Single-fan ultra-short-term output prediction method based on mRMR and TabNet

Abstract

The invention belongs to the technical field of wind power prediction, and particularly relates to a single-fan ultra-short-term output prediction method based on mRMR and TabNet. The method comprises the steps of firstly obtaining historical operation data of a single-machine wind turbine generator; carrying out abnormal value identification and elimination on the original data through a random forest algorithm, and filling missing data by adopting linear interpolation; derived features are constructed on the basis; inputting all the candidate features into a maximum correlation and minimum redundancy feature screening module, and obtaining an optimal feature subset for ultra-short-term prediction according to the correlation between the features and output and the redundancy among the features; constructing a TabNet deep neural network prediction model by using the feature subset, and performing training by inputting single-machine operation features and taking historical ultra-short-term output as a supervision signal to obtain a single-fan ultra-short-term output prediction model; according to the method, the output prediction precision of the single wind turbine generator in a high-volatility scene can be effectively improved, and the interpretability and the feature utilization efficiency of the prediction model are improved.

Description

Technical Field

[0001] This invention relates to the field of wind power generation technology, specifically to a method for predicting the ultra-short-term output of a single wind turbine based on mRMR and TabNet. Background Technology

[0002] With the continuous growth of wind power installed capacity, the stability, dispatchability, and grid connection characteristics of wind farm operation are receiving increasing attention. The output of a single wind turbine is affected by multiple factors such as wind speed, wind direction, air density, temperature, and humidity, exhibiting significant randomness and volatility. Especially on ultra-short time scales, the power changes of wind turbines are highly sudden, and traditional statistical methods, time series models, and some deep learning models are unable to accurately capture the rapid changes in short time, making it difficult for their prediction accuracy to meet the actual dispatch and control requirements.

[0003] Therefore, this invention proposes a single wind turbine ultra-short-term output prediction method based on mRMR and TabNet, which is used to achieve high-precision output prediction of a single wind turbine unit in an ultra-short time scale. Summary of the Invention

[0004] This invention aims to overcome the shortcomings of existing single-wind-turbine ultra-short timescale prediction methods, such as insufficient accuracy, low feature utilization efficiency, and lack of model interpretability. It proposes a single-wind-turbine ultra-short-term output prediction method based on mRMR and TabNet to improve the prediction accuracy of ultra-short timescales and enhance the model's generalization ability and interpretability.

[0005] The technical solution of the present invention is as follows:

[0006] A method for predicting the ultra-short-term power output of a single wind turbine based on mRMR and TabNet includes the following steps:

[0007] Step 1: Obtain historical data related to the output of a single wind turbine unit, including wind speed, wind direction, temperature, humidity, unit capacity, and efficiency of the transmission and power generation systems, and use this data as raw feature data.

[0008] Step 2: Clean the acquired raw feature data using the Isolation Forest algorithm to remove outliers;

[0009] Step 3: Use linear interpolation to fill in the missing data to obtain the preprocessed dataset;

[0010] Step 4: Construct derived features using the preprocessed dataset, including real-time / predicted wind speed, wind speed change rate, historical wind speed sequence, real-time / predicted wind direction, wind direction change rate, temperature, air pressure, humidity, turbulence intensity, and wind shear index.

[0011] Step 5: Input all candidate features into the maximum relevance and minimum redundancy (mRMR) feature screening module. By calculating the correlation between candidate features and the target output and the redundancy between candidate features, an optimal feature subset for ultra-short-term output prediction is obtained.

[0012] Step 6: Construct a TabNet deep neural network model with the obtained optimal feature subset. The model includes an input batch normalization layer, a feature transformation layer, an attention transformation layer, and sequential decision steps. Use historical ultra-short-term output as the supervision label to train the TabNet model, and obtain a single-wind-turbine ultra-short-term output prediction model.

[0013] Step 7: Input the real-time collected data of a single wind turbine into the constructed ultra-short-term output prediction model, and output the output prediction time series value of the single wind turbine on the ultra-short-term scale.

[0014] Further, the specific steps of Step 2 are as follows:

[0015] Step 2.1: Perform normalization processing on the original feature data according to the feature dimension to obtain a normalized data sample set {x1, x2, …, x , k , , , , ,

[0021] ,

[0020] ,

[0019] , (q) ,

[0018] , , ,

[0017] , }, where x i ∈R d ;

[0016] Step 2.2: Randomly select a sub-sample set of size ψ from the normalized data sample set without replacement, and construct an isolation tree iTree based on the sub-sample set. The construction process is as follows: Randomly select a feature dimension q on the sample subset corresponding to the current node, randomly select a splitting value p between the minimum value min(q) and the maximum value max(q) of the feature dimension q, and divide the samples into left and right child nodes according to x (q) <p and x (q) ≥p, and recursively execute until the node contains only 1 sample or reaches the preset maximum tree depth;

[0017] Step 2.3: Repeat the construction of t isolation trees to form an isolation forest model

[0018] Step 2.4: Calculate the path length h(x) of the sample to be detected x in the isolation forest. The path length on a single isolation tree is the number of edges from the root node to the leaf node, and the expected path length in the isolation forest is:

[0019]

[0020] In the formula, h k (x) is the path length of the sample x on the k-th isolation tree;

[0021] Step 2.5 Calculate the anomaly score s(x,n) for sample x, whereby the anomaly score is defined as:

[0022]

[0023] In the formula, n is the number of samples involved in the modeling, and c(n) is the normalization constant, defined as:

[0024]

[0025] Step 2.6 Set the anomaly detection threshold s thr ∈(0,1), when the anomaly score s(x,n)≥s of sample x thr When x is an outlier, it is identified as an outlier and removed. When s(x,n) thr When the sample x is considered a normal value, it is retained.

[0026] Furthermore, step 3 specifically includes:

[0027] 3.1 The data after removing outliers is organized into a time series according to the sampling time order:

[0028] {(t1,v1),(t2,v2),…,(t N ,v N )}

[0029] In the formula, t i For the i-th sampling time, t1 <t2<…<t N v i For this feature at time t i The observed value; if the observation at time i is missing, it is denoted as v. i Missing;

[0030] Step 3.2 Process each feature dimension separately, scanning {v i For each missing position j in the sequence, find the index p of the nearest non-missing sample point to its left and the index q of the nearest non-missing sample point to its right, such that...

[0031] p < j < q

[0032] In the formula, the corresponding time point is t. p With t q The observed values ​​are v p With v q ;

[0033] Step 3.3 For those satisfying t p <t j <t q Missing points (t) j ,v j Linear interpolation is performed using the following formula:​

[0034]

[0035] In the formula, To find the missing observation v j Linear interpolation estimation, For the interval [t] p ,t q The slope on ].

[0036] Furthermore, step 5 specifically includes:

[0037] Step 5.1 Organize the samples after the derived features are constructed into a data matrix.

[0038] X = [x1, x2, ..., x n ] T ∈R n×d

[0039] In the formula, n is the number of samples, d is the dimension of the candidate features, and each row This represents the feature vector of the i-th sample;

[0040] The target vector is composed of the actual single-unit ultra-short timescale output at the corresponding moment:

[0041] y = (y1, y2, ..., y n ) T

[0042] Step 5.2 For the k-th candidate feature x (k) The mutual information between it and the target output y is defined as:

[0043] D k =I(x) (k) ;y)

[0044] in:

[0045]

[0046] In the formula, p(a,b) represents the characteristic x (k) The joint probability of the output y taking the value a and the output y taking the value b, p(a) and p(b) are the corresponding marginal probabilities, which are estimated using the sample frequency;

[0047] Step 5.3 For any two candidate features x (k) With x (j) Calculate their mutual information:

[0048]

[0049] In the formula, p(a,c) represents the characteristic x. (k) The value is a and the feature x(j) The joint probability with value c, and p(a) and p(c) are its marginal probabilities;

[0050] Step 5.4 Select the feature with the largest mutual information with the target output from all candidate features:

[0051]

[0052] Initialize the selected feature set:

[0053] S={k *}

[0054] Remaining set of candidate features:

[0055] Ω={1,2,…,d}\S

[0056] Step 5.5 For each unselected feature k∈Ω, calculate the comprehensive score:

[0057]

[0058] In the formula, D k The mutual information between the k-th feature and the target is given by |S|, where |S| represents the number of selected features. The feature with the highest overall score is added to set S.

[0059]

[0060] Step 5.6 Repeat the above iterative process until the number of selected features reaches the preset dimension upper limit m, and then output the selected feature set obtained at the end of the iteration:

[0061] S={k1,k2,…,k m}

[0062] The corresponding feature set serves as the optimal feature subset for predicting the output of a single wind turbine over an ultra-short timescale.

[0063] Furthermore, step 6 specifically includes:

[0064] Step 6.1: Construct the input matrix from the obtained optimal feature subset:

[0065] Z = [z1, z2, ..., z n ] T

[0066] Perform batch normalization on each feature dimension to make its mean 0 and variance 1:

[0067]

[0068] In the formula: z (k) : The k-th feature, μ k σk represents the mean and standard deviation of the features, and ∈ represents a stable term to avoid a denominator of 0;

[0069] Step 6.2 The feature transformation module consists of several fully connected layers (FC) and nonlinear activation functions, and uses a residual structure:

[0070] H1 = ReLU(W1·BN(Z) + b1)

[0071] H2=ReLU(W2·H1+b2)+H1

[0072] In the formula, W1 is the weight matrix of the first-level linear transformation, b1 is the bias vector corresponding to the first-level linear transformation, W2 is the weight matrix of the second-level linear transformation, and b2 is the bias vector corresponding to the second-level linear transformation.

[0073] The feature transformation output is obtained:

[0074] FT(Z)=H2

[0075] Step 6.3 Utilize the features from the previous decision step to output P (t-1) Calculate the feature selection mask:

[0076] A (t) =ReLU(W a ·P (t-1) +b a )

[0077] In the formula, W a For the linear transformation weight matrix in the attention / feature selection branch, b a This is the bias vector corresponding to the linear transformation.

[0078] The attention distribution is normalized using the Sparsemax activation function to make it sparse:

[0079] M (t) =Sparsemax(A (t) )

[0080] In the formula, M (t) Consistent with the dimensions of the input features, each element represents the selection probability of the corresponding feature;

[0081] Step 6.4 In the t-th decision step, use a mask to perform a dimension-wise weighted selection of the input features:

[0082] Z (t) =M (t) ⊙Z

[0083] In the formula, ⊙ represents element-wise multiplication;

[0084] Then Z (t)Input feature transformation module:

[0085] D (t) =FT(Z) (t) )

[0086] In the formula, D (t) It is the output of the t-th decision step;

[0087] Step 6.5 The final model output of TabNet is accumulated from multiple decision steps:

[0088]

[0089] In the formula, T represents the number of decision steps, and g t These are the trainable step weight parameters;

[0090] Step 6.6 Obtain the predicted value by linearly mapping the cumulative output O:

[0091]

[0092] In the formula, Here, W0 represents the output prediction time series value on an ultra-short time scale, and b0 represents the weight parameter of the output layer.

[0093] Using historical actual output y as the supervision label, the network is trained using the mean squared error loss function:

[0094]

[0095] In the formula, n is the number of samples used to calculate the loss. Let y be the predicted output value for the i-th sample. i Let be the true label value of the i-th sample.

[0096] Iterative optimization is performed using an adaptive learning rate optimization algorithm, and overfitting is prevented by an early stopping strategy on the validation set.

[0097] Compared with the prior art, the beneficial effects of the present invention are:

[0098] 1) This invention introduces an anomaly sample identification and removal mechanism before modeling, which can detect and remove collected data; at the same time, combined with missing value interpolation to make the training samples more continuous and more consistent in time series, reducing the interference of anomalies and missing points on model parameter learning from the source, thereby improving the stability and robustness of ultra-short-term prediction.

[0099] 2) This invention uses mRMR for feature selection, which reduces the redundancy and collinearity between features while ensuring the correlation between features and the prediction target, and reduces the interference of invalid information input on model learning. Compared with the scheme of directly using all features, it can effectively reduce model complexity and overfitting risk, and improve generalization performance under cross-season and cross-wind condition data.

[0100] 3) This invention uses TabNet to build a prediction model, which adaptively filters and weights input features through a multi-decision-step feature selection mechanism, which helps to achieve higher prediction accuracy under complex nonlinear mapping relationships; at the same time, it can output feature selection weight / importance information related to decision-making, making the prediction process interpretable to a certain extent, which is convenient for tracing and analyzing key influencing factors and maintaining the model on the engineering site, and is more conducive to practical application than traditional black box models. Attached Figure Description

[0101] Figure 1 This is a flowchart illustrating the overall workflow of the present invention;

[0102] Figure 2 This is a flowchart of the data preprocessing process;

[0103] Figure 3 This is a flowchart of mRMR feature selection;

[0104] Figure 4 This is a diagram of the TabNet model structure. Detailed Implementation

[0105] The technical solution of the present invention will be further described below with reference to specific embodiments.

[0106] A method for predicting the ultra-short-term power output of a single wind turbine based on mRMR and TabNet includes the following steps:

[0107] Step 1: Obtain historical data related to the output of a single wind turbine unit, including wind speed, wind direction, temperature, humidity; unit capacity, transmission and power generation system efficiency.

[0108] Step 1: Using historical SCADA data from a single grid-connected wind turbine in a wind farm as an example data source, the turbine's rated capacity is 2MW. The sampling period is 1 minute, and the continuous data collection period is from January 1, 2024 to March 31, 2024. The original measurements include: wind speed, wind direction, ambient temperature, humidity, actual active power of the turbine, and operating parameters such as turbine capacity and transmission / generation system efficiency.

[0109] Step 2: The prediction task in this embodiment is: given the available feature inputs at time t, output the predicted sequence of wind turbine active power for the next 5 minutes.

[0110] Step 3: Data partitioning method: Partition the training set / validation set / test set in chronological order = 70% / 15% / 15%, ensuring that the test set time period is not involved in training to meet the time series generalization requirements for engineering online deployment.

[0111] Step 2: Use the random forest algorithm to clean the acquired data and remove outliers.

[0112] Step 1: Normalize the original feature data such as wind speed, wind direction, temperature, humidity, etc. according to the feature dimension to obtain a normalized data sample set {x1, x2, …, x n}, where x i ∈R d ;

[0113] Step 2: Randomly draw a subsample set of size 256 from the normalized data sample set without replacement, and construct an isolation tree iTree based on the subsample set. The construction process is as follows: Randomly select a feature dimension q on the sample subset corresponding to the current node, randomly select a splitting value p between the minimum value min(q) and the maximum value max(1) of this feature dimension, and divide the samples into left and right child nodes according to x (1) <p and x (1) ≥p, and recursively execute until the node contains only 1 sample;

[0114] Step 3: Repeat the construction of 100 isolation trees to form an isolation forest model

[0115] Step 4: Calculate the path length h(x) of the sample x to be detected in the isolation forest. The path length on a single isolation tree is the number of edges from the root node to the leaf node. The expected path length in the isolation forest is:

[0116]

[0117] In the formula, h k (x) is the path length of the sample x on the kth isolation tree;

[0118] Step 5: Calculate the anomaly score s(x, n) of the sample x. The anomaly score is defined as:

[0119]

[0120] In the formula, n is the number of samples participating in the modeling, and c(n) is a normalization constant, defined as:

[0121]

[0122]

[0123] Step 6: Set the anomaly detection threshold s thr =0.65, when the anomaly score s(x,n) ≥ s thr When s(x,n) is considered an outlier, it is removed. thr When the sample x is considered a normal value, it is retained.

[0124] Step 3: Use linear interpolation to fill in the data after the data has been removed, and obtain the preprocessed dataset.

[0125] Step 1: Organize the original features such as wind speed, wind direction, temperature, and humidity, after removing outliers, into a time series according to the sampling time sequence:

[0126] {(t1,v1),(t2,v2),…,(t N ,v N )}

[0127] In the formula, t i For the i-th sampling time, t1 <t2<…<t N v i For this feature at time t i The observed value; if the observation at time i is missing, it is denoted as v. i Missing;

[0128] Step 2: Process each feature dimension separately, scanning {v} i For each missing position j in the sequence, find the index p of the nearest non-missing sample point to its left and the index q of the nearest non-missing sample point to its right, such that...

[0129] p < j < q

[0130] In the formula, the corresponding time point is t. p With t q The observed values ​​are v p With v q ;

[0131] Step 3: For those satisfying t p <t j <t q Missing points (t) j ,v j Linear interpolation is performed using the following formula:

[0132]

[0133] In the formula, To find the missing observation v j Linear interpolation estimation, For the interval [t] p ,t​q The slope on ];

[0134] Step 4: Construct derived features using the preprocessed dataset, including real-time / predicted wind speed, wind speed change rate, historical wind speed sequence, real-time / predicted wind direction, wind direction change rate, temperature, air pressure, humidity, turbulence intensity, and wind shear index.

[0135] Step 5: Input all candidate features into the maximum correlation minimum redundancy (mRMR) feature filtering module. By calculating the correlation between candidate features and target output and the redundancy between candidate features, the optimal feature subset for output prediction on ultra-short time scale is obtained.

[0136] Step 1: Organize the samples after constructing derived features into a data matrix.

[0137] X = [x1, x2, ..., x n ] T ∈R n×d

[0138] In the formula, n is the number of samples, d is the dimension of the candidate features, and each row This represents the feature vector of the i-th sample;

[0139] The target vector is composed of the actual single-unit ultra-short timescale output at the corresponding moment:

[0140] y = (y1, y2, ..., y n ) T

[0141] Step 2: For the k-th candidate feature x (k) The mutual information between it and the target output y is defined as:

[0142] D k =I(x) (k) ;y)

[0143] in:

[0144]

[0145] In the formula, p(a,b) represents the characteristic x (k) The joint probability of the output y taking the value a and the output y taking the value b, p(a) and p(b) are the corresponding marginal probabilities, which are estimated using the sample frequency;

[0146] Step 3: For any two candidate features x (k) With x (j) Calculate their mutual information

[0147]

[0148] In the formula, p(a,c) represents the characteristic x. (k) The value is a and the feature x (j) The joint probability with value c, and p(a) and p(c) are its marginal probabilities;

[0149] Step 4: Select the feature with the highest mutual information with the target output from all candidate features:

[0150]

[0151] Initialize the selected feature set:

[0152] S={k *}

[0153] Remaining set of candidate features:

[0154] Ω={1,2,…,d}\S

[0155] Step 5: For each unselected feature k∈Ω, calculate the overall score:

[0156]

[0157] In the formula, D k The mutual information between the k-th feature and the target is given by |S|, where |S| represents the number of selected features. The feature with the highest overall score is added to set S.

[0158]

[0159] Step 6: Repeat the above iterative process until the number of selected features reaches the preset dimension upper limit m = 15; convert the selected feature set obtained at the end of the iteration.

[0160] S={k1,k2,…,k m}

[0161] The corresponding features are used as the optimal feature subset for predicting the output of a single wind turbine in an ultra-short timescale.

[0162] Step 6: Construct a TabNet deep neural network model using the obtained optimal feature subset. The model includes an input batch normalization layer, a feature transformation layer, an attention transformation layer, and a sequential decision step. Train the TabNet model using historical ultra-short time scale output as a supervision label to obtain a single wind turbine ultra-short time scale output prediction model.

[0163] Step 1: Construct the input matrix from the obtained optimal feature subset:

[0164] Z = [z1, z2, ..., z n ] T

[0165] Perform batch normalization on each feature dimension to make its mean 0 and variance 1:

[0166]

[0167] In the formula: z (k) : The k-th feature, μ k σ k Let be the mean and standard deviation of this feature, and ∈ be a stable term to avoid a denominator of 0.

[0168] Step 2: The feature transformation module consists of several fully connected layers (FC) and nonlinear activation functions, and uses a residual structure.

[0169] H1 = ReLU(W1·BN(Z) + b1)

[0170] H2=ReLU(W2·H1+b2)+H1

[0171] In the formula, W1 is the weight matrix of the first-level linear transformation, b1 is the bias vector corresponding to the first-level linear transformation, W2 is the weight matrix of the second-level linear transformation, and b2 is the bias vector corresponding to the second-level linear transformation.

[0172] The feature transformation output is obtained:

[0173] FT(Z)=H2

[0174] Step 3: Utilize the features from the previous decision step to output P. (t-1) Calculate the feature selection mask:

[0175] A (t) =ReLU(W a ·P (t-1) +b a )

[0176] In the formula, W a For the linear transformation weight matrix in the attention / feature selection branch, b a This is the bias vector corresponding to the linear transformation.

[0177] The attention distribution is normalized using the Sparsemax activation function to make it sparse:

[0178] M (t) =Sparsemax(A (t) )

[0179] In the formula, M (t) Consistent with the dimensions of the input features, each element represents the selection probability of the corresponding feature.

[0180] Step 4: In the t-th decision step, use a mask to perform a dimension-wise weighted selection of the input features:

[0181] Z (t) =M (t) ⊙Z

[0182] In the formula, ⊙ represents element-wise multiplication.

[0183] Features selected with lower values ​​by the mask will not participate in this decision step, and then Z will be... (t) Input feature transformation module:

[0184] D (t) =FT(Z) (t) )

[0185] In the formula, D (t) It is the output of the t-th decision step.

[0186] Step 5: The final output of TabNet is the accumulation of multiple decision steps:

[0187]

[0188] In the formula, T represents the number of decision steps, and g t These are the trainable step weight parameters.

[0189] Step 6: Obtain the predicted time series value by linearly mapping the cumulative output O:

[0190]

[0191] In the formula, For the output power prediction results on an ultra-short time scale, W0 is the weight parameter of the output layer, and b0 is the corresponding bias parameter of the output layer.

[0192] Using historical actual output y as the supervision label, the network is trained using the mean squared error loss function:

[0193]

[0194] In the formula, n is the number of samples used to calculate the loss. Let y be the predicted output value for the i-th sample. i Let be the true label value of the i-th sample.

[0195] The number of decision steps T = 5; the decision attention dimension = 16; the number of feature transformer layers is 3; the sparse selection coefficient is 1.5; the adaptive learning rate optimization algorithm (Adam) is selected, the batch size is 256, the learning rate is 0.001, the maximum number of training rounds is 200, the early stopping strategy patience = 20 is used for iterative optimization, and the validation set early stopping strategy is used to prevent overfitting.

[0196] Step 7: Input the real-time collected data of a single wind turbine into the constructed ultra-short time scale power output prediction model, and output the predicted time series value of the power output of a single wind turbine in the ultra-short time scale.

[0197] Example:

[0198] Step 1: Taking a certain time t in the test set as an example, the model input includes: current wind speed v(t) = 9.8 m / s, Δv(t) = 0.4 m / s; current wind direction θ(t) = 215°, Δθ(t) = -6°; 10-minute turbulence intensity TI = 0.12;

[0199] Wind speed and power sequence segments from the past 60 minutes (as historical features); 15-dimensional feature vector x obtained by mRMR filtering. t .

[0200] The model output is a 5-minute power prediction sequence (in kW):

[0201]

[0202] Corresponding measured power sequence (unit: kW):

[0203] y t =[1005, 1052, 1090, 1132, 1189]

[0204] It can be seen that the predicted sequence is consistent with the measured value in terms of trend and amplitude, and can track minute-level climbing / falling changes.

[0205] Step Two:

[0206] To verify the effectiveness of the present invention, the following comparison methods were selected as the baseline for "prior art / common methods" on the same training and test sets:

[0207] Baseline 1: LSTM sequence model; Baseline 2: XGBoost regression model.

[0208] The evaluation indicators used are MAE, RMSE, and MAPE.

[0209] LSTM: MAE=75.6kW, RMSE=109.2kW, MAPE=7.6%;

[0210] XGBoost: MAE=73.9kW, RMSE=106.8kW, MAPE=7.4%;

[0211] This invention has the following specifications: MAE = 62.7kW, RMSE = 92.4kW, MAPE = 6.1%.

[0212] The above comparison shows that, compared with commonly used machine learning and sequence models, the RMSE of this invention is reduced by 10%-20% on the test set, and it is more accurate in tracking short-term changes in minute-level climbing and fluctuation scenarios, demonstrating the outstanding progress brought about by the combination of "data quality control + feature redundancy removal + deep model with sparse feature selection mechanism".

[0213] This invention provides a method for predicting the ultra-short-term output of a single wind turbine that is structurally clear, highly interpretable, computationally inexpensive, and has superior predictive performance. It is applicable to various scenarios such as wind farm scheduling, electrolytic hydrogen production control, and energy storage system management, and has good engineering application value.

Claims

1. A method for predicting the ultra-short-term output of a single wind turbine based on mRMR and TabNet, characterized in that, Includes the following steps: Step 1: Obtain historical data related to the output of a single wind turbine unit, including wind speed, wind direction, temperature, humidity, unit capacity, and efficiency of the transmission and power generation systems, and use this data as raw feature data. Step 2: Clean the acquired raw feature data using the Isolation Forest algorithm to remove outliers; Step 3: Use linear interpolation to fill in the missing data to obtain the preprocessed dataset; Step 4: Construct derived features using the preprocessed dataset, including real-time / predicted wind speed, wind speed change rate, historical wind speed sequence, real-time / predicted wind direction, wind direction change rate, temperature, air pressure, humidity, turbulence intensity, and wind shear index. Step 5: Input all candidate features into the maximum correlation minimum redundancy mRMR feature filtering module. By calculating the correlation between candidate features and target output and the redundancy between candidate features, the optimal feature subset for output prediction in ultra-short time scale is obtained. Step 6: Construct a TabNet deep neural network model using the obtained optimal feature subset. The model includes an input batch normalization layer, a feature transformation layer, an attention transformation layer, and a sequential decision step. Train the TabNet model using historical ultra-short time scale output as a supervision label to obtain a single wind turbine ultra-short time scale output prediction model. Step 7: Input the real-time collected data of a single wind turbine into the constructed ultra-short time scale power output prediction model, and output the predicted time series value of the power output of a single wind turbine in the ultra-short time scale.

2. The method for predicting the ultra-short-term output of a single wind turbine based on mRMR and TabNet according to claim 1, characterized in that, Step 2 specifically includes: Step 2.1 Normalization is performed on the original feature data according to the feature dimension, and a normalized data sample set {x1, x2, …, x n} is obtained, where x i ∈R d ; Step 2.2 randomly sample a subset of size ψ from the normalized dataset, and build an isolated tree iTree based on the subset, the building process is as follows: randomly select a feature dimension q from the sample subset corresponding to the current node, randomly select a split value p between the minimum value min(q) and the maximum value max(q) of the feature dimension q, and divide the samples into left and right child nodes according to x (q) <p and x (q) ≥p, recursively execute until the node contains only 1 sample or the preset maximum tree depth is reached; Step 2.3 Repeat the construction of t isolated trees to form an isolated forest model. Step 2.4 Calculate the path length h(x) of the sample x to be tested in the isolated forest, where the path length on a single isolated tree is the number of edges from the root node to a leaf node, and the expected path length in the isolated forest is: In the formula, h k (x) is the path length of sample x on the kth isolated tree; Step 2.5 Calculate the anomaly score s(x,n) for sample x, whereby the anomaly score is defined as: In the formula, n is the number of samples involved in the modeling, and c(n) is the normalization constant, defined as: Step 2.6 Set the anomaly detection threshold s thr ∈(0,1), when the anomaly score s(x,n)≥s of sample x thr When x is an outlier, it is identified as an outlier and removed. When s(x,n) thr When the sample x is considered a normal value, it is retained.​ 3. The method for predicting the ultra-short-term output of a single wind turbine based on mRMR and TabNet according to claim 1, characterized in that, Step 3 specifically includes: 3.1 The data after removing outliers is organized into a time series according to the sampling time order: {(t1,v1),(t2,v2),…,(t N ,v N )} In the formula, t i For the i-th sampling time, t1 <t2<…<t N v i For this feature at time t i The observed value; if the observation at time i is missing, it is denoted as v. i Missing; Step 3.2 Process each feature dimension separately, scanning {v i For each missing position j in the sequence, find the index p of the nearest non-missing sample point to its left and the index q of the nearest non-missing sample point to its right, such that... p < j < q In the formula, the corresponding time point is t. p With t p The observed values ​​are v p With v q ; Step 3.3 For those satisfying t p <t j <t q Missing points (t) j ,v j Linear interpolation is performed using the following formula: In the formula, To find the missing observation v j Linear interpolation estimation, For the interval [t] p ,t q The slope on ].

4. The method for predicting the ultra-short-term output of a single wind turbine based on mRMR and TabNet according to claim 1, characterized in that, Step 5 specifically includes: Step 5.1 Organize the samples after the derived features are constructed into a data matrix. X=[x1,x2,...,x n ] T ∈R n×d In the formula, n is the number of samples, d is the dimension of the candidate features, and each row This represents the feature vector of the i-th sample; The target vector is composed of the actual single-unit ultra-short timescale output at the corresponding moment: y=(y1,y2,…,y n ) T Step 5.2 For the k-th candidate feature x (k) The mutual information between it and the target output y is defined as: D k =I(x (k) ;y) in: In the formula, p(a,b) represents the characteristic x (k) The joint probability of the output y taking the value a and the output y taking the value b, p(a) and p(b) are the corresponding marginal probabilities, which are estimated using the sample frequency; Step 5.3 For any two candidate features x (k) With x (j) Calculate their mutual information: In the formula, p(a,c) represents the characteristic x (k) The value is a and the feature x (j) The joint probability with value c, and p(a) and p(c) are its marginal probabilities; Step 5.4 Select the feature with the largest mutual information with the target output from all candidate features: Initialize the selected feature set: S={k * } Remaining set of candidate features: Ω={1,2,…,d}\S Step 5.5 For each unselected feature k∈Ω, calculate the comprehensive score: In the formula, D k The mutual information between the k-th feature and the target is given by |S|, where |S| represents the number of selected features. The feature with the highest overall score is added to set S. Step 5.6 Repeat the above iterative process until the number of selected features reaches the preset dimension upper limit m, and then output the selected feature set obtained at the end of the iteration: S={k1,k2,…,k m } The corresponding feature set serves as the optimal feature subset for predicting the output of a single wind turbine over an ultra-short timescale.

5. The method for predicting the ultra-short-term output of a single wind turbine based on mRMR and TabNet according to claim 1, characterized in that, Step 6 specifically includes: Step 6.1: Construct the input matrix from the obtained optimal feature subset: Z=[z1,z2,...,z n ] T Perform batch normalization on each feature dimension to make its mean 0 and variance 1: In the formula: z (k) : The k-th feature, μ k σ k represents the mean and standard deviation of the features, and ∈ represents a stable term to avoid a denominator of 0; Step 6.2 The feature transformation module consists of several fully connected layers (FC) and nonlinear activation functions, and uses a residual structure: H1 = ReLU(W1·BN(Z) + b1) H2=ReLU(W2·H1+b2)+H1 In the formula, W1 is the weight matrix of the first linear transformation, b1 is the bias vector corresponding to the first linear transformation, W2 is the weight matrix of the second linear transformation, and b2 is the bias vector corresponding to the second linear transformation. The feature transformation output is obtained: FT(Z)=H2 Step 6.3 Utilize the features from the previous decision step to output P (t-1) Calculate the feature selection mask: A (t) =ReLU(W a ·P (t-1) +b a ) In the formula, W a For the linear transformation weight matrix in the attention / feature selection branch, b a The bias vector corresponding to the linear transformation; The attention distribution is normalized using the Sparsemax activation function to make it sparse: M (t) =Sparsemax(A (t) ) In the formula, M (t) Consistent with the dimensions of the input features, each element represents the selection probability of the corresponding feature; Step 6.4 In the t-th decision step, use a mask to perform a dimension-wise weighted selection of the input features: From (t) =M (t) ⊙Z In the formula, ⊙ represents element-wise multiplication; Then Z (t) Input feature transformation module: D (t) =FT(Z (t) ) In the formula, D (t) It is the output of the t-th decision step; Step 6.5 The final model output of TabNet is accumulated from multiple decision steps: In the formula, T represents the number of decision steps, and g t These are the trainable step weight parameters; Step 6.6 Obtain the predicted value by linearly mapping the cumulative output O: In the formula, For the output prediction time series value on an ultra-short time scale, W0 is the weight parameter of the output layer, and b0 is the corresponding bias parameter of the output layer; Using historical actual output y as the supervision label, the network is trained using the mean squared error loss function: In the formula, n is the number of samples used to calculate the loss. Let y be the predicted output value for the i-th sample. i This represents the true label value of the i-th sample. Iterative optimization is performed using an adaptive learning rate optimization algorithm, and overfitting is prevented by an early stopping strategy on the validation set.

Images