Wind power prediction method based on multi-model weight optimization and fusion under different wind speed scenarios

Abstract

This invention discloses a wind power prediction method based on multi-model weight optimization and fusion under different wind speed scenarios, belonging to the field of wind power prediction technology. The method includes: acquiring historical wind power data and historical meteorological data; performing wind speed scenario identification on the historical meteorological data and then clustering it to obtain historical meteorological data under different wind speed scenarios, and constructing a dataset; constructing a multi-algorithm hybrid model based on three prediction models: random forest prediction model, bidirectional long short-term memory neural network prediction model, and support vector regression prediction model; constructing a weight optimization function using an objective function and constraints, and solving for the optimal weights based on KKT conditions to obtain a trained multi-algorithm hybrid model under different wind speed scenarios; after identifying the wind speed scenario from the predicted meteorological data, inputting it into the corresponding multi-algorithm hybrid model under the wind speed scenario, and then outputting the wind power prediction result. Therefore, this invention can achieve wind power prediction.

Description

Technical Field

[0001] This invention relates to a wind power prediction method based on multi-model weight optimization and fusion under different wind speed scenarios, belonging to the technical field of wind power prediction. Background Technology

[0002] With the increasing maturity and sophistication of wind power technology, the installed capacity and scale of wind turbines are constantly expanding. This rapid development of wind power generation places higher demands on power prediction accuracy. Wind speed, as a core driving factor, exhibits dynamic characteristics that lead to strong nonlinear and multimodal power output features in wind turbines. Traditional single prediction models struggle to adapt to the complex variations across the entire wind speed range. Therefore, dividing wind speed scenarios and implementing dynamic fusion of multiple models has become a key path to improving prediction accuracy.

[0003] Commonly used wind power prediction methods include, but are not limited to, Random Forest (RF), Naive Bayes, Support Vector Machine (SVR), Convolutional Neural Network (CNN), Transformer, and Bidirectional Long Short-Term Memory (BiLSTM) algorithms.

[0004] Random forest is a supervised machine learning method based on ensemble learning. It improves prediction accuracy and robustness by constructing multiple decision trees and aggregating the results. Support vector regression, on the other hand, finds the optimal hyperplane to fit the relationship between input features and output power. It can handle nonlinear relationships well and has strong generalization ability for small samples.

[0005] Support Vector Machine (SVR) algorithms implicitly map high-dimensional spaces using kernel functions to handle nonlinear relationships. Commonly used kernel functions include linear kernels, polynomial kernels, and Gaussian kernels (RBF). In wind power prediction, the Gaussian kernel is widely used because it can effectively model the complex nonlinear coupling between meteorological factors such as wind speed and temperature at different altitudes and power. Parameter tuning is a key aspect of SVR: increasing the penalty coefficient can reduce training error but may lead to overfitting; decreasing the insensitive loss parameter can improve regression accuracy but increase computational complexity; and the value that determines the range of influence of the sample needs to be balanced between model complexity and generalization ability. The advantages of SVR are its robustness to small sample data and high-dimensional features, its avoidance of explicit high-dimensional calculations through kernel tricks, and its insensitivity to outliers, making it suitable for power fluctuation scenarios caused by gusts or noise in wind speed data.

[0006] Long Short-Term Memory Neural Network (BiLSTM) models spatiotemporal multi-scale correlations through a bidirectional gating mechanism. Under unstable wind power output conditions such as strong turbulent activity, BiLSTM has a strong time series modeling capability, which can simultaneously consider past and future information, thereby better capturing the time dependence and dynamic characteristics of wind power. It is especially suitable for short-term fluctuation prediction under sudden weather changes.

[0007] While these algorithms have achieved some success in wind power prediction, individual algorithms often face challenges in handling the nonlinear and non-steady-state characteristics of data, potentially leading to insufficient generalization ability. One of the challenges in wind power prediction is the nonlinear and non-steady-state nature of wind speed; the output characteristics of wind power vary significantly under different wind speed scenarios. Each model has its own advantages, but also limitations, making it difficult for a single algorithm to simultaneously handle these complex situations. Summary of the Invention

[0008] The purpose of this invention is to provide a wind power prediction method that optimizes and fuses the weights of multiple models under different wind speed scenarios. By fusing the prediction results of random forest prediction model, support vector regression prediction model and bidirectional long short-term memory neural network prediction model, the accuracy and robustness of wind power prediction are improved, providing strong support for the operation management and scheduling decisions of wind farms.

[0009] To achieve the above-mentioned technical objectives, the present invention will adopt the following technical solution:

[0010] A method for predicting wind power using multi-model weight optimization and fusion under different wind speed scenarios includes the following steps:

[0011] Acquire historical wind power data and historical meteorological data for a specified historical period; historical meteorological data includes multi-dimensional wind field observation data.

[0012] Historical meteorological data is subjected to wind speed scene identification and then clustered to obtain historical meteorological data under different wind speed scenes. The historical meteorological data under different wind speed scenes are then combined with the historical wind power that matches them in time and space to form a dataset for the corresponding wind speed scene. The constructed datasets for different wind speed scenes are then divided into training set, validation set and test set according to a preset ratio. In the dataset, the feature data is historical meteorological data and the label data is historical wind power.

[0013] A multi-algorithm hybrid model is constructed based on three heterogeneous prediction models: random forest prediction model, bidirectional long short-term memory neural network prediction model, and support vector regression prediction model. A weight optimization function is constructed for the three heterogeneous prediction models of the multi-algorithm hybrid model using objective function and constraints. The optimal weight is solved based on KKT conditions to achieve the optimization and fusion of model weights among prediction models under different wind speed scenarios, so as to obtain a well-trained multi-algorithm hybrid model under different wind speed scenarios.

[0014] Based on the trained multi-algorithm hybrid models obtained under different wind speed scenarios, the predicted meteorological data is identified by wind speed scenario and then input into the corresponding multi-algorithm hybrid model under the wind speed scenario to output the wind power prediction result.

[0015] Preferably, wind speed scene identification is performed on historical meteorological data / predicted meteorological data, specifically including the following steps:

[0016] Based on multi-dimensional wind field observation data of historical / forecasted meteorological data at any given time, the horizontal wind energy capture efficiency at the corresponding time was calculated. Vertical disturbance loss factor and wind yaw efficiency ;

[0017] Based on the calculated horizontal wind energy capture efficiency at any given time Vertical disturbance loss factor and wind yaw efficiency Calculate the three-dimensional wind energy coupling efficiency index at the corresponding time.

[0018] By comparing the calculated three-dimensional wind energy coupling efficiency index (TWECI) at any given time with the preset three-dimensional wind energy coupling efficiency index thresholds for different wind speed scenarios, the wind speed scenario at the current time can be identified.

[0019] Preferably, the different wind speed scenarios include: ideal capture scenario, stable wind condition scenario, turbulent mixing scenario, and wind energy inefficient scenario;

[0020] The preset threshold values ​​for the three-dimensional wind energy coupling efficiency index for ideal capture scenarios are (0.8~1.0); for stable wind scenarios, the preset threshold values ​​are (0.5~0.8); for turbulent mixed scenarios, the preset threshold values ​​are (0.3~0.5); and for inefficient wind energy scenarios, the preset threshold values ​​are [0~0.3].

[0021] When identifying wind speed scenarios at any given time, if the current three-dimensional wind energy coupling efficiency index (TWECI) is between 0.8 and 1.0, it indicates that the current wind speed scenario is an ideal capture scenario; if the current three-dimensional wind energy coupling efficiency index (TWECI) is between 0.5 and 0.8, it indicates that the current wind speed scenario is a stable wind condition scenario; if the current three-dimensional wind energy coupling efficiency index (TWECI) is between 0.3 and 0.5, it indicates that the current wind speed scenario is a turbulent mixing scenario; if the current three-dimensional wind energy coupling efficiency index (TWECI) is between 0 and 0.3, it indicates that the current wind speed scenario is a wind energy inefficient scenario.

[0022] Preferably, the horizontal wind energy capture efficiency at any given time Vertical disturbance loss factor and wind yaw efficiency Calculate using the following formulas respectively:

[0023] = ;

[0024] ;

[0025] = ;

[0026] In the above formula: Horizontal wind speed; The rated wind speed of the wind turbine unit; The wind energy utilization coefficient; Indicates vertical wind speed; The current yaw angle of the wind turbine nacelle; Indicates wind direction.

[0027] Preferably, the three-dimensional wind energy coupling efficiency index (TWECI) is calculated using the following formula:

[0028] ;

[0029] In the formula: Indicates the horizontal wind energy capture efficiency; Indicates the vertical disturbance loss factor; This indicates the yaw efficiency due to wind direction.

[0030] Preferably, the multi-algorithm hybrid model is represented by the following formula:

[0031] ;

[0032] In the above formula, This represents the predicted value of the multi-algorithm hybrid model at time t; Let represent the predicted value of the k-th prediction model at time t, where k = 1, 2, 3, corresponding to the random forest prediction model, the bidirectional long short-term memory neural network prediction model, and the support vector regression prediction model, respectively. This represents the model weight of the k-th prediction model in the multi-algorithm hybrid model.

[0033] Preferably, the multi-algorithm hybrid model is constructed through the following steps:

[0034] The constructed datasets are used to train, validate, and test each prediction model to optimize the hyperparameters of the respective prediction models.

[0035] A weight optimization mathematical model is constructed and solved based on KKT conditions to obtain the model weights of each prediction model in the multi-algorithm hybrid model, and a sliding window mechanism is used to dynamically adjust the model weights.

[0036] The objective function of the weight optimization mathematical model is expressed as:

[0037] ;

[0038] In the formula: N represents the total number of times t. This represents the actual power value at time t.

[0039] Preferably, when optimizing the hyperparameters of each prediction model, a single prediction parameter optimization model is defined as follows:

[0040] ;

[0041] ;

[0042] In the formula: Let k be the prediction function of the prediction model. ; Its hyperparameters; Let be its state variable at time t;

[0043] For the random forest prediction model, its hyperparameters Let K be the number of trees and the maximum depth; for a bidirectional long short-term memory neural network prediction model, its hyperparameters are... The hidden layer dimension and learning rate are the hyperparameters for the support vector regression prediction model. Let C be the penalty coefficient and σ be the width of the Gaussian kernel function.

[0044] Preferably, the KKT conditions specifically include:

[0045] Lagrange function:

[0046] ;

[0047] stationarity:

[0048] ;

[0049] Initial feasibility:

[0050] Equality constraints: Inequality constraints: ;

[0051] Duality feasibility: ;

[0052] Complementary relaxation: ;

[0053] in, Let represent the predicted value of the k-th prediction model at time t, where k = 1, 2, 3, corresponding to the random forest prediction model, the bidirectional long short-term memory neural network prediction model, and the support vector regression prediction model, respectively. This represents the model weight of the k-th prediction model in the multi-algorithm hybrid model, obtained by solving the KKT conditions. This represents the actual power value at time t. For the Lagrange multipliers corresponding to the equality constraints, is the KKT multiplier corresponding to the inequality constraint; j represents the traversal index used for internal summation, and like k, represents each prediction model, j=1, 2, 3.

[0054] Another technical objective of this invention is to provide an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the computer program executes the above-described method for wind power prediction by multi-model weight optimization fusion under different wind speed scenarios.

[0055] Based on the above-mentioned technical objectives, the present invention has the following advantages compared with the prior art:

[0056] The wind power prediction method based on multi-model weight optimization and fusion under different wind speed scenarios described in this invention first identifies wind speed scenarios from the acquired historical meteorological data to construct datasets for different wind speed scenarios. Then, it constructs a multi-algorithm hybrid model based on three heterogeneous prediction models: random forest prediction model, bidirectional long short-term memory neural network prediction model, and support vector regression prediction model. A weight optimization function is constructed using the objective function and constraints, and the optimal weights are solved based on KKT conditions. This achieves dynamic optimization and fusion of model weights among the multiple prediction models under different wind speed scenarios, allowing for on-demand fusion and complementary advantages, resulting in a dual leap in overall accuracy and robustness. In a wind farm application, it was found that the multi-algorithm hybrid model under different wind speed scenarios reduced the root mean square error of prediction by up to 22% compared to the single-model prediction. Therefore, the multi-algorithm hybrid model constructed in this invention fully exploits the nonlinearity, temporal sequence, and high-dimensional coupling characteristics of the wind power sequence, achieving complementary advantages under different wind speed scenarios, improving wind power prediction accuracy, and providing strong support for wind farm operation management and scheduling decisions. Attached Figure Description

[0057] Figure 1 This is a flowchart of the wind power prediction method based on multi-model weight optimization and fusion under different wind speed scenarios as described in this invention. Detailed Implementation

[0058] 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, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. Unless otherwise specifically stated, the relative arrangement, expressions, and values ​​of components and steps set forth in these embodiments do not limit the scope of the present invention. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0059] Example 1

[0060] like Figure 1 As shown, the wind power prediction method based on multi-model weight optimization and fusion under different wind speed scenarios of the present invention specifically includes the following steps:

[0061] Step 1: Construct datasets for different wind speed scenarios:

[0062] This invention takes into account the impact of the nonlinearity and non-steady-state nature of wind speed on wind power prediction. Therefore, it uses cluster analysis to divide historical meteorological data into different wind speed scenarios, thereby constructing datasets for different wind speed scenarios to train, test and validate the subsequently established multi-algorithm hybrid model.

[0063] The datasets for different wind speed scenarios are constructed through the following steps:

[0064] Step 1.1: Obtain historical wind power data and historical meteorological data:

[0065] Obtain historical wind power and historical meteorological data for a specified historical period. The historical meteorological data includes multi-dimensional wind field observation data and historical temperature. The multi-dimensional wind field observation data includes historical horizontal wind speed, historical vertical wind speed, and historical wind direction.

[0066] Historical meteorological data and historical wind power are sampling records based on actual wind farm data acquisition and monitoring control systems (SCADA) at set time intervals (e.g., every 15 minutes), including wind speed and wind direction records detected by the wind measurement tower.

[0067] Step 1.2: Perform wind speed scene identification on historical meteorological data:

[0068] This invention, based on multidimensional wind field observation data and addressing the need for wind power prediction, combines three-dimensional wind field characteristics with wind turbine power response, defining a three-dimensional wind energy coupling effectiveness index (TWECI) to identify different wind speed scenarios from historical meteorological data. The specific steps include:

[0069] Step 1.2.1: Based on the multi-dimensional wind field observation data obtained from historical meteorological data at any given time, calculate the horizontal wind energy capture efficiency at the corresponding time. Vertical disturbance loss factor and wind yaw efficiency .

[0070] Horizontal wind energy capture efficiency Calculated using the following formula:

[0071] = ;

[0072] In the above formula, The rated wind speed of the wind turbine unit; Horizontal wind speed; The wind energy utilization coefficient is obtained by querying... Obtained by curve The tip speed ratio is expressed by the following formula:

[0073] ;

[0074] In the above formula, Horizontal wind speed; R is the blade rotation speed; R is the blade rotation radius.

[0075] According to blade momentum theory, when the vertical wind speed exceeds 20% of the horizontal wind speed, the three-dimensional flow separation effect begins to dominate, and the loss of wind energy due to vertical wind speed increases. Therefore, this invention uses the following formula to calculate the vertical disturbance loss factor. :

[0076] ;

[0077] In the above formula: Indicates vertical wind speed; This refers to the horizontal wind speed.

[0078] Wind yaw efficiency Calculated using the following formula:

[0079] = ;

[0080] In the above formula: The current yaw angle of the wind turbine nacelle in the SCADA encoder; Indicates wind direction.

[0081] Step 1.2.2: Based on the calculated horizontal wind energy capture efficiency at any given time. Vertical disturbance loss factor and wind yaw efficiency Calculate the three-dimensional wind energy coupling efficiency index (TWECI) at the corresponding time point. The three-dimensional wind energy coupling efficiency index TWECI is calculated using the following formula:

[0082] ;

[0083] In the formula: Indicates the horizontal wind energy capture efficiency; Indicates the vertical disturbance loss factor; This indicates the yaw efficiency due to wind direction.

[0084] Step 1.2.3: Based on the calculated three-dimensional wind energy coupling efficiency index, compare it with the preset three-dimensional wind energy coupling efficiency index thresholds under different wind speed scenarios to identify the wind speed scenario at the current moment.

[0085] Generally, the closer the three-dimensional wind energy coupling efficiency index is to 1, the better the wind energy coupling efficiency, approaching the ideal power generation scenario; the closer it is to 0, the more unfavorable the power generation scenario. Therefore, this step defines four wind speed scenarios: ideal capture scenario, stable wind condition scenario, turbulent mixing scenario, and inefficient wind energy scenario. The present invention also presets the three-dimensional wind energy coupling efficiency index thresholds for each wind speed scenario as follows: the preset threshold for the three-dimensional wind energy coupling efficiency index for the ideal capture scenario is (0.8~1.0); the preset threshold for the three-dimensional wind energy coupling efficiency index for the stable wind condition scenario is (0.5~0.8); the preset threshold for the three-dimensional wind energy coupling efficiency index for the turbulent mixing scenario is (0.3~0.5); and the preset threshold for the three-dimensional wind energy coupling efficiency index for the inefficient wind energy scenario is [0~0.3].

[0086] Therefore, when identifying wind speed scenarios at any given time, if the current three-dimensional wind energy coupling efficiency index (TWECI) is between (0.8 and 1.0), it indicates that the current wind speed scenario is an ideal capture scenario; if the current three-dimensional wind energy coupling efficiency index (TWECI) is between (0.5 and 0.8), it indicates that the current wind speed scenario is a stable wind condition scenario; if the current three-dimensional wind energy coupling efficiency index (TWECI) is between (0.3 and 0.5), it indicates that the current wind speed scenario is a turbulent mixing scenario; and if the current three-dimensional wind energy coupling efficiency index (TWECI) is between [0 and 0.3], it indicates that the current wind speed scenario is a wind energy inefficient scenario.

[0087] Step 1.3: Based on the identified different wind speed scenarios, construct the corresponding datasets for each wind speed scenario. Specifically:

[0088] Based on the obtained wind speed scenarios, historical meteorological data are clustered to obtain historical meteorological data under different wind speed scenarios.

[0089] The historical meteorological data under different wind speed scenarios are combined with the historical wind power matched in time and space to form a dataset for the corresponding wind speed scenario; in the dataset, the feature data is historical meteorological data (including historical wind speed, etc.) and the label data is historical wind power.

[0090] Step 2: Construct a multi-algorithm hybrid model:

[0091] To fully leverage the heterogeneity of wind power prediction results among the three heterogeneous models—random forest prediction model, bidirectional long short-term memory neural network prediction model, and support vector regression prediction model—this invention constructs a multi-algorithm hybrid model:

[0092] ;

[0093] In the above formula, This represents the predicted value of the multi-algorithm hybrid model at time t; This represents the predicted value of the k-th prediction model at time t; This represents the model weight of the k-th prediction model in the multi-algorithm hybrid model; the value of k is k=1,2,3, which correspond to the random forest prediction model, the bidirectional long short-term memory neural network prediction model, and the support vector regression prediction model, respectively.

[0094] In this invention, a joint parameter adaptive optimization strategy is used to construct a weight optimization function based on the objective function and constraints. The optimal weights are then solved based on KKT conditions to obtain trained multi-algorithm hybrid models for different wind speed scenarios. This achieves dynamic weight optimization and fusion of multiple models under different wind speed scenarios, fully exploiting the nonlinearity, temporal sequence, and high-dimensional coupling characteristics of wind power sequences. This enables collaborative prediction by multiple algorithms under complex wind speed scenarios, improving wind power prediction accuracy and providing strong support for wind farm operation management and scheduling decisions. Specifically, the steps include:

[0095] Step 2.1, Hyperparameter optimization of the single prediction model:

[0096] The datasets for different wind speed scenarios constructed in Step 1 are divided into training set, validation set and test set according to the proportions.

[0097] The constructed training set is used to train each prediction model separately to optimize the hyperparameters of the corresponding prediction model, which can be expressed as:

[0098] ;

[0099] ;

[0100] In the formula, Let k be the prediction function of the prediction model. Its hyperparameters, Its state variables. For the random forest prediction model, its hyperparameters... Let K be the number of trees and the maximum depth; for a bidirectional long short-term memory neural network prediction model, its hyperparameters are... The hidden layer dimension and learning rate are the hyperparameters for the support vector regression prediction model. Let C be the penalty coefficient and σ be the width of the Gaussian kernel function.

[0101] Step 2.2. Construction and solution of the mathematical model for weight optimization based on KKT conditions:

[0102] This step aims to dynamically assign optimal weights to the multi-algorithm hybrid model under different wind speed scenarios. Based on the three single-item prediction models optimized in step 2.1: Random Forest (RF), Bidirectional Long Short-Term Memory (BiLSTM), and Support Vector Regression (SVR), a weight optimization model is constructed, the core of which is to solve a constrained quadratic programming problem.

[0103] First, we define a weighted optimization mathematical model. The goal is to minimize the difference between the weighted combined predicted value and the actual power value on a dataset containing N time points. The mean squared error (MSE) between them. The objective function and constraints are as follows:

[0104] Objective function:

[0105] ;

[0106] in, is the variable to be optimized, representing the weight of the k-th prediction model; It is a known value, representing the predicted power of the k-th prediction model at time t; t is a known value representing the prediction power of the k-th prediction model at time t; k is the model index, k=1, 2, 3, corresponding to the random forest, BiLSTM and SVR models respectively; t is the time point index, t=1, 2, ..., N, where N is the total number of time points for model training or weight optimization, determined by the size of the sliding window.

[0107] Constraints:

[0108] (Equality constraints);

[0109] (Inequality constraints);

[0110] in, It is the predicted value of the k-th model at time t. These are the weights of the k-th model to be solved. The equality constraint that the sum of the weights is 1 ensures the unbiasedness of the combination; the inequality constraint that the weights are non-negative guarantees that each model makes a positive or zero contribution to the final result.

[0111] Because this optimization problem involves both equality and inequality constraints, the traditional Lagrange multiplier method is no longer applicable. Therefore, this invention employs the more universal Karush-Kuhn-Tucker (KKT) conditions for solving the problem. The KKT conditions provide a complete theoretical framework for solving such constrained optimization problems.

[0112] Therefore, we introduce Lagrange multipliers. Corresponding equality constraints are applied, and KKT multipliers are introduced. (k=1, 2, 3) correspond to three inequality constraints. Construct the generalized Lagrangian function:

[0113] ;

[0114] in, It is the constructed generalized Lagrange function; It is a Lagrange multiplier. It is the dual variable associated with the equality constraint; It is a KKT multiplier, which is related to the k-th inequality constraint. The dual variable associated with ≥0.

[0115] According to KKT theory, the constraints for solving this optimization problem are as follows:

[0116] (a). L to The gradient must be zero to maintain stationarity.

[0117] (k=1, 2, 3);

[0118] Here, j is a traversal index used for internal summation, and like k, it represents each prediction model (j=1, 2, 3). The traversal index j is used to distinguish it from the outer derivative variable k.

[0119] (b) The solution must satisfy all original constraints;

[0120] (c) The KKT multipliers corresponding to the inequality constraints must be nonnegative;

[0121] (d) The product of the KKT multiplier and its corresponding inequality constraint is zero, preserving complementary relaxation:

[0122] (k=1, 2, 3);

[0123] Complementary slackness is the core of the KKT conditions, revealing the structure of the optimal solution: if a certain weight If the value in the optimal solution is strictly greater than zero, then the corresponding inequality constraint... ≥ 0 is non-compact (inactive), in which case its KKT multipliers It must be zero. Conversely, if some KKT multiplier... If the value is greater than 0, then the corresponding constraint must be tight (active), that is... =0.

[0124] The solution is as follows:

[0125] Based on the above KKT conditions, we design a solution algorithm. First, we assume that all inequality constraints are non-compact, i.e. =0 (k=1, 2, 3), which means we only consider equality constraints for now. At this point, the stationarity condition simplifies to:

[0126] ;

[0127] This system of equations can be written in a more concise matrix form. Let And define matrix A and vector B as follows:

[0128] ;

[0129] ;

[0130] in It is the element in the k-th row and j-th column of matrix A. It represents the inner product of the prediction sequence of model k and the prediction sequence of model j, reflecting the correlation between the prediction results of the two models. The k-th element of vector B represents the inner product of the actual power sequence and the predicted sequence of model k, reflecting the accuracy of model k's prediction.

[0131] Then the stationarity condition and the equality constraint in the original feasibility are... This forms the following system of linear equations:

[0132] ;

[0133] In the above formula, for the matrix , yes The resulting 3x3 matrix; The weight to be determined , , The resulting 3x1 column vector; It is a 3x1 column vector where all elements are 1. It is a column vector The transpose of is a 1x3 row vector [1, 1, 1]; for the matrix B is composed of elements The resulting 3x1 column vector.

[0134] By solving this system of linear equations, a set of candidate weights can be obtained. Then, check if any of the solutions satisfies... ≥0. If any solution All meet If the value is ≥0, then the candidate weight is considered to be... The optimal solution is one that simultaneously satisfies all KKT conditions. If there exists a certain... If the value is less than 0, then the original feasibility is violated, indicating that initially... The assumption that = 0 does not hold. According to complementary slackness, this weight... The corresponding constraint must be tight, that is... =0. At this point, we will set this weight... The weights are fixed at 0 and removed from the optimization problem. The remaining weights are then used to reconstruct and solve a smaller system of linear equations. This process is repeated iteratively until all the calculated weights satisfy the non-negativity condition.

[0135] Through the rigorous solution process based on the KKT conditions described above, we can ensure that we find the optimal combination of model weights that minimizes the prediction error under the current wind speed scenario. .

[0136] Steps 2 and 3. Dynamic weight allocation mechanism:

[0137] To adapt to the temporal non-stationarity of wind power, a sliding window mechanism is adopted to dynamically adjust the weights. The window length is set to 24 hours (96 15-minute sampling points) to balance computational efficiency and timeliness; the update frequency is to recalculate the weights every hour to ensure that the weights match the current wind speed scenario.

[0138] Step 3: Output the multi-algorithm collaborative wind power prediction results under different wind speed scenarios.

[0139] Based on the multi-algorithm hybrid models for different wind speed scenarios constructed in step two, the predicted meteorological data is subjected to corresponding wind speed scenario identification. After this identification, the data is input into the corresponding multi-algorithm hybrid model for that wind speed scenario to output the wind power prediction result. It should be noted that the specific method for identifying corresponding wind speed scenarios using predicted meteorological data can refer to the method for identifying corresponding wind speed scenarios using historical meteorological data described above, and will not be repeated here.

[0140] Example 2

[0141] The present invention also provides a storage medium, wherein the computer program stored in the storage medium executes the above-described method for wind power prediction by multi-model weight optimization and fusion under different wind speed scenarios when it is run.

[0142] Example 3

[0143] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the above-described method for wind power prediction by multi-model weight optimization fusion under different wind speed scenarios through the execution of the computer program.

[0144] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0145] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0146] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0147] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0148] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0149] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0150] Application examples

[0151] Taking different wind speed scenarios at a certain wind farm as an example, the weight allocation is shown in Table 1:

[0152] Table 1 Weighting under different wind speed scenarios

[0153]

[0154] In Table 1, ω1 represents the weight of the random forest prediction model in the multi-algorithm hybrid model; ω2 represents the weight of the bidirectional long short-term memory neural network prediction model in the multi-algorithm hybrid model; and ω3 represents the weight of the support vector regression prediction model in the multi-algorithm hybrid model.

[0155] In the ideal capture scenario, the random forest prediction model has the highest weight, while in the turbulent mixed scenario and the wind energy inefficient scenario, BiLSTM and SVR dominate, and can better capture the nonlinear process in the changing process.

[0156] Using data from a 52.8MW wind farm, the Random Forest prediction model, as a single model, achieved a root mean square error (RMSE) of 5.21 kWh and a mean absolute error (MAE) of 1.64 kWh, outperforming SVR (RMSE 5.75 kWh, MAE 1.75) but underperforming BiLSTM (RMSE 4.93 kWh, MAE 1.54 kWh). However, after multi-model fusion (a hybrid model combining multiple algorithms), the RMSE decreased to 4.5 kWh, and the MAE to 1.42 kWh, significantly improving prediction accuracy and demonstrating the crucial role of model fusion in prediction results and the effectiveness of dynamic weight optimization.

[0157] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting wind power by multi-model weight optimization and fusion under different wind speed scenarios, characterized in that, Includes the following steps: Acquire historical wind power data and historical meteorological data for a specified historical period; historical meteorological data includes multi-dimensional wind field observation data. Historical meteorological data is subjected to wind speed scene identification and then clustered to obtain historical meteorological data under different wind speed scenes. The historical meteorological data under different wind speed scenes are then combined with the historical wind power that matches them in time and space to form a dataset for the corresponding wind speed scene. The constructed datasets for different wind speed scenes are then divided into training set, validation set and test set according to a preset ratio. In the dataset, the feature data is historical meteorological data and the label data is historical wind power. A multi-algorithm hybrid model is constructed based on three heterogeneous prediction models: random forest prediction model, bidirectional long short-term memory neural network prediction model, and support vector regression prediction model. A weight optimization function is constructed for the three heterogeneous prediction models of the multi-algorithm hybrid model using objective function and constraints. The optimal weight is solved based on KKT conditions to achieve the optimization and fusion of model weights among prediction models under different wind speed scenarios, so as to obtain a well-trained multi-algorithm hybrid model under different wind speed scenarios. Based on the trained multi-algorithm hybrid models under different wind speed scenarios, the predicted meteorological data is identified by wind speed scenario and then input into the corresponding multi-algorithm hybrid model under the wind speed scenario to output the wind power prediction result. The multi-algorithm hybrid model is represented by the following formula: ； In the above formula, This represents the predicted value of the multi-algorithm hybrid model at time t; Let represent the predicted value of the k-th prediction model at time t, where k = 1, 2, 3, corresponding to the random forest prediction model, the bidirectional long short-term memory neural network prediction model, and the support vector regression prediction model, respectively. This represents the model weight of the k-th prediction model in the multi-algorithm hybrid model; The multi-algorithm hybrid model is constructed through the following steps: The constructed datasets are used to train, validate, and test each prediction model to optimize the hyperparameters of the respective prediction models. A weight optimization mathematical model is constructed and solved based on KKT conditions to obtain the model weights of each prediction model in the multi-algorithm hybrid model, and a sliding window mechanism is used to dynamically adjust the model weights. The objective function of the weight optimization mathematical model is expressed as: ； In the formula: N represents the total number of times t. This represents the actual power value at time t.

2. The wind power prediction method based on multi-model weight optimization and fusion under different wind speed scenarios as described in claim 1, characterized in that, Wind speed scene identification is performed using historical / predicted meteorological data, specifically including the following steps: Based on multi-dimensional wind field observation data of historical / forecasted meteorological data at any given time, the horizontal wind energy capture efficiency at the corresponding time was calculated. Vertical disturbance loss factor and wind yaw efficiency ; Based on the calculated horizontal wind energy capture efficiency at any given time Vertical disturbance loss factor and wind yaw efficiency Calculate the three-dimensional wind energy coupling efficiency index at the corresponding time. By comparing the calculated three-dimensional wind energy coupling efficiency index (TWECI) at any given time with the preset three-dimensional wind energy coupling efficiency index thresholds for different wind speed scenarios, the wind speed scenario at the current time can be identified.

3. The wind power prediction method based on multi-model weight optimization and fusion under different wind speed scenarios as described in claim 2, characterized in that, The different wind speed scenarios include: ideal capture scenario, stable wind condition scenario, turbulent mixed scenario, and wind energy inefficient scenario; The preset threshold values ​​for the three-dimensional wind energy coupling efficiency index for ideal capture scenarios are (0.8~1.0); for stable wind scenarios, the preset threshold values ​​are (0.5~0.8); for turbulent mixed scenarios, the preset threshold values ​​are (0.3~0.5); and for inefficient wind energy scenarios, the preset threshold values ​​are [0~0.3]. When identifying wind speed scenarios at any given time, if the current three-dimensional wind energy coupling efficiency index (TWECI) is between 0.8 and 1.0, it indicates that the current wind speed scenario is an ideal capture scenario; if the current three-dimensional wind energy coupling efficiency index (TWECI) is between 0.5 and 0.8, it indicates that the current wind speed scenario is a stable wind condition scenario; if the current three-dimensional wind energy coupling efficiency index (TWECI) is between 0.3 and 0.5, it indicates that the current wind speed scenario is a turbulent mixing scenario; if the current three-dimensional wind energy coupling efficiency index (TWECI) is between 0 and 0.3, it indicates that the current wind speed scenario is a wind energy inefficient scenario.

4. The wind power prediction method based on multi-model weight optimization and fusion under different wind speed scenarios as described in claim 2, characterized in that, Horizontal wind energy capture efficiency at any time Vertical disturbance loss factor and wind yaw efficiency Calculate using the following formulas respectively: = ； ； = ； In the above formula: Horizontal wind speed; The rated wind speed of the wind turbine unit; The wind energy utilization coefficient; Indicates vertical wind speed; The current yaw angle of the wind turbine nacelle; Indicates wind direction.

5. The wind power prediction method based on multi-model weight optimization and fusion under different wind speed scenarios as described in claim 4, characterized in that, The three-dimensional wind energy coupling efficiency index (TWECI) is calculated using the following formula: ； In the formula: Indicates the horizontal wind energy capture efficiency; Indicates the vertical disturbance loss factor; This indicates the yaw efficiency due to wind direction.

6. The wind power prediction method based on multi-model weight optimization and fusion under different wind speed scenarios as described in claim 5, characterized in that, When optimizing the hyperparameters of each prediction model, the single prediction parameter optimization model is defined as follows: ； ； In the formula: Let k be the prediction function of the prediction model. ; Its hyperparameters; Let be its state variable at time t; For the random forest prediction model, its hyperparameters Let K be the number of trees and the maximum depth; for a bidirectional long short-term memory neural network prediction model, its hyperparameters are... The hidden layer dimension and learning rate are the hyperparameters for the support vector regression prediction model. Let C be the penalty coefficient and σ be the width of the Gaussian kernel function.

7. The wind power prediction method based on multi-model weight optimization and fusion under different wind speed scenarios as described in claim 5, characterized in that, KKT conditions specifically include: Lagrange function: ； stationarity: ； Initial feasibility: Equality constraints: Inequality constraints: ; Duality feasibility: ; Complementary relaxation: ; in, Let represent the predicted value of the k-th prediction model at time t, where k = 1, 2, 3, corresponding to the random forest prediction model, the bidirectional long short-term memory neural network prediction model, and the support vector regression prediction model, respectively. This represents the model weight of the k-th prediction model in the multi-algorithm hybrid model, obtained by solving the KKT conditions. This represents the actual power value at time t. For the Lagrange multipliers corresponding to the equality constraints, is the KKT multiplier corresponding to the inequality constraint; j represents the traversal index used for internal summation, and like k, represents each prediction model, j=1, 2, 3.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the computer program is run, it executes the wind power prediction method of multi-model weight optimization fusion under different wind speed scenarios as described in any one of claims 1 to 7.

Images