A power prediction method considering high-dimensional feature selection and physical guidance

Abstract

This invention relates to the field of wind power prediction technology and discloses a power prediction method considering high-dimensional feature selection and physical guidance. The method includes: acquiring historical operation and meteorological data of wind farms according to a preset time step, constructing a high-dimensional spatiotemporal feature sequence set corresponding to a high-dimensional candidate feature set; calculating the priority score of each feature, combining the score to select a core influence feature set and the corresponding core spatiotemporal feature sequence set, predicting power based on the core sequence and obtaining a fitness value; iteratively optimizing the core influence feature set by combining the fitness value and the feature priority score; obtaining spatiotemporal feature sequences at different time steps according to the optimized feature set, constructing a dynamic map of the wind farm corresponding to each time step according to the spatiotemporal feature sequences, and inputting the dynamic map into a preset power prediction model to obtain the final power prediction value. This invention can significantly improve the accuracy, stability, and interpretability of wind farm power prediction.

Description

Technical Field

[0001] This invention relates to the field of wind power prediction technology, and more specifically to a power prediction method that considers high-dimensional feature selection and physical guidance. Background Technology

[0002] Offshore wind farms, especially large-scale cluster wind farms, have widely distributed turbines and complex meteorological conditions. Their output power exhibits nonlinear, non-stationary, and strongly spatiotemporally coupled characteristics. Their power prediction results are an important reference for the intraday dispatch and safe operation of the power system. However, existing prediction methods are difficult to adapt to the complex prediction needs of such wind farms and have many application limitations.

[0003] Existing power prediction models have poor scenario versatility, requiring remodeling or significant parameter adjustments when wind farm layout, climate, or other conditions change. On one hand, traditional algorithms are often designed for fixed feature dimensions, and feature selection methods struggle to adapt to high-dimensional spatiotemporal features, necessitating a redefinition of model inputs when data structures change. On the other hand, physical models and data-driven models are often simply combined, with physical laws existing in a fixed form that cannot be adaptively adjusted, and data-driven models lacking physical constraints and easily affected by training data distribution. The fusion framework of these two approaches is ill-suited to different wind farm layouts and wake conditions, resulting in insufficient generalization ability. Summary of the Invention

[0004] This invention provides a power prediction method that considers high-dimensional feature selection and physical guidance to solve the problems of feature redundancy, insufficient integration of physical mechanisms and data-driven approaches, and low prediction accuracy and stability in existing wind farm power prediction.

[0005] In a first aspect, the present invention provides a power prediction method considering high-dimensional feature selection and physical guidance, the method comprising: According to a preset time step, acquire several sets of historical operation data and historical meteorological data of the wind farm to be predicted, and construct a high-dimensional spatiotemporal feature sequence set corresponding to the high-dimensional candidate feature set. The high-dimensional candidate feature set includes the operation characteristics of each wind turbine, the meteorological characteristics of the wind farm to be predicted, and the time characteristics related to power generation. Based on the high-dimensional spatiotemporal feature sequence set, calculate the priority score of each feature in the high-dimensional candidate feature set. Combine the priority scores of each feature to select the core influencing feature set from the high-dimensional candidate feature set. Select the core spatiotemporal feature sequence set corresponding to the core influencing feature set from the high-dimensional spatiotemporal feature sequence set, and determine the use of the core spatiotemporal feature sequence set to predict the output power of the wind farm to be predicted, and obtain the fitness value. The fitness value and priority scores of each feature are used to iteratively optimize the core influence feature set to obtain the optimized core influence feature set. Based on the optimized core influence feature set, the spatiotemporal feature sequence of the wind farm to be predicted at different time steps in the current time period is obtained. Based on the spatiotemporal feature sequence, a corresponding dynamic graph of the wind farm is constructed for each time step. The dynamic graph contains several nodes and several edges. One node in the dynamic graph corresponds to one wind turbine. The node features are used to characterize the core influencing factors of the wind turbine corresponding to the node. One edge in the dynamic graph connects two wind turbines. The edge features are used to characterize the wake effect between the two wind turbines connected by the edge. The dynamic graph of the wind farm is input into the preset power prediction model to obtain the power prediction value of the wind farm to be predicted.

[0006] This invention provides a power prediction method that considers high-dimensional feature selection and physical guidance. By acquiring historical wind farm operation and meteorological data at preset time steps and constructing a high-dimensional spatiotemporal feature sequence set, it ensures the comprehensiveness and spatiotemporal correlation of the basic data required for power prediction. Furthermore, it calculates the priority score of each feature and filters the core influencing feature set. By quantifying the nonlinear correlation between each feature and power and the redundancy between features, it clarifies the importance of features and eliminates redundant features, reducing the computational load of subsequent prediction models. It iteratively optimizes the core influencing feature set by combining fitness values ​​and feature priority scores. The fitness value measures the predictive performance of the feature set, while the priority score measures the actual importance of the features. This combination achieves dual verification, continuously optimizing feature combinations and improving the adaptability of core features. Furthermore, it constructs a dynamic wind farm graph containing turbine node and wake edge features based on the spatiotemporal feature sequence. Node embedding preserves the core features of the turbine itself, while wake edge features quantify the wake coupling relationship between units, accurately depicting the spatial correlation and wake effect between turbines. This ensures that the feature representation conforms to the physical laws of actual wind farm operation and adapts to the actual physical scenario of power prediction. Finally, the dynamic graph is input into the preset prediction model to obtain the power prediction value, which effectively improves the accuracy and reliability of wind farm power prediction.

[0007] In one optional implementation, the priority score of each feature in the high-dimensional candidate feature set is calculated, including: Based on a high-dimensional spatiotemporal feature sequence set, a random dependency coefficient is used to calculate the nonlinear correlation score between each feature and the power in the high-dimensional candidate feature set, as well as the nonlinear redundancy score between every two features. Based on the nonlinear correlation score and the nonlinear redundancy score, an incremental search strategy is used to perform round-by-round calculations to obtain the comprehensive score of each feature. The priority score of each feature is determined according to the comprehensive score.

[0008] This invention provides a power prediction method that considers high-dimensional feature selection and physical guidance. Based on a high-dimensional spatiotemporal feature sequence set, it uses random dependency coefficients to calculate the nonlinear correlation score between each feature and power, as well as the nonlinear redundancy score between every two features. This accurately quantifies the strength of the nonlinear correlation between features and power, and the degree of redundancy between features, avoiding the problem that traditional linear evaluation methods cannot adapt to the nonlinear characteristics of high-dimensional spatiotemporal features, and ensuring the relevance and accuracy of the evaluation indicators. Furthermore, based on the above two types of scores, an incremental search strategy is used to calculate the comprehensive score of each feature round by round. This can gradually filter and optimize the feature evaluation results, taking into account both the contribution of features to power and the redundancy between features, avoiding bias caused by a single evaluation dimension. Finally, the priority score of each feature is determined according to the comprehensive score, transforming the abstract feature evaluation results into quantifiable importance scores, clarifying the priority ranking of each feature, and providing a reliable judgment standard for the subsequent selection of the core influencing feature set.

[0009] In one optional implementation, a core influencing feature set is obtained by combining the priority scores of each feature from the high-dimensional candidate feature set, including: The priority scores of each feature are normalized to obtain normalized priority scores. Feature populations are set according to the normalized priority scores. Each feature population contains multiple population individual position vectors. The population individual position vectors are binary encoded to represent the selected or unselected state of each feature. The corresponding core influence feature sets are obtained by decoding the individual position vectors of each population in the feature population.

[0010] The power prediction method considering high-dimensional feature selection and physical guidance provided by this invention can unify the feature importance scores of different orders of magnitude into a reasonable range by normalizing the priority scores, avoiding the interference of numerical differences on subsequent population initialization. The binary encoded feature population and position vector are constructed with the normalization results, which can intuitively and standardly represent the selected and unselected states of features, and is suitable for the processing form of optimization algorithms. Furthermore, the core influence feature set is obtained by decoding the position vector, which can quickly generate multiple candidate feature subsets, providing stable and diverse initial schemes for subsequent iterative optimization.

[0011] In one optional implementation, the core influence feature set is iteratively optimized by combining fitness and priority scores of various features to obtain an optimized core influence feature set, including: Based on fitness values, each core influence feature set is divided into multiple alpha wolf feature sets and multiple ordinary feature sets. Using the binary gray wolf optimization algorithm, the encoding of each feature in each ordinary feature set is iteratively updated following the alpha wolf feature set. The priority scores of each feature are combined to correct the iteratively updated encodings of each feature in each ordinary feature set. The alpha wolf feature set and the updated ordinary feature sets are used as new core influence feature sets. The fitness values ​​of each core influence feature set are recalculated, and the process of dividing each core influence feature set into alpha wolf feature sets and ordinary feature sets based on fitness values ​​is repeated until the iteration stops when the convergence condition is met. The optimized core influence feature set is then selected from the current alpha wolf feature set.

[0012] This invention provides a power prediction method that considers high-dimensional feature selection and physical guidance. It divides the feature set into an alpha wolf feature set and a general feature set based on fitness values, enabling rapid identification of the best-performing feature subset in the current iteration and providing clear guidance for subsequent optimization. Furthermore, based on a binary gray wolf optimization algorithm, the optimal solution for the alpha wolf guides general individuals towards high-quality feature combinations, achieving both global search and local optimization. Priority scoring is then used to correct the selected state, preventing blind updates that could lead to the misselection of low-value features or the omission of high-value features. This ensures the algorithm's optimization capabilities while making the feature selection results more closely reflect actual importance. The newly generated feature set is iteratively updated and its fitness is re-evaluated until convergence is met, continuously refining and optimizing feature combinations to ultimately obtain the optimal alpha wolf feature set, improving subsequent modeling efficiency and power prediction stability.

[0013] In one optional implementation, the dynamic map of the wind farm is processed and input into a preset power prediction model to obtain the predicted power value of the wind farm to be predicted, including: Node embedding learning is performed on the dynamic map of the wind farm at each time step to obtain the node embedding vector of each wind turbine node; the node embedding vectors of the same wind turbine at multiple consecutive time steps are arranged in time sequence to obtain the time sequence feature sequence of each unit; the time sequence feature sequence of each unit is input into the preset power prediction model to obtain the power prediction value of each unit; the power prediction values ​​of each unit at the same time step are summed to obtain the power prediction value of the wind farm to be predicted at the corresponding time step.

[0014] This invention provides a power prediction method that considers high-dimensional feature selection and physical guidance. By performing node embedding learning on the dynamic map of a wind farm, it can transform node features and inter-unit correlation information into low-dimensional vectors that are easy for the model to process, improving data adaptability. Furthermore, arranging the node embedding vectors in time sequence fully characterizes the temporal variation of wind turbine operation, enhancing the expression of temporal features. Further, using the model to predict the temporal feature sequences of each unit enables refined calculation of the power of a single wind turbine, improving local prediction accuracy. Summing the predicted values ​​of each wind turbine yields the total power of the entire wind farm, which conforms to the actual physical structure of the wind farm, improving the accuracy and realism of power prediction.

[0015] In one optional implementation, the step of obtaining edge features includes: Based on the wind turbine location coordinates and the incoming wind direction, calculate the downstream wake distance and radial wake distance between the two wind turbines; input the downstream wake distance and radial wake distance into the continuous wake deficit model to calculate the wake influence coefficient between the two wind turbines, and use the wake influence coefficient as the edge feature of the edge connecting the two wind turbines.

[0016] The power prediction method provided by this invention, which considers high-dimensional feature selection and physical guidance, calculates the downstream wake distance and radial wake distance based on the wind turbine location coordinates and the incoming wind direction. This accurately quantifies the relative spatial position relationship between units, providing reliable geometric parameters for wake effect calculation. By inputting the above distances into a continuous wake deficit model to obtain the wake influence coefficient and using it as an edge feature, the wake attenuation effect that actually exists in the wind farm can be transformed into a quantitative indicator that can be learned by the model. This allows the graph structure to simultaneously contain data features and physical mechanism constraints, improving the model's ability to characterize the spatial coupling relationship between wind turbines and making power prediction more consistent with the actual operating characteristics of wind farms.

[0017] Secondly, the present invention provides a power prediction device that considers high-dimensional feature selection and physical guidance, the device comprising: The high-dimensional spatiotemporal feature acquisition module is used to acquire several sets of historical operation data and historical meteorological data of the wind farm to be predicted according to a preset time step, and construct a high-dimensional spatiotemporal feature sequence set corresponding to the high-dimensional candidate feature set. The high-dimensional candidate feature set includes the operation features of each wind turbine, the meteorological features of the wind farm to be predicted, and the time features related to power generation. The priority score calculation module is used to calculate the priority score of each feature in the high-dimensional candidate feature set based on the high-dimensional spatiotemporal feature sequence set. The core impact feature set acquisition module is used to filter the core impact feature set from the high-dimensional candidate feature set by combining the priority scores of each feature; The fitness value calculation module is used to select the core spatiotemporal feature sequence set corresponding to the core influence feature set from the high-dimensional spatiotemporal feature sequence set, and determine the core spatiotemporal feature sequence set to be used to predict the output power of the wind farm to be predicted, and obtain the fitness value. The iterative optimization module is used to iteratively optimize the core influence feature set by combining the fitness value and the priority score of each feature, so as to obtain the optimized core influence feature set. The current feature sequence acquisition module is used to obtain the spatiotemporal feature sequence of the wind farm to be predicted at different time steps in the current time period based on the optimized core influence feature set; The wind farm dynamic graph construction module is used to construct the corresponding wind farm dynamic graph for each time step based on the spatiotemporal feature sequence. The dynamic graph contains several nodes and several edges. One node in the dynamic graph corresponds to one wind turbine. The node features are used to characterize the core influencing factors of the wind turbine corresponding to the node. One edge in the dynamic graph connects two wind turbines. The edge features are used to characterize the wake effect between the two wind turbines connected by the edge. The power prediction module is used to input the dynamic map of the wind farm into the preset power prediction model to obtain the power prediction value of the wind farm to be predicted.

[0018] Thirdly, the present invention provides an electronic device, comprising: a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the computer instructions to perform the power prediction method considering high-dimensional feature selection and physical guidance described in the first aspect or any corresponding embodiment thereof.

[0019] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to perform the power prediction method considering high-dimensional feature selection and physical guidance described in the first aspect or any corresponding embodiment thereof.

[0020] Fifthly, the present invention provides a computer program product, including computer instructions for causing a computer to execute the power prediction method considering high-dimensional feature selection and physical guidance described in the first aspect or any corresponding embodiment thereof. Attached Figure Description

[0021] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0022] Figure 1This is a schematic diagram of an application scenario according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating a power prediction method that considers high-dimensional feature selection and physical guidance according to an embodiment of the present invention. Figure 3 This is a schematic diagram of the PhyGNN-LSTM algorithm for power prediction considering high-dimensional feature selection and physical guidance according to an embodiment of the present invention. Figure 4 This is a schematic diagram illustrating the determination of upstream and downstream relationships between wind turbines using a power prediction method that considers high-dimensional feature selection and physical guidance according to an embodiment of the present invention. Figure 5 This is a structural block diagram of a power prediction device considering high-dimensional feature selection and physical guidance according to an embodiment of the present invention. Figure 6 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] It is understood that before using the technical solutions disclosed in the various embodiments of the present invention, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in the present invention and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.

[0025] As an optional application scenario of this invention, considering the specific application environment architecture or hardware architecture upon which the power prediction method considering high-dimensional feature selection and physical guidance depends, the specific application environment architecture or hardware architecture is described here. Figure 1 As shown, the architecture system may include at least one terminal device and at least one server. Figure 1 The system is illustrated in the example, which includes a computer 101, a mobile terminal 102, and a server 103, and the terminal devices such as the computer 101 and the mobile terminal 102 are connected to the server 103 through a network 110.

[0026] Specifically, the terminal device can be a smartphone, tablet, laptop, PDA, desktop computer, game console, smart TV, smart wearable device, in-vehicle terminal, VR (Virtual Reality) device, AR (Augmented Reality) device, etc. Server 103 can be a standalone physical server, a server cluster, a distributed system, or a cloud server providing cloud services. Network 110 can be a wired or wireless network, examples of which include, but are not limited to, the Internet, corporate intranet, local area network, wide area network, mobile communication network, and combinations thereof.

[0027] According to an embodiment of the present invention, a power prediction method considering high-dimensional feature selection and physical guidance is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0028] This embodiment provides a power prediction method that considers high-dimensional feature selection and physical guidance, which can be used in the aforementioned mobile terminals, such as mobile phones and tablets. Figure 2 This is a flowchart of a power prediction method considering high-dimensional feature selection and physical guidance according to an embodiment of the present invention, as shown below. Figure 2 As shown, the process includes the following steps: Step S201: Acquire several sets of historical operation data and historical meteorological data of the wind farm to be predicted according to the preset time step, and construct a high-dimensional spatiotemporal feature sequence set corresponding to the high-dimensional candidate feature set.

[0029] In one optional embodiment, a multi-time-step input feature is constructed using a sliding time window. For example, with a sampling interval of 15 minutes, historical operating data of the wind turbines and historical meteorological data of the wind farm to be predicted are collected. Based on these data, a high-dimensional spatiotemporal feature sequence set matching the high-dimensional candidate feature set is constructed. The high-dimensional candidate feature set includes three core features: the operating features of each wind turbine, the meteorological features of the wind farm to be predicted, and the time features related to power generation. The operating features include the active power output of the turbine, wind speed, wind direction, blade angle, nacelle position, etc.; the meteorological features include local meteorological elements of the wind farm, such as wind speed, wind direction, temperature, and air pressure; and the time features include time information, such as the specific time, date, and season.

[0030] Step S202: Calculate the priority score of each feature in the high-dimensional candidate feature set based on the high-dimensional spatiotemporal feature sequence set.

[0031] In one optional embodiment, a high-dimensional spatiotemporal feature sequence set is used as the data basis, and each feature in the high-dimensional candidate feature set is quantitatively evaluated to obtain a priority score that can characterize the importance of each feature.

[0032] Step S203: Combine the priority scores of each feature to select the core influence feature set from the high-dimensional candidate feature set.

[0033] In one optional embodiment, the priority score of each feature is used as the screening criterion to select features that have an important impact on wind farm power prediction from the high-dimensional candidate feature set, and these features are integrated to form a core impact feature set.

[0034] Step S204: Select the core spatiotemporal feature sequence set corresponding to the core influence feature set from the high-dimensional spatiotemporal feature sequence set, and determine the core spatiotemporal feature sequence set to be used to predict the output power of the wind farm to be predicted, and obtain the fitness value.

[0035] In one optional embodiment, feature sequences corresponding to the core influence feature set obtained from the initial screening are extracted from the constructed high-dimensional spatiotemporal feature sequence set to form a core spatiotemporal feature sequence set. This core spatiotemporal feature sequence set is then used to predict and calculate the output power of the wind farm, yielding a fitness value. The fitness value is a quantitative indicator used to evaluate the quality of the current core influence feature set.

[0036] Step S205: Combine the fitness value and the priority score of each feature to iteratively optimize the core influence feature set, and obtain the optimized core influence feature set.

[0037] In one optional embodiment, the core influence feature set obtained from the initial screening is adjusted and optimized in multiple rounds by combining the fitness value and the priority score of each feature, and finally the optimized core influence feature set is obtained.

[0038] Step S206: Obtain the spatiotemporal feature sequence of the wind farm to be predicted at different time steps within the current time period based on the optimized core influence feature set.

[0039] In one optional embodiment, the optimized core impact feature set is used as the feature selection criterion. Relevant data of the wind farm to be predicted are collected within the time period when power prediction is required. Feature information under different time steps is extracted according to the preset time step to construct the corresponding spatiotemporal feature sequence.

[0040] Step S207: Construct a corresponding dynamic map of the wind farm for each time step based on the spatiotemporal feature sequence.

[0041] In one optional embodiment, based on the spatiotemporal feature sequences of different time steps in the current time period, a corresponding dynamic graph of the wind farm is constructed for each time step. The dynamic graph contains several nodes and several edges. Each node in the dynamic graph corresponds to one wind turbine, and the node features are used to characterize the core influencing factors of the wind turbine corresponding to the node. An edge in the dynamic graph connects two wind turbines, and the edge features are used to characterize the wake effect between the two wind turbines connected by the edge.

[0042] Step S208: Input the dynamic map of the wind farm into the preset power prediction model to obtain the power prediction value of the wind farm to be predicted.

[0043] In one optional embodiment, the wind farm dynamic map constructed for each time step is input into a pre-built and trained power prediction model. The model processes and analyzes the features and correlation information in the dynamic map and outputs the power prediction value of the wind farm for each time step.

[0044] The power prediction method provided in this embodiment, which considers high-dimensional feature selection and physical guidance, acquires historical wind farm operation and meteorological data at preset time steps and constructs a high-dimensional spatiotemporal feature sequence set, ensuring the comprehensiveness and spatiotemporal correlation of the basic data required for power prediction. Furthermore, it calculates the priority score of each feature and filters the core influencing feature set. By quantifying the nonlinear correlation between each feature and power and the redundancy between features, it clarifies the importance of features and eliminates redundant features, reducing the computational load of subsequent prediction models. The core influencing feature set is iteratively optimized by combining fitness values ​​and feature priority scores. The fitness value measures the predictive performance of the feature set, while the priority score measures the actual importance of the features. This combination achieves dual verification, continuously optimizing feature combinations and improving the adaptability of core features. Furthermore, a dynamic wind farm graph containing turbine node and wake edge features is constructed based on the spatiotemporal feature sequence. Node embedding preserves the core features of the turbine itself, while wake edge features quantify the wake coupling relationship between units, accurately depicting the spatial correlation and wake effect between turbines. This ensures that the feature representation conforms to the physical laws of actual wind farm operation and adapts to the actual physical scenario of power prediction. Finally, the dynamic graph is input into the preset prediction model to obtain the power prediction value, which effectively improves the accuracy and reliability of wind farm power prediction.

[0045] In some optional implementations, step S202 above, calculating the priority score of each feature in the high-dimensional candidate feature set, includes: Step a1: Based on the high-dimensional spatiotemporal feature sequence set, the nonlinear correlation score between each feature and the power and the nonlinear redundancy score between each pair of features in the high-dimensional candidate feature set are calculated using random dependency coefficients.

[0046] In one optional embodiment, based on the constructed high-dimensional spatiotemporal feature sequence set, the nonlinear correlation degree between a single feature in the high-dimensional candidate feature set and the power generation, as well as the nonlinear redundancy degree between any two features, are calculated using random dependency coefficients to obtain the corresponding nonlinear correlation score and nonlinear redundancy score, which are used to quantitatively describe the strength of the correlation between features and power and the amount of repetitive information between features.

[0047] Specifically, for the aforementioned high-dimensional candidate feature set, the maximum relevance minimum redundancy (mRMR) criterion is introduced to calculate a feature priority score for each candidate feature. Assume the original data used for wind power prediction... Total Given candidate features, for a given feature Its feature priority score based on the mRMR criterion is: (1) in Features Priority score, To predict the target variable, For the selected feature set, for The selected features. The size of the selected feature set; for right The F-statistic; Features and A measure of correlation.

[0048] In the above formula The randomized dependence coefficient (RDC) is a nonlinear correlation measure. RDC measures the nonlinear dependence between two features, denoted as . Its core idea is to perform multiple sets of random nonlinear projections on the Copula transformation of the features, and then calculate the maximum canonical correlation and the maximum correlation coefficient between the projected variables. and will As the final RDC score. The expression is: (2) in, This represents the selected optimal feature subset. After obtaining the RDC score, the improved mRMR criterion can be defined as: (3) Step a2: Based on the nonlinear correlation score and the nonlinear redundancy score, an incremental search strategy is used to perform round-by-round calculations to obtain the comprehensive score of each feature.

[0049] In one optional embodiment, the features are evaluated and filtered round by round using an incremental search strategy, and the comprehensive score of each feature is calculated and updated step by step, so that the comprehensive score can simultaneously reflect the contribution of the feature to the power and the degree of redundancy between features.

[0050] Specifically, an incremental search strategy is adopted, and a feature priority queue is constructed step by step according to the improved mRMR criterion. In the round of iteration, maximizing the improved mRMR criterion of equation (3) yields the optimal feature of that round, and moving towards... Add Features, and As a priority score for this feature, it enters... Round iteration, until Search complete.

[0051] Step a3: Determine the priority score for each feature based on the overall score.

[0052] In one optional embodiment, each feature is assigned a quantitative score to indicate its importance based on the overall score of each feature; this is called a priority score.

[0053] The power prediction method considering high-dimensional feature selection and physical guidance provided in this embodiment is based on a high-dimensional spatiotemporal feature sequence set. It uses random dependency coefficients to calculate the nonlinear correlation score between each feature and power, as well as the nonlinear redundancy score between every two features. This accurately quantifies the strength of the nonlinear correlation between features and power, and the degree of redundancy between features, avoiding the problem that traditional linear evaluation methods cannot adapt to the nonlinear characteristics of high-dimensional spatiotemporal features, ensuring the relevance and accuracy of the evaluation indicators. Furthermore, based on the above two types of scores, an incremental search strategy is used to calculate the comprehensive score of each feature round by round. This allows for the gradual screening and optimization of feature evaluation results, taking into account both the contribution of features to power and the redundancy between features, avoiding bias caused by a single evaluation dimension. Finally, the priority score of each feature is determined based on the comprehensive score, transforming the abstract feature evaluation results into quantifiable importance scores, clarifying the priority ranking of each feature, and providing a reliable judgment standard for the subsequent screening of the core influencing feature set.

[0054] In some optional implementations, step S203 above includes: Step b1: Normalize the priority scores of each feature to obtain normalized priority scores.

[0055] In one optional embodiment, the priority scores of each feature are numerically scaled to unify all scores into the same numerical range, resulting in a normalized priority score, which is used to standardize the numerical basis for subsequent population initialization.

[0056] Step b2: Set up feature populations based on normalized priority scores. Each feature population contains multiple population individual position vectors. The population individual position vectors are encoded in binary to represent the selected or unselected state of each feature.

[0057] In one optional embodiment, a population for feature selection is initialized based on a normalized priority score, with each individual in the population corresponding to a position vector. The position vector uses binary encoding, where each bit (0 or 1) indicates whether the corresponding feature is selected. When a feature is selected, it is included in the core influencing feature set and will participate in the subsequent feature input and model training process for wind farm power prediction. When a feature is not selected, it is determined to be redundant or a low-priority feature, and is removed from the candidate feature set, not participating in subsequent feature processing and modeling related to power prediction.

[0058] Specifically, the population individual location vector is .

[0059] in, To search for the number of individuals in the population, i.e., the number of candidate features, Indicates the first One candidate feature was selected. Then it will not be selected. The position vector of an individual in the population is initialized using feature priority scoring, the first... The first wolf Initialization value of the dimensional search space for: (4) Among them For the first Priority scoring of each feature This is the floor function. These are the weighting coefficients. Uniformly distributed.

[0060] Step b3: Based on the position vectors of individuals in each group of the feature population, decode the corresponding core influence feature sets respectively.

[0061] In one optional embodiment, the binary position vector of each individual in the population is decoded, and corresponding features are selected from the high-dimensional candidate feature set according to the position of "1" in the vector to form the core influence feature set corresponding to that individual.

[0062] The power prediction method considering high-dimensional feature selection and physical guidance provided in this embodiment can unify the feature importance scores of different orders of magnitude into a reasonable range by normalizing the priority scores, avoiding the interference of numerical differences on subsequent population initialization. The binary encoded feature population and position vector are constructed with the normalization results, which can intuitively and standardly represent the selected and unselected states of features, and is suitable for the processing form of optimization algorithms. Furthermore, the core influence feature set is obtained by decoding the position vector, which can quickly generate multiple sets of candidate feature subsets, providing stable and diverse initial schemes for subsequent iterative optimization.

[0063] In some optional implementations, step S205 above includes: Step c1: Divide each core influence feature set into multiple alpha wolf feature sets and multiple ordinary feature sets according to the fitness value.

[0064] In one optional embodiment, the fitness value corresponding to each candidate feature set is used as the sorting criterion, and the feature set with better performance is selected as the alpha wolf feature set, while the remaining feature sets are used as ordinary feature sets. This completes the hierarchical division of individuals within the population and provides a guiding target for subsequent iterative updates.

[0065] Specifically, in the first In this iteration, the BGWO evolutionary algorithm based on the improved mRMR first generates the position vector. and the corresponding feature subset Then The data is input into the learning model of the packaging method, and its predictive performance is evaluated using multidimensional time series labeled data.

[0066] The learning model of the packaging method employs extreme gradient boosting (XGBoost) as the learning algorithm in the feature selection optimization iteration process. By iteratively constructing weak learners and optimizing residuals, XGBoost can capture the nonlinear interactions and dependencies between high-dimensional features, thereby quantifying the global contribution of features to wind power prediction and making it suitable for feature selection tasks in high-dimensional data scenarios.

[0067] During the iterative optimization process, the constructed fitness function is used as the evaluation criterion for the quality of feature subsets, thereby achieving a quantitative evaluation of the feature subset selection effect.

[0068] The fitness function is constructed by considering the following aspects: first, the predictive performance of the feature subset after being calculated by the learning model should be as good as possible; second, the dimensionality of the feature subset should be as small as possible. Therefore, the fitness function... for: (5) in Mean-square error (MSE) for wind power prediction using the XGBoost function of the learning algorithm; , These represent the dimensions of the feature subset and the original feature, respectively. , As weight.

[0069] Step c2 involves using the binary gray wolf optimization algorithm to iteratively update the encoding of each feature in each ordinary feature set following the alpha wolf feature set, and then correcting the iteratively updated encoding of each feature in each ordinary feature set by combining the priority score of each feature.

[0070] In one optional embodiment, based on the Binary Gray Wolf Optimization Algorithm (BGWO), the encoding of each feature in each ordinary feature set is iteratively updated following the alpha wolf feature set. During the update process, priority scoring is added to correct the updated encoding of each ordinary feature set, which utilizes priority scoring while retaining the global exploration capability of the evolutionary algorithm.

[0071] Specifically, to increase the likelihood of higher-priority features being selected during computation, the position update rule was improved: features with higher priority scores are more likely to be selected in the population initialization and position update rules of the BGWO algorithm; at the same time, the randomness of the BGWO algorithm itself is preserved, so that it does not rely entirely on feature priority scores. The update rules are as follows: (6), (7), and (8): (6) (7) (8) in, For the first The optimal solution in the nth iteration Dimensional value; For random parameters, This represents the maximum number of iterations. These are the weighting coefficients; Indicates the first The first wolf 3D search space During the next iteration , , The distance set.

[0072] Step c3: Take the alpha wolf feature set and the updated ordinary feature set as the new core influence feature set, recalculate the fitness value of each core influence feature set, and return to the step of dividing each core influence feature set into the alpha wolf feature set and the ordinary feature set according to the fitness value. Stop the iteration when the iteration meets the convergence condition, and select the optimized core influence feature set from the current alpha wolf feature set.

[0073] In an optional embodiment, all feature sets obtained in this iteration are used as the new generation population. The fitness values ​​are re-evaluated in step c1 above, and the alpha wolf feature set and the ordinary feature set are divided again. This process is repeated iteratively until the convergence condition is met. Then, the current optimal alpha wolf feature set is used as the final optimized feature set, completing the entire feature selection process.

[0074] The power prediction method presented in this embodiment, which considers high-dimensional feature selection and physical guidance, divides the alpha wolf feature set and the ordinary feature set according to fitness values. This allows for the rapid identification of the best-performing feature subset in the current iteration, providing clear guidance for subsequent optimization. Furthermore, based on the binary gray wolf optimization algorithm, the optimal solution of the alpha wolf guides ordinary individuals towards high-quality feature combinations, achieving both global search and local optimization. Priority scoring is then used to correct the selected state, preventing blind updates that could lead to the misselection of low-value features or the omission of high-value features. This ensures the algorithm's optimization capabilities while making the feature selection results more closely reflect actual importance. The newly generated feature set is iteratively updated and its fitness is re-evaluated until convergence conditions are met, continuously refining and optimizing feature combinations to ultimately obtain the optimal alpha wolf feature set, improving subsequent modeling efficiency and power prediction stability.

[0075] In some optional implementations, step S208 above includes: Step d1 involves performing node embedding learning on the dynamic graph of the wind farm at each time step to obtain the node embedding vector of each wind turbine node.

[0076] In one optional embodiment, graph embedding technology is used to map the core influencing factors of wind turbine nodes and the correlation between nodes (wake effect) in the dynamic graph to a low-dimensional dense vector space. This transforms the abstract features of the graph structure into numerical vectors that can be recognized and calculated by the predictive model, while retaining the core feature information of the wind turbine nodes and the correlation characteristics between nodes, thus solving the problem that graph structures are difficult to directly input into the model.

[0077] Specifically, wind farm configurations and corresponding meteorological conditions for different time periods are uniformly encoded into a graph structure. A directed graph is defined. This provides a graphical input for the subsequent PhyGNN-LSTM (Physically Guided Graph Neural Network) algorithm, which combines features of wind turbine nodes, edges, and global features.

[0078] The first in the wind farm Node attribute vector of typhoon The set, which includes the node characteristics of each wind turbine node. Set as the corresponding time period Meteorological characteristic data of each wind turbine are used to make the meteorological information of all wind turbine nodes at the same time step their basic attributes. For the node Pointing to node edge attributes A set of data used to characterize the interactions between wind turbines. Used to describe the macroscopic properties of the entire wind farm as a whole. In the graphical representation of this embodiment, The wind speed of the wind farm is set to the global wind speed of the corresponding time period, which enables the model to perceive the overall wind conditions at a unified graph level. Together with the wind speed setting of the node features, it provides PhyGNN-LSTM with comprehensive wind field information.

[0079] In summary, the graphical representation of a wind farm is shown in each time period. In the process, relevant meteorological and geographical information is extracted, and nodes, edges, and global features are dynamically calculated and constructed based on this information, thereby forming a graph structure for input to the PhyGNN-LSTM algorithm.

[0080] Step d2 involves arranging the node embedding vectors of the same wind turbine at multiple consecutive time steps in chronological order to obtain the chronological feature sequence of each unit.

[0081] In one optional embodiment, the power output of the wind farm has temporal correlation. The operating status and feature information of the same wind turbine at different continuous time steps have a sequential dependency relationship. By arranging the node embedding vectors in time sequence, the feature patterns of a single wind turbine changing over time can be fully captured, providing temporal dimension feature support for subsequent single wind turbine power prediction.

[0082] Step d3: Input the time-series characteristic sequence of each unit into the preset power prediction model to obtain the power prediction value of each unit.

[0083] In one optional embodiment, the preset power prediction model can perform in-depth analysis of the time series feature sequence, explore the correlation between the wind turbine time series features and its own output power, and combine the information of the core features of the wind turbine contained in the embedded vector to accurately calculate the power prediction result of each wind turbine at the corresponding time step, thereby achieving accurate prediction of the power of a single wind turbine.

[0084] Specifically, the preset power prediction model is built based on the PhyGNN-LSTM algorithm and mainly consists of two parts. First, the processed graph-structured data is input into the PhyGNN algorithm, which also outputs graph-structured data. Then, the nodes in the graph are mapped as vectors and input into the LSTM algorithm to extract temporal features, enabling power prediction for a single wind turbine in future time periods. Finally, the power of all wind turbines in the wind farm is summed to obtain the predicted power value for the wind farm in future time periods. The PhyGNN-LSTM algorithm is detailed as follows: Figure 3 As shown.

[0085] First, the original dataset for this time period. Feature selection is performed to obtain the selected data. And then convert it into a graphical representation using GNN. ,in accordance with The computing node emits features Received features With edge features . Based on this, the updated edge features are calculated. , in accordance with Dynamically calculate edge weights Finally, the adjusted edge features are obtained. : (9) (10) (11) in, , They are respectively , The parameters; this weighting mechanism can adjust the degree of mutual influence according to the physical relationship between wind turbines, so that the weights have physical meaning. The edge update step corresponds to the calculation of the average deficit factor. The process. The physical guidance function of the wake model is used to calculate the intensity of the interaction between the two wind turbines: (12) in, , Induction coefficient in wake deficit factor The parameters, Rotor radius The parameters, For surface roughness Parameters; and They are respectively Time period , The set of downstream wake distances and the set of radial wake distances. The calculation process and wake deficit factor The calculation process is the same.

[0086] Node update function Using node features , and To update node features, the updated node features for: (13) in, for The parameters. The node update step is to use the aggregation of the average deficit factor to calculate the average wind speed. .

[0087] Global feature update function right Perform an update to obtain the updated global features. : (14) By applying the update function described above, the physically guided GNN layer can be generated from the initial wind farm map. Get the updated image Using it as input to the graph dense layer, we obtain the graph. as follows: (15) in, For graph dense layer functions, Its network parameters.

[0088] Will The nodes in the vector are mapped to obtain , go through The LSTM layer compresses the time series dimension and extracts time series features, outputting the future performance of each wind turbine. Wind power during the period .

[0089] Using predicted power With actual power The sum of the MSEs between them is used as the loss function for model training. The training process employs backpropagation and a gradient descent optimizer, training the PhyGNN and LSTM parts together. During backpropagation, the gradient of the loss function with respect to all trainable parameters of the model, including those in the PhyGNN layers, is calculated. , , Parameters Physical parameters in And the parameters of the LSTM layer. It can be expressed as the following formula: (16) in, This represents the total number of samples.

[0090] Step d4: Sum the power prediction values ​​of each unit at the same time step to obtain the power prediction value of the wind farm to be predicted at the corresponding time step.

[0091] In one optional embodiment, the overall output power of the wind farm is equal to the sum of the output power of all individual wind turbines in the field during the same period. By summing the predicted power values ​​of individual wind turbines, the prediction results of individual wind turbines can be integrated into the overall power prediction result of the wind farm.

[0092] The power prediction method provided in this embodiment, which considers high-dimensional feature selection and physical guidance, transforms node features and inter-unit correlation information into low-dimensional vectors that are easy for the model to process by performing node embedding learning on the dynamic map of the wind farm, thereby improving data adaptability. The node embedding vectors are then arranged chronologically to fully characterize the temporal variation of wind turbine operation, enhancing the expression of temporal features. Furthermore, by using the model to predict the temporal feature sequences of each unit, refined calculation of the power of a single wind turbine can be achieved, improving local prediction accuracy. The summation of the predicted values ​​for each wind turbine yields the total power of the entire wind farm, which conforms to the actual physical structure of the wind farm, improving the accuracy and realism of power prediction.

[0093] In some optional implementations, the step of obtaining edge features includes: Step e1: Calculate the downstream wake distance and radial wake distance between the two fans based on the fan location coordinates and the incoming wind direction.

[0094] In one optional embodiment, in a wind farm, the interaction between the wind and the upstream wind turbine generates a wake effect. Based on the wind turbine spatial coordinates and the incoming wind direction, the relative distance between the two wind turbines is decomposed along the wind direction and the vertical wind direction, respectively, to obtain the downstream wake distance along the flow direction and the radial wake distance vertical to the flow direction.

[0095] Step e2: Input the downstream wake distance and radial wake distance into the continuous wake deficit model, calculate the wake influence coefficient between the two wind turbines, and use the wake influence coefficient as the edge feature of the edge connecting the two wind turbines.

[0096] In one alternative embodiment, the wake effect affects the power of downstream wind turbines, so a quantitative characterization of the power between wind turbines is constructed based on a continuous wake deficit model.

[0097] Specifically, at the downstream wake distance and radial wake distance Below, describe the steady-state wind speed. The wake model is as follows: (17) in, As the tailflow deficit factor, quantified in The amount of wind speed reduction at the location; For incoming wind speed. Specifically, it can be expressed as: (18) in, , and These are the induction coefficient, rotor radius, and surface roughness, respectively. (Wind fan) Due to the fan The resulting average deficit factor for: (19) in, For wind turbine The wind turbine swept across the area, for One point in the middle, For wind turbine The radius of the area swept. , Indicates from the wind turbine center to Downstream of the point, radial wake distance.

[0098] At this time, the fan The power generation is: (20) (twenty one) in, air density, For wind turbine The power coefficient, For wind turbine The aggregation of the average deficit factor.

[0099] Because the wake interaction patterns between wind turbines in a wind farm dynamically change with wind direction, the connectivity relationships in the graph also need to be adaptively reconstructed. (During the time period) In the middle, input data according to the current time period Based on the wind direction characteristics and the fixed geographical coordinates of each wind turbine, the upstream and downstream relationships between wind turbines are dynamically determined, specifically as follows: Figure 4As shown. If and only if the fan With wind turbine It can be determined that any of the following conditions are met. In Within the scope of influence, establish edges : 1) Fan Located in the wind turbine Upstream direction; 2) Euclidean distance between wind turbines Maximum influence distance ; 3) The angle between the line connecting the two fans and the wind direction. Maximum influence angle .

[0100] Once the connections are established, each edge The edge characteristics are specifically defined as a set of key physical quantities: downstream wake distance and radial wake distance The distance value directly reflects the propagation length of the wake in the wind direction and the degree of diffusion perpendicular to the wind direction. It is a key parameter for quantifying the strength of the wake effect, and its value is dynamically calculated based on the wind direction and the actual location of the wind turbine.

[0101] The power prediction method provided in this embodiment, which considers high-dimensional feature selection and physical guidance, calculates the downstream wake distance and radial wake distance based on the wind turbine location coordinates and the incoming wind direction. This accurately quantifies the relative spatial position relationship between units, providing reliable geometric parameters for wake effect calculation. By inputting the above distances into a continuous wake deficit model to obtain the wake influence coefficient and using it as an edge feature, the wake attenuation effect that actually exists in the wind farm can be transformed into a quantitative indicator that can be learned by the model. This allows the graph structure to simultaneously contain data features and physical mechanism constraints, improving the model's ability to characterize the spatial coupling relationship between wind turbines and making power prediction more consistent with the actual operating characteristics of wind farms.

[0102] This embodiment also provides a power prediction device that considers high-dimensional feature selection and physical guidance. This device is used to implement the above embodiments and preferred embodiments, and details already described will not be repeated. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0103] This embodiment provides a power prediction device that considers high-dimensional feature selection and physical guidance, such as... Figure 5 As shown, it includes: The high-dimensional spatiotemporal feature acquisition module is used to acquire several sets of historical operation data and historical meteorological data of the wind farm to be predicted according to a preset time step, and construct a high-dimensional spatiotemporal feature sequence set corresponding to the high-dimensional candidate feature set. The high-dimensional candidate feature set includes the operation features of each wind turbine, the meteorological features of the wind farm to be predicted, and the time features related to power generation. The priority score calculation module is used to calculate the priority score of each feature in the high-dimensional candidate feature set based on the high-dimensional spatiotemporal feature sequence set. The core impact feature set acquisition module is used to filter the core impact feature set from the high-dimensional candidate feature set by combining the priority scores of each feature; The fitness value calculation module is used to select the core spatiotemporal feature sequence set corresponding to the core influence feature set from the high-dimensional spatiotemporal feature sequence set, and determine the core spatiotemporal feature sequence set to be used to predict the output power of the wind farm to be predicted, and obtain the fitness value. The iterative optimization module is used to iteratively optimize the core influence feature set by combining the fitness value and the priority score of each feature, so as to obtain the optimized core influence feature set. The current feature sequence acquisition module is used to obtain the spatiotemporal feature sequence of the wind farm to be predicted at different time steps in the current time period based on the optimized core influence feature set; The wind farm dynamic graph construction module is used to construct the corresponding wind farm dynamic graph for each time step based on the spatiotemporal feature sequence. The dynamic graph contains several nodes and several edges. One node in the dynamic graph corresponds to one wind turbine. The node features are used to characterize the core influencing factors of the wind turbine corresponding to the node. One edge in the dynamic graph connects two wind turbines. The edge features are used to characterize the wake effect between the two wind turbines connected by the edge. The power prediction module is used to input the dynamic map of the wind farm into the preset power prediction model to obtain the power prediction value of the wind farm to be predicted.

[0104] The power prediction apparatus considering high-dimensional feature selection and physical guidance provided in this embodiment of the invention can execute the power prediction method considering high-dimensional feature selection and physical guidance provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects for executing the method. Further functional descriptions of the various modules and units described above are the same as in the corresponding embodiments described above, and will not be repeated here.

[0105] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0106] The following is a detailed reference. Figure 6This diagram illustrates a suitable structural design for implementing an electronic device according to embodiments of the present invention. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 601, which can perform various appropriate actions and processes based on a program stored in read-only memory (ROM) 602 or a program loaded from memory 608 into random access memory (RAM) 603. RAM 603 also stores various programs and data required for the operation of the electronic device. The processor 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0107] Typically, the following devices can be connected to I / O interface 605: input devices 606 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 607 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 608 including, for example, magnetic tapes, hard disks, etc.; and communication devices 609. Communication device 609 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 6 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0108] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 609, or installed from a memory 608, or installed from a ROM 602. When the computer program is executed by the processor 601, it performs the functions defined in the power prediction method considering high-dimensional feature selection and physical guidance according to embodiments of the present invention.

[0109] Figure 6 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.

[0110] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as recordable on a storage medium, or implemented as computer code originally stored on a remote storage medium or a non-transitory machine-readable storage medium and subsequently stored on a local storage medium after being downloaded via a network. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the power prediction method considering high-dimensional feature selection and physical guidance shown in the above embodiments is implemented.

[0111] A portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0112] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A power prediction method considering high-dimensional feature selection and physical guidance, characterized in that, The method includes: According to a preset time step, acquire several sets of historical operation data and historical meteorological data of the wind farm to be predicted, and construct a high-dimensional spatiotemporal feature sequence set corresponding to the high-dimensional candidate feature set. The high-dimensional candidate feature set includes the operation characteristics of each wind turbine, the meteorological characteristics of the wind farm to be predicted, and the time characteristics related to power generation. Based on the high-dimensional spatiotemporal feature sequence set, calculate the priority score of each feature in the high-dimensional candidate feature set; The core influencing feature set is obtained by combining the priority scores of each feature from the high-dimensional candidate feature set; The core spatiotemporal feature sequence set corresponding to the core influence feature set is selected from the high-dimensional spatiotemporal feature sequence set, and the output power of the wind farm to be predicted is determined using the core spatiotemporal feature sequence set to obtain the fitness value. The core influence feature set is iteratively optimized by combining the fitness value and the priority score of each feature to obtain the optimized core influence feature set; Based on the optimized core impact feature set, obtain the spatiotemporal feature sequence of the wind farm to be predicted at different time steps in the current time period; Based on the spatiotemporal feature sequence, a corresponding wind farm dynamic diagram is constructed for each time step. The dynamic diagram contains several nodes and several edges. In the dynamic diagram, one node corresponds to one wind turbine. The node features are used to characterize the core influencing factors of the wind turbine corresponding to the node. In the dynamic diagram, one edge connects two wind turbines. The edge features are used to characterize the wake effect between the two wind turbines connected by the edge. The dynamic graph of the wind farm is input into a preset power prediction model to obtain the predicted power value of the wind farm to be predicted.

2. The method according to claim 1, characterized in that, Calculating the priority score of each feature in the high-dimensional candidate feature set includes: Based on the high-dimensional spatiotemporal feature sequence set, the nonlinear correlation score between each feature and the power and the nonlinear redundancy score between each pair of features in the high-dimensional candidate feature set are calculated using random dependency coefficients. Based on the nonlinear correlation score and the nonlinear redundancy score, an incremental search strategy is used to perform round-by-round calculations to obtain the comprehensive score of each feature. The priority score for each feature is determined based on the overall score.

3. The method according to claim 1, characterized in that, The core influencing feature set is obtained by combining the priority scores of each feature from the high-dimensional candidate feature set, including: The priority scores of each feature are normalized to obtain normalized priority scores; Feature populations are set according to the normalized priority score. Each feature population contains multiple population individual position vectors. The population individual position vectors are encoded in binary to represent the selected or unselected state of each feature. Based on the position vectors of individuals in each population group within the feature population, the corresponding core influence feature sets are decoded to obtain them.

4. The method according to claim 3, characterized in that, The core influence feature set is iteratively optimized by combining the fitness and priority scores of various features to obtain the optimized core influence feature set, including: Based on the fitness value, each core influence feature set is divided into multiple alpha wolf feature sets and multiple ordinary feature sets; Based on the binary gray wolf optimization algorithm, the encoding of each feature in each ordinary feature set is iteratively updated following the alpha wolf feature set, and the encoding of each feature in each ordinary feature set after iterative update is corrected by combining the priority score of each feature. The alpha wolf feature set and the updated ordinary feature set are used as the new core influence feature set. The fitness value of each core influence feature set is recalculated, and the step of dividing each core influence feature set into the alpha wolf feature set and the ordinary feature set according to the fitness value is returned until the iteration stops when the convergence condition is met. The optimized core influence feature set is obtained by screening from the current alpha wolf feature set.

5. The method according to claim 1, characterized in that, The dynamic graph of the wind farm is processed and input into a preset power prediction model to obtain the predicted power value of the wind farm to be predicted, including: Node embedding learning is performed on the dynamic graph of the wind farm at each time step to obtain the node embedding vector of each wind turbine node; Arrange the node embedding vectors of the same wind turbine at multiple consecutive time steps in time sequence to obtain the time sequence feature sequence of each unit. The time-series characteristic sequences of each unit are input into a preset power prediction model to obtain the power prediction value of each unit; The predicted power values ​​of each unit at the same time step are summed to obtain the predicted power value of the wind farm at the corresponding time step.

6. The method according to claim 1, characterized in that, The steps for obtaining edge features include: Calculate the downstream wake distance and radial wake distance between the two wind turbines based on the wind turbine location coordinates and the incoming wind direction; The downstream wake distance and the radial wake distance are input into the continuous wake deficit model to calculate the wake influence coefficient between the two wind turbines, and the wake influence coefficient is used as the edge feature of the edge connecting the two wind turbines.

7. A power prediction device considering high-dimensional feature selection and physical guidance, characterized in that, The device includes: The high-dimensional spatiotemporal feature acquisition module is used to acquire several sets of historical operation data and historical meteorological data of the wind farm to be predicted according to a preset time step, and construct a high-dimensional spatiotemporal feature sequence set corresponding to the high-dimensional candidate feature set. The high-dimensional candidate feature set includes the operation features of each wind turbine, the meteorological features of the wind farm to be predicted, and the time features related to power generation. The priority score calculation module is used to calculate the priority score of each feature in the high-dimensional candidate feature set based on the high-dimensional spatiotemporal feature sequence set. The core impact feature set acquisition module is used to filter the core impact feature set from the high-dimensional candidate feature set by combining the priority scores of each feature; The fitness value calculation module is used to filter the core spatiotemporal feature sequence set corresponding to the core influence feature set from the high-dimensional spatiotemporal feature sequence set, and determine the core spatiotemporal feature sequence set to be used to predict the output power of the wind farm to be predicted, so as to obtain the fitness value. The iterative optimization module is used to iteratively optimize the core influence feature set by combining the fitness value and the priority score of each feature, so as to obtain the optimized core influence feature set. The current feature sequence acquisition module is used to acquire the spatiotemporal feature sequence of the wind farm to be predicted at different time steps in the current time period based on the optimized core influence feature set; The wind farm dynamic map construction module is used to construct a corresponding wind farm dynamic map for each time step based on the spatiotemporal feature sequence. The dynamic map contains several nodes and several edges. In the dynamic map, one node corresponds to one wind turbine. The node features are used to characterize the core influencing factors of the wind turbine corresponding to the node. In the dynamic map, one edge connects two wind turbines. The edge features are used to characterize the wake effect between the two wind turbines connected by the edge. The power prediction module is used to input the dynamic map of the wind farm into a preset power prediction model to obtain the power prediction value of the wind farm to be predicted.

8. An electronic device, characterized in that, include: A memory and a processor are communicatively connected, the memory storing computer instructions, and the processor executing the computer instructions to perform the power prediction method considering high-dimensional feature selection and physical guidance as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing a computer to perform the power prediction method considering high-dimensional feature selection and physical guidance as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, Includes computer instructions for causing a computer to perform the power prediction method considering high-dimensional feature selection and physical guidance as described in any one of claims 1 to 6.

Images