Short-term wind power prediction method considering cold wave meteorological condition

Abstract

The invention discloses a short-term wind power prediction method considering a cold-wave meteorological condition, and belongs to the technical field of short-term wind power prediction, and the prediction method comprises the following specific steps: I, selecting historical wind power data and a meteorological data time sequence of a wind power plant, and preprocessing the inputted wind power plant meteorological data and power data; iI, performing temperature change trend prediction by using a Prophet algorithm, and extracting a cold-wave weather occurrence time period in combination with cold-wave early warning; iII, expanding the sample data in the cold-wave weather period through a DoppelGANger network, and setting network parameters; according to the method, the fidelity of the generated model can be improved, the accuracy of multivariate time sequence prediction is improved, the adaptability of the model in a complex meteorological environment is improved, and the adaptability of the model in the complex meteorological environment is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of short-term wind power prediction, and in particular to a short-term wind power prediction method considering cold wave meteorological conditions. Background Art

[0002] In the context of developing a new power system with new energy as the main supplier, expanding the proportion of clean energy consumption represented by wind power and building a low-carbon, safe and efficient energy system are one of the important measures to achieve my country's "dual carbon" strategic goals. Wind power prediction technology can predict the timing of wind power in the future, effectively reduce the uncertainty brought by wind power generation to the power system, optimize the dispatching level of the power system, reduce the cost of grid operation and control, and improve the absorption capacity of renewable energy. In recent years, affected by climate change with global warming as the main feature, the occurrence of extreme and abnormal weather has shown a significant increase. Among them, cold wave weather represented by strong winds and freezing rain has the characteristics of low probability, high risk and great harm, and has the most significant impact on wind turbines. In the cold wave meteorological scenario, wind power output shows fluctuations of different characteristics on a short-term time scale, which may cause instantaneous power mutations, climbing and "cliff jumping" and other phenomena, seriously affecting the accuracy of wind power prediction.

[0003] At present, wind power forecasting has formed a relatively mature technical system, especially for ultra-short-term and short-term time scale forecasting, its forecasting model is more diverse and the forecasting accuracy is relatively higher. However, at this stage, there is little technical research on wind power forecasting under cold wave weather. And due to the lack of accumulation of historical low-temperature shutdown meteorological data, it is extremely difficult to accurately predict the low-temperature shutdown time of each wind turbine. Therefore, the power forecast of new energy stations currently does not consider the impact of low-temperature shutdown, resulting in significant deviations in the power forecast curve once a cold wave occurs, and the forecasting accuracy is low, which has a great impact on power supply. Therefore, improving the accuracy of wind power forecasting under cold wave meteorological conditions and conducting research on short-term accurate wind power forecasting technology under small sample conditions of cold wave weather are of great practical significance for the reasonable issuance of power system safety warnings and control instructions.

[0004] The problem of short-term wind power prediction under extreme weather scenarios can be divided into the following two research ideas. On the one hand, starting from the mechanism of wind power mutation, combined with data statistical theory, the prediction method of wind power mutation events is studied, and then a short-term power prediction strategy is formulated. Under extreme weather conditions, the output power of wind farms is greatly affected, because the wind turbine is equipped with a protection system. When the external ambient temperature or wind speed reaches a certain critical value, the protection mechanism will be triggered, causing the power to drop sharply or even disconnect from the grid, which is manifested as a significant decrease on the power curve. In order to solve this problem, many prediction studies have been conducted on power ramp events. However, the wind power loss caused by extreme weather scenarios is not exactly the same as the wind power ramp event. For example, cold wave weather is a continuous process. The current wind turbine electrical control system usually takes additional measures to ensure that the wind turbine can still operate in an extremely low temperature environment to avoid complete shutdown. This situation means that the wind power output continues to exist at a relatively low level for a period of time. However, in contrast, wind power ramping is manifested as a steep rise and fall in wind power in a short period of time. Therefore, under extreme weather conditions, the simple wind power ramp time prediction method is not completely applicable. On the other hand, the fluctuation laws of the strongly correlated factors of the weather system (wind speed, temperature, precipitation, etc.) and the wind power time series on a short-term time scale are studied, and the correlation and matching characteristics between the wind power fluctuation process and the meteorological data fluctuation process are further explored. The modeling method based on the fluctuation process is a new prediction technology developed under the development trend of big data in recent years. It mainly utilizes the potential persistence law of wind power series on short-term and ultra-short-term scales, and combines and divides the series by analyzing and mining the evolution law of similar historical fluctuation states, so as to realize the judgment of the processing state of the near future period. In summary, it is urgent to propose a prediction method for extreme meteorological scenarios. This method needs to accurately identify the period of extreme meteorological occurrence and conduct detailed research on this period, so as to improve the adaptability of the prediction model in extreme scenarios.

[0005] Small sample learning has become a hot topic in the field of artificial intelligence research in recent years. In recent years, the research on small sample learning has proposed corresponding solutions based on the two perspectives of model and data. From the model level, a method based on transfer learning has been proposed, which can quickly learn new concepts or skills even with a small amount of sample data; from the data level, a data enhancement method is adopted, that is, a large number of samples are generated using a small number of original samples, which specifically includes data generation and feature enhancement. Although the above methods solve the problem of sample expansion of sequence data to a certain extent, they do not capture the complex associations between temporal features and different sequences, as well as the long-term correlations within a single time series. Summary of the invention

[0006] The purpose of the present invention is to solve the defects in the prior art and to propose a short-term wind power prediction method considering cold wave meteorological conditions.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] A short-term wind power prediction method considering cold wave meteorological conditions, the specific steps of the prediction method are as follows:

[0009] Ⅰ. Select the historical wind power data and meteorological data time series of the wind farm power station, and pre-process the input wind farm meteorological data and power data;

[0010] Ⅱ. Use the Prophet algorithm to predict the temperature change trend, combine it with the cold wave warning, and extract the period when the cold wave weather will occur;

[0011] III. Expand the sample data of cold wave weather period through DoppelGANger network and set the network parameters;

[0012] IV. Decouple the meteorological data and power data to generate subsequences using the MLP generator, feed the meteorological data to the time series generator at each time step, and preprocess the power data based on the generated meteorological data;

[0013] V. Normalize each data feature and use the maximum and minimum values ​​of each time series as a random variable to generate samples;

[0014] VI. Take the generated subsequences and pseudo attributes as input and generate the target sequence through RNN batch processing;

[0015] VII. Establish a short-term wind power forecasting model based on the SF-TSN-TFT improved neural network, make separate forecasts for normal weather and cold wave weather periods, and combine the forecast results to form a preliminary forecast result;

[0016] VIII. Through multiple error correction strategies, different types of residuals are predicted and multiple corrections are completed, and multiple iterative corrections are performed. When the residuals are unpredictable or the maximum number of iterations is reached, the iterative process of error correction is terminated;

[0017] IX. Add the preliminary prediction results to the corrected residual sequence to obtain the final wind power prediction results.

[0018] As a further solution of the present invention, the specific steps of pre-processing the input wind farm meteorological data and power data in step I are as follows:

[0019] S1.1: The change point grouping algorithm is used to identify the shutdown data and some scattered abnormal data in the wind power curve data set. Then, the quartile algorithm is used to divide a set of data into 4 parts in ascending order, each of which accounts for 25% of the entire data set. The lower quartile is recorded as Q1, the median is recorded as Q2, and the upper quartile is recorded as Q3;

[0020] S1.2: Calculate the difference between the upper quartile and the lower quartile and record it as the interquartile range I QR , according to Q1-1.5I QR and Q3+1.5I QR Get the inner limit of outliers [F L , F U ] and remove the data beyond this range;

[0021] S1.3: Use the three-moment interpolation technique to construct a cubic spline interpolation function, and use the cubic spline interpolation method to fill in missing values ​​in the data. When the missing values ​​are less than 20%, the mean or median can be used to fill in the continuous variables, and the categorical variables are not filled in or are treated as a single category, or the mode is used to fill in. When the missing values ​​are between 20% and 80%, the mean or median can be used to fill in the continuous variables, and an indicator dummy variable is created for each variable with missing values. When the missing values ​​are greater than 80%, an indicator dummy variable is created for each variable with missing values ​​and is used in subsequent modeling. At the same time, the original variable is no longer used.

[0022] As a further solution of the present invention, the specific steps of using the Prophet algorithm to predict the temperature change trend in step II, combined with the cold wave warning, to extract the cold wave weather occurrence period are as follows:

[0023] S2.1: Establish trend functions, period functions and special node accidental factor functions respectively, and decompose the input historical time series through the established groups of functions, decomposing the series into four parts: trend term, period term, special time node accidental factor term and random fluctuation term;

[0024] S2.2: Model the trend items, cycle items, accidental factors items at special time nodes and random fluctuation items obtained by decomposition, establish an additive regression model, and predict the temperature change trend;

[0025] S2.3: Based on the rate change amplitude of the trend change point and combined with the time of occurrence of historical cold wave events, the rate change amplitude is defined as α. When the condition |α|≥10 is met, it indicates that the temperature has undergone a short-term strong change, and the complete cold wave weather period in which the short-term strong change occurred is extracted.

[0026] As a further solution of the present invention, the specific steps of expanding the sample data of the cold wave weather period by the DoppelGANger network in step III are as follows:

[0027] S3.1: The generative model that captures the data distribution is used as the generator G, and the discriminative model that estimates the probability of the training data samples is used as the discriminator D. The GAN model is constructed based on the generator G and the discriminator D;

[0028] S3.2: The random noise data z is input into the generator G, and the corresponding wind power data G(z) is generated. The real power data x and the generated data G(z) are input into the discriminator for discrimination. Then the discrimination result of the discriminator is fed back to the generator and the discriminator respectively. The two update their own parameters according to the feedback result L(G, D) for optimization. The feedback result L(G, D) is the loss parameter of the model, and its specific expression is as follows:

[0029]

[0030] In the formula, as well as Respectively represent the expected values ​​of the distribution of the true power data x and the noise data z; G(·) and D(·) represent the data output by the generator and the discriminator respectively;

[0031] S3.3: When training a generative adversarial network, the generator and the discriminator are connected into a network, and the hidden relationships between different sequences and the temporal dependencies within the sequences are comprehensively considered through the DoppelGANger network. At the same time, each group of data is normalized separately. The generator and the discriminator are continuously and alternately updated based on the processed data until the training is completed.

[0032] As a further solution of the present invention, the specific steps of using the MLP generator to decouple the meteorological data and the power data to generate subsequences in step IV are as follows:

[0033] S4.1: Take the bottom layer of the MLP generator as the input layer, input multiple n-dimensional vectors, then fully connect the hidden layer with the input layer, input the input layer vector X into the hidden layer, and obtain the output representation f(W1X+b1) of the hidden layer, where W1 represents the weight, b1 represents the bias, and function f represents the tanh function tanh(a)=(e a -e -a ) / (e a +e -a );

[0034] S4.2: After the output layer receives the output of the hidden layer, it generates the output softmax(W2X1+b2) of that output layer through softmax regression to obtain each group of subsequences after the meteorological data and power data are decoupled, where X1 represents the output f(W1X+b1) of the hidden layer.

[0035] As a further solution of the present invention, the short-term wind power prediction model in step VII specifically includes a TFT network model, a multi-head attention mechanism and a spatiotemporal normalization module; The TFT network model specifically includes a gating mechanism, a feature selection network, a static covariate encoder and a timing processing module; As a further solution of the present invention, the specific calculation formula of the gating mechanism is as follows: In the formula, and Both represent the middle layer; represents the normalization of the standard layer; An indicator representing weight sharing; represents the exponential linear unit activation function, when hour, The activation function will be used as the identity function when hour, The activation function produces a constant output and performs linear processing; represents the gated linear unit, and its specific calculation formula is as follows: In the formula, represents the activation function; and Represent weight and bias respectively; represents the Hadamard product of the elements; The feature selection network selects relevant input variables at each time step. Different feature inputs use separate variable selection networks with different weights. There are three types of input variables, namely static features, dynamic time-varying features, and dynamic time-invariant features. Assume Representative Features in The conversion input at the moment is: In the formula, represent The flattened vector of all historical input feature vectors at the moment; the feature selection weights are provided to the and external context vectors through the GRN and Softmax layers respectively. The weights are calculated as follows: In the formula, Represents a feature selection weight vector; Represents the context vector obtained from the static covariate encoder. For static variables, since it has access to static information, is omitted; at each time step, by giving each Enter your own GRN to use an additional nonlinear processing layer: In the formula, Represents the feature vector after the variable is processed; at all times The weights are shared in the feature vector, and then weighted according to the variable selection weights of the processed features. The results after weighted calculation are as follows: In the formula, Represents the feature vector No. elements; The static covariate encoder integrates static features into the network to regulate temporal dynamics by encoding the context vector; The time series processing module learns long-term and short-term temporal relationships from historical and measured time series variable inputs; The multi-head attention mechanism multiplies the importance coefficients of each data and then adds them together. The calculation of the full attention mechanism adopts the request-primary key-value mode, where the attention mechanism is established as follows: In the formula, represents the normalization function. Based on the current attention mechanism, different heads are used for different representation subspaces through the multi-head attention mechanism. Its specific expression is as follows: In the formula, , as well as Respectively represent , and The weight of a specific head; Represents a linear combination of all heads The series output of The multi-head attention mechanism is adjusted to achieve the sharing of the values ​​of each head, and the additive aggregation of all heads is adopted. The specific calculation formula of additive aggregation is as follows: In the formula, Represents the weight value shared by all heads; For the final linear mapping; The spatiotemporal normalization module specifically includes a time normalization module and a space normalization module, which are used to refine high-frequency and local components respectively; The time normalization module is used to refine the high-frequency component from the input mixed signal, and the high-frequency component and the low-frequency component are respectively expressed by the following expressions: In the formula, and Respectively represent The time series in High-frequency and low-frequency components at each moment; and Respectively represent The time series in The local high-frequency and low-frequency components at the moment; and Respectively represent The global high-frequency component at the moment; If the rate of change of the low-frequency component is much smaller than that of the high-frequency component, that is, each low-frequency component is approximately equal to a constant over a period of time, then the high-frequency component can be expressed as: In the formula, Represents a relatively small constant that maintains numerical stability; is observable; and Respectively represent The low frequency influence on the time series The mean and standard deviation of and Represent a pair of vectors that can be learned; The spatial normalization module is used to refine the local component composed of the local high-frequency component and the local low-frequency component, and the local and global components are summarized and expressed by the following expressions: In the formula, and Respectively represent The global and local components at the moment; Substituting the estimated values ​​of the four unobservable variables into the formula, the local component is expressed as .

[0076] As a further solution of the present invention, the specific steps of predicting different types of residuals and completing multiple corrections by using multiple error correction strategies in step VIII are as follows:

[0077] S5.1: Assume that the prediction residual and prediction result of the vth training data set in the cth revision are and There are M training sets in total. The prediction residuals and prediction results of the test data in the cth correction are and

[0078] S5.2: According to the different main error sources of different error types, explore the main influencing factors of different errors, and conduct correlation tests between error sequences and meteorological variables. The specific calculation formula for the correlation test is as follows:

[0079]

[0080] In the formula, x i Represents the sequence attribute, that is, the relevant influencing factor; y i represents the target prediction sequence, i.e., the error sequence;

[0081] S5.3: Detect the predictability of the residual. If it is predictable, the ARIMA model is trained using the residual. If it is unpredictable, the error correction process is stopped and the trained ARIMA model is used for prediction to obtain the corrected prediction results of the training set and the test set.

[0082] S5.4: In the new round of correction, the correction shrinkage rate r is introduced to control the progress of error correction, and in the iterative correction process, the formula Calculate the residual. When the residual is unpredictable or reaches the maximum number of iterations, terminate the iterative process of error correction and output the prediction results after multiple corrections.

[0083]

[0084] Compared with the prior art, the present invention has the following beneficial effects:

[0085] This short-term wind power prediction method considering cold wave meteorological conditions expands the scarce samples in the cold wave period by using the sequence data sample expansion method based on DoppelGANger, which can improve the fidelity of the generated model and improve it based on the traditional GAN ​​network. Then, a short-term wind power prediction method based on the improved SF-TSN-TFT is proposed, which makes full use of the strong feature extraction ability and interpretability of the TFT model, and fully mines the static data such as wind power and meteorological factors under historical cold wave weather. On this basis, the spatiotemporal normalization module is introduced to fully extract the hidden dynamic characteristics of meteorological factors, thereby improving the accuracy of multivariate time series prediction, further capturing the complex dynamic characteristics of influencing factors in different time periods, constructing the SF-TSN-TFT prediction model, and realizing "end-to-end" prediction, thereby improving the adaptability of the model in complex meteorological environments. Finally, multiple error corrections are performed on different fluctuation processes of the error sequence to further improve the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0087] Figure 1 A flowchart of a short-term wind power prediction method considering cold wave meteorological conditions proposed by the present invention;

[0088] Figure 2 A schematic diagram of the logic framework of a short-term wind power prediction method considering cold wave meteorological conditions proposed by the present invention;

[0089] Figure 3 This is a DoppelGANger principle architecture diagram of a short-term wind power prediction method considering cold wave meteorological conditions proposed by the present invention;

[0090] Figure 4 This is a flow chart of the prediction algorithm of the improved SF-TSN-TFT model of a short-term wind power prediction method considering cold wave meteorological conditions proposed by the present invention;

[0091] Figure 5 This is a flow chart of a multiple error correction method based on different error types for a short-term wind power forecasting method considering cold wave meteorological conditions proposed by the present invention. DETAILED DESCRIPTION

[0092] Reference Figure 1-5 A short-term wind power prediction method considering cold wave meteorological conditions is proposed. The specific steps of the prediction method are as follows:

[0093] The historical wind power data and meteorological data time series of the wind farm power station are selected, and the input wind farm meteorological data and power data are preprocessed.

[0094] Specifically, the change point grouping algorithm is used to identify the shutdown data and some scattered abnormal data in the wind power curve data set. Then, the quartile algorithm is used to divide a set of data into 4 parts in ascending order, each accounting for 25% of the entire data set. The lower quartile is recorded as Q1, the median is recorded as Q2, and the upper quartile is recorded as Q3. The difference between the upper quartile and the lower quartile is recorded as the interquartile range I QR , according to Q1-1.5I QR and Q3+1.5I QR Get the inner limit of outliers [F L , F U], and the data beyond this range are eliminated, the cubic spline interpolation function is constructed using the three-moment interpolation technique, and the missing values ​​in the data are filled by the cubic spline interpolation method. When the missing values ​​are less than 20%, the mean or median can be used to fill in the continuous variables, and the categorical variables are not filled or are treated as a single category, or the mode is used to fill in. When the missing values ​​are between 20% and 80%, the mean or median can be used to fill in the continuous variables, and an indicator dummy variable is created for each variable with missing values. When the missing values ​​are greater than 80%, an indicator dummy variable is created for each variable with missing values ​​and is used in subsequent modeling, and the original variables are no longer used.

[0095] The Prophet algorithm is used to predict temperature change trends, combined with cold wave warnings, to extract the time period when cold wave weather will occur.

[0096] Specifically, trend functions, periodic functions and special node accidental factor functions are established respectively, and the input historical time series is decomposed through the established groups of functions, and the series is decomposed into four parts: trend term, periodic term, special time node accidental factor term and random fluctuation term. The trend term, periodic term, special time node accidental factor term and random fluctuation term obtained by decomposition are modeled respectively, and an additive regression model is established to predict the temperature change trend. According to the rate change amplitude of the trend change point and combined with the time of occurrence of historical cold wave events, the rate change amplitude is defined as α. When the condition |α|≥10 is met, it indicates that the temperature has undergone a short-term strong change, and the complete cold wave weather period when the short-term strong change occurs is extracted.

[0097] The DoppelGANger network is used to expand the sample data during the cold wave weather period and set the network parameters.

[0098] Specifically, refer to Figure 2 as well as Figure 3 It can be seen that the generative model that captures the data distribution is used as the generator G, and the discriminative model that estimates the probability from the training data samples is used as the discriminator D. The GAN model is constructed based on the generator G and the discriminator D. The random noise data z is input into the generator G, and the corresponding wind power data G(z) is generated. The real power data x and the generated data G(z) are input into the discriminator for discrimination. Then the discrimination results of the discriminator are fed back to the generator and the discriminator respectively. The two update their own parameters according to the feedback results L(G, D) for optimization. The feedback result L(G, D) is the loss parameter of the model. When training the generative adversarial network, the generator and the discriminator are connected into a network, and the hidden relationship between different sequences and the time dependency within the sequence are comprehensively considered through the DoppelGANger network. At the same time, each group of data is normalized separately. The generator and the discriminator are continuously and alternately updated based on the processed data until the training is completed.

[0099] It should be further explained that the specific expression of L(G,D) is as follows:

[0100]

[0101] In the formula, as well as They represent the expected values ​​of the distribution of the true power data x and the noise data z respectively; G(·) and D(·) represent the data output by the generator and the discriminator respectively.

[0102] The meteorological data and power data are decoupled to generate subsequences using an MLP generator. The meteorological data is fed to the time series generator at each time step, and the power data is preprocessed based on the generated meteorological data.

[0103] Specifically, the bottom layer of the MLP generator is used as the input layer, and multiple n-dimensional vectors are input. Then, the hidden layer is fully connected to the input layer, the input layer vector X is input into the hidden layer, and the output representation f(W1X+b1) of the hidden layer is obtained, where W1 represents the weight, b1 represents the bias, and the function f represents the tanh function tanh(a)=(e a -e -a ) / (e a +e -a ), after the output layer receives the output of the hidden layer, it generates the output softmax(W2X1+b2) of that output layer through softmax regression to obtain each group of subsequences after the meteorological data and power data are decoupled, where X1 represents the output f(W1X+b1) of the hidden layer.

[0104] Each data feature is normalized, and the maximum and minimum values ​​of each time series are used as a random variable to generate samples.

[0105] It should be further explained that the data is normalized by global min and max and stored as global constants, and then training is performed on the normalized data. For example, wind power is a measured value, denoted by f i , the meteorological influencing factors are metadata, denoted as A i (such as wind speed / direction, temperature, precipitation, etc.), the mode collapse problem can occur in both measured data and metadata, and the standard normalization method is optimized to normalize each time series sample separately and store min / max as "fake" metadata. i ,A i ) is to train Therefore, the GAN learns to generate these two pseudo metadata, defining the min / max of each time series respectively, and then rescaling the measurements to the actual range.

[0106] The generated subsequences and pseudo attributes are taken as input and the target sequence is generated through RNN batch processing.

[0107] It should be further explained that the basic structure of RNN includes input layer, hidden layer and output layer. Among them, the state of the hidden layer will be updated with the time step and serve as one of the inputs of the next time step.

[0108] In addition, it should be noted that the hidden state update formula of RNN is:

[0109]

[0110] Where: x t> is the input of the current time step, a <t>< / t> is the hidden state of the current time step, a <t-1>< / t-1> is the hidden state of the previous time step, W ax and W aa is the weight matrix, b a is the bias term and tanh() is the activation function.

[0111] The output formula of RNN is:

[0112]

[0113] Where: is the output of the current time step, W ya is the weight matrix, b y is the bias term;

[0114] Compared with the single sample generation of traditional generative adversarial networks, in order to reduce the number of RNN passes, a batch generation method is proposed, that is, in each pass of the RNN, S samples are generated (for example, historical wind power data of S consecutive sampling points with a data sampling point every 15 minutes), where S is an adjustable parameter, which effectively reduces the number of RNN passes.

[0115] A short-term wind power forecasting model based on the improved SF-TSN-TFT neural network is established to make separate forecasts for normal weather and cold wave weather periods, and the forecast results are combined to form preliminary forecast results.

[0116] It should be further explained that the reference Figure 4 It can be seen that the short-term wind power prediction model specifically includes the TFT network model, the multi-head attention mechanism, and the spatiotemporal normalization module; The TFT network model specifically includes a gating mechanism, a feature selection network, a static covariate encoder, and a temporal processing module; As a further solution of the present invention, the specific calculation formula of the gating mechanism is as follows: In the formula, and Both represent the middle layer; represents the normalization of the standard layer; An indicator representing weight sharing; represents the exponential linear unit activation function, when hour, The activation function will be used as the identity function when hour, The activation function produces a constant output and performs linear processing; represents the gated linear unit, and its specific calculation formula is as follows: In the formula, represents the activation function; and Represent weight and bias respectively; represents the Hadamard product of the elements; Feature selection network, selects relevant input variables at each time step, and different feature inputs use separate variable selection networks with different weights. There are three types of input variables, namely static features, dynamic time-varying features, and dynamic time-invariant features. Assume Representative Features in The conversion input at the moment is: In the formula, represent The flattened vector of all historical input feature vectors at the moment; the feature selection weights are provided to the and external context vectors through the GRN and Softmax layers respectively. The weights are calculated as follows: In the formula, Represents a feature selection weight vector; Represents the context vector obtained from the static covariate encoder. For static variables, since it has access to static information, is omitted; at each time step, by giving each Enter your own GRN to use an additional nonlinear processing layer: In the formula, Represents the feature vector after the variable is processed; at all times The weights are shared in the feature vector, and then weighted according to the variable selection weights of the processed features. The results after weighted calculation are as follows: In the formula, Represents the feature vector No. elements; The static covariate encoder incorporates static features into the network to regulate temporal dynamics by encoding the context vector; The time series processing module learns long-term and short-term temporal relationships from historical and measured time series variable inputs; The multi-head attention mechanism multiplies the importance coefficients of each data and then adds them together. The calculation of the full attention mechanism adopts the request-primary key-value mode. The expression for establishing the attention mechanism is as follows: In the formula, represents the normalization function. Based on the current attention mechanism, different heads are used for different representation subspaces through the multi-head attention mechanism. Its specific expression is as follows: In the formula, , as well as Respectively represent , and The weight of a specific head; Represents a linear combination of all heads The series output of The multi-head attention mechanism is adjusted to achieve the sharing of the values ​​of each head, and the additive aggregation of all heads is adopted. The specific calculation formula of additive aggregation is as follows: In the formula, Represents the weight value shared by all heads; For the final linear mapping; The spatiotemporal normalization module specifically includes a time normalization module and a space normalization module, which are used to refine high-frequency and local components respectively; Among them, the time normalization module is used to refine the high-frequency component from the input mixed signal, and the high-frequency component and the low-frequency component are expressed by the following expressions respectively: In the formula, and Respectively represent The time series in High-frequency and low-frequency components at each moment; and Respectively represent The time series in The local high-frequency and low-frequency components at the moment; and Respectively represent The global high-frequency component at the moment; If the rate of change of the low-frequency component is much smaller than that of the high-frequency component, that is, each low-frequency component is approximately equal to a constant over a period of time, then the high-frequency component can be expressed as: In the formula, Represents a relatively small constant that maintains numerical stability; is observable; and Respectively represent The low frequency influence on the time series The mean and standard deviation of and Represent a pair of vectors that can be learned; The spatial normalization module is used to refine the local component composed of the local high-frequency component and the local low-frequency component, and the local and global components are summarized and expressed by the following expressions: In the formula, and Respectively represent The global and local components at the moment; Substituting the estimated values ​​of the four unobservable variables into the formula, the local component is expressed as .

[0157] Through multiple error correction strategies, different types of residuals are predicted and multiple corrections are completed, and multiple iterative corrections are performed. When the residuals are unpredictable or the maximum number of iterations is reached, the iterative process of error correction is terminated.

[0158] Specifically, refer to Figure 5 It can be seen that, assuming that the prediction residual and prediction result of the vth training data set in the cth correction are and There are M training sets in total. The prediction residuals and prediction results of the test data in the cth correction are and According to the different main error sources of different error types, the main influencing factors of different errors are explored, and the correlation test between the error sequence and the meteorological variables is carried out to detect the predictable state of the residual. If it is predictable, the ARIMA model is trained using the residual. If it is unpredictable, the error correction process is stopped, and the ARIMA model obtained after training is used for prediction to obtain the corrected training set prediction results and test set prediction results. In the new round of correction, the correction shrinkage rate r is introduced to control the progress of error correction, and in the iterative correction process, the formula is used Calculate the residual. When the residual is unpredictable or reaches the maximum number of iterations, terminate the iterative process of error correction and output the prediction results after multiple corrections.

[0159]

[0160] In addition, it should be noted that the specific calculation formula for the correlation test is as follows:

[0161]

[0162] In the formula, x i Represents the sequence attribute, that is, the relevant influencing factor; y i Represents the target prediction sequence, that is, the error sequence.

[0163] The preliminary prediction results are added to the corrected residual sequence to obtain the final wind power prediction results.

Claims

1. A short-term wind power prediction method considering cold wave meteorological conditions, characterized in that: The specific steps of the prediction method are as follows: Ⅰ. Select the historical wind power data and meteorological data time series of the wind farm power station, and pre-process the input wind farm meteorological data and power data; Ⅱ. Use the Prophet algorithm to predict the temperature change trend, combine it with the cold wave warning, and extract the period when the cold wave weather will occur; III. Expand the sample data of cold wave weather period through DoppelGANger network and set the network parameters; IV. Decouple the meteorological data and power data to generate subsequences using the MLP generator, feed the meteorological data to the time series generator at each time step, and preprocess the power data based on the generated meteorological data; V. Normalize each data feature and use the maximum and minimum values ​​of each time series as a random variable to generate samples; VI. Take the generated subsequences and pseudo attributes as input and generate the target sequence through RNN batch processing; VII. Establish a short-term wind power forecasting model based on the SF-TSN-TFT improved neural network, make separate forecasts for normal weather and cold wave weather periods, and combine the forecast results to form a preliminary forecast result; VIII. Through multiple error correction strategies, different types of residuals are predicted and multiple corrections are completed, and multiple iterative corrections are performed. When the residuals are unpredictable or the maximum number of iterations is reached, the iterative process of error correction is terminated; IX. Add the preliminary prediction results to the corrected residual sequence to obtain the final wind power prediction results.

2. A short-term wind power prediction method considering cold wave meteorological conditions according to claim 1, characterized in that: The specific steps of preprocessing the input wind farm meteorological data and power data in step I are as follows: S1.1: The change point grouping algorithm is used to identify the shutdown data and some scattered abnormal data in the wind power curve data set. Then, the quartile algorithm is used to divide a set of data into 4 parts in ascending order, each of which accounts for 25% of the entire data set. The lower quartile is recorded as Q1, the median is recorded as Q2, and the upper quartile is recorded as Q3; S1.2: Calculate the difference between the upper quartile and the lower quartile and record it as the interquartile range I QR , according to Q1-1.5I QR and Q3+1.5I QR Get the inner limit of outliers [F L , F U ] and remove the data beyond this range; S1.3: Use the three-moment interpolation technique to construct a cubic spline interpolation function, and use the cubic spline interpolation method to fill in missing values ​​in the data. When the missing values ​​are less than 20%, the mean or median can be used to fill in the continuous variables, and the categorical variables are not filled in or are treated as a single category, or the mode is used to fill in. When the missing values ​​are between 20% and 80%, the mean or median can be used to fill in the continuous variables, and an indicator dummy variable is created for each variable with missing values. When the missing values ​​are greater than 80%, an indicator dummy variable is created for each variable with missing values ​​and is used in subsequent modeling. At the same time, the original variable is no longer used.

3. A short-term wind power prediction method considering cold wave meteorological conditions according to claim 2, characterized in that: The specific steps of using the Prophet algorithm to predict the temperature change trend and extracting the period of cold wave weather in step II are as follows: S2.1: Establish trend functions, period functions and special node accidental factor functions respectively, and decompose the input historical time series through the established groups of functions, decomposing the series into four parts: trend term, period term, special time node accidental factor term and random fluctuation term; S2.2: Model the trend items, cycle items, accidental factors items at special time nodes and random fluctuation items obtained by decomposition, establish an additive regression model, and predict the temperature change trend; S2.3: Based on the rate change amplitude of the trend change point and combined with the time of occurrence of historical cold wave events, the rate change amplitude is defined as α. When the condition |α|≥10 is met, it indicates that the temperature has undergone a short-term strong change, and the complete cold wave weather period in which the short-term strong change occurred is extracted.

4. The short-term wind power prediction method considering cold wave meteorological conditions according to claim 1 is characterized in that: The specific steps of expanding the sample data of the cold wave weather period through the DoppelGANger network described in step III are as follows: S3.1: The generative model that captures the data distribution is used as the generator G, and the discriminative model that estimates the probability of the training data samples is used as the discriminator D. The GAN model is constructed based on the generator G and the discriminator D; S3.2: The random noise data z is input into the generator G, and the corresponding wind power data G(z) is generated. The real power data x and the generated data G(z) are input into the discriminator for discrimination. Then the discrimination result of the discriminator is fed back to the generator and the discriminator respectively. The two update their own parameters according to the feedback result L(G, D) for optimization. The feedback result L(G, D) is the loss parameter of the model, and its specific expression is as follows: In the formula, as well as Respectively represent the expected values ​​of the distribution of the true power data x and the noise data z; G(·) and D(·) represent the data output by the generator and the discriminator respectively; S3.3: When training a generative adversarial network, the generator and the discriminator are connected into a network, and the hidden relationships between different sequences and the temporal dependencies within the sequences are comprehensively considered through the DoppelGANger network. At the same time, each group of data is normalized separately. The generator and the discriminator are continuously and alternately updated based on the processed data until the training is completed.

5. The short-term wind power prediction method considering cold wave meteorological conditions according to claim 1 is characterized in that: The specific steps of using the MLP generator to decouple the meteorological data and power data to generate subsequences in step IV are as follows: S4.1: Take the bottom layer of the MLP generator as the input layer, input multiple n-dimensional vectors, then fully connect the hidden layer with the input layer, input the input layer vector X into the hidden layer, and obtain the output representation f(W1X+b1) of the hidden layer, where W1 represents the weight, b1 represents the bias, and function f represents the tanh function tanh(a)=(e a -e -a ) / (e a +e -a ); S4.2: After the output layer receives the output of the hidden layer, it generates the output softmax(W2X1+b2) of that output layer through softmax regression to obtain each group of subsequences after the meteorological data and power data are decoupled, where X1 represents the output f(W1X+b1) of the hidden layer.

6. The short-term wind power prediction method considering cold wave meteorological conditions according to claim 1 is characterized in that: The short-term wind power prediction model described in step VII specifically includes a TFT network model, a multi-head attention mechanism, and a spatiotemporal normalization module.

7. The short-term wind power prediction method considering cold wave meteorological conditions according to claim 1 is characterized in that: The specific steps of predicting different types of residuals and completing multiple corrections through multiple error correction strategies described in step VIII are as follows: S5.1: Assume that the prediction residual and prediction result of the vth training data set in the cth revision are and There are M training sets in total. The prediction residuals and prediction results of the test data in the cth correction are and S5.2: According to the different main error sources of different error types, explore the main influencing factors of different errors, and conduct correlation tests between error sequences and meteorological variables. The specific calculation formula for the correlation test is as follows: In the formula, x i Represents the sequence attribute, that is, the relevant influencing factor; y i represents the target prediction sequence, i.e., the error sequence; S5.3: Detect the predictability of the residual. If it is predictable, the ARIMA model is trained using the residual. If it is unpredictable, the error correction process is stopped and the trained ARIMA model is used for prediction to obtain the corrected prediction results of the training set and the test set. S5.4: In the new round of correction, the correction shrinkage rate r is introduced to control the progress of error correction, and in the iterative correction process, the formula Calculate the residual. When the residual is unpredictable or reaches the maximum number of iterations, terminate the iterative process of error correction and output the prediction results after multiple corrections.

Images