Wind turbine generator short-term power prediction method considering wind speed and operation state

Abstract

The invention discloses a wind turbine generator short-term power prediction method considering wind speed and operation states, and relates to the technical field of wind power generation, and the method comprises the steps: collecting the historical operation data of a wind turbine generator, constructing a training set under different operation states, and training a Bi-TCN network-based unit power prediction model under different operation states; and selecting a model corresponding to the operation state according to the current state of the target unit, and calculating a short-term power prediction value of the target unit according to the collected short-term wind speed prediction value of the target unit and historical wind speed power data of the previous day. The influence of the wind speed and the operation state can be considered, and the accuracy of short-term power prediction of the wind turbine generator is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power generation, and more specifically, to a short-term power prediction method for a wind turbine considering wind speed and operating status. Background Art

[0002] Traditional power prediction methods in wind power generation often only predict the future average power of the unit based on the power curve and the numerical weather forecast wind speed, without considering the wind speed difference at the hub of different wind turbines and the influence of the unit operating status. However, during the operation of wind turbines, various uncertain factors such as time-varying wind speed and status are faced, resulting in poor short-term power prediction accuracy, difficult to meet the short-term power scheduling requirements of wind farms within a day, and causing unnecessary economic losses.

[0003] Therefore, how to consider wind speed and the operating status of wind turbines to improve the prediction accuracy of short-term power is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0004] In view of this, the present invention provides a short-term power prediction method for a wind turbine considering wind speed and operating status, which can consider the influence of wind speed and the operating status of the wind turbine on the power generation of the unit and improve the accuracy of short-term power prediction of the wind turbine.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A short-term power prediction method for a wind turbine considering wind speed and operating status, comprising the following steps:

[0007] Step 1: Collect historical operation data of the wind turbine under different operating statuses, and construct training sets under different operating statuses; the operating statuses of the wind turbine include normal status, attention status, and abnormal status;

[0008] Step 2: Construct a unit power prediction model based on a bidirectional temporal convolutional network (Bi-TCN network), and use the training set to train the unit power prediction models under different operating statuses;

[0009] Step 3: Detect the current status of the target unit, and select the trained unit power prediction model corresponding to the operating status according to the current status;

[0010] Collect the short-term wind speed prediction value of the target unit in the future and the historical wind speed power data of the previous day, input them into the selected unit power prediction model, and obtain the short-term power prediction value of the target unit; process the short-term wind speed prediction value into a future wind speed sequence and input it into the unit power prediction model; obtain the short-term wind speed prediction value from the weather forecast; the historical wind speed power data includes the wind speed of the previous day and the power of the target unit.

[0011] Preferably, the process of constructing the training set in step 1 includes:

[0012] Step 11: Collect historical operation data of multiple units under different operation states, including historical SCADA monitoring data and historical state data, and perform preprocessing;

[0013] Step 12: Use the abnormal state detection method to identify the preprocessed historical state data of each unit, and generate historical state sequences for the historical state data under the same operation state identified respectively in chronological order;

[0014] Step 13: Concatenate the historical state sequences under different operation states and the corresponding preprocessed historical SCADA monitoring data to generate the training set under different operation states.

[0015] Preferably, the preprocessing includes anomaly rejection, missing value filling, standardization, etc.; Use the abnormal state detection method to perform identification to generate state detection results, and label the historical state data with the labels of the corresponding operation states according to the state detection results.

[0016] The technical effect of the above technical solution is that, adopting a divide-and-conquer strategy, different unit power prediction models are trained respectively based on the training sets under different operation states to achieve accurate prediction.

[0017] Preferably, the unit power prediction model based on the Bi-TCN network includes an input layer, a first residual block, a second residual block, a fusion layer, a fully connected layer, and an output layer;

[0018] The training set is input into the input layer. The input layer constructs a new sample set containing multi-time scale parameters based on the training set according to the sliding time window, and inputs the new sample set into the first residual block, the second residual block, and the fusion layer respectively;

[0019] The outputs of the first residual block and the second residual block are transmitted to the fusion layer. The fusion layer concatenates and fuses the output of the first residual block, the output of the second residual block, and the new sample set to obtain the fusion features;

[0020] The fully connected layer calculates the short-term power prediction value according to the fusion features;

[0021] The output layer outputs the short-term power prediction value.

[0022] Preferably, the first residual block includes a first dilated convolutional layer, a batch normalization layer, a first dropout layer, a second dilated convolutional layer, a first input sequence normalization layer, a second dropout layer, a third dilated convolutional layer, a second input sequence normalization layer, and a third dropout layer connected in sequence;

[0023] The second residual block includes a first transpose layer, a first dilated convolutional layer, a batch normalization layer, a first dropout layer, a second dilated convolutional layer, a first input sequence normalization layer, a second dropout layer, a third dilated convolutional layer, a second input sequence normalization layer, a third dropout layer, and a second transpose layer connected in sequence.

[0024] Preferably, the filters in the first dilated convolutional layer, the second dilated convolutional layer, and the third dilated convolutional layer are represented as f:{0,...,k s -1}→R, and one convolution operation corresponds to one jump operation of the input sequence. The expression for the convolution operation is:

[0025]

[0026] where F(x) is the result of the dilated convolution on the element x in the input X; k s represents the filter size; x - d*i is the indicator in the convolution calculation direction; f(i f ) represents the i-th filter; d represents the dilation factor; * d represents the dilated convolution operation.

[0027] The technical effect of the above technical solution is that it can increase the receptive field of the TCN network by using a larger filter k s and dilation factor d; the size of a dilated convolutional layer is (k s -1)d, and the larger the dilation size, the more helpful it is for the output and better expression of a wide range of inputs, thereby increasing the receptive field; the dilation factor d = 1 is a dilated convolution equivalent to a regular convolution, and it usually grows exponentially with the network depth. The receptive field of the Bi-TCN network depends on the stacking number of residual blocks. When the filter size ks = 3, the dilation factor d = 4, and the stacking number of residual blocks is n s = 1, the receptive field size of the TCN model is 3 * 4 * 1 = 12, and the residual block contains an optional skip convolution connection layer.

[0028] Preferably, the residual block is another component of the TCN. The fully connected layer is represented as:

[0029] o = Activation[x + F(x)]

[0030] where o represents the output of the fully connected layer, F(x) represents the residual mapping learned by the dilated convolutional layer, and the weight normalization of the batch normalization layer and the input sequence normalization layer is applied to the convolutional filters within the first residual block or the second residual block; Activation represents the activation function of the fusion layer, and the ReLU function is used.

[0031] Preferably, the dropout layer adopts a regularization method to discard some random outputs in the network according to a preset dropout rate. The number of neurons to be discarded is determined by the dropout rate, and the value range of the dropout rate is [0, 1], which represents the probability that the neurons in the output layer are discarded.

[0032] Preferably, the power within t - 24 to t hours and the wind speed within t - 24 to t + 4 hours under different operating states are used as the model inputs, and the unit power at t + 4 hours is used as the model output to train the unit power prediction model; the unit power prediction model is expressed as:

[0033]

[0034] Among them, represents the predicted power output at each iteration during the training process from t + 1 to t + τ, where τ = 4; y[t - t1:t] represents the historical power sequence from the current t moment to t1 moments before t, where t1 = 24; x1[t - t1:t] represents the historical wind speed sequence from the current t moment to t1 moments before t; x2[t + 1:t + τ] represents the future wind speed sequence from t + 1 to t + τ; θ represents the model parameters; f i represents the unit power prediction model in state i; τ = w·r, r is the step size of the sliding time window in the input layer (r = 1,..., s), for the initial window, r is equal to 0, and w represents the window width of the sliding time window in the input layer; considering that the time scale of short - term power prediction is 4 hours, the window width is set to 15 minutes and the step size is set to 16 here; the historical state sequence includes the historical power sequence and the historical wind speed sequence;

[0035] Calculate the average value of the predicted power output for all iterations to obtain the short - term power prediction value of the output layer It is expressed as:

[0036]

[0037] Among them, represents the predicted power at time t; s represents the s - th time window.

[0038] In the input layer, input samples are constructed by sliding sampling through a sliding time window in the training set.

[0039] Preferably, during the training process of the unit power prediction model, the AdaBelief optimizer is used to optimize the model parameters of the Bi - TCN network; the expression of the AdaBelief optimizer is:

[0040]

[0041] m t = β1m t-1+(1 - β1)g t

[0042] s t = β1ν t-1 +(1 - β2)(g t - m t ) 2

[0043]

[0044] where f t (θ t ) represents the minimum loss function optimized at the t-th step; θ t represents the parameter of the loss function at time t; g t represents the gradient of the minimum loss function f t (θ t-1 ) with respect to θ t ; m t represents the predicted gradient value at time t; s t and ν t respectively represent the exponential moving averages of (g t - m t ) 2 and g t 2 ; is the partial derivative symbol; β1 and β2 are smoothing parameters. During the model training process, the training set is divided into a training data set and a test data set according to a ratio of 7:3, and the unit output power for the next 4 hours is predicted.

[0045] Preferably, the current operating data of the wind turbine is collected, including the current SCADA monitoring data and the current status data, the current status of the unit is judged according to the current SCADA monitoring data, and the unit power prediction model corresponding to the status is selected;

[0046] The short-term wind speed prediction value of the unit and the historical wind speed power data of the previous day are input into the selected unit power prediction model, and the short-term power prediction value for the next 4 hours is output.

[0047] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses a short-term power prediction method for a wind turbine considering wind speed and operating status. First, the historical operation data of the unit is divided into training sets according to normal status, attention status, and abnormal status respectively, and training sets for the unit power prediction model under different operating statuses are constructed; then, the power within t - 24 to t hours and the wind speed within t - 24 to t + 4 hours under different operating statuses are used as model inputs, and the unit power at t + 4 is used as the model output to train the Bi-TCN power prediction models under three operating statuses; finally, the current status of the unit is detected, the Bi-TCN model corresponding to the status is selected, and based on the short-term wind speed prediction value of the unit and the historical wind speed power data of the previous day, the power of the unit in the next 4 hours is predicted. The specific beneficial effects include:

[0048] (1) The Bi-TCN network consists of two time convolutional layers in opposite directions, which can extract past and future information respectively, thus effectively capturing the bidirectional dependencies in the time series. In contrast, traditional time convolutional networks can only judge the current state based on past time feature information. Bi-TCN can better utilize the front and back associations in the time series data and improve the prediction accuracy;

[0049] (2) The Bi-TCN network combines the local feature extraction ability of CNN and the sequence modeling ability of RNN. Its convolutional kernel can perform convolutional operations on the input sequence to extract local features, and then effectively capture the local features and global dependencies in the wind power sequence, which helps to more comprehensively and accurately understand and predict the change law of the wind turbine power;

[0050] (3) Wind power is affected by various meteorological factors such as wind speed, wind direction, temperature, humidity, and air pressure. The Bi-TCN network can take these related meteorological factors as multivariate inputs, comprehensively consider the influence of various factors on wind power, and thus more accurately predict the wind turbine power. Compared with the model with single-variable input, it can better cope with complex actual situations and improve the accuracy and reliability of prediction;

[0051] (4) The predicted short-term power of the unit can be compared with the benchmark power of the target unit to analyze the short-term power fluctuation characteristics of the unit, so as to formulate a reasonable short-term power scheduling plan and maintenance strategy for the wind power operation enterprise, and improve the grid connection competitiveness of the wind power enterprise from the aspect of wind turbine management. Brief Description of the Drawings

[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on the provided drawings.

[0053] Figure 1 Schematic diagram of the short-term power prediction method for a wind turbine considering wind speed and operating state provided by the present invention;

[0054] Figure 2 Schematic diagram of the short-term power prediction structure of a wind turbine considering wind speed and operating state in the embodiment provided by the present invention;

[0055] Figure 3 Schematic diagram of the Bi-TCN network structure provided by the present invention;

[0056] Figure 4 Schematic diagram of the iterative training process of the unit power prediction model based on a sliding time window in the embodiment provided by the present invention;

[0057] Figure 5 Schematic diagram of the training loss function of the unit power prediction model under different states in the embodiment provided by the present invention;

[0058] Figure 6 Schematic diagram of the power prediction results of the unit power prediction model for Unit 22 under different states in the embodiment provided by the present invention;

[0059] Figure 7 Schematic diagram of the comparison of the short-term power prediction results of Unit 22 by each model provided by the present invention;

[0060] Figure 8 Schematic diagram of the local comparison results of the power prediction during the period from T4 to T5 provided by the present invention. Detailed implementation manners

[0061] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0062] The embodiments of the present invention disclose a short-term power prediction method for a wind turbine considering wind speed and operating state, as Figure 1 shown, including the following steps:

[0063] S1: Collect the historical operation data of the wind turbine under different operation states, and construct the training sets under different operation states; the operation states of the wind turbine include normal state, attention state, and abnormal state;

[0064] S2: Construct a unit power prediction model based on the Bi-TCN network, and use the training sets to train the unit power prediction models under different operation states;

[0065] S3: Detect the current state of the target unit, and select the trained unit power prediction model corresponding to the operation state according to the current state; collect the short-term wind speed prediction values of the target unit in the future and the historical wind speed-power data of the previous day, and input them into the selected unit power prediction model to obtain the short-term power prediction value of the target unit.

[0066] Further, obtain the short-term wind speed prediction values from the weather forecast; process the short-term wind speed prediction values into future wind speed sequences and input them into the unit power prediction model; the historical wind speed-power data includes the wind speed of the previous day and the power of the target unit. Obtain the meteorological predicted wind speed according to the weather forecast.

[0067] Further, use the meteorological predicted wind speed as the input of the benchmark unit, predict the power of the benchmark unit as the reference power, and construct the power curve of the benchmark unit; according to the power curve of the benchmark unit, the corresponding power of the benchmark unit can be found according to the meteorological predicted wind speed.

[0068] Further, the process of constructing the training sets in S1 includes:

[0069] S11: Collect the historical operation data of multiple units under different operation states, including historical SCADA monitoring data and historical state data, and perform preprocessing;

[0070] S12: Use the abnormal state detection method to identify the preprocessed historical state data of each unit, and generate historical state sequences for the historical state data under the same identified operation state in chronological order;

[0071] S13: Concatenate the historical state sequences under different operation states and the corresponding preprocessed historical SCADA monitoring data to generate the training sets under different operation states.

[0072] Further, the preprocessing includes abnormal value elimination, missing value filling, standardization, etc.; use the abnormal state detection method to perform identification to generate state detection results, and label the historical state data with the corresponding operation state labels according to the state detection results. Adopt the divide-and-conquer strategy to train different unit power prediction models based on the training sets under different operation states to achieve accurate prediction.

[0073] Further, the unit power prediction model based on the Bi-TCN network includes an input layer, a first residual block, a second residual block, a fusion layer, a fully connected layer, and an output layer; the training set is input into the input layer, and the input layer constructs a new sample set containing multi-time scale parameters based on the training set according to a sliding time window, and inputs the new sample set into the first residual block, the second residual block, and the fusion layer respectively; the outputs of the first residual block and the second residual block are transmitted to the fusion layer, and the fusion layer splices and fuses the output of the first residual block, the output of the second residual block, and the new sample set to obtain a fusion feature; the fully connected layer calculates the short-term power prediction value according to the fusion feature; the output layer outputs the short-term power prediction value.

[0074] Further, the first residual block includes a first dilated convolutional layer, a batch normalization layer, a first dropout layer, a second dilated convolutional layer, a first input sequence normalization layer, a second dropout layer, a third dilated convolutional layer, a second input sequence normalization layer, and a third dropout layer connected in sequence;

[0075] The second residual block includes a first flipping layer, a first dilated convolutional layer, a batch normalization layer, a first dropout layer, a second dilated convolutional layer, a first input sequence normalization layer, a second dropout layer, a third dilated convolutional layer, a second input sequence normalization layer, a third dropout layer, and a second flipping layer connected in sequence. The data is collected and the training set is divided according to the operating state. The process of the unit power prediction model processing the data in the training set is as Figure 3 shown.

[0076] Further, the filters in the first dilated convolutional layer, the second dilated convolutional layer, and the third dilated convolutional layer are represented as f:{0,...,k s -1}→R. One convolution operation corresponds to one jump operation of the input sequence. The expression for the convolution operation is:

[0077]

[0078] where F(x) is the result of the dilated convolution on the element x in the input X; k s represents the filter size; x - d*i is the indicator of the convolution calculation direction; f(i f ) represents the i-th filter; d represents the dilation factor; * d represents the dilated convolution operation. The receptive field of the TCN network can be increased by using a larger filter k s and dilation factor d; the size of a dilated convolutional layer is (k s -1)d. The larger the dilation size, the more helpful it is for the output and better expression of a wide range of inputs, thereby increasing the receptive field; the dilation factor d = 1 is a dilated convolution equivalent to a regular convolution, and usually increases exponentially with the network depth. The receptive field of the Bi-TCN network depends on the stacking number of the residual blocks. When the filter size ks = 3, dilation factor d = 4, and the stacking number of residual blocks is n s When n = 1, the receptive field size of the TCN model is 3 * 4 * 1 = 12, and the residual block contains an optional skip convolution connection layer.

[0079] Furthermore, the residual block is another component of the TCN. The fully connected layer is expressed as:

[0080] o = Activation[x + F(x)]

[0081] where o represents the output of the fully connected layer, F(x) represents the residual mapping learned by the dilated convolution layer, and the weight normalization of the batch normalization layer and the input sequence normalization layer is applied to the convolution filter within the first residual block or the second residual block; Activation represents the activation function of the fusion layer, and the ReLU function is adopted.

[0082] Furthermore, the dropout layer adopts a regularization method to discard some random outputs in the network according to a preset dropout rate. The number of neurons to be discarded is determined by the dropout rate, and the value range of the dropout rate is [0, 1], which represents the probability that the output layer neurons are discarded.

[0083] Furthermore, as Figure 4 shown, the power from t - 24 to t hours and the wind speed from t - 24 to t + 4 hours under different operating states are used as the model inputs, and the unit power at t + 4 hours is used as the model output to train the unit power prediction model; the unit power prediction model is expressed as:

[0084]

[0085] where represents the predicted power output at each iteration during the training process from t + 1 to t + τ, τ = 4; y[t - t1:t] represents the historical power sequence from the current t moment to t1 moments before t, t1 = 24; x1[t - t1:t] represents the historical wind speed sequence from the current t moment to t1 moments before t; x2[t + 1:t + τ] represents the future wind speed sequence from t + 1 to t + τ; θ represents the model parameters; f i represents the unit power prediction model in state i; τ = w·r, r is the step size of the sliding time window of the input layer (r = 1,..., s), for the initial window, r is equal to 0, w represents the window width of the sliding time window of the input layer; considering that the time scale of short - term power prediction is 4 hours, the window width is set to 15 minutes and the step size is set to 16 here; the historical state sequence includes the historical power sequence and the historical wind speed sequence;

[0086] Calculate the average value of the predicted power output at all iterations to obtain the short - term power prediction value of the output layer Expressed as:

[0087]

[0088] Wherein, represents the predicted power at time t; s represents the s-th time window.

[0089] Furthermore, in the input layer, input samples are constructed by sliding sampling in the training set through a sliding time window, as shown in Figure 4 (a) in which shows the process of constructing input samples by the sliding time window, Figure 4 (b) in which shows the moving process of the window. The product of the moving window width and its moving step determines the duration of the predicted power, Figure 4 (c) in which describes the mapping relationship of the prediction model, Figure 4 (d) in which describes the output form.

[0090] Furthermore, during the training process of the unit power prediction model, the AdaBelief optimizer is used to optimize the model parameters of the Bi-TCN network; the expression of the AdaBelief optimizer is:

[0091]

[0092] m t = β1m t-1 + (1 - β1)g t

[0093] s t = β1ν t-1 + (1 - β2)(g t - m t ) 2

[0094]

[0095] Wherein, f t (θ t ) represents the minimum loss function optimized at the t-th step; θ t represents the parameter of the loss function at time t; g t represents the gradient of the minimum loss function f t (θ t-1 ) with respect to θ t at the (t - 1)-th step; m t represents the predicted value of the gradient at time t; s t and ν t respectively represent (g t - m t ) 2 and g t 2Exponential Moving Average; is the symbol of partial derivative; β1 and β2 are smoothing parameters. During the model training process, the training set is divided into a training data set and a test data set according to a ratio of 7:3, and the unit output power for the next 4 hours is predicted.

[0096] Furthermore, collect the current operating data of the wind turbine, including the current SCADA monitoring data and the current status data, judge the current status of the unit according to the current SCADA monitoring data, and select the unit power prediction model under the corresponding status;

[0097] Input the short-term wind speed prediction value of the unit and the historical wind speed power data of the previous day into the selected unit power prediction model, and output the short-term power prediction value for the next 4 hours.

[0098] On the other hand, in a specific embodiment, a short-term power prediction method for a wind turbine considering wind speed and operating status is applied to a certain wind farm, such as Figure 2 shown. Collect the experimental data set:

[0099] The data is from the SCADA system of a certain wind farm, with an installed capacity of 50MW, the rated power of each unit is 2.0MW, and the benchmark unit in this wind farm is No. 4; all units in this wind farm are of the same model, batch, and manufacturer; take the operating data of Unit 22 as an example to verify the feasibility of the short-term power prediction model.

[0100] Select the operating data of Unit 22 from January 1, 2020 to December 31, 2020 for verification. The sampling time frequency of the SCADA system is 15 minutes. Among them, there are 35041 sampling points for air density, with 95 missing samples, 34609 sampling points for SCADA power data, with 427 missing samples, and 34126 sampling points for SCADA wind speed data, with 1010 missing samples.

[0101] The specific process of short-term power prediction for this wind farm includes the following steps:

[0102] S1. Divide the unit historical operating data into training sets according to normal status, attention status, and abnormal status respectively, and construct the training sets of the unit power prediction models under different operating statuses;

[0103] Concatenate various types of data, merge the time and date fields and convert them into an index timestamp, and classify the training dataset based on the historical status data of the unit; first, remove the data outside the cut-in wind speed and cut-out wind speed, leaving 33,291 sampling points, and then preprocess the outliers and missing values of each type of data. Among them, the amount of normal status data is 20,352, the amount of attention status data is 9,285, and the amount of abnormal status data is 3,654. For each type of status data, divide the training set and test set according to a 7:3 ratio;

[0104] S2. Use the power from t-24 to t hours and the wind speed from t-24 to t+4 hours under different operating states as the model inputs, and the power of the unit at t+4 as the model output to train the unit power prediction models under three operating states;

[0105] S2.1. Use the historical power and wind speed under different states and the wind speed in the period to be predicted as the model inputs, and the power in the period to be predicted as the model output, and divide the labels for the dataset to form a two-dimensional matrix with 33,291 rows and 3 columns, and the label is a one-dimensional matrix with 33,291 rows and 1 column;

[0106] Since the output of the prediction model is the power within the next 4 hours, the length of the time window in the Bi-TCN network should be greater than 4 hours, that is, 16 sampling points. The time sliding window value Lookback is set to N. The sliding time window takes the first N-1 records as the first element of the input, and takes the Nth record in the label matrix as the label. The sliding window slides down one record in the dataset, and finally forms a three-dimensional matrix; the number of elements in this matrix is the total number of records in the original dataset minus the sliding window plus 1. Each element is a two-dimensional matrix within a sliding window. The final shape of the three-dimensional matrix is (33,291, N, 2), and the shape of the label set is (33,291, 1);

[0107] Construct new training sets using a sliding window for each type of status respectively. The process of constructing the training set is as Figure 4 shown. For the convenience of drawing, the Lookback value in the figure is set to 3;

[0108] The length of the input sequence is a key variable affecting the prediction accuracy. To analyze the influence of the input variable time scale on the unit power prediction model of the present invention, set the input variables of the unit power prediction model considering the unit status to different lengths and calculate their prediction accuracies respectively;

[0109] Set the input lengths of the model to 4 hours, 12 hours, 24 hours, and 36 hours respectively, and the output length of the model is 4 hours (step = 16 steps). The short-term power prediction accuracies under different input sequence lengths are shown in Table 1;

[0110] Table 1 Unit Status - Model Prediction Performance for Different Input Sequence Lengths

[0111]

[0112] As can be seen from the table, as the time series increases, more information on the short - term dependence of each parameter is included, so the identification accuracy is higher. However, the model prediction accuracy is not directly proportional to the input sequence. When the input sequence length exceeds 24 hours, the prediction error no longer decreases significantly. To ensure that the number of samples in the constructed dataset reaches a certain quantity, the input sequence length cannot be too large. If it is greater than 48 hours, the number of the constructed dataset will decrease significantly, resulting in insufficient model training samples. Therefore, the input sequence length is set to 24 hours;

[0113] S2.2. Train the unit power prediction model;

[0114] Taking the operation data of Unit 22 as an example, after preprocessing the training set and validation set in different operating states, input them into the constructed unit power prediction model, and perform cyclic iterative training 200 times. Calculate its loss value with the minimum of the MSE function as the goal as Figure 5 shown, where Figure 5 (a), (b), (c), and (d) in it represent the model training loss function curves in the normal state, attention state, abnormal state, and state - not - considered state respectively. The abscissa represents the number of iterations, and the ordinate represents the loss function value (MSE);

[0115] As the number of iterations increases, the loss function values of the unit power prediction model (Bi - TCN model) of the present invention on the training set and the test set both decrease rapidly. In the normal state, after 8 iterations of calculation, the loss function values on the two datasets start to decrease smoothly. After 50 iterations, the loss function values tend to be stable, indicating that the model has converged. In the attention and abnormal states, after about 20 iterations of calculation, the loss function values on the two datasets tend to decrease smoothly; The convergence speed of the prediction model without considering the unit state is relatively slow. After 80 iterations, the loss function value tends to be stable, and the loss function value is significantly greater than the prediction result considering the state. The loss function values of the Bi - TCN model on the training set and the validation set are very close, indicating that the Bi - TCN has strong generalization ability and no overfitting; Since the short - term power prediction model is executed every 15 minutes to dynamically update the prediction result, the single - calculation time should be less than 5 minutes. The calculation time of the model of the present invention is 43 seconds, which is much less than the standard - required time interval;

[0116] Input the validation set data into the well-trained model to predict the future short-term power of the target unit and compare it with the actual power data. The prediction accuracies of the model on the training set and the test set are shown in Table 2. The mean absolute error and the root mean square error are used to measure the prediction error of the model. As the unit state deteriorates, the prediction accuracy of the model slightly decreases, but it always remains within the range of less than 30 kW. The short-term power prediction error without considering the unit state is between the attention state and the abnormal state.

[0117] Table 2 Prediction Errors of Bi-TCN Model under Different States

[0118]

[0119]

[0120] S3. Detect the current state of the unit, select the unit power prediction model corresponding to the state, and predict the power of the unit in the next 4 hours based on the short-term wind speed prediction value of the unit and the historical wind speed and power data of the previous day.

[0121] According to the requirements of the short-term power dispatch within a day in the wind farm, predict the power within the next 4 hours every 15 minutes in a rolling manner. Take Unit 22 as an example to carry out ultra-short-term power prediction. Detect the operating state of the unit at the current moment, and select the power prediction model under different training sets according to different states. Use the historical wind speed, historical power data, and future wind speed prediction value as inputs to predict the output power of the target unit in the short term. Among them, Table 3 shows the state detection results of Unit 22 during the T1 to T10 short-term power dispatch cycles before the generator rear bearing failure shutdown. Each time period is 4 hours long and contains 16 sampling points. There are a total of 160 sampling points from T1 to T10.

[0122] Table 3 States of Unit 22 during T1 to T10

[0123]

[0124]

[0125] Figure 6 In (a), it is the power prediction result of Unit 22 during the T1 to T10 dispatch cycles. The abscissa represents the sample number, and the ordinate represents the predicted power. Figure 6 In (b), the abscissa represents the measured power value, and the ordinate represents the predicted power value, indicating that the discrete points of the predicted power value and the measured value are evenly distributed around the regression line, verifying the stable effectiveness of the model output result.

[0126] Further, compare the predicted short-term power of the unit with the benchmark power of the target unit, analyze the short-term power fluctuation characteristics of the unit, and formulate a reasonable short-term power scheduling plan and maintenance strategy for the wind power operation enterprise; the process of calculating the short-term power fluctuation characteristics of the unit is as follows:

[0127] S1: Collect the historical wind speed power data of the benchmark unit to construct the average wind speed power curve of the benchmark unit, collect the current meteorological predicted wind speed (NWP wind speed) of the benchmark unit, find the benchmark power corresponding to the NWP wind speed according to the average wind speed power curve, and input the NWP wind speed and the benchmark power into the unit power prediction model under normal conditions to predict the power of the benchmark unit; it includes the following steps:

[0128] S11: Combine the bin analysis method and the cubic spline interpolation method to jointly construct the average wind speed power curve of the benchmark unit;

[0129] S111: Use the bin method to divide the wind speed interval, determine the wind speed center point corresponding to each interval, calculate the average wind speed and average power in each interval, and then use the cubic spline interpolation method to fit the points in each interval into an average power curve;

[0130] S112: The calculation formulas for the average wind speed and average power of the benchmark unit during the statistical period are as follows:

[0131]

[0132] In the formula, V i is the standardized average wind speed of the i-th wind speed interval; V i,j is the standardized wind speed of data j in the i-th wind speed interval; is the standardized average output power of the i-th wind speed interval; P i,j is the standardized average output power of wind speed j in the i-th wind speed interval; N i is the number of wind speed power data in the i-th wind speed interval;

[0133] S113: The average wind speed and average output power of each interval can be calculated using the formula in S412. For these interval points, the cubic spline interpolation method is used to model the power curve, and the average wind speed power curve is fitted as f0;

[0134] The complete B-spline curve B(t) of order (p + 1) can be expressed as:

[0135]

[0136] b j represents the control points affecting the shape of the power curve, which are taken as the center points of each wind speed interval here; n c represents the number of wind speed center points; N j,p(t) is the basis function of the (p + 1)-th order B-spline. When , its value is 1; when p = 0 and t ∈ [t j , t j+1 ), its value is 0; when p ≠ 0, the basis function can be expressed as:

[0137]

[0138] where t j is the j-th internal node of the p-th order B-spline curve B(t); the equation of the third-order B-spline curve is:

[0139]

[0140] f0(t) represents the average wind speed power at time t; N j,3 (t)b represents the basis function of the third-order B-spline at time t; b j represents the wind speed center point of the j-th interval; q represents the index of the node interval where the parameter t is located, which is used to locate the basis function and control points required for the current wind speed interval;

[0141] S12: Construct the average wind speed power curve. The output power of the reference unit at time t is expressed as:

[0142] P t-theory = f0[V H (t)]

[0143] where f0 is the average wind speed power curve, and V H (t) is the meteorological predicted wind speed of the reference unit;

[0144] S13: Collect the meteorological predicted wind speed of the current reference unit, and determine the output power as the reference power according to the average wind speed power curve;

[0145] S14: Input the meteorological predicted wind speed and the reference power into the unit power prediction model under normal conditions to obtain the predicted power of the reference unit;

[0146] S2: Compare the short-term power prediction value of the target unit with the power of the reference unit to analyze the short-term power fluctuation characteristics of the unit;

[0147] The power fluctuation index mainly measures the relative deviation between the output power of the wind turbine and the theoretical power in the short term in the future. Since it is difficult to obtain the theoretical power curve of the wind turbine, the present invention uses the power of the reference unit in the same period as the reference power, and measures the power fluctuation characteristics by the deviation between the output power of the target unit and the power of the reference unit; the calculation process of the power fluctuation is:

[0148]

[0149] In the formula, P t is the predicted value of the wind turbine power at time t; P t-theory is the predicted value of the reference machine power corresponding to the numerically weather-predicted wind speed at time t; P o is the rated power of the wind turbine; n is the number of power sample points within the period to be predicted;

[0150] S3: Formulate a short-term power scheduling plan and maintenance strategy according to the short-term power fluctuation characteristics of the unit.

[0151] Furthermore, in order to verify the effect of the power prediction method of the present invention, the SVM model, ARIMA model, LSTM model, TCN model, Bi-TCN model, unit status-TCN model are compared with the unit power prediction model (unit status-Bi-TCN model) proposed by the present invention;

[0152] Compare the prediction accuracies of each model, and use the mean absolute error (MAE) and root mean square error (RMSE) as evaluation indicators;

[0153] Each model is independently tested 30 times. The time length of the input variable is 24 hours, and the time step of the output power is 4 hours. Figure 7 During the time periods T1 to T4, the status of the wind turbine did not change and was in a normal state. Considering that the prediction result of the model of the present invention almost overlaps with the true power value, where the abscissa represents the sample number and the ordinate represents the predicted power value (KW); At time T5, the operating status of the wind turbine changed. The prediction results of the models that did not consider the unit status deviated significantly from the true value, and the prediction error gradually increased. However, the prediction model considering the unit status could still track the change trend of the true power; During the time periods T5 to T7, the unit was in a state of attention. Although the status of the unit did not change during these three time periods, the status index value gradually decreased, indicating that the unit status gradually declined after time T5; The deviation between the prediction model that did not consider the unit operating status and the true power value gradually increased. The prediction model considering the unit operating status still had a good prediction effect at time T7; During the time periods T8 to T10, the unit was in an abnormal state. The predicted values of the single benchmark models all deviated from the true value. The prediction result of Bi-TCN was slightly improved, but since the model was directly trained without considering the unit status, it could not accurately predict the short-term power under abnormal conditions.

[0154] Figure 8 is Figure 7Comparison of local prediction results during the time period from T4 to T5. When the unit changes from the normal state to the attention state, the prediction deviations of the ARIMA and SVM models increase significantly; the LSTM and TCN models can handle the long-term and short-term dependencies of time series, and their prediction performance is slightly better than that of the ARIMA and SVM models; the Bi-TCN model without considering the unit state has improved prediction deviation under the action of the flipping layer, but the prediction effect is still poor between the 57th and 68th sample points; while the model of the present invention can accurately predict the power change trend and has the best prediction effect in the attention state;

[0155] Table 4 shows the state detection results of the unit during 10 short-term power scheduling cycles from T1 to T10 before the abnormal oil temperature of the gearbox of Unit 22.

[0156] Table 4 Operating state of Unit 22 during the period from T1 to T10

[0157]

[0158]

[0159] In this specification, each embodiment is described in a progressive manner. The key points of each embodiment are the differences from other embodiments. The same or similar parts among the embodiments can be referred to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0160] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but will be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A short-term power prediction method for wind turbines considering wind speed and operating status, characterized in that, It includes the following steps: Step 1: Collect the historical operation data of the wind turbine under different operation states, and construct the training sets under different operation states; Step 2: Construct a unit power prediction model based on a bidirectional temporal convolutional network, and use the training sets to train the unit power prediction models under different operation states; Step 3: Detect the current state of the target unit, and select the trained unit power prediction model corresponding to the operation state according to the current state; Collect the short-term wind speed prediction values of the target unit in the future and the historical wind speed-power data of the previous day, and input them into the selected unit power prediction model to obtain the short-term power prediction value of the target unit.

2. A short-term power prediction method for a wind turbine considering wind speed and operating status according to claim 1, characterized in that, The process of constructing the training set in Step 1 includes: Step 11: Collect the historical operation data of multiple units under different operation states, including historical SCADA monitoring data and historical state data, and perform preprocessing; Step 12: Use the abnormal state detection method to identify the preprocessed historical state data of each unit, and generate historical state sequences in chronological order for the historical state data under the same identified operation state; Step 13: Concatenate the historical state sequences under different operation states and the corresponding preprocessed historical SCADA monitoring data to generate the training sets under different operation states.

3. A short-term power prediction method for a wind turbine considering wind speed and operating status according to claim 1, characterized in that The unit power prediction model based on the bidirectional temporal convolutional network includes an input layer, a first residual block, a second residual block, a fusion layer, a fully connected layer, and an output layer; The training sets are input into the input layer. The input layer constructs a new sample set containing multi-time scale parameters based on the training sets according to a sliding time window, and inputs the new sample set into the first residual block, the second residual block, and the fusion layer respectively; The outputs of the first residual block and the second residual block are transmitted to the fusion layer. The fusion layer concatenates and fuses the output of the first residual block, the output of the second residual block, and the new sample set to obtain the fusion features; The fully connected layer calculates the short-term power prediction value according to the fusion features; The output layer outputs the short-term power prediction value.

4. A short-term power prediction method for a wind turbine considering wind speed and operating status according to claim 3, characterized in that, The first residual block includes a first dilated convolutional layer, a batch normalization layer, a first dropout layer, a second dilated convolutional layer, a first input sequence normalization layer, a second dropout layer, a third dilated convolutional layer, a second input sequence normalization layer, and a third dropout layer connected in sequence; The second residual block includes a first flipping layer, a first dilated convolutional layer, a batch normalization layer, a first dropout layer, a second dilated convolutional layer, a first input sequence normalization layer, a second dropout layer, a third dilated convolutional layer, a second input sequence normalization layer, a third dropout layer, and a second flipping layer connected in sequence.

5. A short-term power prediction method for a wind turbine considering wind speed and operating status according to claim 4, characterized in that The filters in the first dilated convolutional layer, the second dilated convolutional layer, and the third dilated convolutional layer are represented as \(f:\{0,\ldots,k s - 1\}\to\mathbb{R}\), and one convolution operation corresponds to one jump operation on the input sequence. The expression for performing the convolution operation is: Among them, F(x) is the result of performing dilated convolution on the element x in the input X; k s represents the filter size; x - d*i is an indicator of the convolution calculation direction; f(i f ) represents the i-th filter; d represents the dilation factor; * d represents the dilated convolution operation.

6. A short-term power prediction method for a wind turbine considering wind speed and operating status according to claim 3, characterized in that The fully connected layer is expressed as: o = Activation[x + F(x)] where, o represents the output of the fully connected layer, F(x) represents applying the residual mapping learned by the dilated convolutional layer, and the weight normalization of the batch normalization layer and the input sequence normalization layer to the convolutional filters within the first residual block or the second residual block; Activation represents the activation function of the fusion layer, and the ReLU function is adopted.

7. A short-term power prediction method for a wind turbine considering wind speed and operating conditions according to claim 2, characterized in that Taking the power within t - 24 to t hours and the wind speed within t - 24 to t + 4 hours under different operating states as model inputs, and the unit power at t + 4 hours as the model output, a unit power prediction model is trained; the unit power prediction model is expressed as: Among them, represents the predicted power output at each iteration during the training process from time t+1 to time t+τ, where τ = 4; y[t-t1:t] represents the historical power sequence from the current time t to t1 time instants before time t, where t1 = 24; x1[t-t1:t] represents the historical wind speed sequence from the current time t to t1 time instants before time t; x2[t+1:t+τ] represents the future wind speed sequence from time t+1 to time t+τ; θ represents the model parameters; f i represents the unit power prediction model in state i; τ = w·r, where r is the step size of the sliding time window in the input layer (r = 1,..., s), and for the initial window, r is equal to 0, and w represents the window width of the sliding time window in the input layer; the historical state sequence includes the historical power sequence and the historical wind speed sequence; The short-term power prediction value of the output layer is obtained by calculating the average value of the predicted power of all iterative outputs Expressed as: Among them, represents the predicted power at time t; s represents the s-th time window.

8. A short-term power prediction method for a wind turbine considering wind speed and operating status according to claim 1, characterized in that During the training process of the unit power prediction model, the AdaBelief optimizer is used to optimize the model parameters of the Bi-TCN network; the expression of the AdaBelief optimizer is: m t = β1m t-1 + (1 - β1)g t s t = β1ν t-1 + (1 - β2)(g t - m t ) 2 Among them, f t (θ t ) represents the minimum loss function optimized at the t-th step; θ t represents the parameter of the loss function at time t; g t represents the minimum loss function f t (θ t-1 ) with respect to θ t gradient; m t represents the predicted value of the gradient at time t; s t and ν t respectively represent (g t - m t ) 2 and g t 2 exponential moving average; is the partial derivative symbol; β1 and β2 are smoothing parameters.

Images