dynamic semi-deterministic factors such as gearbox temperature (60℃), no power curtailment orders, and full-load grid operation, a rule-based engine model can determine if the gearbox temperature is normal based on preset rules, issuing an early warning if it exceeds a threshold. Data-driven models, such as neural networks learning the relationship between these factors and wind power from historical data, can be used for prediction. For meteorological uncertainties such as turbulence intensity (0.3), wind speed fluctuation (2m / s), and cloud cover (10%), a deep learning model (LSTM) can capture the changing patterns of these factors over time. Statistical models, including the ARIMA model optimized with grey relational analysis, can filter out highly correlated factors and optimize model parameters for prediction. The model architecture information for different types of factors is sequentially linked and combined. First, the physical mechanism model processes the static deterministic factors and outputs intermediate results. Then, the intermediate results are input together with the dynamic semi-deterministic factors into the rule engine model and the data-driven model to obtain new intermediate results. Finally, the new intermediate results and meteorological uncertainty factors are input into the deep learning model and the statistical model. By combining the output results of each model, an intelligent hybrid model architecture is constructed. The intelligent hybrid model architecture includes the model architecture information of various types of influencing factors.

[0046] By classifying the factors influencing wind power and selecting appropriate models for different types of factors, the advantages of various models can be fully utilized. Physical mechanism models can accurately describe the physical processes of static deterministic factors, rule engine models and data-driven models can flexibly handle changes in dynamic semi-deterministic factors, and deep learning models and statistical models can effectively cope with the randomness of meteorological uncertainties. By sequentially combining the information from multiple model architectures to construct an intelligent hybrid model architecture, the complementary advantages of different models are achieved, improving the adaptability of the intelligent hybrid model to complex wind power influencing factors. This enhances the accuracy and stability of wind power prediction, providing a reliable basis for wind farm operation and scheduling.

[0047] An attention mechanism is introduced to perform multi-scale feature extraction and wind power prediction training on the historical power database of the wind farm based on the set of wind power influencing factors and the intelligent hybrid model architecture, resulting in multiple types of ultra-short-term wind power factor prediction models.

[0048] Furthermore, a multi-class ultra-short-term wind power factor prediction model is obtained, including: performing model variable analysis on the wind power influencing factor set according to the intelligent hybrid model architecture to determine the multi-class power prediction input and output factor variable set; introducing an attention mechanism to extract multi-scale features from the historical power database of the wind farm based on the multi-class power prediction input and output factor variable set to obtain a multi-class factor associated multi-scale feature dataset; and using the intelligent hybrid model architecture to train the multi-class factor associated multi-scale feature dataset for wind power prediction to obtain the multi-class ultra-short-term wind power factor prediction model.

[0049] Furthermore, the intelligent hybrid model architecture is used to train the multi-factor-associated multi-scale feature dataset for wind power prediction, resulting in a multi-class ultra-short-term wind power factor prediction model. This includes: presetting an ultra-short-term prediction time window; performing time-series labeling on the multi-factor-associated multi-scale feature dataset according to the ultra-short-term prediction time window to obtain a multi-factor-associated multi-scale time-series dataset; labeling the input and output variables of the multi-factor-associated multi-scale time-series dataset to obtain a multi-factor-associated multi-scale time-series sample set; and using the intelligent hybrid model architecture to train the multi-factor-associated multi-scale time-series sample set for wind power prediction, resulting in a multi-class ultra-short-term wind power factor prediction model.

[0050] Specifically, based on the intelligent hybrid model architecture, model variable analysis is performed on the set of factors influencing wind power. This involves determining the required input and output variables according to different model types, including physical mechanism models, rule engine models, data-driven models, deep learning models, and statistical models, forming multiple sets of input and output factor variables for power prediction. Furthermore, the input data undergoes z-score standardization or min-max normalization. For example, static deterministic factors such as wind speed, wind direction, and air density are used as basic power calculation input variables; equipment status, pitch angle, and power curtailment orders are used as dynamic correction input variables; and turbulence intensity and sudden wind speed changes are used as random fluctuation prediction input variables. The output variable is uniformly set as the predicted wind power value. By analyzing the variables in the set of factors influencing wind power, the role of each variable in the model input and output is clarified, ensuring that the model can accurately capture the key factors affecting wind power.

[0051] An attention mechanism is introduced to learn feature weights for input variables in a multi-class power prediction input-output factor variable set. This attention mechanism automatically assigns weights by calculating the contribution of different input features to the target prediction result, enabling the model to focus on features that significantly impact power changes. Specifically, after mapping the multi-class input variables in the multi-class power prediction input-output factor variable set to feature vectors, attention weights are calculated using query vectors, key vectors, and value vectors. These weights are then used to weight and fuse the original features. Multi-scale feature extraction is performed by combining data at different time granularities, such as 5 minutes, 15 minutes, and 1 hour, thereby constructing a multi-class factor-related multi-scale feature dataset containing short-term fluctuation and trend features. This multi-class factor-related multi-scale feature dataset provides rich input information for ultra-short-term wind power prediction, enabling the simultaneous capture of power fluctuations caused by sudden wind speed changes, equipment status changes, and meteorological uncertainties. This achieves high-precision prediction of power for the next few minutes to hours, improving the stability and reliability of the prediction.

[0052] After obtaining the multi-factor-related multi-scale feature dataset, time series samples are constructed. First, a short-term forecasting time window is preset, which can be set from a few minutes to several hours to meet the short-term power forecasting needs of actual wind farm operations, for example, set to 15 minutes or 30 minutes. The multi-factor-related multi-scale feature dataset is then time-series labeled according to the short-term forecasting time window, i.e., continuous historical data is organized into time series input samples using a sliding time window method, forming a multi-factor-related multi-scale time series dataset. Then, the multi-factor-related multi-scale time series dataset is labeled with input and output variables: input variables include the characteristics of various influencing factors at historical moments, such as meteorological, equipment, and scheduling data from the past 12 time steps; output variables are the wind power values ​​within the future short-term time window.

[0053] An intelligent hybrid model architecture is adopted to train a multi-scale time series sample set associated with multiple factors. Among them, a multi-class ultra-short-term wind power factor prediction model is constructed through hierarchical training: for static deterministic factors, such as wind speed, wind direction, air density, air pressure and terrain features, a physical mechanism model is used to calculate the base power of the wind farm. The physical mechanism model realizes the power output through power curve formula or computational fluid dynamics (CFD) simulation, without the need for gradient training based on historical data. Secondly, for dynamic semi-deterministic factors, such as unit equipment status, blade pitch angle, and power curtailment commands, a rule engine model is used for decision output. The rule engine includes a rule base, a fact base, and an inference engine, which generates power adjustment information by matching preset rules with real-time input. Meanwhile, for trainable data-driven models, such as multilayer perceptrons (MLP), the inputs are dynamic semi-deterministic factors and historical power data. The mean squared error (MSE) is used as the loss function for training, the optimization algorithm is Adam, the initial learning rate is set to 0.001, the batch size is 64, the maximum number of training rounds is 200, and an early stopping strategy is adopted. Training is stopped when the validation set loss decreases by less than 0.0001 within 10 consecutive rounds to ensure convergence stability. Furthermore, for meteorological uncertainties such as turbulence intensity, sudden wind speed changes, and cloud cover, a deep learning model is used for time-series feature extraction and power prediction training. The training also employs the MSE loss function and the Adam optimization algorithm, with an initial learning rate of 0.001, a batch size of 64, and 200 training epochs, supplemented by an early stopping strategy and gradient pruning to prevent gradient explosion. For statistical models, such as the ARIMA model combined with grey relational analysis for optimization, the power and meteorological sequences are first stabilized, and the autoregression order, differencing order, and moving average order are determined. Model parameters are fitted using historical data, and grey relational analysis is used to correct prediction biases under extreme weather conditions. Through hierarchical training, the intelligent hybrid model architecture generates prediction outputs for different combinations of wind power influencing factors, and forms a set of multiple ultra-short-term wind power factor prediction models based on the type of influencing factors.

[0054] For example, following an intelligent hybrid model architecture, the set of factors influencing wind power is analyzed, identifying wind speed, wind direction, air density, gearbox temperature, and turbulence intensity as input variables, and wind power as the output variable. After introducing an attention mechanism, multi-scale feature extraction is performed on the historical power database. For wind speed data, hourly time windows are used to extract features, yielding hourly average, maximum, and minimum values; daily time windows are used to extract features, revealing daily trends in wind speed. Similarly, multi-scale feature extraction is performed on other influencing factors to obtain a multi-factor-related multi-scale feature dataset. A 15-minute ultra-short-term prediction time window is preset. The multi-factor-related multi-scale feature dataset is then time-series labeled according to this time window, mapping historical feature data to wind power values ​​within the next 15 minutes, resulting in a multi-factor-related multi-scale time-series dataset. Finally, input and output variable labels are applied to this dataset to obtain a multi-factor-related multi-scale time-series sample set. An intelligent hybrid model architecture is used to train a multi-scale time series sample set with multiple factors. After training, a multi-class ultra-short-term wind power factor prediction model is obtained. This multi-class ultra-short-term wind power factor prediction model can predict the wind power value within the next 15 minutes based on factors such as wind speed, wind direction, air density, gearbox temperature, and turbulence intensity at the current moment.

[0055] By dynamically assigning weights to multiple influencing factors through an attention mechanism and training the model with multi-scale time features and intelligent hybrid models, the model can simultaneously capture the short-term fluctuation characteristics and trend changes of wind power. This enables the construction of multiple ultra-short-term wind power factor prediction models for different factors, thereby improving the accuracy and stability of ultra-short-term wind power prediction.

[0056] Dynamic weighted decision-making is performed on the various ultra-short-term wind power factor prediction models to generate an ultra-short-term wind power prediction hybrid model, and wind power closed-loop prediction is performed through the ultra-short-term wind power prediction hybrid model.

[0057] Furthermore, generating a hybrid model for ultra-short-term wind power prediction includes: calculating in real time the model prediction error parameters of the multiple types of ultra-short-term wind power factor prediction models within a preset adjacent time step threshold; determining the dynamic decision coefficients of the hybrid model based on the model prediction error parameters; and using the dynamic decision coefficients of the hybrid model to perform dynamic weighted decision-making on the multiple types of ultra-short-term wind power factor prediction models to generate a hybrid model for ultra-short-term wind power prediction.

[0058] Specifically, during the prediction process, the model prediction error parameters of various ultra-short-term wind power factor prediction models are calculated in real time within a preset adjacent time step threshold. The preset adjacent time step threshold refers to a pre-defined time range, such as selecting data from the past 5 time steps as the calculation basis. For each ultra-short-term wind power factor prediction model, within the preset adjacent time step threshold, the predicted value at each time step is compared with the actual value, and model prediction error parameters such as mean square error and mean absolute error are calculated. The mean square error is used to measure the average of the squares of the differences between the predicted value and the actual value, and the mean absolute error is the average of the absolute values ​​of the differences between the predicted value and the actual value. The model prediction error parameters reflect the prediction accuracy of each model for wind power changes under the current environment.

[0059] Based on the model prediction error parameters, the dynamic decision coefficients of each model are dynamically calculated to determine the dynamic decision coefficients of the hybrid model. For example, using the error reciprocal weighting method, the reciprocal of the prediction error parameter of each model is first calculated, and then the reciprocals of the error parameters of all models are normalized so that the sum of all decision coefficients is 1. The normalized result is the dynamic decision coefficient of the hybrid model of each ultra-short-term wind power factor prediction model. Among them, the smaller the error of the ultra-short-term wind power factor prediction model, the larger its corresponding decision coefficient.

[0060] Based on the dynamic decision coefficients of the hybrid model, a dynamic weighted decision is made on multiple ultra-short-term wind power factor prediction models. The prediction result of each ultra-short-term wind power factor prediction model is multiplied by its corresponding dynamic decision coefficient. Then, the weighted prediction results of all ultra-short-term wind power factor prediction models are summed to obtain the final ultra-short-term wind power prediction hybrid model.

[0061] By integrating the advantages of multiple ultra-short-term wind power factor prediction models through dynamic weighted decision-making, different ultra-short-term wind power factor prediction models may have different prediction performance under different conditions. Dynamic weighted decision-making can adjust the weights of each ultra-short-term wind power factor prediction model in real time according to the recent prediction error of the ultra-short-term wind power factor prediction model, so that the prediction results are more accurate and stable. The generated ultra-short-term wind power prediction hybrid model can dynamically adapt to the performance of each ultra-short-term wind power factor prediction model under different time and environmental conditions, realize high-precision prediction of ultra-short-term wind power, and thus improve the accuracy and stability of ultra-short-term wind power prediction.

[0062] Furthermore, the closed-loop prediction of wind power through the ultra-short-term wind power prediction hybrid model includes: predicting wind power through the ultra-short-term wind power prediction hybrid model, obtaining the wind power prediction deviation value, and performing feedback optimization and closed-loop prediction on the ultra-short-term wind power prediction hybrid model based on the wind power prediction deviation value.

[0063] Specifically, the current set of factors influencing wind power is input into the ultra-short-term wind power prediction hybrid model. The ultra-short-term wind power prediction hybrid model makes wind power predictions through dynamic weighted decision-making to obtain the current wind power prediction value. The wind power prediction value is compared with the actual measured wind power to calculate the wind power prediction deviation value, which is the absolute value of the difference between the wind power prediction value and the actual measured wind power.

[0064] Based on the calculated wind power prediction deviation, a feedback optimization is performed on the ultra-short-term wind power prediction hybrid model. This involves using the wind power prediction deviation as a feedback signal to dynamically adjust the decision weights of each sub-model within the ultra-short-term wind power prediction hybrid model, enabling it to correct prediction errors under specific environmental or conditions. The feedback optimization mechanism can be implemented through a recursive approach or a sliding window method. This accumulates prediction deviations from previous time steps and uses them to update weights or adjust model parameters, while maintaining the predictive capability for the latest input data, thus forming a closed-loop prediction.

[0065] By performing closed-loop prediction through continuous iteration, the hybrid model for ultra-short-term wind power prediction can adaptively adjust under different wind conditions, equipment status, and meteorological conditions, thereby better coping with the randomness and volatility of wind power and further improving the accuracy and stability of ultra-short-term wind power prediction.

[0066] By using closed-loop prediction, the deviation between the prediction model and the actual wind power can be captured in a timely manner. The feedback mechanism can be used to dynamically adjust and optimize the hybrid model of ultra-short-term wind power prediction. This allows the hybrid model to continuously adapt to the complex operating environment of wind farms and the dynamic changes in wind power, effectively improving the accuracy and reliability of predictions. This provides more accurate decision-making basis for the scheduling, operation and management of wind farms, and helps to improve the absorption capacity of wind power and the stability of the power grid.

[0067] Example 2, based on the same inventive concept as the ultra-short-term wind power prediction method based on the intelligent hybrid model in the previous examples, such as... Figure 2 As shown, this application provides an ultra-short-term wind power prediction device based on an intelligent hybrid model, wherein the ultra-short-term wind power prediction device based on the intelligent hybrid model includes: The influencing factor extraction module 11 is used to construct a historical power database of wind farms, extract influencing factors from the historical power database of wind farms, and obtain a set of wind power influencing factors; the model architecture design module 12 is used to design an intelligent hybrid model architecture based on the set of wind power influencing factors, the intelligent hybrid model architecture including model architecture information of various types of influencing factors; the prediction model acquisition module 13 is used to introduce an attention mechanism to perform multi-scale feature extraction and wind power prediction training on the historical power database of wind farms based on the set of wind power influencing factors and the intelligent hybrid model architecture, and obtain multiple types of ultra-short-term wind power factor prediction models; the wind power prediction module 14 is used to perform dynamic weighted decision-making on the multiple types of ultra-short-term wind power factor prediction models, generate an ultra-short-term wind power prediction hybrid model, and perform closed-loop prediction of wind power through the ultra-short-term wind power prediction hybrid model.

[0068] Furthermore, the influencing factor extraction module 11 is also used to: perform data cleaning and preprocessing on the data from various sources in the historical power database of the wind farm according to the data source application standard to obtain an available wind farm power dataset; obtain the types of wind power influencing factors, which include meteorological factors, equipment factors, time factors, grid dispatching factors, and terrain spatial factors; perform spatiotemporal alignment and influencing factor extraction on the available wind farm power dataset according to the types of wind power influencing factors to obtain a preliminary wind power influencing factor set; and perform principal component analysis and screening on the preliminary wind power influencing factor set to obtain a wind power influencing factor set.

[0069] Furthermore, the influencing factor extraction module 11 is also used to: standardize the preliminary wind power influencing factor set to obtain a standard wind power influencing factor feature set; perform decentralization and feature decomposition on the standard wind power influencing factor feature set to obtain eigenvalues ​​and eigenvectors of the covariance matrix; select principal component information based on the eigenvalues; and based on the principal component information, project the standard wind power influencing factor feature set into the eigenvectors for dimensionality reduction and merge them to obtain a wind farm power dimensionality reduction factor set; and perform backtesting verification and screening on the wind farm power dimensionality reduction factor set to obtain a wind power influencing factor set.

[0070] Furthermore, the model architecture design module 12 is also used to: classify the set of wind power influencing factors to obtain wind farm influencing factor types, which include static deterministic factors, dynamic semi-deterministic factors, and meteorological uncertain factors; determine the wind farm factor type adaptation model set according to the wind farm influencing factor types; perform specific model architecture analysis on the wind farm influencing factor types according to the wind farm factor type adaptation model set to obtain multi-type influencing factor model architecture information; and sequentially combine the multi-type influencing factor model architecture information to construct an intelligent hybrid model architecture.

[0071] Furthermore, the model architecture design module 12 is also used to: perform model adaptability analysis on the static deterministic factors to obtain the model types adapted to the deterministic factors, wherein the model types adapted to the deterministic factors include physical mechanism models; obtain the model types adapted to the semi-deterministic factors based on the dynamic semi-deterministic factors, wherein the model types adapted to the semi-deterministic factors include rule engine models and data-driven models; obtain the model types adapted to the uncertainties based on the meteorological uncertainties, wherein the model types adapted to the uncertainties include deep learning models and statistical models; and determine the wind farm factor type adaptation model set based on the model types adapted to the deterministic factors, the model types adapted to the semi-deterministic factors, and the model types adapted to the uncertainties.

[0072] Furthermore, the prediction model acquisition module 13 is also used to: perform model variable analysis on the wind power influencing factor set according to the intelligent hybrid model architecture to determine the multi-class power prediction input and output factor variable set; introduce an attention mechanism to extract multi-scale features from the wind farm historical power database based on the multi-class power prediction input and output factor variable set to obtain a multi-class factor associated multi-scale feature dataset; and use the intelligent hybrid model architecture to train the multi-class factor associated multi-scale feature dataset for wind power prediction to obtain a multi-class ultra-short-term wind power factor prediction model.

[0073] Furthermore, the prediction model acquisition module 13 is also used to: preset an ultra-short-term prediction time window; perform time-series identification on the multi-factor-related multi-scale feature dataset according to the ultra-short-term prediction time window to obtain a multi-factor-related multi-scale time-series dataset; identify input and output variables on the multi-factor-related multi-scale time-series dataset to obtain a multi-factor-related multi-scale time-series sample set; and use the intelligent hybrid model architecture to train the multi-factor-related multi-scale time-series sample set for wind power prediction to obtain a multi-type ultra-short-term wind power factor prediction model.

[0074] Furthermore, the wind power prediction module 14 is also used to: calculate the model prediction error parameters of the multi-type ultra-short-term wind power factor prediction model in real time within a preset adjacent time step threshold; determine the dynamic decision coefficient of the hybrid model based on the model prediction error parameters; and use the dynamic decision coefficient of the hybrid model to perform dynamic weighted decision-making on the multi-type ultra-short-term wind power factor prediction model to generate an ultra-short-term wind power prediction hybrid model.

[0075] Furthermore, the wind power prediction module 14 is also used to: predict wind power through the ultra-short-term wind power prediction hybrid model, obtain wind power prediction deviation value, and perform feedback optimization and closed-loop prediction on the ultra-short-term wind power prediction hybrid model based on the wind power prediction deviation value.

[0076] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The ultra-short-term wind power prediction method and specific examples based on the intelligent hybrid model in the foregoing embodiment 1 are also applicable to the ultra-short-term wind power prediction device based on the intelligent hybrid model in this embodiment. Through the foregoing detailed description of the ultra-short-term wind power prediction method based on the intelligent hybrid model, those skilled in the art can clearly understand the ultra-short-term wind power prediction device based on the intelligent hybrid model in this embodiment. Therefore, for the sake of brevity, it will not be described in detail here.

[0077] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0078] Obviously, those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A method for ultra-short-term wind power prediction based on an intelligent hybrid model, characterized in that, The method includes: A historical power database of wind farms is constructed, and influencing factors are extracted from the historical power database of wind farms to obtain a set of wind power influencing factors; Based on the set of factors affecting wind power, an intelligent hybrid model architecture is designed, which includes model architecture information for each type of influencing factor. An attention mechanism is introduced to perform multi-scale feature extraction and wind power prediction training on the historical power database of the wind farm based on the set of wind power influencing factors and the intelligent hybrid model architecture, resulting in multiple ultra-short-term wind power factor prediction models. Dynamic weighted decision-making is performed on the various ultra-short-term wind power factor prediction models to generate an ultra-short-term wind power prediction hybrid model, and wind power closed-loop prediction is performed through the ultra-short-term wind power prediction hybrid model.

2. The ultra-short-term wind power prediction method based on an intelligent hybrid model as described in claim 1, characterized in that, The set of factors affecting wind power output was obtained, including: According to the data source application standards, the data from each source in the historical power database of the wind farm are cleaned and preprocessed to obtain the usable wind farm power dataset. The types of factors affecting wind power are obtained, including meteorological factors, equipment factors, time factors, grid dispatching factors, and terrain and spatial factors. According to the types of wind power influencing factors, the available wind farm power dataset is spatiotemporally aligned and influencing factors are extracted to obtain a preliminary wind power influencing factor set. Principal component analysis and screening were performed on the preliminary set of factors affecting wind power to obtain the final set of factors affecting wind power.

3. The ultra-short-term wind power prediction method based on an intelligent hybrid model as described in claim 2, characterized in that, Principal component analysis and screening were performed on the preliminary set of wind power influencing factors to obtain the wind power influencing factor set, which includes: The preliminary set of wind power influencing factors is standardized to obtain a standard set of wind power influencing factor features; The standard wind power influencing factor feature set is decentralized and decomposed to obtain the eigenvalues ​​and eigenvectors of the covariance matrix; Principal component information is selected based on the eigenvalues, and the standard wind power influencing factor feature set is dimensionality-reduced and projected onto the feature vector based on the principal component information for merging, to obtain the wind farm power dimensionality-reduced factor set. The set of factors affecting wind farm power dimensionality reduction was backtested and screened to obtain the set of factors affecting wind power.

4. The ultra-short-term wind power prediction method based on an intelligent hybrid model as described in claim 1, characterized in that, Design an intelligent hybrid model architecture, including: The set of factors affecting wind power is classified to obtain the types of wind farm influencing factors, which include static deterministic factors, dynamic semi-deterministic factors, and meteorological uncertain factors. Based on the types of wind farm influencing factors, determine the wind farm factor type adaptation model set; Based on the wind farm factor type adaptation model set, a specific model architecture analysis is performed on the wind farm influencing factor type to obtain multi-type influencing factor model architecture information; The information of the multi-type influencing factor model architecture is sequentially linked and combined to construct an intelligent hybrid model architecture.

5. The ultra-short-term wind power prediction method based on an intelligent hybrid model as described in claim 4, characterized in that, Based on the types of wind farm influencing factors, a set of wind farm factor type adaptation models is determined, including: A model fit analysis is performed on the static deterministic factors to obtain the model types that the deterministic factors fit, including physical mechanism models. Based on the dynamic semi-deterministic factors, the semi-deterministic factor adaptation model type is obtained, which includes rule engine model and data-driven model; Based on the aforementioned meteorological uncertainties, the uncertainty factor adaptation model types are obtained, including deep learning models and statistical models. Based on the deterministic factor adaptation model type, the semi-deterministic factor adaptation model type, and the uncertain factor adaptation model type, determine the wind farm factor type adaptation model set.

6. The ultra-short-term wind power prediction method based on an intelligent hybrid model as described in claim 1, characterized in that, Multiple ultra-short-term wind power factor prediction models were obtained, including: Based on the intelligent hybrid model architecture, model variable analysis is performed on the set of wind power influencing factors to determine multiple sets of power prediction input and output factor variables. An attention mechanism is introduced to extract multi-scale features from the historical power database of the wind farm based on the set of multi-class power prediction input and output factor variables, thereby obtaining a multi-class factor-related multi-scale feature dataset. The intelligent hybrid model architecture is used to train the multi-scale feature dataset associated with the multi-type factors for wind power prediction, thereby obtaining a multi-type ultra-short-term wind power factor prediction model.

7. The ultra-short-term wind power prediction method based on an intelligent hybrid model as described in claim 6, characterized in that, The intelligent hybrid model architecture is used to train the multi-factor-related multi-scale feature dataset for wind power prediction, resulting in a multi-class ultra-short-term wind power factor prediction model, including: A preset ultra-short-term prediction time window is used to perform time-series labeling on the multi-factor-related multi-scale feature dataset according to the ultra-short-term prediction time window, thereby obtaining a multi-factor-related multi-scale time-series dataset. Input and output variable identification is performed on the multi-factor-associated multi-scale time series dataset to obtain a multi-factor-associated multi-scale time series sample set. The intelligent hybrid model architecture is used to train the wind power prediction on the multi-scale time series sample set associated with the multi-type factors, so as to obtain a multi-type ultra-short-term wind power factor prediction model.

8. The ultra-short-term wind power prediction method based on a smart hybrid model as described in claim 1, characterized in that, Generate a hybrid model for ultra-short-term wind power prediction, including: Real-time calculation of the model prediction error parameters of the various ultra-short-term wind power factor prediction models within a preset adjacent time step threshold; Based on the model prediction error parameters, determine the dynamic decision coefficients of the hybrid model; The dynamic decision coefficients of the hybrid model are used to dynamically weight the multi-type ultra-short-term wind power factor prediction models to generate an ultra-short-term wind power prediction hybrid model.

9. The ultra-short-term wind power prediction method based on an intelligent hybrid model as described in claim 1, characterized in that, Wind power closed-loop prediction is performed using the aforementioned ultra-short-term wind power prediction hybrid model, including: Wind power prediction is performed using the ultra-short-term wind power prediction hybrid model to obtain wind power prediction deviation values. Based on the wind power prediction deviation values, the ultra-short-term wind power prediction hybrid model is then optimized using feedback and closed-loop prediction.

10. A short-term wind power prediction device based on an intelligent hybrid model, characterized in that, The step of implementing the ultra-short-term wind power prediction method based on a smart hybrid model according to any one of claims 1 to 9, wherein the ultra-short-term wind power prediction device based on a smart hybrid model comprises: The influencing factor extraction module is used to construct a historical power database of wind farms, extract influencing factors from the historical power database of wind farms, and obtain a set of influencing factors for wind power. The model architecture design module is used to design an intelligent hybrid model architecture based on the set of wind power influencing factors. The intelligent hybrid model architecture includes model architecture information for each type of influencing factor. The prediction model acquisition module is used to introduce an attention mechanism to perform multi-scale feature extraction and wind power prediction training on the historical power database of the wind farm based on the set of wind power influencing factors and the intelligent hybrid model architecture, so as to obtain multiple types of ultra-short-term wind power factor prediction models. The wind power prediction module is used to perform dynamic weighted decision-making on the multiple ultra-short-term wind power factor prediction models, generate an ultra-short-term wind power prediction hybrid model, and perform closed-loop prediction of wind power through the ultra-short-term wind power prediction hybrid model.

Images