A data-driven LSTM-SA neural network method for ultra-short-term power prediction of wind turbine operation and maintenance.

Abstract

This invention discloses a wind turbine operation and maintenance data-driven LSTM-SA neural network ultra-short-term power prediction method, comprising: obtaining and preprocessing operation and maintenance data of the target wind turbine; using the lasso algorithm and Pearson correlation coefficient method for feature selection, and then training an FNN model to obtain the functional relationship between wind turbine power generation and the selected data; based on the FNN model, using the chain rule to obtain the rate of change of power generation; using the rate of change of power generation and historical power values ​​together as a training set to train an LSTM network, and using an attention mechanism to optimize the LSTM network model to obtain an LSTM-SA neural network model; and realizing the prediction of the output power of the target wind turbine at the next moment. This method has advantages such as higher prediction accuracy, faster convergence speed, and better stability, and can solve the problem of ultra-short-term power prediction with limited wind power training data but high accuracy requirements.

Description

Technical Field

[0001] This invention relates to the field of wind turbine power prediction technology, and in particular to a wind turbine operation and maintenance data-driven LSTM-SA neural network ultra-short-term power prediction method. Background Technology

[0002] Wind energy is an important renewable energy source that helps my country bridge its energy gap and achieve its dual-carbon goals on schedule. In recent years, my country's wind power generation scale has continued to expand, ranking first in the world. However, influenced by factors such as wind resources, wind power generation is characterized by volatility, intermittency, and disorder, which can adversely affect grid stability during grid connection. To reduce wind power fluctuations and improve wind energy utilization, accurate wind power forecasting and appropriate control strategies are needed.

[0003] Methods for predicting wind power output include physical methods, statistical methods, and artificial intelligence methods. Physical methods convert numerical weather prediction results into wind resource information under actual environmental conditions based on the topography of wind farms, and then calculate the predicted power output based on the performance parameters of wind turbines. Statistical methods establish statistical relationships between meteorological elements in numerical weather prediction and actual power output by studying historical data of wind farms, thereby predicting wind power output. Both of these methods rely on accurate weather forecasts and accurate statistical relationships, resulting in relatively low precision.

[0004] Artificial intelligence-based power prediction methods do not require the establishment of physical models, giving them a significant advantage in fitting high-dimensional nonlinear sample spaces. However, most AI prediction methods are directly driven by data, exhibiting black-box characteristics and poor physical interpretability. The literature "Research on Wind Power Prediction Model Based on Improved LSTM" proposes optimizing the neural network model to improve prediction accuracy. The literature "Wind Power Combination Prediction Model Based on Temporal Decomposition and Machine Learning" proposes using mode decomposition for data augmentation, improving model prediction accuracy by enhancing data quality. The literature "Short-Term Wind Power Prediction Based on Convolutional Neural Network-Long Short-Term Memory Network with Attention Mechanism" uses CNN to augment data and employs an attention mechanism to optimize model parameter weights, achieving higher prediction accuracy.

[0005] However, these methods only use a single data variable to build a model for power prediction, without considering the relationship between other variables and power during actual wind turbine operation. This results in drawbacks such as low prediction accuracy and limited practicality. The paper "Ultra short-term probability prediction of wind power based on LSTM network and condition normal distribution" uses historical wind speed and power data for power prediction, establishing a normal distribution model for prediction errors. The paper "Wind Turbine Data Analysis and LSTM-Based Prediction in SCADA System" predicts four power-related characteristics based on the relationship between wind and energy: active power, wind speed, wind direction, and theoretical power. However, it does not address how to simultaneously predict wind turbine power using all four quantities.

[0006] The literature [BANSAL JC, FARSWAN P. Wind farm layout using biogeography based optimization[J]. Renewable Energy] significantly improves the model's prediction accuracy by adding meteorological factors and the wind turbine's own status to the data. However, the addition of meteorological data leads to a larger sample space, increased computational complexity, and more bloated model parameters, resulting in decreased efficiency and poorer performance of LSTM training.

[0007] Therefore, how to overcome the shortcomings of the aforementioned literature on wind turbine power prediction and provide a faster and more accurate prediction method has become an urgent problem for practitioners in the field. Summary of the Invention

[0008] The purpose of this invention is to provide a wind turbine operation and maintenance data-driven LSTM-SA neural network ultra-short-term power prediction method, which can solve the problem of ultra-short-term power prediction with limited wind power training data and high accuracy requirements.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] This invention provides a wind turbine operation and maintenance data-driven LSTM-SA neural network ultra-short-term power prediction method, comprising the following steps:

[0011] Obtain the operation and maintenance data of the target wind turbine, and preprocess the operation and maintenance data; the operation and maintenance data includes: wind speed, blade deflection angle, yaw angle, humidity and temperature;

[0012] The preprocessed wind turbine operation and maintenance data were used to perform feature filtering using the Lasso algorithm and Pearson correlation coefficient method to obtain the filtered data that has a significant impact on wind power changes.

[0013] Based on the screened data, an FNN model is trained to obtain the functional relationship between the wind turbine's power generation and the screened data.

[0014] Based on the FNN model obtained through training, the rate of change of power generation is obtained using the chain rule.

[0015] The rate of change of the power generation and the historical power values ​​are used together as the training set to train the LSTM network. The attention mechanism is used to optimize the LSTM network model to obtain the LSTM-SA neural network model.

[0016] The next moment's output power of the target wind turbine is predicted based on the LSTM-SA neural network model.

[0017] Furthermore, the operation and maintenance data of the target wind turbine is obtained, and the operation and maintenance data is preprocessed, including:

[0018] The operation and maintenance data of the target wind turbine are collected at preset intervals periodically.

[0019] Remove null and outlier values ​​from the operation and maintenance data and perform normalization processing.

[0020] Furthermore, the preprocessed wind turbine operation and maintenance data are used for feature selection using the lasso algorithm to filter out operation and maintenance data with feature coefficients that are approximately 0 or equal to 0.

[0021] Furthermore, the formula for the lasso algorithm is as follows:

[0022]

[0023]

[0024] (1) In the formula, s represents the harmonic parameter; y i Let be the objective function value, be the active power output of the wind turbine at time i, and α be the value of y. i The estimated value, x ij Let be the state characteristic value of the wind turbine at time i, β be the coefficient of the characteristic value x, and argmin mean minimizing the value of the right-hand side. Values.

[0025] Furthermore, the preprocessed wind turbine operation and maintenance data were used for feature selection using the Pearson correlation coefficient method, retaining operation and maintenance data with correlation coefficients close to 1.

[0026] Furthermore, the formula for the Pearson correlation coefficient method is as follows:

[0027]

[0028] (2) In the formula, x i For eigenvalues, y is the average value of this characteristic value. i For the target value, This represents the average of the target values.

[0029] Furthermore, based on the trained FNN model, the rate of change of power generation is obtained using the chain rule, including:

[0030] The partial derivative of power P with respect to wind speed S is obtained using an FNN:

[0031]

[0032] (11) In the formula, F(S,θ) represents an implicit function; S represents wind speed; θ represents blade deflection angle; P' represents the derivative of power generation P;

[0033] Using the backpropagation algorithm, the partial derivative of the loss function Loss with respect to the global direction is obtained, along with the partial derivatives of the loss function with respect to the wind speed S and P. Then, P' is obtained through the chain rule, which serves as the rate of change of power generation.

[0034] Furthermore, when training the LSTM network:

[0035] The gradient descent optimization algorithm uses Adam, the loss function is MSE, the LSTM-SA is set to 2 hidden layers with 6 neurons, the LSTM layer is 1 layer, and the LSTM layer is followed by the SA layer and the fully connected layer.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] This invention provides a wind turbine operation and maintenance data-driven LSTM-SA neural network ultra-short-term power prediction method. It uses wind turbine operation and maintenance data to extract the relationship between power generation and the objective environment, derives the formula for the derivative of the neural network with respect to the input value, and uses this formula to obtain this relationship. Historical power data is processed using an LSTM neural network, and the hidden layers of the LSTM are optimized using a self-attention mechanism, improving the model's accuracy and generalization ability. Compared with traditional deep learning models such as LSTM, GRU, and RNN, the LSTM-SA model has higher prediction accuracy and generalization ability, and higher convergence efficiency, resulting in more accurate wind power prediction. Attached Figure Description

[0038] Figure 1Flowchart of an LSTM-SA neural network-driven ultra-short-term power prediction method for wind turbine operation and maintenance data;

[0039] Figure 2 This is a schematic diagram of the gating mechanism of LSTM;

[0040] Figure 3 This is a schematic diagram of the calculation process of the self-attention model;

[0041] Figure 4 This is a diagram of the LSTM Self-attention model structure.

[0042] Figure 5 Flowchart of an LSTM-SA neural network ultra-short-term power prediction algorithm driven by wind turbine operation and maintenance data;

[0043] Figure 6 A three-dimensional scatter plot of wind speed, blade angle, and power;

[0044] Figure 7 A two-dimensional scatter plot of wind speed, blade angle, and power;

[0045] Figure 8 A graph showing the decreasing trend of the FNN loss function;

[0046] Figure 9 This is a schematic diagram of FNN regression prediction;

[0047] Figure 10 This is a diagram of the LSTM-SA structure;

[0048] Figure 11 This describes the descent process of the LSTM-SA loss function.

[0049] Figure 12 The image shows the prediction results of LSTM-SA.

[0050] Figure 13 A comparison chart of results from different models. Detailed Implementation

[0051] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0052] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," "outer," "front end," "rear end," "both ends," "one end," and "the other end," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0053] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installed," "equipped with," "connected," etc., should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0054] Wind turbines generate a large amount of sensor data. It's necessary to identify key variables related to power and combine their changes with historical data to improve the accuracy of power predictions. Blindly increasing the data dimensionality will lead to resource waste and negatively impact prediction speed and effectiveness. Therefore, referring to... Figure 1 As shown, this invention provides a wind turbine operation and maintenance data-driven LSTM-SA neural network ultra-short-term power prediction method, comprising:

[0055] S10. Obtain the operation and maintenance data of the target wind turbine, and preprocess the operation and maintenance data to meet the input format of the neural network; the operation and maintenance data includes: wind speed, blade deflection angle, yaw angle, humidity and temperature;

[0056] S20. Use the lasso algorithm and Pearson correlation coefficient method to perform feature screening on the preprocessed wind turbine operation and maintenance data to obtain the screening data that has a significant impact on wind power changes; remove irrelevant data.

[0057] S30. Based on the screened data, train an FNN model to obtain the functional relationship between the wind turbine power generation and the screened data;

[0058] S40. Based on the FNN model obtained through training, the rate of change of power generation is obtained using the backpropagation algorithm and the chain rule.

[0059] S50. The rate of change of the power generation and the historical power value are used together as the training set to train the LSTM network. The attention mechanism is used to optimize the LSTM network model to obtain the LSTM-SA neural network model.

[0060] S60. Predict the output power of the target wind turbine at the next moment based on the LSTM-SA neural network model.

[0061] This invention uses an operation and maintenance data-driven Long Short-Term Memory (LSTM-SA) neural network prediction algorithm to predict ultra-short-term wind power. First, features such as wind speed, blade deflection angle, yaw angle, humidity, and temperature from wind turbine operation and maintenance are selected using the Pearson correlation coefficient method and the Lasso algorithm to determine the correlation between wind speed, blade deflection angle, and power. Then, a fully connected neural network (FNN) is trained to establish a mapping model between wind speed, blade deflection angle, and power, and the power change rate is calculated based on wind speed and blade deflection angle. Finally, based on historical power data and the power change rate, an LSTM neural network is used to obtain the predicted output power of the wind turbine at the next moment. To improve training efficiency, a self-attention mechanism is used to optimize the hidden layers of the LSTM model. Test results show that, compared with other prediction algorithms, it has advantages such as higher prediction accuracy, faster convergence speed, and better stability, and can solve the problem of ultra-short-term power prediction with limited wind power training data but high accuracy requirements.

[0062] 1. Explain the model construction principles in steps S20-S50 above:

[0063] 1.1 In step S20, the Lasso and Pearson coefficient methods are used to perform feature selection:

[0064] Lasso regression, or Lasso (Least Absolute Shrinkage and Selection Operator, Tibshirani (1996)) regression method, is essentially a compression estimation. It refines the model through a penalty function. When there are many features, it can compress or set the coefficients of some features to zero. The less correlated a feature is, the more likely its coefficient is to be set to zero. The remaining non-zero coefficients correspond to features with greater correlation, thereby achieving feature selection.

[0065] In general regression models, features are typically considered independent. To eliminate the interference of different units of measurement of these features on the regression results, all independent variables need to be standardized. That is, the mean of the independent variables should be 0, and the variance should be 1. The Lasso regression formula for a general regression model is:

[0066]

[0067]

[0068] (1) In the formula, s represents the harmonic parameter; y i Let be the objective function value, be the active power output of the wind turbine at time i, and α be the value of y. i The estimated value, x ij Let be the state characteristic value of the wind turbine at time i, β be the coefficient of the characteristic value x, and argmin mean minimizing the value of the right-hand side. Values.

[0069] Where s≥0 are harmonic parameters, and for any s, there is an estimate of α. By adjusting s, the overall regression coefficient can be reduced. Let s 0 =∑ j |β j |, when s≥s 0 When the coefficients of certain features decrease, or even approach or equal to 0, these irrelevant or minimally related independent variables will be eliminated, thereby improving the accuracy and interpretability of the regression model.

[0070] The Pearson correlation coefficient is a method for measuring vector similarity. The output ranges from -1 to +1, where 0 represents no correlation, negative values ​​represent negative correlation, and positive values ​​represent positive correlation. The formula for the Pearson correlation coefficient is as follows:

[0071]

[0072] It is generally believed that the closer Person is to 1, the higher the correlation.

[0073] In this formula, x i For eigenvalues, y is the average value of this characteristic value. i For the target value, This represents the average of the target values.

[0074] Using the Lasso and Person algorithms to filter wind power data for operation and maintenance can reveal the features that play a major role in changes in wind power output.

[0075] 1.2 In step S30, the FNN principle

[0076] FNN is a basic type of BP neural network. Its key feature is the use of the Backpropagation (BP) algorithm to adjust model weights. The core of this algorithm is that when the network output deviates from the actual value, this difference is propagated from the back to the front, causing the weights of each layer to be adjusted in the direction of reducing the deviation (gradient descent), thus achieving global adjustment. Through continuous forward calculation to calculate the error and backward calculation to calculate the derivative, the model parameters are adjusted to reduce the deviation to the expected range.

[0077] According to the Universal Approximation Theorem (Cybenko (1989)), FNNs have a strong fitting ability and can fit any continuous nonlinear function. The relationship between operation and maintenance data and wind power is a high-dimensional nonlinear function, which can be fitted with high accuracy using FNNs.

[0078] 1.3 In step S50, the principle of LSTM Self-attention...

[0079] The methods mentioned above provide ideas and approaches for building the first-stage regression model. In order to build the second-stage prediction model, this invention chose LSTM Self-attention.

[0080] LSTM is a variant of Recurrent Neural Network (RNN) that, while processing long-term sequences, avoids long-range dependencies by introducing a gating mechanism. A schematic diagram of the LSTM neuron gating mechanism is shown below. Figure 2 As shown.

[0081] The "gate" in LSTM is a soft gate, with a value between (0,1), which represents filtering information at a certain ratio. Figure 2 The LSTM recurrent cell structure contains a forget gate f t Input gate i t Output gate o t This allows for the selective preservation of information. (Forget gate) t Control the internal state c of the previous moment t-1 How much information needs to be discarded; input gate i t Controls how much information from the current input needs to be saved; output gate o t Control the internal state c at the current moment t How much information can be output to the external state h? t The formula for calculating the three doors is:

[0082] f t =σ(W f [ht-1 ,x t ]+b f (3)

[0083] i t =σ(W it [h t-1 ,x t ]+b i (4)

[0084] o t =σ(W ot [h t-1 ,x t ]+b o (5)

[0085] Where σ is the Logistic function (sigmoid function), x t h is the input at the current moment. t-1 This is the output from the previous moment.

[0086] Expressed in matrix form as follows:

[0087] c' t =σ(W c [h t-1 ,x t ]+b c (6)

[0088]

[0089] c t =f t ·c t-1 +i t ·c' t (8)

[0090] h t =o t ·tanh(c t (9)

[0091] Where, x t is the current input, and W and b are network parameters.

[0092] Standard LSTM uses a traditional encoder-decoder structure. Regardless of length, the input data sequence to the LSTM is encoded into a fixed-length vector representation. While LSTM's memory function can preserve long-term states, it struggles with handling large, high-dimensional datasets in practical applications. During training, the model may ignore important temporal information or focus on less important information, leading to degraded performance and reduced prediction accuracy. To address these shortcomings, this paper incorporates a self-attention mechanism into the hidden state layer of the LSTM. This mechanism allows the model to assess the importance of input temporal information at different times, improving the convergence speed of the LSTM without increasing computational complexity. The self-attention model employs a query-key-value (QKV) model. The specific calculation process is as follows:

[0093] For the external output h obtained from formula (9) t This is linearly mapped to three different spaces:

[0094] Q = W q h t

[0095] K = W k h t

[0096] V = W v h t (10)

[0097] Q, K, and V are matrices composed of the query vector, key vector, and value vector, respectively.

[0098] For each query vector q n ∈Q, the output vector is obtained using the key-value pair attention mechanism.

[0099] h n =att((K,V),q n )

[0100] Let h n Replace h t The output is sent to external space to participate in the next round of calculations, and the calculation process is as follows: Figure 3 As shown.

[0101] In summary, the model structure of LSTM Self-attention is as follows: Figure 4 As shown, the FC layer is a fully connected layer.

[0102] 1.4 Algorithm Design for Taking Partial Derivatives of FNN with Respect to Input Layer Variables

[0103] FNNs can fit nonlinear continuous functional relationships. The data objects addressed in this paper are wind speed S, blade deflection angle θ, and power P, which cannot change abruptly and are all continuous. Since continuity implies differentiability, theoretically, the partial derivative of power P with respect to wind speed S can be obtained using an FNN.

[0104]

[0105] (11) In the formula, S represents wind speed; θ represents blade deflection angle; P' represents the derivative of power generation P; in the above formula, F(S,θ) is an implicit function and cannot be directly differentiated. However, when using the backpropagation algorithm, the loss function Loss will take a partial derivative with respect to the global direction. The partial derivatives of the loss function with respect to wind speed S and P can be obtained. Let the loss function be the mean squared error loss function MSE:

[0106]

[0107] Where i is the data sequence number and n is the total number of data.

[0108] P i Since P is a constant, we can use the chain rule to find P':

[0109]

[0110]

[0111] remember For S grad Then we have:

[0112]

[0113] n is the amount of data, but in formula (14) n = 1 is taken, so the rate of change and trend of power P under a certain S can be obtained, which provides a specific basis for predicting the next change of P.

[0114] 1.5 Operation and Maintenance Data-Driven Predictive Model Structure

[0115] In summary, the flow of the wind turbine operation and maintenance data-driven LSTM-SA neural network ultra-short-term power prediction algorithm provided by this invention is as follows: Figure 5 As shown, by analyzing wind turbine operation and maintenance data, the correlation between wind speed, blade deflection angle and power is determined. Then, an FNN model is established, and the power change rate is calculated from the wind speed and blade deflection angle. Finally, the power change rate and historical power values ​​are used together to train the LSTM neural network, and a self-attention mechanism is adopted to improve the training effect.

[0116] Finally, in step S60, the output power of the target wind turbine at the next moment is predicted based on the LSTM-SA neural network model; the prediction accuracy is high.

[0117] 2. Case Analysis

[0118] 2.1 Dataset Description

[0119] The experimental data used in this embodiment is from the Baidu KDD CUP 2022 competition (https: / / aistudio.baidu.com / aistudio / competition / detail / 152 / 0 / task-definition). Wind turbine No. 5 from the Longyuan Wind Farm was selected, and data was sampled every 10 minutes, resulting in a total of 26,470 data points. The meanings of the data names are shown in Table 1.

[0120] Feature name Feature meaning Wspd(m / s) Wind speed collected by an anemometer Wdir(·) wind direction Etmp (°C) Temperature of the surrounding environment of the fan Itmp (°C) turbine temperature Ndir(·) yaw angle of the wind turbine Pad(·) The angle of blade deflection Prtv(kW) reactive power Patv(kW) Active power

[0121] Table 1

[0122] 2.2 Data Preprocessing

[0123] 2.2.1 Normalization Processing

[0124] In this dataset, different features have different dimensions, and some of these dimensions differ significantly in value and physical meaning, which may negatively impact the training results. Therefore, it is necessary to perform maximum-minimum normalization on the historical data to eliminate these dimensions and compress the data to the range [0,1]. The maximum-minimum normalization formula is:

[0125]

[0126] Where, x i Let be any number in the dataset. Let x be the mean of the dataset, max(x) be the maximum value of the dataset, and min(x) be the minimum value of the dataset.

[0127] 2.2.2 Data Filtering Using Lasso and Person Algorithms

[0128] In this dataset, the presence of numerous features, especially those with low relevance to the labels, often negatively impacts model performance, leading to reduced prediction accuracy. Therefore, it is necessary to analyze the data features, remove worthless features, and reduce data dimensionality.

[0129] In this embodiment, the correlation coefficient method and the lasso algorithm were used to screen the features, and it was found that the correlation between wind speed S, blade turning angle θ, and power P was higher. The correlation values ​​are shown in Table 2.

[0130] Feature Name Lasso algorithm Person Law Wind speed Wspd 147.772 0.964 Blade deflection angle Pad -534.477 0.033 Ambient temperature Etmp -20.740 -0.027 Fan internal temperature Itmp 21.604 0.153 Wind direction Wdir 0.714 -0.069 Yaw angle Ndir 0.006 -0.182

[0131] Table 2 Correlation Comparison

[0132] like Figure 6 The image shows a three-dimensional scatter plot of wind speed, blade angle, and power. Figure 7 The figure shows a two-dimensional scatter plot of wind speed, blade angle, and power.

[0133] 2.3 Model Evaluation Indicators

[0134] To evaluate the predictive performance of the model, this embodiment uses the root mean square error (RMSE) and the coefficient of determination R0. 2 This serves as a standard for evaluating the predictive performance of the model.

[0135]

[0136]

[0137] Where y i The true value of the target wind turbine's output power. The predicted value of the target wind turbine output power. Let n be the average output power of the wind turbines in the operation and maintenance data test set, and n be the number of samples. MSE reflects the difference between model predictions and actual values. 2 MSE reflects the proportion of the total variance of the dependent variable that can be explained by the independent variables through the regression relationship. The smaller the MSE, the better the model fit, while R0... 2 The closer the value is to 1, the better the model fit.

[0138] 2.4 Construction and Training of the FNN Wind Power Fitting Model

[0139] Using [St,θt] as the input space and [Pt] as the output space, the ratio of training set, validation set, and test set is set to 8:1:1. A 2×500×50×1 FNN is constructed, with Adam (adaptive learning rate gradient descent) as the optimization algorithm and MSE (mean squared error loss) as the loss function. Training is performed with early stopping enabled, i.e., iteration stops when the error between the loss of the training set and the loss of the test set is less than 5%. The gradient descent trend of the FNN is as follows. Figure 8 As shown.

[0140] The obtained FNN is denoted as F(S) t ,θ t ) = P t On the test set, F(S) t ,θ t The corrected determination coefficient R) 2 The mean square error (RMSE) is 0.984.

[0141] F(S t ,θ t Regression prediction results are as follows Figure 9 As shown.

[0142] 2.5. Construction and training of the prediction model.

[0143] 2.5.1 Construction and Partitioning of the Dataset

[0144] Historical power generation data of wind turbine No. 5 of Longyuan Wind Farm were selected. For the operation and maintenance data at each moment, formula (14) was used to calculate the rate of change of power generation with respect to wind speed P' at that moment. t The power P at the corresponding time t and rate of change P' t Perform vector concatenation [P' t ,P t ].

[0145] Calculated P' t Examples are shown in Table 3 below:

[0146]

[0147]

[0148] Table 3

[0149] 3.5.2 Training of the LSTM-SA prediction model

[0150] Analysis of current wind power generation and historical data reveals that the current power output is more significantly influenced by the historical power output over the previous six time steps. Therefore, the X values ​​from the previous six time steps are selected. i =[P' i ,P i The current power generation Pt serves as the input space, and the current power generation Pt serves as the output space, denoted as […].

[0151] LSTMSA(X t-6 X t-5 X t-4 X t-3 X t-2 X t-1 ) = P t ,

[0152] The dataset contains 18,926 data points. The first 80% is selected as the training set, the middle 10% as the validation set, and the last 10% as the test set. The Adam gradient descent optimization algorithm is used, and the MSE loss function is chosen. The LSTM-SA layer is configured with two hidden layers (6 neurons each), one LSTM layer, followed by an SA layer and a fully connected layer. Figure 10 As shown.

[0153] The loss function descent process of LSTM-SA on the training and validation sets is as follows: Figure 11 As shown, training stops when the error between the loss function value of the training set and the loss function value of the validation set is less than 5%.

[0154] 2.6 Prediction Results and Analysis

[0155] 2.6.1 Evaluation Criteria for Prediction Results

[0156] For example, according to Article 11 of the "Detailed Rules for Grid Connection Assessment of Hunan Power Grid" (hereinafter referred to as the "Detailed Rules") issued by the Hunan Regulatory Office of the National Energy Administration on February 18, 2022, wind farms should establish scientific power prediction models, including ultra-short-term wind power prediction models for the next 15 minutes to 4 hours. The accuracy of ultra-short-term power prediction should not be less than 87%, and the accuracy calculation formula is as follows:

[0157]

[0158] In the above formula, Cap represents the available capacity of the wind farm.

[0159] Additionally, the coefficient of determination R is used. 2 The coefficient of determination R is used to measure the regression accuracy of the model. 2 The calculation method is shown in formula (16).

[0160] 2.6.2 Results of the Prediction Model

[0161] Comparing the predicted results output by the LSTM-SA prediction model with the actual power results, for example... Figure 12 Table 4 shows the scoring results of the evaluation function.

[0162] <![CDATA[R 2 ]]> ACC 0.96544 98.8％

[0163] Table 4

[0164] It can be seen that the LSTM-SA prediction results based on wind power operation and maintenance data are close to the actual wind power curve, and the prediction accuracy meets the requirements of the "Detailed Rules".

[0165] 2.6.3 Comparison with different models

[0166] The prediction performance of the LSTM-SA model based on wind power operation and maintenance data is compared with that of traditional LSTM networks, GRU networks, and RNN networks. Figure 13 Table 5 shows the evaluation index values ​​for different models to compare their prediction results.

[0167] <![CDATA[R 2 ]]> ACC LSTM 0.92523 96.8％ GRU 0.92506 94.4％ RNN 0.92588 94.1％

[0168] Table 5

[0169] from Figure 13 As shown in Table 5, within the same wind power operation and maintenance test set, the LSTM-SA model most closely approximates the actual wind power curve. Compared with traditional LSTM, GRU, and RNN, the LSTM-SA model reduces errors by approximately 2.0%, 4.4%, and 4.7%, respectively.

[0170] Compared to traditional prediction methods, deep learning methods offer advantages in data analysis and high abstraction. However, deep learning methods sometimes overlook the impact of the actual objective environment on wind power generation, resulting in poor interpretability. This invention proposes a wind turbine operation and maintenance data-driven LSTM-SA neural network ultra-short-term power prediction method. It uses wind turbine operation and maintenance data to extract the relationship between power generation and the objective environment, deriving the formula for the derivative of the neural network with respect to the input value. This formula is used to obtain this relationship. Historical power data is processed using an LSTM neural network, and the hidden layers of the LSTM are optimized using a self-attention mechanism, improving the model's accuracy and generalization ability. Compared with traditional deep learning models such as LSTM, GRU, and RNN, the LSTM-SA model exhibits higher prediction accuracy and generalization ability, as well as higher convergence efficiency. This demonstrates the practicality of the operation and maintenance data-driven LSTM-SA wind power prediction algorithm in wind power prediction problems.

[0171] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A wind turbine operation data driven LSTM-SA neural network ultra-short-term power prediction method, characterized in that, The method comprises the following steps: Obtaining the operation and maintenance data of a target wind turbine, and preprocessing the operation and maintenance data; the operation and maintenance data comprises wind speed, blade deflection angle, yaw angle, humidity and temperature; Using a lasso algorithm and a Pearson correlation coefficient method to perform feature screening on the preprocessed wind turbine operation and maintenance data, and obtaining screened data that has a greater impact on wind power change; Training an FNN model based on the screened data, and obtaining a functional relationship between wind turbine power generation and the screened data; Using a back propagation algorithm and a chain rule to obtain a change rate of power generation based on the trained FNN model; Using the change rate of power generation and historical power values as a training set to train an LSTM network, optimizing the LSTM network model by using an attention mechanism, and obtaining an LSTM-SA neural network model; Predicting the next moment output power of the target wind turbine based on the LSTM-SA neural network model; Wherein, using the chain rule to obtain the change rate of power generation based on the trained FNN model comprises: Obtaining the partial derivative of power P with respect to wind speed S by FNN: (11) where F(S, θ) represents an implicit function; S represents a wind speed; θ represents a blade pitch angle; P ' represents a derivative of the power generation P; Using the back propagation algorithm, the loss function Loss is globally differentiated to obtain the partial derivative of the loss function with respect to the wind speed S and the partial derivative of the loss function with respect to P; through the chain rule, the P ' , as the rate of change of the generated power.

2. The fan operation and maintenance data-driven LSTM-SA neural network ultra-short-term power prediction method according to claim 1, characterized in that, Obtaining the operation and maintenance data of a target wind turbine, and preprocessing the operation and maintenance data comprises: Collecting the operation and maintenance data of the target wind turbine at a periodic interval and a preset time; Removing null values and abnormal values of the operation and maintenance data, and performing normalization processing.

3. The wind turbine operation data-driven LSTM-SA neural network ultra-short-term power prediction method according to claim 1, characterized in that, Using a lasso algorithm to perform feature selection on the preprocessed wind turbine operation and maintenance data, and screening out operation and maintenance data with a feature coefficient close to 0 or equal to 0.

4. The wind turbine operation data-driven LSTM-SA neural network ultra-short-term power prediction method according to claim 2, characterized in that, The formula of the lasso algorithm is as follows: (1) where s represents a harmonic parameter; y i is a target function value, P(i) is an active power outputted by the wind turbine at i moment, a is an estimated value of y i , x ij is a state characteristic value of the wind turbine at i moment, β is a coefficient of the characteristic value x, and argmin means a value that makes the value of the right formula minimum .

5. The fan operation data-driven LSTM-SA neural network ultra-short-term power prediction method according to claim 1, characterized in that, Using a Pearson correlation coefficient method to perform feature selection on the preprocessed wind turbine operation and maintenance data, and retaining operation and maintenance data with a correlation coefficient close to 1.

6. The wind turbine operation data-driven LSTM-SA neural network ultra-short-term power prediction method according to claim 5, characterized in that, The formula of the Pearson correlation coefficient method is as follows: (2) where x i is the characteristic value, is the average of the characteristic value, y i is the target value, is the average of the target value.

7. The wind turbine operation data-driven LSTM-SA neural network ultra-short-term power prediction method according to claim 1, characterized in that, When training the LSTM network: The gradient descent optimization algorithm uses Adam, the loss function selects MSE, the LSTM-SA hidden layer is set to 2 layers, the hidden layer neuron is set to 6, the LSTM layer is set to 1 layer, and the LSTM layer is connected to the SA layer and the full connection layer.

Images