A wind power prediction method based on adaptive multi-source feature selection

Abstract

This invention discloses a wind power prediction method based on adaptive multi-source feature selection, belonging to the field of wind power prediction technology. Addressing the shortcomings of existing methods, this method collects multi-source heterogeneous data and constructs an initial feature set. Then, it decomposes the prediction task into multiple related but differently distributed sub-tasks based on season or wind farm geographical location. Within each sub-task, an adaptive feature selection strategy based on population information entropy is employed, dynamically adjusting the binarization threshold and flip probability in real time, and guiding the flip direction based on feature importance. Between sub-tasks, a cross-task value assessment and knowledge transfer strategy is used to evaluate the value of knowledge sources, select the optimal knowledge source for weighted aggregation, and adaptively adjust the learning rate. After iterative optimization, the optimal feature subset for each sub-task is output, and LSTM prediction models are trained accordingly. This invention effectively reduces feature dimensionality, significantly improves prediction accuracy, and adapts to different seasons and wind farm characteristics.

Description

Technical Field

[0001] This invention belongs to the field of wind power prediction technology, specifically relating to a wind power prediction method based on adaptive multi-source feature selection. Background Technology

[0002] Accurate wind power forecasting is crucial for grid dispatching, renewable energy integration, and electricity market transactions. Short-term (24-hour) wind power forecasting typically relies on multi-source information, including numerical weather prediction (NWP) data, historical power data, and turbine status data. However, the raw features collected from actual wind farms have extremely high dimensionality and exhibit significant redundancy and nonlinear coupling, specifically manifested in the following ways: First, the dimensionality of NWP grid point data explodes. A single wind farm typically contains dozens of grid points (such as a 7×7 grid), and each grid point contains meteorological elements such as wind speed, wind direction, temperature, humidity, air pressure, and cloud cover at multiple altitude levels, resulting in hundreds of dimensions for NWP features alone. The correlation coefficient of wind speed sequences between adjacent grid points is often higher than 0.9, and a large number of features carry redundant information.

[0003] Second, the data is mixed with heterogeneous features from multiple sources. In addition to NWP, it is also necessary to integrate historical power sliding windows (hourly power, statistical values, etc. for the previous 24 hours), wind turbine status (pitch angle, speed, gearbox temperature, etc.), observations from adjacent meteorological stations, time features (periodic coding for hours / weeks / months, holiday markers), and derived features based on historical wind speed (rate of change, moving average, etc.). The total feature dimension usually exceeds 800 dimensions, while the sample size is limited by the number of historical years (usually only 8760 hours / year), resulting in a severe "curse of dimensionality".

[0004] Third, the importance of features adapts to seasonal and wind speed ranges. In low wind speed ranges, subtle changes in wind direction have a significant impact on power; in high wind speed ranges, the unit may reach its rated power, and the importance of features other than wind speed decreases. The distribution of meteorological features differs significantly between different seasons (such as the spring windy season and the summer light windy season), but the underlying physical laws (such as the wind speed-power conversion curve) have cross-seasonal commonalities.

[0005] Fourth, the need for collaborative forecasting across multiple wind farms. Multiple geographically proximate wind farms share large-scale meteorological processes, but due to differences in terrain and turbine models, feature selection results contain both shared elements (such as wind speed being consistently crucial) and differing elements (such as turbulence characteristics caused by terrain). Existing methods either model each wind farm independently, ignoring knowledge transfer, or use a uniform feature set, which cannot adapt to the characteristics of each wind farm.

[0006] For the high-dimensional feature selection problem, existing technologies mainly fall into three categories: filtering methods (such as Pearson correlation coefficient and maximum information coefficient) are computationally efficient but ignore the interaction between features; wrapping methods (such as genetic algorithms and particle swarm optimization) evaluate subset quality through classifier feedback, but single-task search is prone to getting trapped in local optima and has high computational cost; embedded methods integrate feature selection into model training, but lack the ability to adapt to adaptive changes. In recent years, evolutionary multi-task optimization has been introduced into high-dimensional feature selection, improving search efficiency by decomposing the original problem into multiple related sub-tasks and sharing knowledge. However, existing EMTO methods are all geared towards general classification tasks (such as gene expression data) and are not specifically designed for the multi-source heterogeneous features, seasonal variation characteristics, and multi-wind farm scenarios of wind power prediction. At the same time, the feature selection parameters are fixed, the knowledge transfer source is randomly selected, and there is a lack of prior protection mechanisms for the importance of features (such as wind speed-related features should be preferentially retained).

[0007] In summary, there is an urgent need for a wind power prediction feature selection method that can integrate multi-source heterogeneous features, adaptively adjust screening parameters, and realize value-perceived knowledge transfer between multiple tasks, in order to solve the technical challenges of high dimensionality, strong redundancy, nonlinear coupling, and cross-task commonality mining. Summary of the Invention

[0008] To address the technical challenges of existing methods in solving high-dimensionality, strong redundancy, nonlinear coupling, and cross-task commonality mining, this invention provides a wind power prediction method based on adaptive multi-source feature selection.

[0009] To achieve the above objectives, the present invention employs the following technical solutions: A wind power prediction method based on adaptive multi-source feature selection, characterized by the following steps: Step 1: Multi-source heterogeneous data acquisition and original feature set construction: Acquire raw data from multiple wind farms within the target area. The raw data includes numerical weather forecast grid data, historical power data of each wind farm, wind turbine status data, observation data from adjacent meteorological stations, time feature data, and derived features based on historical wind speed. The six types of raw data are combined to form an original feature set, which has a dimension of no less than 800. Step 1 includes: Step 1.1: The numerical weather forecast grid data includes wind speed, wind direction, temperature, humidity, air pressure and cloud cover at multiple altitude levels. The number of grid points is no less than 49, the number of meteorological elements is no less than 10, and the data constitutes a feature of no less than 490 dimensions. Step 1.2: The historical power data of each wind farm includes hourly power for the previous 24 hours, statistical power values ​​for the previous 24 hours, and power at the same time for the previous 7 days, forming a feature of no less than 35 dimensions. Simultaneously, the actual power value of each wind farm at the corresponding time is used as the wind power label. ; Step 1.3: The wind turbine status data includes the pitch angle, rotational speed, nacelle temperature, gearbox temperature, generator speed and active power of each wind turbine. The number of wind turbines is not less than 50, forming a feature of not less than 300 dimensions. Step 1.4: The observation data of the adjacent meteorological stations includes the observation values ​​of three adjacent meteorological stations. Each meteorological station includes wind speed, wind direction, temperature, humidity and air pressure, forming a feature of no less than 15 dimensions. Step 1.5: The time feature data includes sine / cosine codes for hours, days of the week, and months, as well as holiday identifiers and season identifiers, forming a feature of no less than 11 dimensions; Step 1.6: The derived feature data based on historical wind speed includes wind speed statistics and wind speed change rates for the previous 6 hours, 12 hours, and 24 hours, constituting features of no less than 10 dimensions.

[0010] Step 2: Multi-task decomposition based on linear and nonlinear correlation: Divide the original feature set into multiple sub-tasks, each sub-task corresponding to a feature selection sub-problem. Each feature selection sub-problem shares the same original feature space but has different feature distributions. Specifically, the original feature set can be divided into four seasonal sub-tasks according to the season: spring, summer, autumn, and winter, with each seasonal sub-task corresponding to a feature selection sub-problem; or into multiple wind farm sub-tasks according to the geographical location of the wind farm, with each wind farm sub-task corresponding to a feature selection sub-problem. Step 2 includes: Step 2: Multi-task decomposition based on linear and nonlinear correlation: Divide the original feature set into four seasonal sub-tasks according to the season: spring, summer, autumn and winter. Each seasonal sub-task corresponds to a feature selection sub-problem; or divide it into multiple wind farm sub-tasks according to the geographical location of the wind farm. Each wind farm sub-task corresponds to a feature selection sub-problem; each feature selection sub-problem shares the same original feature space but has different feature distributions. Step 2.1: Calculate the linear correlation weights; for each original feature , To determine the total number of features, the Pearson correlation coefficient (PCC) was used to calculate its correlation with the wind power label. The degree of linear correlation is expressed by the formula: ; in, For the first The sample mean of each feature, This represents the label mean. Indicates the linear correlation weights; The value range is [0,1], and the larger the value, the stronger the linear correlation. Indicates the total number of samples; Represents the sample index, indicating the first... Each sampling time; Indicates the first In the nth sample The values ​​of each feature; This represents the wind power label corresponding to the i-th sample; Step 2.2: Calculate the nonlinear correlation weights; evaluate the nonlinear relationship using the maximum information coefficient (MIC); for the original features Wind power label Divide the scatter plot into different grids and calculate the maximum value of the normalized mutual information: ; in, Indicates mutual information, Indicates grid division, The constraint function representing mesh generation is typically taken as... , The total number of samples, Represents the non-linear correlation weights. A higher value indicates a stronger non-linear correlation; Representing original features Wind power label The number of partitions in the grid partitioning; Step 2.3: Adjust the linear correlation weights respectively and nonlinear correlation weights The two weighted sequences are independently processed as follows: features are sorted from highest to lowest weight value to obtain a descending sequence. Then, an inflection point detection algorithm is applied to this descending sequence to find the inflection point (i.e., the point of maximum curvature) on the weight curve. Features above the inflection point constitute a highly correlated subset, and features below the inflection point constitute a low-correlation subset.

[0011] This results in four subsets: : A highly correlated subset based on PCC weights; : A highly correlated subset based on MIC weights; : A subset of low-correlation subsets based on PCC weights; : Low-correlation subsets based on MIC weights; Step 2.4: Construct four subtasks, generating four complementary optimization tasks based on the highly relevant subset and the low-relevance subset: Task 1: Optimize using only a subset of highly correlated PCC cases; the optimization space is... (That is, perform feature search within this subset); Task 2: Randomly add a small number of low-correlation features to the high-correlation subset of PCC, with an optimization space of... (That is, adding some low-correlation features of PCC to the high-correlation subset of PCC to form a hybrid search space), where From A small number of features randomly sampled from the data; Task 3: Using only the highly correlated subset of MIC, the optimization space is... (That is, perform feature search within this subset); Task 4: Randomly add a small number of low-correlation features to the high-correlation subset of MIC, with an optimization space of... (That is, adding some low-MIC features to the high-MIC subset to form a hybrid search space) From A small number of features randomly sampled from the data.

[0012] In this way, the four sub-tasks focus on linearly highly correlated features, linearly mixed features, nonlinearly highly correlated features, and nonlinearly mixed features, respectively. Through subsequent knowledge transfer mechanisms, they learn from each other and can fully explore the complex relationship between features and power.

[0013] Step 3: Population initialization and encoding: For each feature selection subproblem, initialize a particle population, where the position of each particle is a binary vector with a dimension equal to the total number of original features, used to indicate whether the corresponding feature is selected; at the same time, initialize the velocity vector of each particle; Step 3.1: Particle Encoding. For each feature selection sub-problem, initialize a particle swarm. The position of each particle is a D-dimensional continuous position vector (D is the total number of original features), denoted as... ,in , Indicates the first One feature was selected. Indicates the first One feature was not selected; each particle also has a velocity vector of the same dimension. ,in, , This represents the preset maximum speed, used to limit the single-step update range of particles and prevent particles from exceeding the search space.

[0014] Step 3.2: Population initialization. For each subtask, set the population size to [value]. Each binary feature selection vector is randomly initialized to 0 or 1 with a probability of 0.5; each velocity dimension is uniformly and randomly initialized in the interval [−1,1]; at the same time, the individual optimal position (initially itself) and the global optimal position of the population for each particle are recorded; Step 3.3: Definition of the fitness function; the fitness function is used to evaluate the quality of a feature subset, taking into account both prediction accuracy and the number of features; for binary feature selection vectors Define fitness: ; in, This represents the classification accuracy of the feature-trained classifier on the validation set. The number of selected features, This indicates the selection status of the j-th feature; a value of 1 indicates that the feature is selected, and a value of 0 indicates that the feature is not selected. α is a balance coefficient, with a value range of [0,1]. Step 4: Adaptive feature selection based on population information entropy: Within each feature selection sub-problem, the frequency of each feature being selected in the population is calculated in real time to obtain the population information entropy. The binarization threshold and feature flipping probability are adaptively adjusted according to the population information entropy value. When flipping, important features are prioritized for protection based on the correlation between the feature and wind power. Step 4 includes: Step 4.1: Calculate the frequency of selection for each feature in the current population, and then calculate the population information entropy; count the selection frequency of all particles in the current population for each feature dimension, assuming the population size is... For the first Dimensional features, selected frequency: ; in, For the first The first particle The dimension can take the value (0 or 1); The formula for calculating the information entropy of the entire population is as follows: ; in, Represents the total number of features. Indicates the first The frequency with which a feature is selected in the current population. This represents a preset small constant used to avoid zero values ​​in logarithmic operations; The population information entropy The value range of is [0,1]. When the population information entropy When the population information entropy is determined to be in a high-entropy state, it is considered to be in a high-entropy state. It is then determined to be a low-entropy state; Step 4.2: Adaptively adjust the binarization threshold. When the population information entropy is higher than the preset threshold, reduce the feature binarization threshold and balance the probability of feature addition and deletion to encourage the exploration of more features; when the population information entropy is lower than the preset threshold, increase the feature binarization threshold and increase the probability of feature deletion to promote feature simplification. The formula for adaptively adjusting the binarization threshold is: ; ; in, Indicates the first Binarization threshold for each feature, Indicates the first The importance weights of each feature and Indicates the preset coefficient; when the population information entropy hour Decrease when population information entropy hour Increase; Step 4.3: Adaptively adjust the flipping probability. In the flipping operation, the features to be flipped are selected based on their importance weights, with wind speed-related features having a higher importance weight than other features; the flipping probability is adjusted during the flipping operation. The adaptive adjustment formula is: ; in, To preset the base flip probability, This represents the current iteration number. This represents the total number of iterations. and For preset coefficients; when Delete features when Add features at the time; The flipping operation is guided by feature importance during the flipping process; when the random number rand < When flipping, a vector is selected based on the current binary characteristics. Reverse the state: If the current binary feature selection vector If a bit with a value of 1 exists, the "delete" operation is performed first: select the feature with the lowest importance from all selected features. Flip it to 0; If the current binary feature selection vector If all values ​​are 0, then perform the "Add" operation: select the most important feature from all unselected features. Flip it to 1; If both 1 and 0 are present, the probability of deletion and addition is determined by the current flip probability. Decision, when Delete features when Add features at the time; Step 4.4: Fitness Comparison and Acceptance; Perform the above flipping operation on each winning particle. If a new binary feature selection vector is obtained after flipping... Calculate its fitness (The specific formula is the fitness function in step 3.3); if Then use the new binary feature selection vector. Replace the original particle's position and update the particle's consecutive position representation; otherwise, retain the original particle.

[0015] Step 5: Cross-task value assessment and adaptive knowledge transfer: A value-aware knowledge transfer strategy is adopted to perform cross-task knowledge transfer among the multiple feature selection sub-problems. The transfer value of the winning particles of other sub-tasks is evaluated from three dimensions: potential, diversity, and relevance. The winning particles with the highest value are selected for weighted aggregation, and the learning rate is adaptively adjusted to update the particles to be optimized in the current sub-task. Step 5 includes: Step 5.1: Starting from the particle to be optimized in the current feature selection subproblem, evaluate the value of the winning particle in other feature selection subproblems. The value is calculated by combining three dimensions: potential, diversity and correlation. Potential is based on the fitness difference between particles, diversity is based on the Hamming distance of the feature selection pattern, and correlation is based on the cosine similarity of the meteorological feature weights between feature selection subproblems. The formula for calculating the potential is: ; in, This indicates the fitness of the particle to be optimized. This indicates the fitness of the candidate winning particle. The fitness function is defined in step 3.3. Indicates a preset small constant; The formula for calculating the diversity is: ; in, The binary feature selection vector representing the candidate winning particle is in the th order. Dimension value, The binary feature selection vector of the particle to be optimized is represented in the th... The value of the dimension; The formula for calculating the correlation is: ; in, This represents the feature weight vector for the current feature selection subproblem. This represents the feature weight vector of the feature selection subproblem to which the candidate winning particle belongs; Calculation of feature weight vectors for each subtask. For each subtask, each feature is calculated independently based on the corresponding training samples. Weights: This forms the feature weight vector for the subtask.

[0016] The comprehensive formula for calculating value is: ; in, , , All are preset weighting coefficients, satisfying + + =1; Step 5.2: Select the winning particle with the highest value as the knowledge source; Step 5.3: Aggregate the feature selection vectors of the selected knowledge sources by value weighting to obtain aggregated knowledge; The weights of the weighted aggregation The calculation formula is: ; in, Indicates the number of selected knowledge sources. Indicates the first The value score of each knowledge source Indicates the first The value score of each knowledge source Indicates the preset scaling factor; Step 5.4: Adaptively adjust the learning rate based on the consistency between the current iteration progress and the value of the selected knowledge source, and update the position of the particle to be optimized using the aggregated knowledge; The formula for calculating the adaptive learning rate η is: ; in, This indicates the preset base learning rate. Indicates the preset scaling factor. Indicates the current iteration number. Indicates the total number of iterations. The standard deviation of the value score of the selected knowledge source is represented; the learning rate is higher in the early stage of the iteration or when the value of the selected knowledge sources differs greatly, and lower in the later stage of the iteration or when the value of the selected knowledge sources tends to be consistent.

[0017] The update method for the particles to be optimized is as follows: ; ; in, This indicates the speed before the update. Indicates the position before the update. Indicates the updated speed. This indicates the updated position, where best_local is the optimal particle position in the current feature selection subproblem. This represents aggregated knowledge. , , A random number within the interval [0,1]. To adaptively adjust the learning rate.

[0018] Step 5.5: For each particle to be optimized, if the updated fitness is better than the original fitness, then accept the update; otherwise, retain the original particle.

[0019] Step 6: Iterative optimization and output of the optimal feature subset: Iteratively execute steps 4 and 5 until the preset number of iterations or convergence condition is reached, and output the corresponding optimal feature subset for each feature selection subproblem; Step 7: Prediction Model Training and Online Prediction: Input the optimal feature subsets of each seasonal task or each wind farm task into the prediction model to train the wind power prediction model for the corresponding season or wind farm, which is then used for online prediction.

[0020] The prediction model mentioned in step 7 is a long short-term memory network model.

[0021] A wind power prediction system based on adaptive multi-source feature selection, the system comprising a data acquisition module, a task partitioning module, a feature selection optimization module, a cross-task knowledge transfer module, and a prediction module; The data acquisition module is used to acquire the raw data described in step 1; The task partitioning module is used to divide the raw data into multiple feature selection sub-problems according to season or wind farm geographical location; The feature selection optimization module is used to perform the adaptive feature filtering based on population information entropy described in step 4 within each feature selection sub-problem; The cross-task knowledge transfer module is used to perform the cross-task knowledge transfer described in step 5. The prediction module is used to train a prediction model and perform wind power prediction by selecting the optimal feature subset output from each feature sub-problem.

[0022] The feature selection optimization module includes a population initialization unit, a fitness evaluation unit, an information entropy calculation unit, a threshold adaptive adjustment unit, and an importance-aware flipping unit; the cross-task knowledge transfer module includes a value evaluation unit, a knowledge source selection unit, an aggregated knowledge calculation unit, a learning rate adaptive unit, and a particle update unit.

[0023] Compared with the prior art, the present invention has the following advantages: First, it achieves deep fusion of multi-source heterogeneous data. This invention is the first to integrate six major categories of information—numerical weather forecast grid data, historical power, unit status, meteorological station observations, time characteristics, and wind speed-derived characteristics—into a unified feature selection framework, with a total dimension exceeding 800. This more realistically reflects the actual data environment of wind farms and overcomes the limitations of existing methods that rely on a single or limited number of data sources.

[0024] Second, the feature selection parameters are adaptive. The adaptive feature selection strategy based on population information entropy proposed in this invention can perceive the diversity status of the population in real time and adaptively adjust the binarization threshold and flip probability accordingly, solving the problem of exploration and development imbalance caused by the use of fixed parameters (such as threshold 0.6 and flip probability 0.7) in traditional methods. At the same time, the introduction of feature importance priors protects key features such as wind speed, enhancing the physical interpretability of the prediction model.

[0025] Third, value-driven cross-task knowledge transfer. This invention comprehensively evaluates the transfer value of knowledge sources from three dimensions: potential, diversity, and relevance, and adopts a relationship-aware selection mechanism to avoid negative transfer that may be caused by random transfer. The adaptive learning rate further balances global exploration and local development, making it particularly suitable for feature selection problems with both commonalities and individual characteristics in multi-seasonal and multi-wind farm scenarios.

[0026] Fourth, the task decomposition method is flexible. Users can choose to decompose the data by season (to adapt to seasonal changes in meteorological distribution) or by the geographical location of wind farms (to adapt to collaborative forecasting of multiple wind farms) based on the actual characteristics of the data, making it highly scalable and practical. Attached Figure Description

[0027] Figure 1 This is an overall flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the particle swarm feature selection process provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of multi-task decomposition based on linear and nonlinear correlation provided in an embodiment of the present invention; Figure 4 A schematic diagram of an adaptive feature selection strategy based on population information entropy provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of a cross-task value assessment and adaptive knowledge transfer strategy provided in an embodiment of the present invention; Figure 6 This is a bar chart comparing the number of features selected by the method of this invention and the comparison algorithm in four seasonal tasks.

[0028] Figure 7 This is a comparison of the fitness convergence curves of the method of this invention and the comparison algorithm on the spring task. Detailed Implementation

[0029] To gain a deeper understanding of this invention, we will provide a comprehensive and detailed description. However, this invention has various implementations and is not limited to the specific examples listed herein. These examples are presented to enhance a full understanding of the disclosure of this invention.

[0030] This embodiment focuses on three geographically close wind farms in a certain region, each with 50 wind turbines. The historical data spans one year (8760 hours), with a sampling interval of 1 hour. Numerical weather forecast (NWP) data uses a 7×7 grid, covering the area surrounding the wind farms. The original feature dimensions total 861, specifically composed as follows: 490 dimensions of NWP grid data (49 grid points × 10 meteorological elements), 35 dimensions of historical power data, 300 dimensions of turbine status data (50 turbines × 6 status variables), 15 dimensions of observation data from adjacent meteorological stations (3 meteorological stations × 5 observations), 11 dimensions of time feature data, and 10 dimensions of derived features based on historical wind speed. The purpose of this invention is to automatically select the optimal feature subset from these 861 features for predicting wind power output over the next 24 hours.

[0031] A wind power prediction method based on adaptive multi-source feature selection, the overall process is as follows: Figure 1 As shown, it includes the following steps: Step 1: Multi-source heterogeneous data acquisition and original feature set construction: Acquire raw data from multiple wind farms within the target area. The raw data includes numerical weather forecast grid data, historical power data, unit status data, observation data from adjacent meteorological stations, time feature data, and derived features based on historical wind speed. The six types of raw data are combined to form an original feature set, which has a dimension of no less than 800. First, raw data is obtained from the wind farm data acquisition and monitoring (SCADA) system, the historical NWP database of the meteorological service provider, and weather station observation records. Specifically, this includes the following six types of data: (1) Numerical weather forecast grid data: Centered on the target wind farm, a total of 49 grid points of 7×7 were selected. Each grid point recorded wind speed and direction at heights of 10 meters, 50 meters, and 100 meters (6 dimensions in total), temperature and humidity at height of 2 meters (2 dimensions), surface air pressure (1 dimension), and total cloud cover (1 dimension). Each grid point contained 10 meteorological elements, totaling 490 dimensions.

[0032] (2) Historical power data of each wind farm: For each wind farm, the hourly power values ​​of the past 24 hours are collected (24 dimensions), and the maximum, minimum, average and standard deviation of power in these 24 hours are calculated (4 dimensions), as well as the power values ​​at the same time (i.e. the same hour) in the past 7 days (7 dimensions), for a total of 35 dimensions.

[0033] (3) Wind turbine status data: For each wind turbine, six status variables are collected: pitch angle, rotor speed, nacelle temperature, gearbox temperature, generator speed, and active power. A total of 300 dimensions are collected for 50 turbines.

[0034] (4) Observation data of adjacent meteorological stations: Select the three meteorological stations closest to the wind farm. Each meteorological station provides five observation values: wind speed, wind direction, temperature, humidity and air pressure, for a total of 15 dimensions.

[0035] (5) Time characteristic data: In order to capture the periodic changes in wind power, the hours (0~23), the days of the week (1~7), and the months (1~12) are sine / cosine encoded (each 2-dimensional), plus holiday identifiers (0 / 1, 1-dimensional), and seasonal unique thermal encoding (spring, summer, autumn, winter, a total of 4 dimensions), for a total of 11 dimensions.

[0036] (6) Derivative characteristics based on historical wind speed: calculate the mean, standard deviation, maximum and minimum wind speed of the previous 6 hours (4 dimensions), the mean and standard deviation of wind speed of the previous 12 hours (2 dimensions), the mean and standard deviation of wind speed of the previous 24 hours (2 dimensions), and the rate of change of wind speed between time t and time t-1, and between time t and time t-3 (2 dimensions), for a total of 10 dimensions.

[0037] The above six types of features are concatenated into an original feature vector with a total dimension of 861, denoted as . Where N is the number of samples (8760 in this example), and D=861. Simultaneously, record the corresponding wind power label. .like Figure 2 As shown, these six categories of features constitute the multi-source heterogeneous original feature set of this invention.

[0038] Step 2: Multi-task decomposition based on linear and nonlinear correlations: The original feature set is divided into multiple sub-tasks, each corresponding to a feature selection sub-problem. Each feature selection sub-problem shares the same original feature space but has different feature distributions. Specifically, the original feature set is divided into four seasonal sub-tasks: spring, summer, autumn, and winter, with each seasonal sub-task corresponding to a feature selection sub-problem; or into multiple wind farm sub-tasks based on the geographical location of the wind farm, with each wind farm sub-task corresponding to a feature selection sub-problem. Due to the high dimensionality of the original features and the presence of mixed linear and nonlinear correlations, this invention decomposes the feature selection problem into multiple complementary sub-tasks to improve search efficiency and fully mine feature information. This embodiment uses a seasonal division method, dividing the annual data into four sub-tasks: spring (March-May), summer (June-August), autumn (September-November), and winter (December-February). Each sub-task shares the same original feature space, but the sample distribution and feature importance differ.

[0039] The specific decomposition process is as follows (see...). Figure 3 ): Step 2.1: Calculate the linear correlation weights; for each original feature , To determine the total number of features, the Pearson correlation coefficient (PCC) was used to calculate its correlation with the wind power label. The degree of linear correlation is expressed by the formula: ; in, For the first The sample mean of each feature, This represents the label mean. Indicates the linear correlation weights; The value range is [0,1], and the larger the value, the stronger the linear correlation. Indicates the total number of samples; Represents the sample index, indicating the first... Each sampling time; Indicates the first In the nth sample The values ​​of each feature; This represents the wind power label corresponding to the i-th sample; Step 2.2: Calculate the nonlinear correlation weights; evaluate the nonlinear relationship using the maximum information coefficient (MIC); for the original features Wind power label Divide the scatter plot into different grids and calculate the maximum value of the normalized mutual information: ; in, Indicates mutual information, Indicates grid division, The constraint function representing mesh generation is typically taken as... , The total number of samples, Represents the non-linear correlation weights. A higher value indicates a stronger non-linear correlation; Representing original features Wind power label The number of partitions in the grid partitioning; Step 2.3: Adjust the linear correlation weights respectively and nonlinear correlation weights The two weighted sequences are independently processed as follows: features are sorted from highest to lowest weight value to obtain a descending sequence. Then, an inflection point detection algorithm is applied to this descending sequence to find the inflection point (i.e., the point of maximum curvature) on the weight curve. Features above the inflection point constitute a highly correlated subset, and features below the inflection point constitute a low-correlation subset.

[0040] This results in four subsets: : A highly correlated subset based on PCC weights; : A highly correlated subset based on MIC weights; : A subset of low-correlation subsets based on PCC weights; : Low-correlation subsets based on MIC weights; Step 2.4: Construct four subtasks, generating four complementary optimization tasks based on the highly relevant subset and the low-relevance subset: Task 1: Optimize using only a subset of highly correlated PCC cases; the optimization space is... (That is, perform feature search within this subset); Task 2: Randomly add a small number of low-correlation features to the high-correlation subset of PCC, with an optimization space of... (That is, adding some low-correlation features of PCC to the high-correlation subset of PCC to form a hybrid search space), where From A small number of features randomly sampled from the data; Task 3: Using only the highly correlated subset of MIC, the optimization space is... (That is, perform feature search within this subset); Task 4: Randomly add a small number of low-correlation features to the high-correlation subset of MIC, with an optimization space of... (That is, adding some low-MIC features to the high-MIC subset to form a hybrid search space) From A small number of features randomly sampled from the data.

[0041] In this way, the four sub-tasks focus on linearly highly correlated features, linearly mixed features, nonlinearly highly correlated features, and nonlinearly mixed features, respectively. Through subsequent knowledge transfer mechanisms, they learn from each other and can fully explore the complex relationship between features and power.

[0042] Step 3: Population initialization and encoding: For each feature selection subproblem, initialize a particle population, where the position of each particle is a binary vector with a dimension equal to the total number of original features, used to indicate whether the corresponding feature is selected; at the same time, initialize the velocity vector of each particle; Step 3.1: Particle Encoding. For each feature selection sub-problem, initialize a particle swarm. The position of each particle is a D-dimensional continuous position vector (D is the total number of original features), denoted as... ,in , Indicates the first One feature was selected. Indicates the first One feature was not selected; each particle also has a velocity vector of the same dimension. ,in, , This represents the preset maximum speed, used to limit the single-step update range of particles and prevent particles from exceeding the search space.

[0043] Step 3.2: Population initialization. For each subtask, set the population size to [value]. Each binary feature selection vector is randomly initialized to 0 or 1 with a probability of 0.5; each velocity dimension is uniformly and randomly initialized in the interval [−1,1]; at the same time, the individual optimal position (initially itself) and the global optimal position of the population for each particle are recorded; Step 3.3: Definition of the fitness function; the fitness function is used to evaluate the quality of a feature subset, taking into account both prediction accuracy and the number of features; for binary feature selection vectors Define fitness: ; in, This represents the classification accuracy of the feature-trained classifier (in this embodiment, a 5-Nearest Neighbor classifier, or 5-NN) on the validation set. Since wind power prediction is a regression problem, in practical applications, the power can be discretized into several intervals, or the root mean square error (RMSE) can be directly used as the accuracy metric. To maintain versatility, this invention uses a 5-NN classifier to evaluate the discriminative power of the feature subset. α represents the number of selected features; α is the balance coefficient, ranging from [0,1]. In this embodiment, α=0.9 is used, which prioritizes prediction accuracy while applying moderate pressure to encourage feature simplification. For regression tasks, ACC can be replaced with the normalized accuracy index 1−NRMSE, where NRMSE is the ratio of root mean square error to power range.

[0044] Step 4: Adaptive feature selection based on population information entropy: Within each feature selection sub-problem, the frequency of each feature being selected in the population is calculated in real time to obtain the population information entropy. The binarization threshold and feature flipping probability are adaptively adjusted according to the population information entropy value. When flipping, important features are prioritized for protection based on the correlation between the feature and wind power. In each iteration of each subtask, the fitness of all particles in the current population is first calculated, and they are sorted from highest to lowest fitness. The first half of the particles are designated as "winners," and the second half as "losers." For the winners, an adaptive feature selection strategy (such as...) is used. Figure 4 (As shown) will be updated; for the losers, they will be updated in step 5 through cross-task knowledge transfer.

[0045] This step describes in detail the process of updating the winner.

[0046] Step 4 includes: Step 4.1: Calculate the frequency of selection for each feature in the current population, and then calculate the population information entropy; count the selection frequency of all particles in the current population for each feature dimension, assuming the population size is... For the first Dimensional features, selected frequency: ; in, For the first The first particle The dimension can take the value (0 or 1); The formula for calculating the information entropy of the entire population is as follows: ; in, Represents the total number of features. Indicates the first The frequency with which a feature is selected in the current population. This represents a preset small constant used to avoid zero values ​​in logarithmic operations; The population information entropy The value range of is [0,1]. When the population information entropy When the population information entropy is determined to be in a high-entropy state, it is considered to be in a high-entropy state. It is then determined to be a low-entropy state; Step 4.2: Adaptively adjust the binarization threshold. When the population information entropy is higher than the preset threshold, reduce the feature binarization threshold and balance the probability of feature addition and deletion to encourage the exploration of more features; when the population information entropy is lower than the preset threshold, increase the feature binarization threshold and increase the probability of feature deletion to promote feature simplification. The formula for adaptively adjusting the binarization threshold is: ; ; in, Indicates the first Binarization threshold for each feature, Indicates the first The importance weights of each feature and Indicates the preset coefficient; when the population information entropy hour Decrease when population information entropy hour Increase; Step 4.3: Adaptively adjust the flipping probability. In the flipping operation, the features to be flipped are selected based on their importance weights, with wind speed-related features having a higher importance weight than other features; the flipping probability is adjusted during the flipping operation. The adaptive adjustment formula is: ; in, To preset the base flip probability, This represents the current iteration number. This represents the total number of iterations. and For preset coefficients; when Delete features when Add features at the time; The flipping operation is guided by feature importance during the flipping process; when the random number rand < When flipping, a vector is selected based on the current binary characteristics. The state selection determines which dimension to flip: If the current binary feature selection vector If a bit with a value of 1 exists (i.e., a feature is already selected), the "delete" operation will be performed first: select the feature with the lowest importance from all selected features. Flip it to 0; If the current binary feature selection vector If all values ​​are 0, then perform the "Add" operation: select the most important feature from all unselected features. Flip it to 1; If both 1 and 0 are present, the probability of deletion and addition is determined by the current flip probability. Decision, when Delete features when Add features at the time; this embodiment uses direct comparison: if If the value is greater than 0.5 and a selected feature exists, then delete the least important feature; if... If the value is less than 0.5 and there are unselected features, then the most important feature is added; if both conditions are met or neither is met, then features are randomly selected with equal probability.

[0047] Step 4.4: Fitness Comparison and Acceptance; Perform the above flipping operation on each winning particle. If a new binary feature selection vector is obtained after flipping... Calculate its fitness (The specific formula is the fitness function in step 3.3); if Then use the new binary feature selection vector. Replace the original particle's position and update the particle's consecutive position representation; otherwise, retain the original particle.

[0048] The above steps are performed independently for each winning particle, resulting in an updated set of winners. For example... Figure 4 As shown, this strategy adaptively adjusts the screening intensity by sensing population diversity and consciously protects important features, thus achieving adaptive and guided feature selection.

[0049] Step 5: Cross-task value assessment and adaptive knowledge transfer: A value-aware knowledge transfer strategy is adopted to perform cross-task knowledge transfer among the multiple feature selection sub-problems. The transfer value of the winning particles of other sub-tasks is evaluated from three dimensions: potential, diversity, and relevance. The winning particles with the highest value are selected for weighted aggregation, and the learning rate is adaptively adjusted to update the particles to be optimized in the current sub-task. Step 5 includes: Step 5.1: Starting from the particle to be optimized in the current feature selection subproblem, evaluate the value of the winning particle in other feature selection subproblems. The value is calculated by combining three dimensions: potential, diversity and correlation. Potential is based on the fitness difference between particles, diversity is based on the Hamming distance of the feature selection pattern, and correlation is based on the cosine similarity of the meteorological feature weights between feature selection subproblems. The formula for calculating the potential is: ; in, This indicates the fitness of the particle to be optimized. This indicates the fitness of the candidate winning particle. The fitness function is defined in step 3.3. Indicates a preset small constant; The formula for calculating the diversity is: ; in, The binary feature selection vector representing the candidate winning particle is in the th order. Dimension value, The binary feature selection vector of the particle to be optimized is represented in the th... The value of the dimension; The formula for calculating the correlation is: ; in, This represents the feature weight vector for the current feature selection subproblem. This represents the feature weight vector of the feature selection subproblem to which the candidate winning particle belongs; Calculation of feature weight vectors for each subtask. For each subtask, each feature is calculated independently based on the corresponding training samples. Weights: This forms the feature weight vector for the subtask.

[0050] The comprehensive formula for calculating value is: ; in, , , All are preset weighting coefficients, satisfying + + =1; Step 5.2: Select the winning particle with the highest value as the knowledge source; Step 5.3: Aggregate the feature selection vectors of the selected knowledge sources by value weighting to obtain aggregated knowledge; The weights of the weighted aggregation The calculation formula is: ; in, Indicates the number of selected knowledge sources. Indicates the first The value score of each knowledge source Indicates the first The value score of each knowledge source Indicates the preset scaling factor; Step 5.4: Adaptively adjust the learning rate based on the consistency between the current iteration progress and the value of the selected knowledge source, and update the position of the particle to be optimized using the aggregated knowledge; The formula for calculating the adaptive learning rate η is: ; in, This indicates the preset base learning rate. Indicates the preset scaling factor. Indicates the current iteration number. Indicates the total number of iterations. The standard deviation of the value score of the selected knowledge source is represented; the learning rate is higher in the early stage of the iteration or when the value of the selected knowledge sources differs greatly, and lower in the later stage of the iteration or when the value of the selected knowledge sources tends to be consistent.

[0051] The update method for the particles to be optimized is as follows: ; ; in, This indicates the speed before the update. Indicates the position before the update. Indicates the updated speed. This indicates the updated position, where best_local is the optimal particle position in the current feature selection subproblem. This represents aggregated knowledge. , , A random number within the interval [0,1]. To adaptively adjust the learning rate.

[0052] Step 5.5: For each particle to be optimized, if the updated fitness is better than the original fitness, then accept the update; otherwise, retain the original particle.

[0053] When the current subtask This step ends after all losers in the population have completed the above updates. Note the following boundary conditions: If no winning particle has been produced in other subtasks (e.g., in the first iteration), cross-task migration is temporarily disabled, and the loser updates solely based on its own best particle. =0; If any candidate particle has a negative or zero potential value (i.e., the candidate particle is not better than the current loser), it will be excluded from the value ranking and will not be considered. If the total number of available candidate particles after screening is less than K, then all available particles are selected, and the K value of the weighted aggregation is adjusted accordingly.

[0054] like Figure 5 As shown, this cross-task knowledge transfer strategy effectively avoids negative transfer that may be caused by random transfer through multi-dimensional value assessment and relation-aware selection, while balancing exploration and development by using an adaptive learning rate.

[0055] Step 6: Iterative optimization and output of the optimal feature subset Repeat steps 4 and 5 until the termination condition is met. In this embodiment, the maximum number of iterations is set to T=100. Furthermore, if the change in population entropy H is less than 0.01 over 10 consecutive iterations, the population is considered to have converged and can be terminated early.

[0056] After each iteration, record the current best individual (the particle with the lowest fitness) in each subtask. Upon termination of the iteration, each subtask outputs its optimal subset of features. (k=1, 2, 3, 4), and the corresponding feature index list.

[0057] Step 7: Predictive Model Training and Online Prediction After obtaining the optimal feature subsets for the four seasonal tasks, they are used to train the corresponding wind power prediction models. This embodiment uses a Long Short-Term Memory (LSTM) network as the predictor because it is good at processing time series data.

[0058] Step 7.1: Data Preparation. For subtasks Extracting the optimal feature subset from the original data The corresponding feature sets form the training set, validation set, and test set. Divided chronologically: the first 70% is the training set, the middle 15% is the validation set, and the last 15% is the test set.

[0059] Step 7.2: LSTM Model Configuration. The LSTM model structure is as follows: the input layer dimension equals the number of selected features; the hidden layer contains two LSTM units, each with 50 neurons; the output layer is a fully connected layer, outputting the wind power for the next 24 hours (24 dimensions). Mean squared error (MSE) is used as the loss function, with the Adam optimizer, an initial learning rate of 0.001, a batch size of 64, 100 training epochs, and an early stopping mechanism (stopping if the validation set loss does not decrease for 10 consecutive epochs).

[0060] Step 7.3: Model Training. Train the LSTM model using the training and validation sets. After training, evaluate the prediction performance on the test set, using metrics including root mean square error (RMSE) and mean absolute percentage error (MAPE).

[0061] Step 7.4: Online Prediction. In actual deployment, the season (spring, summer, autumn, or winter) is determined based on the current date, and the optimal feature subset for the corresponding subtask and the pre-trained LSTM model are selected. Feature values ​​belonging to this feature subset from real-time data are collected, input into the model, and the hourly wind power prediction values ​​for the next 24 hours are output. For multi-wind farm scenarios, if a wind farm-based decomposition method is adopted, the corresponding model and feature subset are selected according to the wind farm to be predicted.

[0062] Step 8: Experimental Verification and Result Analysis To verify the effectiveness of the method proposed in this invention, this embodiment conducts experimental verification on a self-built multi-wind farm dataset.

[0063] Step 8.1: The self-built dataset contains one year of historical data (sampling interval of 1 hour) from three wind farms in a certain region, divided into four subsets according to season: spring (March-May), summer (June-August), autumn (September-November), and winter (December-February). Each seasonal subset is further divided into training set (70%), validation set (15%), and test set (15%) in chronological order. The specific size of the dataset is shown in Table 1.

[0064] Table 1. Partitioning of the self-built dataset

[0065] The original feature set is constructed according to step 1, with a total dimension of 861.

[0066] Step 8.2: Compare the algorithm and parameter settings The following mainstream feature selection algorithms were selected for comparison: CSO, LDW-PSO, MTPSO, MFCSO, MBDPSO, and AAMCSO. The population size for all algorithms was set to 200, the maximum number of iterations to 100, and the balance coefficient α in the fitness function to 0.9. The adaptive parameters of the method in this invention are set as follows: =0.2, =0.1, =0.6, =0.15, =0.05, =0.5, =0.3.

[0067] Step 8.3: Experimental Results and Analysis Step 8.3.1: Comparison of the number of features Figure 6 The bar chart shows the number of features selected by each algorithm on the four seasonal tasks. It can be seen that the method of this invention selects the fewest features (average approximately 92 dimensions) on all seasonal tasks, reducing the number by more than 42% compared to other algorithms. This indicates that the adaptive feature selection strategy based on population information entropy of this invention can effectively eliminate redundant and irrelevant features.

[0068] Step 8.3.2: Convergence Performance Comparison Figure 7 The fitness convergence curves of various algorithms on the spring task are presented. The method of this invention reaches stability at approximately 40 generations and has the lowest final fitness value (approximately 0.072), while other algorithms converge more slowly but have higher final fitness values. This verifies that the cross-task value-aware knowledge transfer strategy can accelerate convergence and improve the quality of solutions.

[0069] Step 8.3.3: Changes in population information entropy To verify the adjustment effect of the adaptive feature selection strategy proposed in this invention, the average values ​​of key parameters at different iteration stages were recorded on the spring task, as shown in Table 2. In the early iteration stage (generations 1-20), the population information entropy... The value is relatively high (approximately 0.82), indicating that the algorithm is in the exploratory phase, and the average binarization threshold is [value missing]. Low (approximately 0.45), flip probability A value close to 0.55 ensures a near-balance between the probabilities of feature addition and deletion, which is beneficial for maintaining population diversity. As iterations continue (generations 21-60), the population information entropy... The value gradually decreases to around 0.50, the threshold and flip probability begin to rise, and the algorithm gradually shifts towards development. Entering the later iteration stage (generations 61-100), the population information entropy... When the threshold drops below 0.30 and rises above 0.65, the flip probability increases to 0.75. At this point, the algorithm tends to delete features, promoting feature subset simplification. This data trend is consistent with the theoretical design, proving the effectiveness of the adaptive adjustment mechanism.

[0070] Table 2. Variation of adaptive parameters with iteration stage in the spring task.

[0071] Step 8.3.4: Comparison of Prediction Accuracy Table 3 shows the prediction accuracy of each algorithm on the test sets for the four seasons (evaluated using LSTM models). The method of this invention achieves the highest accuracy in all seasons, with an average accuracy of 93.72%, which is about 6.5 percentage points higher than the other algorithms. This is especially true in spring (windy season) and winter (cold season).

[0072] Table 3. Prediction accuracy (%) of different algorithms on the test set

[0073] A wind power prediction system based on adaptive multi-source feature selection, the system comprising a data acquisition module, a task partitioning module, a feature selection optimization module, a cross-task knowledge transfer module, and a prediction module; The data acquisition module is used to acquire the raw data described in step 1; The task partitioning module is used to divide the raw data into multiple feature selection sub-problems according to season or wind farm geographical location; The feature selection optimization module is used to perform the adaptive feature filtering based on population information entropy described in step 4 within each feature selection sub-problem; The cross-task knowledge transfer module is used to perform the cross-task knowledge transfer described in step 5. The prediction module is used to train a prediction model and perform wind power prediction by selecting the optimal feature subset output from each feature sub-problem.

[0074] The feature selection optimization module includes a population initialization unit, a fitness evaluation unit, an information entropy calculation unit, a threshold adaptive adjustment unit, and an importance-aware flipping unit; the cross-task knowledge transfer module includes a value evaluation unit, a knowledge source selection unit, an aggregated knowledge calculation unit, a learning rate adaptive unit, and a particle update unit.

[0075] Contents not described in detail in this specification are prior art known to those skilled in the art. Although illustrative specific embodiments of the invention have been described above to facilitate understanding by those skilled in the art, it should be understood that the invention is not limited to the scope of the specific embodiments. Various modifications are readily apparent to those skilled in the art as long as they fall within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of this invention are protected.

Claims

1. A wind power prediction method based on adaptive multi-source feature selection, characterized in that, Includes the following steps: Step 1: Multi-source heterogeneous data acquisition and original feature set construction: Acquire raw data from multiple wind farms within the target area. The raw data includes numerical weather forecast grid data, historical power data of each wind farm, wind turbine status data, observation data from adjacent meteorological stations, time feature data, and derived features based on historical wind speed. The above six types of raw data constitute the original feature set, and the dimension of the original feature set is not less than 800. Step 2: Multi-task decomposition based on linear and nonlinear correlation: Divide the original feature set into multiple sub-tasks, each sub-task corresponding to a feature selection sub-problem. Each feature selection sub-problem shares the same original feature space but has different feature distributions. Specifically, the original feature set can be divided into four seasonal sub-tasks according to the season: spring, summer, autumn, and winter, with each seasonal sub-task corresponding to a feature selection sub-problem; or into multiple wind farm sub-tasks according to the geographical location of the wind farm, with each wind farm sub-task corresponding to a feature selection sub-problem. Step 3: Population initialization and encoding: For each feature selection subproblem, initialize a particle population, where the position of each particle is a binary vector with a dimension equal to the total number of original features, used to indicate whether the corresponding feature is selected; at the same time, initialize the velocity vector of each particle; Step 4: Adaptive feature selection based on population information entropy: Within each feature selection sub-problem, the frequency of each feature being selected in the population is calculated in real time to obtain the population information entropy. The binarization threshold and feature flipping probability are adaptively adjusted according to the population information entropy value. When flipping, important features are prioritized for protection based on the correlation between the feature and the wind power tag. Step 5: Cross-task value assessment and adaptive knowledge transfer: A value-aware knowledge transfer strategy is adopted to perform cross-task knowledge transfer among the multiple feature selection sub-problems. The transfer value of the winning particles of other sub-tasks is evaluated from three dimensions: potential, diversity, and relevance. The winning particles with the highest value are selected for weighted aggregation, and the learning rate is adaptively adjusted to update the particles to be optimized in the current sub-task. Step 6: Iterative optimization and output of the optimal feature subset: Iteratively execute steps 4 and 5 until the preset number of iterations or convergence condition is reached, and output the corresponding optimal feature subset for each feature selection subproblem; Step 7: Prediction Model Training and Online Prediction: Input the optimal feature subsets of each seasonal task or each wind farm task into the prediction model to train the wind power prediction model for the corresponding season or wind farm, which is then used for online prediction.

2. The wind power prediction method based on adaptive multi-source feature selection according to claim 1, characterized in that, Step 1 includes: Step 1.1: The numerical weather forecast grid data includes wind speed, wind direction, temperature, humidity, air pressure and cloud cover at multiple altitude levels. The number of grid points is no less than 49, the number of meteorological elements is no less than 10, and the data constitutes a feature of no less than 490 dimensions. Step 1.2: The historical power data of each wind farm includes hourly power for the previous 24 hours, statistical power values ​​for the previous 24 hours, and power at the same time for the previous 7 days, forming a feature of no less than 35 dimensions. Simultaneously, the actual power value of each wind farm at the corresponding time is used as the wind power label. ; Step 1.3: The wind turbine status data includes the pitch angle, rotational speed, nacelle temperature, gearbox temperature, generator speed and active power of each wind turbine. The number of wind turbines is not less than 50, forming a feature of not less than 300 dimensions. Step 1.4: The observation data of the adjacent meteorological stations includes the observation values ​​of three adjacent meteorological stations. Each meteorological station includes wind speed, wind direction, temperature, humidity and air pressure, forming a feature of no less than 15 dimensions. Step 1.5: The time feature data includes sine / cosine codes for hours, days of the week, and months, as well as holiday identifiers and season identifiers, forming a feature of no less than 11 dimensions; Step 1.6: The derived feature data based on historical wind speed includes wind speed statistics and wind speed change rates for the previous 6 hours, 12 hours, and 24 hours, constituting features of no less than 10 dimensions.

3. The wind power prediction method based on adaptive multi-source feature selection according to claim 2, characterized in that, Step 2 includes: Step 2.1: Calculate the linear correlation weights; for each original feature , Given the total number of original features, the Pearson correlation coefficient is used to calculate the value of each original feature. With wind power label The degree of linear correlation is expressed by the formula: ； in, For the first The sample mean of each feature, The label mean; Indicates the linear correlation weights; Indicates the total number of samples; This represents the sample index, and represents the i-th sampling time. This represents the value of the j-th feature in the i-th sample; This represents the wind power label corresponding to the i-th sample; Step 2.2: Calculate the nonlinear correlation weights; evaluate the nonlinear relationship using the maximum information coefficient; for the original features... Wind power label Divide the scatter plot into different grids and calculate the maximum value of the normalized mutual information: ； in, Indicates mutual information, Indicates grid division, Indicates the non-linear correlation weights; The constraint function represents the mesh generation. The total number of samples, Representing original features Wind power label The number of partitions in the grid partitioning; Step 2.3: Feature subset partitioning; weighting the linear correlations respectively. and linear correlation weights The two weighted sequences are independently processed as follows: the features are sorted from high to low according to their weight values ​​to obtain a descending sequence; then, an inflection point detection algorithm is applied to the descending sequence to find the inflection point on the weight curve; features above the inflection point constitute a highly correlated subset, and features below the inflection point constitute a low-correlation subset. This results in four subsets: : A highly correlated subset based on PCC weights; : A highly correlated subset based on MIC weights; : A subset of low-correlation subsets based on PCC weights; : Low-correlation subsets based on MIC weights; Step 2.4: Construct four subtasks, generating four complementary optimization tasks based on the highly relevant subset and the low-relevance subset: Task 1: Optimize using only a subset of highly correlated PCC features, i.e., perform feature search within this subset. The optimization space is... ; Task 2: Randomly add some low-correlation features from PCC to the high-correlation subset to form a hybrid search space. The optimization space is... ,in From A small number of features randomly sampled from the data; Task 3: Use only the highly relevant subset of MIC, i.e., perform feature search within this subset, with an optimization space of... ; Task 4: Randomly add some low-relevance features of the MIC to the high-relevance subset of MIC to form a hybrid search space. The optimization space is... ,in, From A small number of features randomly sampled from the data.

4. The wind power prediction method based on adaptive multi-source feature selection according to claim 3, characterized in that, Step 3 includes: Step 3.1: Particle Encoding. For each feature selection subproblem, initialize a particle swarm, with each particle's position defined as a unique identifier. 3D continuous position vector, The total number of original features is denoted as . ,in , Indicates the first One feature was selected. Indicates the first One feature was not selected; each particle also has a velocity vector of the same dimension. ,in, , This indicates the preset maximum speed; Step 3.2: Population initialization. For each subtask, set the population size to [value]. Each binary feature selection vector is randomly initialized to 0 or 1 with a probability of 0.5; each velocity dimension is uniformly and randomly initialized in the interval [−1,1]; at the same time, the individual optimal position and the global optimal position of each particle are recorded. Step 3.3: Define the fitness function; for binary feature selection vectors Define fitness: ; in, Represents the total number of original features. This represents the classification accuracy of the feature-trained classifier on the validation set. The number of selected features, α represents the selection status of the j-th feature, where 1 indicates that the feature is selected and 0 indicates that the feature is not selected; α is the balance coefficient, with a value range of [0,1].

5. The wind power prediction method based on adaptive multi-source feature selection according to claim 4, characterized in that, Step 4 includes: Step 4.1: Calculate the frequency of selection for each feature in the current population, and then calculate the population information entropy; count the selection frequency of all particles in the current population for each feature dimension, assuming the population size is... For the first Dimensional features, selected frequency: ; in, For the first The first particle Dimensional values; The formula for calculating the information entropy of the entire population is as follows: ; in, Represents the total number of original features. Indicates the first The frequency with which a feature is selected in the current population. Indicates a preset small constant; The population information entropy The value range is [0,1]. When the population information entropy When the population information entropy is determined to be in a high-entropy state, it is considered to be in a high-entropy state. It is then determined to be a low-entropy state; Step 4.2: Adaptively adjust the binarization threshold. When the population information entropy is higher than the preset threshold, decrease the feature binarization threshold and balance the probability of feature addition and deletion; when the population information entropy is lower than the preset threshold, increase the feature binarization threshold and increase the probability of feature deletion. The formula for adaptively adjusting the binarization threshold is: ; ; in, Indicates the first Binarization threshold for each feature, Indicates the first The importance weights of each feature and Indicates the preset coefficient; when the population information entropy >0.5 Decrease when population information entropy <0.5 Increase; Step 4.3: Adaptively adjust the flipping probability. In the flipping operation, the features to be flipped are selected based on their importance weights, with wind speed-related features having a higher importance weight than other features; the flipping probability is adjusted during the flipping operation. The adaptive adjustment formula is: ; in, To preset the base flip probability, This represents the current iteration number. This represents the total number of iterations. and For preset coefficients; when Delete features when Add features at the time; The flipping operation is guided by feature importance during the flipping process; when the random number rand < When flipping, a vector is selected based on the current binary characteristics. Reverse the state: If the current binary feature selection vector If a bit with a value of 1 exists, the "delete" operation is performed first: select the feature with the lowest importance from all selected features. Flip it to 0; If the current binary vector If all values ​​are 0, then perform the "Add" operation: select the most important feature from all unselected features. Flip it to 1; If both 1 and 0 are present, the probability of deletion and addition is determined by the current flip probability. Decision, when Delete features when Add features at the time; Step 4.4: Fitness Comparison and Acceptance; Perform the above flipping operation on each winning particle. If a new binary feature selection vector is obtained after flipping... Calculate its fitness if Then use the new binary vector Replace the original particle's position and update the particle's consecutive position representation; otherwise, retain the original particle.

6. The wind power prediction method based on adaptive multi-source feature selection according to claim 5, characterized in that, Step 5 includes: Step 5.1: Starting from the particle to be optimized in the current feature selection subproblem, evaluate the value of the winning particle in other feature selection subproblems. The value is calculated by combining three dimensions: potential, diversity and correlation. Potential is based on the fitness difference between particles, diversity is based on the Hamming distance of the feature selection pattern, and correlation is based on the cosine similarity of the meteorological feature weights between feature selection subproblems. The formula for calculating the potential is: ; in, This indicates the fitness of the particle currently being optimized. This indicates the fitness of the candidate winning particle. The fitness function is defined in step 3.

3. Indicates a preset small constant; The formula for calculating the diversity is: ; in, The binary feature selection vector representing the candidate winning particle is in the th order. Dimension value, The binary feature selection vector of the particle to be optimized is represented in the th... The value of the dimension; The formula for calculating the correlation is: ; in, This represents the feature weight vector for the current feature selection subproblem. This represents the feature weight vector of the feature selection subproblem to which the candidate winning particle c belongs; Calculation of feature weight vectors for each subtask: For each subtask, calculate each original feature independently based on the training samples corresponding to that subtask. Weights: This forms the feature weight vector for the subtask; The comprehensive formula for calculating value is: ; in, , , All are preset weighting coefficients, satisfying ; Step 5.2: Select the winning particle with the highest value as the knowledge source; Step 5.3: Aggregate the feature selection vectors of the selected knowledge sources by value weighting to obtain aggregated knowledge; The weights of the weighted aggregation The calculation formula is: ; in, Indicates the number of selected knowledge sources. Indicates the first The value score of each knowledge source Indicates the first The value score of each knowledge source Indicates the preset scaling factor; Step 5.4: Adaptively adjust the learning rate based on the consistency between the current iteration progress and the value of the selected knowledge source, and update the position of the particle to be optimized using the aggregated knowledge; The formula for calculating the adaptive learning rate η is: ; in, This indicates the preset base learning rate. Indicates the preset scaling factor. Indicates the current iteration number. Indicates the total number of iterations. This represents the standard deviation of the value score of the selected knowledge source; The update method for the particles to be optimized is as follows: ; ； in, This indicates the speed before the update. Indicates the position before the update. Indicates the updated speed. Indicates the updated position. Select the optimal particle position in the subproblem for the current feature. This represents aggregated knowledge. , , A random number within the interval [0,1]. To adaptively adjust the learning rate; Step 5.5: For each particle to be optimized, if the updated fitness is better than the original fitness, then accept the update; otherwise, retain the original particle.

7. The wind power prediction method based on adaptive multi-source feature selection according to claim 6, characterized in that, The prediction model mentioned in step 7 is a long short-term memory network model.

8. A wind power prediction system based on adaptive multi-source feature selection, the system being used to implement the wind power prediction method based on adaptive multi-source feature selection as described in any one of claims 1-7, characterized in that, The system includes a data acquisition module, a task partitioning module, a feature selection optimization module, a cross-task knowledge transfer module, and a prediction module. The data acquisition module is used to acquire the raw data described in step 1; The task partitioning module is used to divide the raw data into multiple feature selection sub-problems according to season or wind farm geographical location; The feature selection optimization module is used to perform the adaptive feature filtering based on population information entropy described in step 4 within each feature selection sub-problem; The cross-task knowledge transfer module is used to perform the cross-task knowledge transfer described in step 5. The prediction module is used to train a prediction model and perform wind power prediction by selecting the optimal feature subset output from each feature sub-problem.

9. A wind power prediction system based on adaptive multi-source feature selection according to claim 8, characterized in that, The feature selection optimization module includes a population initialization unit, a fitness evaluation unit, an information entropy calculation unit, a threshold adaptive adjustment unit, and an importance-aware flipping unit; the cross-task knowledge transfer module includes a value evaluation unit, a knowledge source selection unit, an aggregated knowledge calculation unit, a learning rate adaptive unit, and a particle update unit.

Images