A multi-model cooperative fusion wind turbine power intelligent prediction method and device

Abstract

The application discloses a kind of wind turbine power intelligent prediction method and device of multi-model synergic fusion, belong to wind turbine operation control and new energy prediction technical field.The method includes obtaining wind turbine data, wind turbine data is divided into several wind processes according to fixed sample length, each wind process index is calculated, and effective wind process set is obtained according to each wind process index;Obtain the feature vector of each effective wind process, obtain the similarity matrix using the feature vector, and divide each effective wind process into several working condition clusters based on spectral clustering algorithm;Each working condition cluster sub-model is constructed to each working condition cluster, and working condition cluster sub-model stack is formed;Global prediction model is constructed based on all effective wind processes, and global prediction model stack is formed;For the target wind process to be predicted, working condition cluster sub-model prediction value and global prediction model prediction value are obtained using working condition cluster sub-model and global prediction model, and power prediction value is obtained using dynamic weighting mode.

Description

Technical Field

[0001] This invention belongs to the field of wind turbine operation control and new energy prediction technology, specifically involving a method and device for intelligent prediction of wind turbine power through multi-model collaborative fusion. Background Technology

[0002] With the large-scale grid connection of renewable energy, the randomness and volatility of wind power have become increasingly prominent. To ensure the safe and stable operation of the power grid and improve the wind power absorption rate, it is necessary to obtain the most accurate short-term power forecast results possible at the grid dispatching and unit control levels. Traditional power forecasting methods can be broadly divided into two categories: physical model methods and data-driven methods. The former relies on meteorological field modeling and unit aerodynamic mechanisms, which are complex to model and highly dependent on parameters. The latter utilizes historical measurement data to establish statistical or intelligent models, such as time series models, support vector machines, and neural networks, which have been widely used in practical engineering.

[0003] In recent years, deep learning methods, especially Long Short-Term Memory (LSTM) networks and Convolutional Neural Networks (CNNs) suitable for time-series modeling, have been introduced into wind power prediction. Related research shows that compared to traditional shallow models, deep neural networks can better characterize the temporal correlations and nonlinear coupling characteristics between SCADA (Supervisory Control and Data Acquisition) signals such as wind speed, engine speed, current, and active power, thus significantly improving prediction accuracy. However, most existing deep learning prediction methods typically use a single global model to train all operating data, assuming that the operating behavior of the unit under different wind conditions, loads, and control strategies can be uniformly described by the same model. This approach is prone to underfitting certain operating conditions and overfitting others when wind conditions change drastically and operating conditions are complex, leading to a significant decrease in the model's predictive performance for certain wind processes.

[0004] On the other hand, SCADA data from wind turbines contains a large number of invalid samples, including those from shutdowns, low wind speeds, or near-zero power. Directly incorporating these samples into modeling would not only dilute the training effect on critical operating ranges but also introduce significant distribution bias, impacting the model's generalization ability. While existing research has proposed removing some invalid data using simple thresholds or wind speed range-based methods, these often only consider a single physical quantity (such as wind speed or power) and lack a systematic assessment of "near-zero processes," making it difficult to balance data utilization with effective data removal.

[0005] Furthermore, in highly non-stationary wind farm environments, a single model often struggles to maintain stable predictive performance across all operating conditions. To improve robustness, some scholars have proposed integrating or weighted fusion of multiple models. However, existing fusion strategies often employ fixed weights or set weights solely based on offline evaluation results, failing to reflect the real-time performance of the model under current operating conditions and lacking the ability to adaptively select the locally optimal model for different conditions. Some works have introduced ideas based on similar wind process matching or multi-model voting, but these require maintaining large-scale historical sample databases or complex matching mechanisms, resulting in significant computational burden and implementation difficulties, hindering their widespread application in engineering fields.

[0006] Therefore, it is necessary to propose a wind turbine power prediction method with a relatively simple structure, which can make full use of SCADA multi-source information and has the ability to adapt to operating conditions and select dynamic models, so as to improve the robustness and engineering feasibility of the model while ensuring prediction accuracy. Summary of the Invention

[0007] This invention addresses the aforementioned problems in existing technologies by providing a multi-model collaborative fusion method and device for intelligent wind turbine power prediction. This method, based on fully utilizing multi-source SCADA monitoring data from wind turbines, divides the original operating data into wind processes and removes invalid wind processes. It constructs statistical features of wind processes and implements operating condition clustering, training deep learning prediction sub-models on different operating condition clusters, while simultaneously establishing a global prediction model covering all wind processes. Furthermore, it introduces a dynamic model selection and fusion mechanism based on historical prediction errors, adaptively selecting or weighting the optimal model between the cluster model stack and the global model stack. This significantly improves the accuracy and robustness of short-term wind turbine power prediction without significantly increasing online computational complexity.

[0008] A multi-model collaborative fusion method for intelligent prediction of wind turbine power, the method comprising:

[0009] Acquire wind turbine data and divide the data into several wind processes according to a fixed sample length. Calculate the various wind process indices and obtain the effective wind process set based on these indices. ;

[0010] Obtain the feature vectors of each effective wind process, and use the feature vectors to obtain the similarity matrix. Based on the spectral clustering algorithm, each effective wind process is divided into several operating condition clusters;

[0011] Construct sub-models for each operating condition cluster and form a stack of sub-models for each operating condition cluster; construct a global prediction model based on all effective wind processes and form a stack of global prediction models.

[0012] For the target wind process to be predicted, the predicted values ​​of the working condition cluster model and the global prediction model are obtained using the working condition cluster model and the global prediction model, and the power prediction value is obtained by dynamic weighting.

[0013] Furthermore, the wind turbine data includes wind speed. Impeller speed Generator speed Three-phase current of the power grid and active power To obtain a time-series dataset, missing data in wind turbine data is interpolated using the preceding or following normal data point. , The feature vector formed by wind turbine data. .

[0014] Furthermore, the calculation of each wind process index yields an effective wind process set based on these indexes. include:

[0015] For each wind process Define a near-zero indicator function:

[0016] ;

[0017] in For preset wind speed threshold and , For preset power threshold and The near-zero value of the wind process is obtained based on the near-zero indicator function. The near-zero percentage of each wind process is calculated as follows:

[0018] ;

[0019] in The sample size for the wind process is as follows, and the percentage of operational processes is:

[0020] ;

[0021] And the average wind speed is:

[0022] ;

[0023] make Indicates the near-zero value of the wind process. The length of the longest consecutive segment of "1"s, if it satisfies:

[0024] ;

[0025] This wind process is then considered invalid and discarded, resulting in the set of valid wind processes:

[0026] ;

[0027] in The threshold is near zero. The longest near-zero length threshold, To minimize the percentage of operation, The lower bound of wind speed and .

[0028] Furthermore, the feature vectors of each effective wind process are obtained, and a similarity matrix is ​​obtained using the feature vectors. include:

[0029] For any effective wind process Obtain statistical features of the wind process, and then obtain the feature vector of the wind process based on these statistical features. eigenvectors of all wind processes After robust standardization and dimensionality reduction, a radial basis function is used:

[0030] ;

[0031] Construct a similarity matrix ,in The kernel parameters are the radial basis function kernel similarity function; the statistical features of the wind process include average wind speed. Wind speed variance Average power Power variance Average rate of change of wind speed during wind process Power fluctuation rate Wind power ratio One or more of them.

[0032] Furthermore, the division of each wind process into several operating condition clusters based on the spectral clustering algorithm includes:

[0033] Construct degree matrix ,in:

[0034] ;

[0035] Construct the normalized Laplace matrix:

[0036] ;

[0037] Solve smallest The eigenvectors corresponding to each eigenvalue form the feature matrix. ,Will Treating the row vectors of the first sample as new samples, performing k-means clustering on them, and then... The individual wind processes were classified into cluster labels as follows: Within the operating condition cluster, among which This is a similarity matrix.

[0038] Furthermore, operating condition cluster sub-models are constructed for each operating condition cluster, forming a cluster model stack including:

[0039] For any cluster of operating conditions The time series data corresponding to all wind processes within the cluster are sorted by length. Sample construction is performed using a sliding window, for sample sequence number Define the input sequence

[0040] ;

[0041] The corresponding prediction is for the active power at the next sampling time immediately following this sequence:

[0042] ;

[0043] Will Flattened along the feature dimensions and standardized, the data is fed into the Long Short-Term Memory (LSTM) network model. The sub-model of this cluster's LSM network model is denoted as... ;

[0044] Assume that the Long Short-Term Memory (LSTM) network model unit is at time step The input is The hidden state is The memory state is Its state update equation can be written as:

[0045] ;

[0046] in For clusters The parameter set of the Long Short-Term Memory network model;

[0047] Utilize the hidden state of the last time step The predicted output is obtained after linear mapping:

[0048] ;

[0049] Training intra-cluster sub-models using mean squared error as the loss function:

[0050] ;

[0051] By minimizing Obtain the optimal parameters Sub-models of all clusters Together they form the cluster model stack;

[0052] The construction of a global prediction model based on all effective wind processes, and the formation of a global model stack, includes:

[0053] Set up effective wind processes All wind process data are spliced ​​together without clustering, using a length of [length missing]. Sample construction is performed using a sliding window, for sample sequence number Define the input sequence:

[0054] ;

[0055] The corresponding prediction is for the active power at the next sampling time immediately following this sequence:

[0056] ;

[0057] Using an LSTM model with the same structure as the operating condition cluster A global power prediction model is established, and its parameter training loss function is:

[0058] ;

[0059] in For global model output, To represent the true active power, after training, the global model... Together with the cluster model stack, it forms a multi-model prediction system.

[0060] Furthermore, before training the Long Short-Term Memory network, the wind processes within the operating condition cluster and the entire wind process are compared with the input samples and target data. Standardize them separately:

[0061] ;

[0062] ;

[0063] in , The mean-covariance correlation matrix is ​​calculated from the training set. , The mean and standard deviation of the power; the predicted output obtained from training. The predicted power in original units is obtained by inverse standardization:

[0064] .

[0065] Furthermore, for the target wind process to be predicted, at each time step Above, the corresponding working condition cluster sub-model is used simultaneously. With global prediction model Obtain the predicted value and ;

[0066] Let the historical error window length be... At any moment Construct the nearest Mean absolute error of the step:

[0067] ;

[0068] ;

[0069] in This is the actual power; according to and A model performance index is constructed, and dynamic model fusion is performed based on this index to obtain the final predicted power. .

[0070] Furthermore, at any moment Calculate the fusion weights:

[0071] ;

[0072] ;

[0073] in The temperature coefficient is used; the final predicted power is:

[0074] .

[0075] A multi-model collaborative fusion intelligent wind turbine power prediction device includes a processor and a memory. The memory stores programs or instructions that can run on the processor. When the programs or instructions are executed by the processor, they implement the steps of the method described above.

[0076] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0077] The adaptive power prediction method for wind turbines based on operating condition clustering and dynamic model selection provided by this invention first divides the original SCADA time series into several wind processes through a multi-index comprehensive judgment mechanism. Then, it uses indicators such as the near-zero proportion of wind speed and active power, the longest continuous near-zero duration, the operating proportion, and the average wind speed to accurately remove invalid wind processes such as shutdown and low wind speed. This reduces the interference of invalid data on model training from the source and improves the utilization rate of effective samples and the generalization ability of the prediction model.

[0078] Furthermore, this invention constructs statistical features of wind processes based on effective wind processes, uses spectral clustering algorithms to divide wind processes into several operating condition clusters with similar operating characteristics, and independently trains corresponding deep learning sub-models such as LSTM within each operating condition cluster. At the same time, it constructs a global prediction model covering all wind processes, thereby forming a multi-model prediction system that combines operating condition cluster sub-models with the global prediction model, effectively balancing local fitting accuracy and overall consistency under different wind conditions and power levels.

[0079] Finally, this invention proposes a dynamic model selection and fusion mechanism based on historical prediction errors. By evaluating the prediction errors of the most recent few steps online, it adaptively performs weighted fusion between the operating condition cluster sub-model and the global model, ensuring that the prediction output always tends towards the model with the smaller error under the current operating condition. This effectively suppresses the performance degradation of a single model under special operating conditions and significantly improves the accuracy and robustness of short-term power prediction for wind turbines. Compared with existing technologies that rely solely on a single global model or use fixed-weight fusion, this invention does not require the introduction of complex online similarity retrieval or large-scale sample matching processes. It can be implemented simply by adding a lightweight operating condition clustering and dynamic fusion module to the existing prediction model. The structure is simple and easy to implement in actual wind turbine controllers or edge computing platforms. Attached Figure Description

[0080] Figure 1 This is a flowchart of the wind turbine adaptive power prediction method according to the present invention.

[0081] Figure 2 This is a schematic diagram of the overall structure of the method in an embodiment of the present invention;

[0082] Figure 3 This is a diagram showing the wind process clustering results according to an embodiment of the present invention;

[0083] Figure 4 This is another wind process clustering result diagram according to an embodiment of the present invention;

[0084] Figure 5 This is another wind process clustering result diagram according to an embodiment of the present invention;

[0085] Figure 6 This is another wind process clustering result diagram according to an embodiment of the present invention;

[0086] Figure 7 This is a power prediction result diagram of an embodiment of the present invention during a wind test process;

[0087] Figure 8 For the embodiments of the present invention in Figure 7 The diagram shows the dynamic weight changes during the test wind process.

[0088] Figure 9This is a power prediction result diagram of an embodiment of the present invention during a test of another wind process;

[0089] Figure 10 For the embodiments of the present invention in Figure 9 The diagram shows the dynamic weight changes during the test wind process. Detailed Implementation

[0090] The present invention will now be described in detail through exemplary embodiments. However, it should be understood that, without further description, the methods, structures, and features of one embodiment may be advantageously incorporated into other embodiments.

[0091] This invention provides a method for intelligent prediction of wind turbine power through multi-model collaborative fusion, comprising the following steps.

[0092] S101: Acquire wind turbine data and divide the wind turbine data into several wind processes according to a fixed sample length. Calculate the various wind process indices and obtain the effective wind process set based on these indices. .

[0093] The wind turbine data includes data from the wind turbine's operation, typically collected by a wind turbine monitoring and data acquisition system (SCADA), with a sampling period of [missing information]. Discrete time is denoted as .

[0094] The wind turbine data includes wind speed. Impeller speed Generator speed Three-phase current of the power grid and active power In this embodiment, the wind turbine unit includes multiple or all of the wind speed components. Impeller speed Generator speed Three-phase current of the power grid and active power This includes data on various aspects of wind turbine operation, enabling a more accurate assessment of wind turbine power. The wind turbine data is used to construct the original data vector:

[0095] ;

[0096] Abnormal situations may occur during data acquisition or transmission, resulting in missing wind turbine data. To address the missing data in the wind turbine data, a time-series dataset is obtained by interpolating the missing data using the preceding or following normal data point.

[0097] .

[0098] When missing data exists in the wind turbine data, forward interpolation is preferred, using the nearest historical normal value to interpolate the missing data. If no forward normal value is available, backward interpolation is used, using the nearest backward normal value. Through forward or backward interpolation, all time points are considered. superior It is defined.

[0099] After obtaining the wind turbine data, the data is divided into several wind processes according to a fixed sample length. Specifically, the fixed sample length is... According to a fixed sample length With step size For time series datasets By dividing the window into several intervals, we obtain several wind process intervals:

[0100] ;

[0101] in For the first The number of samples for each wind process; when the data of undivided wind turbine units is sufficient to constitute a complete wind process during the wind turbine data partitioning process. When the data from undivided wind turbine units is insufficient to constitute a complete wind process, Incomplete wind processes may not be included in subsequent data processing.

[0102] Calculate various wind process indicators, including near-zero values. Nearly zero percentage Operating percentage Average wind speed The length of the longest consecutive segment of "1"s in near-zero values. The calculation methods for each indicator are as follows.

[0103] For each wind process Define a near-zero indicator function:

[0104] ;

[0105] in For preset wind speed threshold and , For preset power threshold and The near-zero value of the wind process is obtained based on the near-zero indicator function. The near-zero percentage of each wind process is calculated as follows:

[0106] ;

[0107] in The sample size for the wind process is as follows, and the percentage of operational processes is:

[0108] ;

[0109] And the average wind speed is:

[0110] ;

[0111] Indicates the near-zero value of the wind process. The length of the longest consecutive segment of "1"s, if it satisfies:

[0112] ;

[0113] in The threshold is near zero. The longest near-zero length threshold, To minimize the percentage of operation, The lower bound of wind speed and . Adjustments can be made based on unit capacity, wind resource conditions at the site, and communication bandwidth constraints to achieve a trade-off between prediction accuracy and computational complexity.

[0114] This wind process is then considered invalid and discarded, resulting in the set of valid wind processes:

[0115] .

[0116] A wind process is considered invalid and removed if it simultaneously meets the following criteria: a high percentage of near-zero wind speeds, an excessively long duration of consecutive near-zero wind speeds, a low percentage of winds in operation, and an average wind speed below vmin. Specific thresholds can be set according to different wind turbine types and wind resource conditions.

[0117] S102: Obtain the feature vectors of each effective wind process, and use the feature vectors to obtain the similarity matrix. Based on the spectral clustering algorithm, each effective wind process is divided into several operating condition clusters.

[0118] For any effective wind process eigenvectors It consists of statistical features of the wind process. The process involves acquiring these statistical features, obtaining feature vectors for the wind process based on these features, and then using these feature vectors to derive a similarity matrix. .

[0119] For example, the statistical characteristics of the wind process include the mean wind speed. Wind speed standard deviation Wind speed variance Average power Power variance Average rate of change of wind speed during wind process Power fluctuation rate Wind power ratio One or more of the following. Preferably, the feature vector. From average wind speed Average rate of change of wind speed during wind process Wind speed standard deviation Average power Power fluctuation rate The composition of wind power ratio is as follows:

[0120] .

[0121] eigenvectors of all wind processes After robust standardization and dimensionality reduction, a radial basis function is used:

[0122] ;

[0123] Construct a similarity matrix ,in represents the kernel parameters of the radial basis kernel similarity function.

[0124] Specifically, for any effective wind process The statistical characteristics of each wind process are calculated as follows:

[0125] ;

[0126] ;

[0127] ;

[0128] ;

[0129] ;

[0130] ;

[0131] ;

[0132] ;

[0133] In the formula To prevent tiny constants with a denominator of zero.

[0134] Using the obtained similarity matrix Based on the spectral clustering algorithm, each wind process is divided into several operating condition clusters. Specifically, a degree matrix is ​​constructed. ,in:

[0135] ;

[0136] Construct the normalized Laplace matrix:

[0137] ;

[0138] Solve smallest The eigenvectors corresponding to each eigenvalue form the feature matrix. ,Will Treating the row vectors of the first sample as new samples, performing k-means clustering on them, and then... The individual wind processes were classified into cluster labels as follows: Within the operating condition cluster, among which This is a similarity matrix. Therefore, the wind process is divided into... The operating condition cluster, recorded as the... The cluster number to which this wind process belongs is:

[0139] ;

[0140] A cluster model generation module is established for each working condition cluster, which is used for subsequent training of the corresponding deep learning sub-models by cluster.

[0141] S103: Construct sub-models for each operating condition cluster and form a stack of sub-models for each operating condition cluster; construct a global prediction model based on all effective wind processes and form a stack of global prediction models.

[0142] For any cluster of operating conditions The time series data corresponding to all wind processes within the cluster are sorted by length. Sample construction is performed using a sliding window, for sample sequence number Define the input sequence

[0143] ;

[0144] The corresponding prediction target is the active power at the next time step:

[0145] ;

[0146] Will Flattened along the feature dimensions and standardized, the data is fed into the Long Short-Term Memory (LSTM) network model. The sub-model of this cluster's LSM network model is denoted as... The standardization process can be z-score standardization or other standardization methods, which are not limited here. Standardization can eliminate differences in the dimensions and numerical ranges of different features, thus enabling better training of neural network models.

[0147] Assume that the Long Short-Term Memory (LSTM) network model unit is at time step The input is The hidden state is The memory state is Its state update equation can be written as:

[0148] ;

[0149] in For clusters The parameter set of the Long Short-Term Memory Network model.

[0150] Utilize the hidden state of the last time step The predicted output is obtained after linear mapping:

[0151] ;

[0152] in For output layer weights, For output layer bias, Indicates a cluster of operating conditions , Indicates that for the first Using samples, train intra-cluster sub-models with mean squared error as the loss function:

[0153] ;

[0154] in For predicted values, The true value is obtained. Minimize using stochastic gradient descent or adaptive optimization algorithms. Obtain the optimal parameters Sub-models of all clusters Together they form the cluster model stack.

[0155] A global prediction model is constructed based on all effective wind processes, forming a global model stack. Specifically, the effective wind processes are set... All wind process data are spliced ​​together without clustering, using a length of [length missing]. Sample construction is performed using a sliding window, for sample sequence number Define the input sequence:

[0156] ;

[0157] The corresponding prediction is for the active power at the next sampling time immediately following this sequence:

[0158] ;

[0159] Using an LSTM model with the same structure as the operating condition cluster A global power prediction model is established, and its parameter training loss function is:

[0160] ;

[0161] in For global model prediction output, To represent the true active power, after training, the global model... Together with the cluster model stack, it forms a multi-model prediction system.

[0162] In another embodiment of the present invention, the wind processes within the operating condition cluster and the entire wind process are compared with the input samples and target data before long short-term memory network training. Standardize them separately:

[0163] ;

[0164] ;

[0165] in , The mean-covariance correlation matrix is ​​calculated from the training set. , The mean and standard deviation of the power are given.

[0166] Predicted output obtained from training The predicted power in original units is obtained by inverse standardization:

[0167] .

[0168] S104: For the target wind process to be predicted, the predicted values ​​of the working condition cluster model and the global prediction model are obtained using the working condition cluster model and the global prediction model, and the power prediction value is obtained by using a dynamic weighting method.

[0169] The operating condition cluster sub-model stack is connected in series with the global prediction model stack. For the target wind process to be predicted, at each time step... Above, the sub-models of the corresponding working condition clusters are used simultaneously. With global model Obtain the predicted value and Let the length of the historical error window be... At any moment Construct the nearest Mean absolute error of the step:

[0170] ;

[0171] ;

[0172] in This is the actual power; according to and Construct a model performance index, and use this index to perform dynamic model fusion to obtain the final predicted value. Historical error window length Adjustments can be made based on unit capacity, wind resource conditions at the site, and communication bandwidth constraints to achieve a trade-off between prediction accuracy and computational complexity.

[0173] At any moment Calculate the fusion weights:

[0174] ;

[0175] ;

[0176] in This is the temperature coefficient.

[0177] The final predicted value is:

[0178] .

[0179] Taking the 2MW wind turbine unit B-35 of the Taipingli Phase II wind farm in Hunan Province as the research object, this paper describes the adaptive power prediction method based on operating condition clustering and dynamic model selection described in this invention.

[0180] The SCADA sampling cycle for this unit is 5 minutes, with 288 sampling points per 24 hours. After removing days with obvious errors and severe missing data, a total of 369 valid wind processes were obtained as the training set and 180 wind processes as the test set. Each wind process corresponds to 24 hours of continuous operating data for this unit.

[0181] At each sampling time The following 7 features are selected as model inputs: wind speed Impeller speed (spindle speed) Generator speed Three-phase current of the power grid Active power Let the eigenvectors be:

[0182] ;

[0183] After filling in the missing values ​​for each of the above quantities using forward and backward padding, the original sample sequence is formed. The window length for wind process segmentation is taken as... Step length The near-zero threshold is determined by wind speed. ,power Near-zero percentage threshold The longest consecutive near-zero ratio is 0.4 (i.e.) Minimum operating percentage Lower limit of average wind speed The wind process characteristics were selected from six quantities: average wind speed, wind speed change rate, wind speed standard deviation, average power, power fluctuation rate, and wind power ratio. The number of operating condition clusters was obtained through spectral clustering. The LSTM prediction model has an input dimension of 7 and a time window length of [missing information]. Number of hidden layer units The network has 1 layer and the batch size is [missing information]. Number of training rounds Learning rate The length of the historical error window in the dynamic fusion part is taken as... Softmax temperature coefficient The structural block diagram of the method proposed in this invention is shown in the figure below. Figure 2 middle.

[0184] The adaptive power prediction method for wind turbines based on operating condition clustering and dynamic model selection described in this invention was used to conduct experiments on the measured data of the 2 MW wind turbine B-35 in Taipingli Phase II, Hunan Province. The results are shown in Table 1 and below. Figures 3 to 7 Table 1 presents a comparison of the prediction model performance on two typical test wind processes: For test wind process 64 in cluster 0, the fusion model slightly outperforms the global model in terms of MAE and RMSE, with the coefficient of determination increasing from 0.94 to 0.95 and SMAPE slightly decreasing. This indicates that under the condition that the global model has already fitted well, the fusion strategy will not destroy its overall performance and can further bring a slight improvement in accuracy. For test wind process 62 in cluster 1, the MAE of the fusion model decreased from 19.58 to 14.52, the RMSE decreased from 37.83 to 35.33, and the coefficient of determination increased from 0.69 to 0.73. This shows that under the condition that the global model's prediction performance is relatively poor, the fusion model can significantly improve the power prediction accuracy.

[0185] Figures 3-6 The segmentation results of the wind process show that there are obvious distinctions between different working condition clusters in the feature space, which provides a basis for subsequent cluster-based modeling. Figure 7 and Figure 9Comparison curves of power prediction results during the test wind process are presented respectively. The prediction trajectory of the fusion model is closer to the actual power curve as a whole, especially in the range where wind speed and power fluctuate greatly, and it can track the power change trend better. Figure 8 and Figure 10 The dynamic weight changes for the two corresponding wind processes are shown: during periods of stable conditions and good global model prediction, the global model weight is close to 1; while during periods of local conditions deviating from the global distribution or with strong volatility, the weights of sub-models within the cluster increase significantly. This indicates that the fusion strategy can adaptively switch between the two types of models based on historical prediction errors, avoiding performance degradation of a single model under specific conditions. (See Table 1 and...) Figures 3 to 10 The results show that the method described in this invention effectively distinguishes and models different operating conditions on actual wind turbine data, and significantly improves the accuracy and robustness of short-term power prediction through dynamic model selection and fusion.

[0186] Table 1 Performance Comparison of Prediction Models

[0187]

[0188] In another embodiment, the present invention also provides a multi-model collaborative fusion wind turbine power intelligent prediction device, including a processor and a memory. The memory stores programs or instructions that can run on the processor. When the programs or instructions are executed by the processor, they implement the steps of the method described above, which will not be repeated here.

[0189] The above embodiments are used to explain the present invention, but not to limit the present invention. Any modifications and changes made to the present invention within the spirit and scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A method for intelligent prediction of wind turbine power through multi-model collaborative fusion, characterized in that, The method includes: Acquire wind turbine data and divide the data into several wind processes according to a fixed sample length. Calculate the various wind process indices and obtain the effective wind process set based on these indices. ; Obtain the feature vectors of each effective wind process, and use the feature vectors to obtain the similarity matrix. Based on the spectral clustering algorithm, each effective wind process is divided into several operating condition clusters; Construct sub-models for each operating condition cluster and form a stack of sub-models for each operating condition cluster; construct a global prediction model based on all effective wind processes and form a stack of global prediction models. For the target wind process to be predicted, the predicted values ​​of the operating condition cluster model and the global prediction model are obtained using the operating condition cluster model and the global prediction model, respectively. A dynamic weighting method is then used to obtain the predicted power value. The wind turbine data includes wind speed. Impeller speed Generator speed Three-phase current of the power grid and active power To obtain a time-series dataset, missing data in wind turbine data is interpolated using the preceding or following normal data point. , The feature vector formed by wind turbine data. The calculation of each wind process index yields an effective wind process set based on these indexes. include: For each wind process Define a near-zero indicator function: ； in For preset wind speed threshold and , For preset power threshold and The near-zero value of the wind process is obtained based on the near-zero indicator function. The near-zero percentage of each wind process is calculated as follows: ； in The sample size for the wind process is as follows, and the percentage of operational processes is: ； And the average wind speed is: ； make Indicates the near-zero value of the wind process. The length of the longest consecutive segment of "1"s in the middle, if it satisfies: ； This wind process is then considered invalid and discarded, resulting in the set of valid wind processes: ； in The threshold is near zero. The longest near-zero length threshold, To minimize the percentage of operation, The lower bound of wind speed and .

2. The intelligent prediction method for wind turbine power based on multi-model collaborative fusion as described in claim 1, characterized in that... The feature vectors of each effective wind process are obtained, and a similarity matrix is ​​obtained using the feature vectors. include: For any effective wind process Obtain statistical features of the wind process, and then obtain the feature vector of the wind process based on these statistical features. eigenvectors of all wind processes After robust standardization and dimensionality reduction, a radial basis function is used: ； Construct a similarity matrix ,in The kernel parameters are the radial basis function kernel similarity function; the statistical features of the wind process include average wind speed. Wind speed variance Average power Power variance Average rate of change of wind speed during wind process Power fluctuation rate Wind power ratio One or more of them.

3. The intelligent prediction method for wind turbine power based on multi-model collaborative fusion according to claim 2, characterized in that... The spectral clustering algorithm divides each wind process into several operating condition clusters, including: Construct degree matrix ,in: ； Construct the normalized Laplace matrix: ； Solve smallest The eigenvectors corresponding to each eigenvalue form the feature matrix. ,Will Treating the row vectors of the first sample as new samples, performing k-means clustering on them, and then... The individual wind processes were classified into cluster labels as follows: Within the operating condition cluster, among which This is a similarity matrix.

4. The intelligent prediction method for wind turbine power based on multi-model collaborative fusion as described in claim 1, characterized in that... Constructing sub-models for each operating condition cluster and forming a cluster model stack includes: For any cluster of operating conditions The time series data corresponding to all wind processes within the cluster are sorted by length. Sample construction is performed using a sliding window, for sample sequence number Define the input sequence: ； The corresponding prediction is for the active power at the next sampling time immediately following this sequence: ； Will Flattened along the feature dimensions and standardized, the data is fed into the Long Short-Term Memory (LSTM) network model. The sub-model of this cluster's LSM network model is denoted as... ; Assume that the Long Short-Term Memory (LSTM) network model unit is at time step The input is The hidden state is The memory state is Its state update equation can be written as: ； in For clusters The parameter set of the Long Short-Term Memory network model; Utilize the hidden state of the last time step The predicted output is obtained after linear mapping: ； Training intra-cluster sub-models using mean squared error as the loss function: ； By minimizing Obtain the optimal parameters Sub-models of all clusters Together they form the cluster model stack; The construction of a global prediction model based on all effective wind processes, and the formation of a global model stack, includes: Set up effective wind processes All wind process data are spliced ​​together without clustering, using a length of [length missing]. Sample construction is performed using a sliding window, for sample sequence number Define the input sequence: ； The corresponding prediction is for the active power at the next sampling time immediately following this sequence: ； Using an LSTM model with the same structure as the operating condition cluster A global power prediction model is established, and its parameter training loss function is: ； in For global model output, To represent the true active power, after training, the global model... Together with the cluster model stack, it forms a multi-model prediction system.

5. The intelligent prediction method for wind turbine power based on multi-model collaborative fusion according to claim 4, characterized in that... Before training the Long Short-Term Memory network, the wind processes within the operating condition cluster and the entire wind process are compared with the input samples. and target Standardize them separately: ； ； in , The mean-covariance correlation matrix is ​​calculated from the training set. , The mean and standard deviation of the power; the predicted output obtained from training. The predicted power in original units is obtained by inverse standardization: 。 6. The intelligent prediction method for wind turbine power based on multi-model collaborative fusion according to claim 1, characterized in that... For the target wind process to be predicted, at each time step Above, the corresponding working condition cluster sub-model is used simultaneously. With global prediction model Obtain the predicted value and ; Let the historical error window length be... At any moment Construct the nearest Mean absolute error of the step: ； ； in This is the actual power; according to and A model performance index is constructed, and dynamic model fusion is performed based on this index to obtain the final predicted power. .

7. The intelligent prediction method for wind turbine power based on multi-model collaborative fusion according to claim 6, characterized in that, At any moment Calculate the fusion weights: ； ； in The temperature coefficient is used; the final predicted power is: 。 8. A multi-model collaborative fusion intelligent power prediction device for wind turbines, characterized in that, It includes a processor and a memory, the memory storing a program or instructions that can run on the processor, the program or instructions being executed by the processor to implement the steps of the method as described in any one of claims 1-7.

Images