A wind farm theoretical power prediction method based on power cut-off identification and a terminal device

Abstract

The application discloses a wind farm theoretical power prediction method based on power limitation identification and a terminal device, belongs to the technical field of wind power prediction, and can solve the problem that the existing wind power prediction method underestimates the power generation capacity of a wind farm, thereby affecting the adaptability and prediction accuracy of a prediction model. The method comprises the following steps: S1, constructing a meteorological feature input set; S2, determining a power limitation data set and a non-power limitation abnormal operation data set from historical operation data of the wind farm; S3, dividing wind turbines of the wind farm into a first relevant group and a second relevant group, and respectively constructing corresponding theoretical power models; S4, correcting power of the power limitation data set by using the theoretical power model, and correcting power of the non-power limitation abnormal operation data set by using an interpolation method to obtain a wind farm corrected power sequence; and S5, inputting the wind farm corrected power sequence and the meteorological feature input set into a power prediction model to predict the theoretical power of the wind farm. The application is used for power prediction of the wind farm.

Description

Technical Field

[0001] This invention relates to a method and terminal equipment for predicting the theoretical power of wind farms based on power curtailment identification, belonging to the field of wind power prediction technology. Background Technology

[0002] Wind power, as a clean and sustainable energy source, is playing an increasingly prominent role in my country's energy system. However, unlike traditional dispatchable energy sources such as thermal and hydropower, wind power output is highly dependent on natural weather conditions and is difficult to control artificially. Therefore, highly reliable wind power forecasting technology is crucial for ensuring the safe and stable operation of the power grid.

[0003] The accuracy of wind power prediction is highly dependent on the quality of input data. However, in actual operation, wind farm data acquisition is often affected by factors such as data acquisition device failures, transmission anomalies, and power curtailment, resulting in a large number of outliers and missing values ​​in the power data, which seriously affects the reliability of prediction. Especially under large-scale wind power grid-connected operation, wind farms are generally subject to power curtailment due to grid dispatch constraints. During curtailment periods, the actual output power of wind turbines is artificially suppressed under conditions of sufficient wind resources, causing the wind speed-power relationship to deviate from the wind turbine power curve. However, this deviation is not caused by equipment failure or measurement error, but reflects the grid dispatch's active control over wind turbine output. Existing wind power prediction methods usually treat such power deviations, along with failures, shutdowns, or measurement anomalies, as abnormal data and discard or simply correct them, failing to effectively distinguish between curtailment and non-curtailment anomaly states. This easily leads to the misjudgment of curtailment data with clear physical significance as invalid anomalies, resulting in a systematic underestimation of the wind farm's generating capacity, and consequently affecting the adaptability of the prediction model to real operating conditions and the prediction accuracy.

[0004] Meanwhile, current wind power forecasting methods are still insufficient in the collaborative utilization of multi-source heterogeneous data. They fail to fully integrate information such as refined observation data from meteorological towers, numerical weather prediction, and wind farm operation data, making it difficult for models to fully capture the key changing characteristics and complex environmental dependencies of wind power sequences. Summary of the Invention

[0005] This invention provides a method and terminal equipment for predicting the theoretical power of wind farms based on power curtailment identification, which can solve the problem that existing wind power prediction methods underestimate the generating capacity of wind farms, thereby affecting the adaptability of prediction models to real operating conditions and the accuracy of prediction.

[0006] On the one hand, the present invention provides a method for predicting the theoretical power of wind farms based on power curtailment identification, the method comprising:

[0007] S1. Construct a meteorological feature input set based on historical operation data and numerical weather prediction data of the wind farm;

[0008] S2. Identify abnormal data in the historical operation data to obtain a power rationing dataset and a non-power rationing abnormal operation dataset.

[0009] S3. Divide the wind turbines of the wind farm into a first correlation group and a second correlation group according to the correlation of wind turbine wind speed, and construct a first theoretical power model corresponding to the first correlation group and a second theoretical power model corresponding to the second correlation group respectively.

[0010] S4. The theoretical power of the power curtailment data belonging to the first relevant group in the power curtailment dataset is corrected using the first theoretical power model; the theoretical power of the power curtailment data belonging to the second relevant group in the power curtailment dataset is corrected using the second theoretical power model; and the theoretical power of the non-power curtailment abnormal operation dataset is corrected using interpolation, to obtain the wind farm corrected power sequence.

[0011] S5. Input the corrected power sequence of the wind farm, the meteorological feature input set, and the wind farm's wind tower observation data into the power prediction model to obtain the theoretical power prediction sequence of the wind farm.

[0012] Optionally, S1 specifically includes:

[0013] Correlation analysis was performed on various meteorological variables in numerical weather prediction data and wind power in historical operation data of wind farms. Meteorological variables with a correlation degree that met the preset requirements were used to form a meteorological feature input set.

[0014] Optionally, S2 specifically includes:

[0015] A dynamic power threshold is set based on the standard power curve of the wind turbines in the wind farm.

[0016] Based on the dynamic power threshold and the characteristic parameters of the wind turbine, abnormal data are identified in all data points of the historical operation data to obtain a power curtailment dataset and a non-power curtailment abnormal operation dataset.

[0017] Optionally, step S3, which divides the wind turbines of the wind farm into a first correlation group and a second correlation group according to the correlation of wind turbine wind speed, specifically includes:

[0018] A multicollinearity test was performed on the historical wind speed sequences of each wind turbine in the historical operating data. Based on the test results, the wind turbines of the wind farm were divided into a first correlation group and a second correlation group.

[0019] Optionally, the construction of the first theoretical power model corresponding to the first related group in S3 specifically includes:

[0020] Determine the first input parameters for the first relevant group; the first input parameters include the median wind speed, the number of operating units, and the median power.

[0021] A first theoretical power model is constructed based on the first input parameters.

[0022] Optionally, the construction of the second theoretical power model corresponding to the second related group in S3 specifically includes:

[0023] Determine the second input parameters for the second relevant group; the second input parameters include the wind speed sequence of each fan;

[0024] A second theoretical power model is constructed based on the second input parameters.

[0025] Optionally, the power prediction model is a CNN-LSTM-lightGBM combined prediction model; S5 specifically includes:

[0026] The model input parameters are respectively input into the CNN-LSTM layer and the lightGBM layer of the combined prediction model to obtain the first-stage power prediction results and the second-stage power prediction results; the model input parameters include the wind farm corrected power sequence, the meteorological feature input set, and the wind farm's wind tower observation data;

[0027] The power prediction results of the first stage and the power prediction results of the second stage are weighted and summed to obtain the theoretical power prediction sequence of the wind farm.

[0028] Optionally, the weighting coefficients of the first-stage power prediction results and the second-stage power prediction results are determined by the inverse error weighting method.

[0029] Optionally, prior to S5, the method further includes:

[0030] The corrected power sequence of the wind farm and the meteorological feature input set are normalized respectively.

[0031] On the other hand, the present invention provides a terminal device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the wind farm theoretical power prediction method based on power curtailment identification as described above.

[0032] The beneficial effects that this invention can produce include:

[0033] The wind farm theoretical power prediction method based on power curtailment identification provided by this invention achieves multi-source information complementarity by integrating historical wind farm operation data, meteorological tower observation data, and numerical weather prediction data. It also effectively improves the effectiveness and prediction stability of the model input features by screening meteorological features with high explanatory power for power prediction based on two-dimensional correlation analysis. Furthermore, it achieves accurate identification of power curtailment data through dynamic power threshold boundary detection based on the standard power curve of the wind turbine, and identifies and filters non-curtailment abnormal operation data by combining the physical constraints of the wind turbine power curve. Based on the multicollinearity test results, it performs cluster modeling of wind turbine units and constructs groups suitable for strong correlation. The theoretical power model for weakly correlated groups is corrected using a theoretical power inversion method for the identified power curtailment data. For abnormal operation data without power curtailment, a combination of linear interpolation and nearest neighbor interpolation is used for data recovery and filling. This avoids interference from power curtailment and abnormal operation data with model training and prediction results. At the same time, a combined prediction model integrating CNN, LSTM, attention mechanism and lightGBM is constructed, and an error reciprocal weighted fusion strategy based on historical prediction error is adopted to fully explore the temporal characteristics of wind power series. Finally, it realizes ultra-short-term power prediction for wind farms and meets the practical application needs of grid dispatch and wind farm operation management. Attached Figure Description

[0034] Figure 1 A flowchart of a wind farm theoretical power prediction method based on power curtailment identification provided in an embodiment of the present invention;

[0035] Figure 2 A flowchart of a wind farm power curtailment data correction method provided in an embodiment of the present invention;

[0036] Figure 3 This is a schematic diagram of the wind turbine power limitation data before correction, provided in an embodiment of the present invention.

[0037] Figure 4 This is a schematic diagram showing the corrected power limitation data for wind turbines provided in an embodiment of the present invention;

[0038] Figure 5 This is a schematic diagram of the prediction process of the combined prediction model provided in an embodiment of the present invention;

[0039] Figure 6 Power prediction curves for different prediction models provided in embodiments of the present invention;

[0040] Figure 7 Power prediction curves for different prediction methods provided in embodiments of the present invention. Detailed Implementation

[0041] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0042] This invention provides a method for predicting the theoretical power of wind farms based on power curtailment identification, such as... Figure 1 As shown, the method includes:

[0043] S1. Construct a meteorological feature input set based on the historical operation data of the wind farm and numerical weather prediction data.

[0044] Specifically, this includes: conducting correlation analysis on various meteorological variables in numerical weather forecast data of wind farms and wind power in historical operation data of wind farms, and forming a meteorological feature input set by meteorological variables whose correlation degree meets the preset requirements.

[0045] In practical implementation, the first step is to acquire multi-source data related to wind power. This multi-source data includes at least historical operational data of the wind farm, meteorological tower observation data, and numerical weather prediction data. The historical operational data of the wind farm includes the output power of individual wind turbines and the corresponding historical wind speed and direction sequences for continuous time steps, as well as the overall wind farm output power for continuous time steps. This data characterizes the power generation performance and environmental response of individual wind turbines and the entire wind farm over historical periods. The meteorological tower observation data includes observation data of various meteorological variables for continuous time steps, providing a record of actual meteorological conditions in the wind farm area. The numerical weather prediction data includes predicted data of various meteorological variables for continuous time steps. Meteorological variables include wind speed, wind direction, temperature, humidity, and air pressure at different heights, enabling a quantitative description of meteorological conditions from multiple dimensions.

[0046] In this embodiment, considering that wind power is influenced by multiple meteorological factors, there may be significant linear relationships or complex nonlinear relationships between different meteorological variables and wind power. To avoid the one-sided results of feature selection due to relying solely on a single correlation analysis method, before constructing the power prediction model, meteorological variables in the numerical weather prediction data are screened based on two-dimensional correlation analysis to obtain the meteorological feature input set for the power prediction model. Specifically, based on historical wind farm operation data, the degree of correlation between each meteorological variable in the numerical weather prediction and wind power is evaluated from both linear and nonlinear correlation dimensions. This embodiment uses a method combining the Pearson Correlation Coefficient (PCC) and the Maximum Information Coefficient (MIC) to comprehensively screen meteorological features.

[0047] The Pearson correlation coefficient is used to measure the degree of linear correlation between meteorological variables and wind power, and it is calculated as follows:

[0048] ;

[0049] In the formula, Represents a sequence of meteorological variables. This represents the wind power sequence corresponding to the time step. Indicates the covariance of the two. and These represent the standard deviations of meteorological variables and wind power, respectively. The Pearson correlation coefficient is used to... The absolute value of the value represents the degree of linear influence of meteorological variables on changes in wind power.

[0050] Meanwhile, to characterize the potential nonlinear relationship between meteorological variables and wind power, a maximum mutual information coefficient is introduced to evaluate their nonlinear correlation. The maximum mutual information coefficient is based on information theory principles, and its core lies in characterizing the information dependency between two variables through mutual information.

[0051] Among them, meteorological variables With wind power Mutual information between them is defined as:

[0052] ;

[0053] In the formula, Representing meteorological variables With wind power The joint probability distribution, and These represent their marginal probability distributions. Mutual information measures the amount of information shared between variables; the higher the value, the stronger the correlation between them.

[0054] Based on this, to eliminate the influence of different sample sizes and grid partitioning methods on the mutual information value, the mutual information can be... Normalization is performed to comprehensively characterize the nonlinear relationship between meteorological variables and wind power.

[0055] After calculating the Pearson correlation coefficient between each meteorological variable and the total wind power output, and mutual information Then, meteorological variables whose correlation degree meets the preset requirements are combined into a meteorological feature input set.

[0056] S2. Identify abnormal data in historical operation data to obtain power rationing dataset and non-power rationing abnormal operation dataset.

[0057] Specifically, this includes: first, setting a dynamic power threshold based on the standard power curve of the wind turbines in the wind farm; then, identifying abnormal data for all data points in the historical operation data based on the dynamic power threshold and the characteristic parameters of the wind turbines, to obtain a power curtailment dataset and a non-power curtailment abnormal operation dataset.

[0058] In this embodiment, to accurately identify the power-limited and non-power-limited abnormal operating states of the wind turbine during operation, a dynamic power threshold boundary is constructed based on the standard power curve parameters provided by the wind turbine manufacturer to identify abnormal data in the historical operating data of the wind turbine. Unlike common power data cleaning methods that simply remove or correct abnormal power data using uniform rules, this embodiment does not uniformly regard all operating data deviating from the standard power curve as invalid anomalies. Instead, it distinguishes and identifies the causes of abnormal data based on the physical constraints of the wind turbine's standard power curve. When the wind turbine has sufficient wind resources, but its measured power is lower than the theoretical power output corresponding to the power curve and meets the continuity criterion, the operating state is determined to be a power-limited operating state, and the corresponding data is the power-limited data. When the wind turbine power and wind speed clearly violate the physical constraints of the power curve and do not have the characteristics of power-limited operation, it is determined to be a non-power-limited abnormal operating state.

[0059] Specifically, for any wind turbine, its standard power curve and cut-in wind speed Rated wind speed Cut-off wind speed and rated power Closely related. Assuming at time step... Below, the measured wind speed and measured power of the fan are respectively and .

[0060] During the identification of abnormal operation data without power curtailment, the following criteria are used to identify historical operation data, taking into account the physical constraints of the wind turbine power curve:

[0061] when or When this occurs, the data is determined to be abnormal operation data not related to power rationing.

[0062] when and When the data exceeds 3% to 5% of the maximum power generation capacity of a single wind turbine, it is determined to be abnormal operation data without power curtailment.

[0063] when and When the data exceeds 3% to 5% of the maximum power generation capacity of a single wind turbine, it is determined to be abnormal operation data without power curtailment.

[0064] when At that time, if If two or more consecutive time steps are both 0, the fan will stop at the corresponding time step, and the data will be judged as abnormal operation data not due to power outage.

[0065] Identifying power-limited operation requires constructing a dynamic threshold that reflects the minimum power level of the wind turbine during normal operation. Based on the aforementioned power curves and characteristic parameters, a piecewise dynamic power threshold function is constructed within different wind speed ranges to characterize the minimum reasonable output power of the wind turbine under normal operating conditions. Its general form can be expressed as:

[0066] ;

[0067] in, Indicates the first Wind speed threshold nodes This represents the power ratio coefficient within the corresponding wind speed range. Different wind turbine models have different wind speed nodes and ratio coefficients set according to their factory power curve parameters, thereby forming a dynamic power threshold boundary that matches the wind turbine model.

[0068] In the process of identifying power rationing data, non-power rationing abnormal data in historical operation data are first removed, and then the remaining data is used to make the following power rationing judgment.

[0069] Define single time step conditions :

[0070] ;

[0071] If there exist at least two consecutive time steps that satisfy the condition These time steps are then classified as power rationing data; if a certain time step satisfies However, if the above-mentioned persistence condition is not met, i.e., it is an isolated single point, then it is judged as random non-power-limited abnormal data.

[0072] By employing the aforementioned multi-condition joint judgment method, the power curtailment status and non-power curtailment abnormal operation status of wind turbines can be effectively identified under different wind speed conditions. The identified power curtailment data and non-power curtailment abnormal operation data are not directly discarded, but rather marked and temporarily stored. In subsequent theoretical power calculations and data correction stages, the power data for the corresponding time period is inverted and filled in, thereby avoiding the adverse impact of abnormal data on model training and prediction results.

[0073] S3. Divide the wind turbines of the wind farm into a first correlation group and a second correlation group according to the correlation of the wind turbine wind speed, and construct a first theoretical power model corresponding to the first correlation group and a second theoretical power model corresponding to the second correlation group respectively.

[0074] The above-mentioned division of wind turbines in a wind farm into a first correlation group and a second correlation group based on the correlation of wind turbine wind speeds includes: performing multicollinearity tests on the historical wind speed sequences of each wind turbine in historical operating data, and dividing the wind turbines in the wind farm into a first correlation group and a second correlation group based on the test results.

[0075] By performing multicollinearity tests on historical wind speed data of wind turbines, the wind turbines in the wind farm are divided into the first correlation group (also known as the strong correlation group) and the second correlation group (also known as the weak correlation group).

[0076] Assuming there are a total of Typhoon machines, at the same time Below, the wind speed sequence is formed by the hub height wind speed of each wind turbine. Its expression is as follows:

[0077] .

[0078] Because wind turbines are spatially concentrated within a wind farm, there may be a significant linear correlation between the wind speed sequences of different turbines. When the wind speed of a particular turbine... When the wind speed can be approximated by a linear combination of the wind speeds of the other fans, the following relationship exists:

[0079] ;

[0080] in, For regression coefficients, If the term is a residual, it indicates the presence of multicollinearity in the wind turbine speed data. Multicollinearity leads to instability in the parameters of the power fitting model, thus affecting the accuracy of theoretical power inversion.

[0081] To avoid the influence of collinearity in wind speed data, a collinearity test based on the variance expansion coefficient is first performed on the historical wind speed series of the wind turbines. Variance expansion coefficient The formula used to evaluate the correlation between the wind speed of each fan and the wind speed of other fans is as follows:

[0082] ;

[0083] in, Indicates the first The regression determination coefficient between the wind speed of the j-th fan and the wind speeds of all other fans can be obtained by fitting a regression equation with the wind speed sequence of the j-th fan as the dependent variable and the wind speed sequences of all other fans as independent variables. It is generally believed that when... When there is no multicollinearity; when At that time, there is strong multicollinearity; when At that time, severe multicollinearity exists. If If the wind turbine is classified as part of a strongly correlated group, it will be placed in a weakly correlated group; otherwise, it will be classified as part of a weakly correlated group. This method effectively identifies redundant information between wind speeds and avoids the impact of multicollinearity on the model.

[0084] The aforementioned construction of the first theoretical power model corresponding to the first relevant group specifically includes:

[0085] First, determine the first input parameters for the first relevant group; the first input parameters include the median wind speed, the number of operating units, and the median power; then, construct the first theoretical power model based on the first input parameters.

[0086] For the first correlation group (i.e., the strongly correlated group), due to the high redundancy of its wind speed information, using the group's wind speed statistical features instead of individual machine wind speeds as model input can reduce noise interference and improve model stability. Let the time step be... The fleet includes The wind speed sequence of the operating wind turbines is as follows:

[0087] ;

[0088] Then the median wind speed of the aircraft group The definition is as follows:

[0089] when When it is an odd number, ;

[0090] when When it is even, ;

[0091] The wind speed sequence has been sorted in ascending order of numerical value.

[0092] Group power statistics were used instead of individual unit power as model input; in the strongly correlated group data after removing power curtailment data and abnormal operation data without power curtailment, the median wind speed was determined. A similar method was used to determine the median power. .

[0093] Based on this, the median wind speed Median power and the number of machines started at the corresponding time step Using these as input features, a first theoretical power model based on a BP neural network is constructed. The mapping relationship can be expressed as:

[0094] .

[0095] The above-mentioned construction of the second theoretical power model corresponding to the second related group specifically includes:

[0096] First, determine the second input parameters for the second relevant group; the second input parameters include the wind speed sequence of each wind turbine; then construct the second theoretical power model based on the second input parameters.

[0097] For the second correlation group (i.e., the weak correlation group), the correlation between wind speeds is relatively weak. Retaining the wind speed of individual turbines helps improve the accuracy of theoretical power inversion. Therefore, the wind speed sequences of each turbine are used directly. As input, a second theoretical power model based on a BP neural network is established. The mapping relationship can be expressed as:

[0098] .

[0099] S4. For the power curtailment data belonging to the first relevant group in the power curtailment dataset, the theoretical power is corrected using the first theoretical power model. For the power curtailment data belonging to the second relevant group in the power curtailment dataset, the theoretical power is corrected using the second theoretical power model. For the non-power curtailment abnormal operation dataset, the theoretical power is corrected using the interpolation method, thus obtaining the corrected power sequence of the wind farm.

[0100] refer to Figure 2 As shown, theoretical power inversion and filling are performed on data marked as being under power curtailment (i.e., power curtailment data) using the first and second theoretical power models. The power curtailment data is corrected by restoring the maximum generating power of wind turbines under no scheduling restrictions to recover the actual generating capacity of the wind farm. The correction process involves reversing the theoretical power based on the actual wind speed and power of each wind turbine, combined with the characteristic model of its respective turbine group, and then filling the corrected power value into the corresponding power curtailment data in the original data. A comparison of wind turbine power curtailment data before and after correction can be found in [reference needed]. Figure 3 and Figure 4 As shown.

[0101] For data on non-power-limited abnormal operation status (i.e., non-power-limited abnormal operation data), an appropriate interpolation correction strategy is adopted according to its abnormality type. In this embodiment, an interpolation correction method combining linear interpolation and nearest neighbor interpolation is adopted.

[0102] In the process of processing abnormal operation data without power restrictions, for outliers, if they are in the middle range, linear interpolation is used for correction; if they are in the start and end range, nearest neighbor interpolation is used for correction.

[0103] After correcting all power curtailment data and abnormal operation data for non-power curtailment using theoretical power, the corrected power sequence for each wind turbine is obtained. Then, by summing the corrected power sequences of all wind turbines in the wind farm, the corrected power sequence of the wind farm can be obtained.

[0104] S5. Input the corrected power sequence of the wind farm, the meteorological feature input set, and the wind farm's meteorological tower observation data into the power prediction model to obtain the theoretical power prediction sequence of the wind farm.

[0105] Prior to S5, the method further includes:

[0106] The corrected power sequence of the wind farm and the meteorological feature input set are normalized respectively.

[0107] To improve the training stability and convergence speed of the model, Min-Max normalization is used to map the obtained wind farm corrected power sequence and the selected meteorological feature input set to the [0,1] interval. This ensures that the numerical range of all input features is uniform, thereby improving the training efficiency and stability of the power prediction model.

[0108] In this invention, the power prediction model employs a combined CNN-LSTM-lightGBM prediction model; S5 specifically includes:

[0109] The model input parameters are respectively input into the CNN-LSTM layer and the lightGBM layer of the combined prediction model to obtain the first-stage power prediction results and the second-stage power prediction results; the model input parameters include the wind farm corrected power sequence, meteorological feature input set and wind farm anemometer tower observation data;

[0110] The power prediction results of the first stage and the power prediction results of the second stage are weighted and summed to obtain the theoretical power prediction sequence of the wind farm.

[0111] refer to Figure 5As shown, in the first-stage prediction model, a CNN is first introduced to extract features from the input time-series data (i.e., the model input parameters). Through local connectivity and parameter sharing mechanisms, the CNN can effectively uncover the local variation patterns of meteorological features and power sequences within the time window while reducing the size of the model parameters.

[0112] Let the input feature sequence within the time window be:

[0113] ;

[0114] in, Indicates the first Multidimensional input feature vectors at each time step This represents the length of the time window. The feature extraction process of the convolutional layer can be represented as:

[0115] ;

[0116] in, For convolution kernel parameters, For bias terms, This is a non-linear activation function. Through convolution operations, the model can capture the influence patterns of meteorological features such as wind speed and direction on power changes over short timescales.

[0117] To reduce feature dimensionality and suppress the risk of overfitting, a pooling operation is introduced after the convolutional layer to downsample local features, thereby enhancing the model's generalization ability.

[0118] Feature sequences output by convolutional layers The data is then further fed into an LSTM layer for time-series modeling. By introducing memory units and gating structures, LSTM effectively alleviates the gradient vanishing or gradient exploding problems that exist in traditional recurrent neural networks in long sequence modeling, and can characterize the long-term dependence of wind power evolution over time.

[0119] The state update process of an LSTM cell is as follows:

[0120] ;

[0121] in, , , Represent The outputs of the forget gate, input gate, and output gate are constantly updated; This represents the Sigmoid activation function, used to map the gated output to the 0~1 range; express The state of the memory cells of the LSTM at any given time is used to store historical timing information; express The hidden layer output of the LSTM at each time step; , , and These represent the weight parameter matrices for the corresponding gating structure and candidate memory units, respectively. , , and These represent the corresponding bias parameters; This represents element-wise multiplication. This represents the hyperbolic tangent activation function. Through a gating mechanism, the network can dynamically update and filter information to capture key temporal patterns.

[0122] To further enhance the model's ability to identify information at key time steps, an attention mechanism is introduced after the LSTM layer output to weight the hidden states at each time step. The calculation process is as follows:

[0123] ;

[0124] in, For the first Attention weights at each time step For trainable attention weight vectors, This represents the weighted temporal features. Through an attention-weighted mechanism, the model can autonomously focus on moments that contribute more to the prediction, suppressing the influence of redundant information. Finally, the output from the fully connected layer yields the first-stage power prediction result. .

[0125] Simultaneously, the same model input parameters are fed into the lightGBM model based on gradient boosting decision trees for training. By constructing multiple regression trees and performing additive ensemble, lightGBM effectively characterizes the complex nonlinear relationship between meteorological features and wind power, as well as the higher-order interactions between features. Its prediction results are expressed as follows:

[0126] ;

[0127] in, Indicates the first A tree of return, This represents the number of regression trees. Compared to deep learning models, LightGBM is insensitive to feature scale and exhibits better stability and generalization ability under conditions of small to medium sample sizes. The second-stage power prediction results... This can be used as the result of the first stage power prediction. An effective supplement.

[0128] When performing weighted fusion, the weighting coefficients of the first-stage power prediction results and the second-stage power prediction results are determined by the inverse error weighting method.

[0129] Based on the prediction error of the two-stage model on historical samples, the prediction results are fused using an inverse error weighting method. Let the first-stage and second-stage models be at the same time step... The prediction errors are respectively and The corresponding expression is as follows:

[0130] ;

[0131] The corresponding weighting coefficients are respectively and The expression is as follows:

[0132] ;

[0133] The final fusion prediction result is as follows:

[0134] ;

[0135] This yields an ultra-short-term theoretical power prediction sequence for the wind farm over the next 4 hours, calculated in 15-minute increments. .

[0136] According to the latest industry standards in wind power prediction, the root mean square error (RMSE), mean absolute error (MAE), and correlation coefficient r are selected as evaluation indicators for model prediction performance. For RMSE and MAE, the smaller the value, the lower the prediction error of the model; for r, the closer the value is to 1, the better the model's fit. The calculation formulas for the three evaluation indicators are as follows:

[0137] ;

[0138] In the formula, for Predicted power at time, for Actual power at time, for The operating capacity of the wind farm at all times. For the total number of samples, This represents the average actual power during the error statistics period. This represents the average predicted power over the error statistics period.

[0139] As shown in Tables 1 and 2 below, by identifying and correcting anomalies such as power curtailment in historical wind power data, and further incorporating multi-source meteorological information for fusion modeling, the ultra-short-term power prediction method proposed in this invention exhibits significant advantages in both prediction accuracy and stability. Specifically, as shown in Tables 1 and 2... Figure 6 As shown, under the same network structure, the CNN-LSTM model using corrected power data exhibits lower RMSE and MAE scores compared to the model using uncorrected power data, while improving the correlation coefficient r. This indicates that power data correction effectively mitigates the adverse effects of abnormal fluctuations and measurement errors on model training and prediction results. Furthermore, by introducing LightGBM for feature modeling and fusion, the CNN-LSTM-lightGBM model further optimizes the RMSE, MAE, and correlation coefficient evaluation metrics. The model based on corrected power data achieves the lowest prediction error and the highest correlation, verifying the ability of the proposed method to characterize power change trends under complex wind conditions. Further, as shown in Table 2 and... Figure 7 As shown, under different input information configurations, prediction methods with single historical power or limited meteorological features have relatively high error levels. With the increase in the dimensions of meteorological features in Numerical Weather Prediction (NWP), the model's prediction accuracy gradually improves, indicating that multiple meteorological elements can provide more sufficient prior information for theoretical power changes. When multi-source data fusion is used, the prediction model's RMSE and MAE both reach optimal levels, and the correlation coefficient is significantly improved, indicating that multi-source information fusion can effectively enhance the model's comprehensive perception of wind farm operating status and wind resource changes.

[0140] Table 1 Evaluation Indicators for Theoretical Power Prediction of Different Models

[0141] Table 2 Evaluation Indicators for Theoretical Power Prediction Using Different Methods

[0142] Another embodiment of the present invention provides a wind farm theoretical power prediction system based on power curtailment identification, comprising:

[0143] The feature filtering module constructs a meteorological feature input set based on the historical operation data and numerical weather prediction data of the wind farm;

[0144] The anomaly identification module identifies anomalies in historical operational data to obtain power rationing datasets and non-power rationing anomaly operational datasets.

[0145] The abnormal data processing module is used to divide the wind turbines of the wind farm into a first correlation group and a second correlation group according to the correlation of the wind turbine wind speed, and to construct a first theoretical power model corresponding to the first correlation group and a second theoretical power model corresponding to the second correlation group, respectively. For the power curtailment data belonging to the first correlation group in the power curtailment data, the first theoretical power model is used to correct the theoretical power, and for the power curtailment data belonging to the second correlation group in the power curtailment data, the second theoretical power model is used to correct the theoretical power. For the non-power curtailment abnormal operation data, the interpolation method is used to correct the theoretical power, so as to obtain the corrected power sequence of the wind farm.

[0146] The power prediction module is used to input the corrected power sequence of the wind farm, the meteorological feature input set, and the wind farm's meteorological tower observation data into the power prediction model to obtain the theoretical power prediction sequence of the wind farm.

[0147] Another embodiment of the present invention provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the wind farm theoretical power prediction method based on power curtailment identification as described above.

[0148] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0149] 1. A multi-source collaborative modeling method is proposed, which integrates historical operation data of wind farms, observation data from meteorological towers, and numerical weather prediction data. By complementing multi-source information and aligning time, the overall characterization capability of wind farm operation status and meteorological evolution characteristics is improved, providing a reliable data foundation for ultra-short-term power prediction.

[0150] 2. A two-dimensional meteorological feature screening method based on Pearson correlation coefficient and maximum mutual information coefficient is proposed. This method evaluates the correlation between meteorological variables and wind power from both linear and nonlinear dimensions, reduces redundant feature interference, and improves the effectiveness and prediction stability of model input features.

[0151] 3. A method combining dynamic power threshold detection based on the standard power curve of wind turbines with wind turbine grouping modeling is proposed to achieve accurate identification of the power curtailment status of wind turbines. By combining multicollinearity test, wind turbines are divided into strongly correlated groups and weakly correlated groups, and differentiated theoretical power models are constructed to inversely correct the power curtailment data, so as to avoid the power curtailment data from interfering with the model training and prediction results.

[0152] 4. A combined prediction method integrating CNN, LSTM, attention mechanism and lightGBM model is proposed. The reciprocal weighted fusion strategy based on historical prediction error is adopted to adaptively integrate the prediction results of different sub-models, fully explore the temporal characteristics of wind power sequence, and realize high-precision prediction of wind farm ultra-short-term power every 15 minutes in the next 4 hours, which meets the practical application needs of grid dispatch and wind farm operation management.

[0153] The above descriptions are merely a few embodiments of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any modifications or alterations made by those skilled in the art without departing from the scope of the technical solution of the present invention using the disclosed technical content are equivalent to equivalent implementation cases and fall within the scope of the technical solution.

Claims

1. A wind farm theoretical power prediction method based on power cut identification, characterized in that, The method includes: S1. Construct a meteorological feature input set based on historical operation data and numerical weather prediction data of the wind farm; S2. Identify abnormal data in the historical operation data to obtain a power rationing dataset and a non-power rationing abnormal operation dataset. S3. Divide the wind turbines of the wind farm into a first correlation group and a second correlation group according to the correlation of wind turbine wind speed, and construct a first theoretical power model corresponding to the first correlation group and a second theoretical power model corresponding to the second correlation group respectively. S4. The theoretical power of the power curtailment data belonging to the first relevant group in the power curtailment dataset is corrected using the first theoretical power model; the theoretical power of the power curtailment data belonging to the second relevant group in the power curtailment dataset is corrected using the second theoretical power model; and the theoretical power of the non-power curtailment abnormal operation dataset is corrected using interpolation, to obtain the wind farm corrected power sequence. S5. Input the corrected power sequence of the wind farm, the meteorological feature input set, and the wind farm's wind tower observation data into the power prediction model to obtain the theoretical power prediction sequence of the wind farm.

2. The method of claim 1, wherein, S1 specifically includes: Correlation analysis was performed on various meteorological variables in numerical weather prediction data and wind power in historical operation data of wind farms. Meteorological variables with a correlation degree that met the preset requirements were used to form a meteorological feature input set.

3. The method of claim 1, wherein, S2 specifically includes: A dynamic power threshold is set based on the standard power curve of the wind turbines in the wind farm. Based on the dynamic power threshold and the characteristic parameters of the wind turbine, abnormal data are identified in all data points of the historical operation data to obtain a power curtailment dataset and a non-power curtailment abnormal operation dataset.

4. The method of claim 1, wherein, The step S3, which divides the wind turbines of the wind farm into a first correlation group and a second correlation group based on the correlation of wind turbine wind speed, specifically includes: A multicollinearity test was performed on the historical wind speed sequences of each wind turbine in the historical operating data. Based on the test results, the wind turbines of the wind farm were divided into a first correlation group and a second correlation group.

5. The method of claim 1, wherein, The construction of the first theoretical power model corresponding to the first related group in S3 specifically includes: Determine the first input parameters for the first relevant group; the first input parameters include the median wind speed, the number of operating units, and the median power. A first theoretical power model is constructed based on the first input parameters.

6. The method according to claim 1, characterized in that, The construction of the second theoretical power model corresponding to the second related group in S3 specifically includes: Determine the second input parameters for the second relevant group; the second input parameters include the wind speed sequence of each fan; A second theoretical power model is constructed based on the second input parameters.

7. The method of claim 1, wherein, The power prediction model is a CNN-LSTM-lightGBM combined prediction model; S5 specifically includes: The model input parameters are respectively input into the CNN-LSTM layer and the lightGBM layer of the combined prediction model to obtain the first-stage power prediction results and the second-stage power prediction results; the model input parameters include the wind farm corrected power sequence, the meteorological feature input set, and the wind farm's wind tower observation data; The power prediction results of the first stage and the power prediction results of the second stage are weighted and summed to obtain the theoretical power prediction sequence of the wind farm.

8. The method of claim 7, wherein, The weighting coefficients of the first-stage power prediction results and the second-stage power prediction results are determined by the inverse error weighting method.

9. The method of claim 1, wherein, Prior to S5, the method further includes: The corrected power sequence of the wind farm and the meteorological feature input set are normalized respectively.

10. A terminal device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the wind farm theoretical power prediction method based on power rationing identification as described in any one of claims 1 to 9.

Images