A short-term wind power prediction method considering cold wave weather conditions

Abstract

The application discloses a kind of short-term wind power prediction methods under considering cold wave meteorological conditions, belong to short-term wind power prediction technical field, the specific steps of the prediction method are as follows: I, select wind farm power station historical wind power data and meteorological data time sequence, and pre-process input wind farm meteorological data and power data;II, temperature change trend is predicted using Prophet algorithm, combined with cold wave early warning, extract cold weather period of time;III, through DoppelGANger network, cold weather period sample data is expanded, and network parameter is set;The application can improve the fidelity of the generated model, improve the accuracy of multivariate time series prediction, improve the adaptability of the model in complex meteorological environment, improve the adaptability of the model in complex meteorological environment.

Description

Technical Field

[0001] This invention relates to the field of short-term wind power prediction technology, and in particular to a short-term wind power prediction method that takes into account cold wave meteorological conditions. Background Technology

[0002] Against the backdrop of developing a new power system with new energy sources as the main supply, expanding the proportion of clean energy consumption, represented by wind power, and building a low-carbon, safe, and efficient energy system are important measures to achieve my country's "dual-carbon" strategic goals. Wind power forecasting technology can predict the timing of wind power output in the future, effectively reducing the uncertainty brought by wind power generation to the power system, optimizing the power system dispatch level, reducing grid operation and control costs, and improving the absorption capacity of renewable energy. In recent years, affected by climate change characterized by global warming, the occurrence of extreme and abnormal weather has shown a significant increasing trend. Among them, cold waves, represented by strong winds and freezing rain, are characterized by low probability, high risk, and great harm, and have the most significant impact on wind turbine units. Under cold wave meteorological scenarios, wind power output exhibits fluctuations with different characteristics on a short-term time scale, which may cause instantaneous power mutations, ramp-up, and "cliff-like" phenomena, seriously affecting the accuracy of wind power forecasting.

[0003] Currently, wind power forecasting has formed a relatively mature technical system, especially for ultra-short-term and short-term timescale forecasts, where forecasting models are more diverse and prediction accuracy is relatively higher. However, there is currently limited research on wind power forecasting under cold wave weather conditions. Furthermore, due to the lack of accumulated historical meteorological data on low-temperature shutdowns, accurately predicting the shutdown time of each wind turbine under low-temperature conditions is extremely difficult. Therefore, current power forecasts for renewable energy power plants do not consider the impact of low-temperature shutdowns, leading to significant deviations in power forecast curves and low prediction accuracy during cold waves, which significantly impacts power supply security. Therefore, improving the accuracy of wind power forecasting under cold wave weather conditions and conducting research on short-term accurate wind power forecasting technology under small sample conditions during cold waves is of significant practical importance for power system safety early warning and the rational issuance of control commands.

[0004] For the problem of short-term wind power prediction under extreme weather scenarios, research approaches can be divided into two categories. One approach starts with the mechanism of sudden wind power changes, combining statistical data to study prediction methods for sudden wind power events, and then formulates short-term power prediction strategies. Under extreme weather conditions, the output power of wind farms is greatly affected because wind turbines have protection systems. When the external ambient temperature or wind speed reaches a certain critical value, the protection mechanism is triggered, leading to a sharp drop in power or even grid disconnection, which is reflected in a significant decrease on the power curve. To address this issue, many studies have been conducted on predicting power ramp-up events. However, the wind power loss caused by extreme weather scenarios is not entirely the same as wind power ramp-up events. For example, cold waves are a continuous process, and current wind turbine electrical control systems typically take additional measures to ensure that the turbines can continue to operate in extremely low temperatures, avoiding complete shutdown. This situation means that wind power output remains at a relatively low level for a period of time. However, in contrast, wind power ramp-up is characterized by a sharp increase and decrease in wind power within a short period. Therefore, simple wind power ramp-up time prediction methods are not entirely applicable under extreme weather conditions. On the other hand, this study investigates the strong correlation factors of weather systems (wind speed, temperature, precipitation, etc.) and the fluctuation patterns of wind power time series on short-term time scales, further exploring the correlation and matching characteristics between wind power fluctuation processes and meteorological data fluctuation processes. The fluctuation process-based modeling method is a novel prediction technique developed in recent years under the trend of big data development. It mainly utilizes the potential persistent patterns of wind power sequences on short-term and ultra-short-term scales, analyzes and mines the evolution patterns of similar historical fluctuation states, combines and divides the sequences, and thus realizes the judgment of the processing state in the near future. In summary, there is an urgent need to propose a prediction method for extreme weather scenarios. This method needs to accurately identify the periods of extreme weather occurrence and conduct refined research on these periods to improve the adaptability of the prediction model under extreme scenarios.

[0005] Few-shot learning has become a hot topic in artificial intelligence research in recent years. Research on few-shot learning has focused on two main aspects: models and data, proposing corresponding solutions. From the model level, transfer learning-based methods have been proposed, enabling the rapid learning of new concepts or skills using only a small amount of sample data. From the data level, data augmentation methods are employed, which generate a large number of samples from the original small sample set; specifically, this includes data generation and feature enhancement. While these methods have addressed the problem of expanding the sample size of sequence data to some extent, they fail to capture temporal features, complex relationships between different sequences, and long-term correlations within a single time series. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by proposing a short-term wind power prediction method that takes into account cold wave meteorological conditions.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

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

[0009] I. Select historical wind power data and meteorological data time series of wind farm power stations, and preprocess the input wind farm meteorological data and power data;

[0010] II. Use the Prophet algorithm to predict temperature change trends, and combine it with cold wave warnings to extract the time periods when cold waves occur;

[0011] III. Expand the sample data for cold wave periods using the DoppelGANger network and set the network parameters;

[0012] IV. Use an MLP generator to decouple meteorological data and power data to generate subsequences. Feed 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 random variables to generate samples;

[0014] VI. Using the generated subsequences and pseudo-attributes as input, batch process them using an RNN to generate the target sequence;

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

[0016] VIII. Through a multi-error correction strategy, different types of residuals are predicted and multiple corrections are performed. Multiple iterations of correction are carried out. When the residuals become unpredictable or the maximum number of iterations is reached, the error correction iteration process 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 aspect of the present invention, the specific steps for preprocessing the input wind farm meteorological data and power data in step I are as follows:

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

[0020] S1.2: Calculate the difference between the upper quartile and the lower quartile, denoted as the interquartile range I. QR According to Q1-1.5I QR and Q3+1.5I QR Obtain the inner limit of outliers [F] L F U ], and remove data that is outside the range;

[0021] S1.3: Construct a cubic spline interpolation function using the three-moment interpolation technique, and fill in missing values ​​in the data using the cubic spline interpolation method. When the missing values ​​are less than 20%, the mean or median can be used to fill in the missing values ​​for continuous variables, while categorical variables are not filled in or are treated as a single category, or the mode is used to fill in the missing values. When the missing values ​​are between 20% and 80%, the mean or median can be used to fill in the missing values ​​for 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 participates in subsequent modeling, while the original variables are no longer used.

[0022] As a further aspect of the present invention, the specific steps for using the Prophet algorithm to predict temperature change trends and combining it with cold wave warnings to extract the time periods of cold wave weather in step II are as follows:

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

[0024] S2.2: Model the trend term, periodic term, accidental factor term at special time nodes, and random fluctuation term obtained from the decomposition, respectively, establish an additive regression model, and predict the temperature change trend.

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

[0026] As a further aspect of the present invention, the specific steps for augmenting the sample data of the cold wave period using the DoppelGANger network in step III are as follows:

[0027] S3.1: Use the generative model that captures the data distribution as the generator G, and the discriminative model that estimates the probability from the training data samples as the discriminator D. Construct a GAN model based on the generator G and the discriminator D.

[0028] S3.2: Input random noise data z into generator G to generate corresponding wind power data G(z). Input the real power data x and the generated data G(z) together into discriminator for discrimination. Then, the discrimination result of discriminator is fed back to generator and discriminator respectively. Both update their own parameters for optimization based on the feedback result L(G,D), where 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 G(·) and D(·) represent the expected values ​​of the distributions 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.

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

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

[0033] S4.1: The bottom layer of the MLP generator is used as the input layer, with multiple n-dimensional vectors as input. Then, the hidden layer is fully connected to the input layer. The input vector X is input into the hidden layer, and the output representation f(W1X+b1) is obtained, where W1 represents the weights, b1 represents the bias, and the function f represents the tanh function tanh(a) = (e^(-b1 / a)). a -e -a ) / (e a +e -a ), where a = W1X + b1;

[0034] S4.2: After receiving the output from the hidden layer, the output layer generates a softmax(W2X1+b2) for that output layer through softmax regression, where W2 represents the weight of the output layer and b2 represents the bias of the output layer, in order to obtain the subsequences of each group after the decoupling of meteorological data and power data, where X1 represents the output of the hidden layer f(W1X+b1).

[0035] As a further aspect 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.

[0036] The TFT network model specifically includes a gating mechanism, a feature selection network, a static covariate encoder, and a timing processing module.

[0037] As a further aspect of the present invention, the specific calculation formula for the gating mechanism is as follows:

[0038] GRN ω (a,c)=LayerNorm(a+GLU ω (η1))

[0039] η1=W 1,ω η2+b 1,ω

[0040] η2=ELU(W 2,ω a+W 3,ω a+b 2,ω )

[0041] In the formula, η1 and η2 both represent intermediate layers; LayerNorm represents the normalization of the standard layer; ω represents a weight sharing index; ELU represents the exponential linear unit activation function, when W 2,ω a+W 3,ω a+b 2,ω When >> 0, the ELU activation function will be used as the identity function, when W 2,ω a+W 3,ω a+b 2,ω When << 0, the ELU activation function produces a constant output, performing linear processing; GLU represents the gated linear unit, and its specific calculation formula is as follows:

[0042] GLU ω (γ)=σ(W 4,ω γ+b 4,ω )⊙(W 5,ω γ+b 5,ω )

[0043] In the formula, σ(·) represents the activation function; W (·) With b (·)⊙ represents the weight and bias respectively; ⊙ represents the Hadamard product of the elements.

[0044] The feature selection network selects relevant input variables at each time step. Different feature inputs are selected using individual variables with different weights. There are three types of input variables: static features, dynamic time-varying features, and dynamic time-invariant features. Assuming ξ represents the transformed input of the j-th feature at time t, then:

[0045]

[0046] In the formula, Ξ t This represents a flat vector representing all historical input feature vectors at time t; feature selection weights are provided to the external context vectors through GRN and Softmax layers, respectively, and the weights are calculated as follows:

[0047]

[0048] In the formula, v χt c represents a feature selection weight vector; s The context vector c represents the information obtained from the static covariate encoder. For static variables, since it has already accessed static information, it is the context vector c. s Omitted; at each time step, by giving each ξ t Input your own GRN to use an additional non-linear processing layer:

[0049]

[0050] In the formula, ξ t The feature vector represents the processed variables; the weights are shared across all time points t, and then weights are selected based on the processed features. The weighted result is as follows:

[0051]

[0052] In the formula, Represents the feature vector v χt The j-th element;

[0053] The static covariate encoder integrates static features into the network and adjusts temporal dynamics by encoding the context vector;

[0054] The time series processing module learns long-term and short-term time relationships from historical and measured time series variable inputs;

[0055] The multi-head attention mechanism is based on multiplying the importance coefficients of each data point and then summing them. The full attention mechanism is calculated using a request-primary key-value pattern, and the expression for establishing the attention mechanism is as follows:

[0056] Attention(Q,K,V)=A(Q,K)V

[0057]

[0058] In the formula, A() represents the normalization function. Based on the current attention mechanism, different heads are used for different representation subspaces through a multi-head attention mechanism. Its specific form is as follows:

[0059]

[0060] In the formula, as well as W represents the weights of specific heads Q, K, and V, respectively; H Represents all head H of the linear combination h Serial output;

[0061] The multi-head attention mechanism is adjusted to achieve value sharing among all heads, and additive aggregation of all heads is adopted, with the specific calculation formula for additive aggregation as follows:

[0062] InterpretableMultiHead(Q,K,V)=HW H

[0063]

[0064] In the formula, W V W represents the weight value shared by all heads. H Used for the final linear mapping;

[0065] The spatiotemporal normalization module specifically includes a time normalization module and a spatial normalization module, which are used to refine high-frequency and local components, respectively.

[0066] The time normalization module is used to refine the high-frequency components from the input mixed signal, and the high-frequency components and low-frequency components are represented by the following expressions:

[0067]

[0068] In the formula, and These represent the high-frequency and low-frequency components of the i-th time series at time t, respectively. and These represent the local high-frequency and low-frequency components of the i-th time series at time t, respectively. and These represent the global high-frequency components at time t;

[0069] If the rate of change of the low-frequency components is much smaller than that of the high-frequency components, that is, each low-frequency component is approximately equal to a constant over a period of time, then the high-frequency components are represented as:

[0070]

[0071] In the formula, ε represents a relatively small constant that maintains numerical stability; C i,t It is observable; and Representing C under low-frequency influence on the i-th time series respectively i,t The mean and standard deviation; and Each represents a pair of learnable vectors;

[0072] The spatial normalization module is used to refine the local components composed of local high-frequency components and local low-frequency components, and the summarized local and global components are represented by the following expressions:

[0073]

[0074] In the formula, and Let represent the global and local components at time t, respectively; substituting the estimates of the four unobservable variables into the formula, we obtain the representation of the local components as follows:

[0075] As a further aspect of the present invention, the specific steps for predicting and performing multiple corrections on different types of residuals using a multiple error correction strategy in step VIII are as follows:

[0076] S5.1: Assume that the prediction residual and prediction result of the v-th training dataset in the c-th correction are... and There are M training sets in total. The prediction residuals and prediction results of the test data in the c-th correction are: and

[0077] S5.2: Based on the different main sources of error for different error types, explore the main influencing factors of different errors, and conduct a correlation test between the error series and meteorological variables. The specific calculation formula for the correlation test is as follows:

[0078]

[0079] In the formula, x i Represents sequence attributes, i.e., relevant influencing factors; y i This represents the target prediction sequence, i.e., the error sequence; as well as represents the mean of variables x and y respectively; n represents the total number of data points; r represents the Pearson correlation coefficient;

[0080] S5.3: Detect the predictable state of the residual. If it is predictable, the ARIMA model uses the residual for training. If it is unpredictable, stop the error correction process and use the trained ARIMA model for prediction to obtain the corrected training set prediction results and test set prediction results.

[0081] S5.4: In the new round of correction, a correction shrinkage rate r1 is introduced to control the progress of error correction, and the formula is used during the iterative correction process. Calculate the residuals, where M represents the total number of elements in the training dataset. When K = 1, 2, ..., M, it indicates that the model is corrected using the training set. T represents the total number of elements in the test dataset. When K = 1, 2, ..., T, it indicates that the final prediction is made on the test set. When the residuals become unpredictable or the maximum number of iterations is reached, the error correction iteration process is terminated, and the prediction results after multiple corrections are output. Among them, forecasting SF-TSN-TFT This represents the uncorrected result after the initial prediction of the model.

[0082] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0083] This short-term wind power prediction method considering cold wave meteorological conditions expands the scarce samples during the cold wave period using a sequence data sample augmentation method based on DoppelGANger, which improves the fidelity of the generated model. It also improves upon the traditional GAN ​​network and proposes an improved SF-TSN-TFT short-term wind power prediction method. This method fully utilizes the strong feature extraction capability and interpretability of the TFT model to fully mine static data such as wind power and meteorological factors during historical cold waves. Based on this, a spatiotemporal normalization module is introduced to fully extract the hidden dynamic features of meteorological factors, thereby improving the accuracy of multivariate time series prediction. Furthermore, it captures the complex dynamic features of influencing factors at different time periods, constructs an SF-TSN-TFT prediction model, and achieves end-to-end prediction, thus improving the model's adaptability in complex meteorological environments. Finally, multiple error corrections are applied to different fluctuation processes of the error sequence to further improve prediction accuracy. Attached Figure Description

[0084] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0085] Figure 1This is a flowchart of a short-term wind power prediction method considering cold wave meteorological conditions proposed in this invention.

[0086] Figure 2 This is a schematic diagram of the logical framework of a short-term wind power prediction method considering cold wave meteorological conditions proposed in this invention.

[0087] Figure 3 This is a schematic diagram of the DoppelGANger principle architecture of a short-term wind power prediction method considering cold wave meteorological conditions proposed in this invention.

[0088] Figure 4 The flowchart shows the improved SF-TSN-TFT model prediction algorithm for short-term wind power prediction under cold wave meteorological conditions proposed in this invention.

[0089] Figure 5 This is a flowchart of a multiple error correction method based on different error types for a short-term wind power prediction method considering cold wave meteorological conditions proposed in this invention. Detailed Implementation

[0090] Reference Figure 1-5 A short-term wind power prediction method considering cold wave weather conditions is proposed. The specific steps of this prediction method are as follows:

[0091] Select historical wind power data and meteorological data time series of wind farm power stations, and preprocess the input wind farm meteorological data and power data.

[0092] Specifically, a variable-point grouping algorithm is used to identify outage data and some scattered outliers in the wind power curve dataset. Then, a quartile algorithm is used to divide the data into four parts in ascending order, each representing 25% of the entire dataset. The lower quartile is denoted as Q1, the median as Q2, and the upper quartile as Q3. The difference between the upper and lower quartiles is calculated and denoted as the interquartile range I. QR According to Q1-1.5I QR and Q3+1.5I QR Obtain the inner limit of outliers [F] L F UThe data was removed from the range, and a cubic spline interpolation function was constructed using the three-moment interpolation technique. Missing values ​​in the data were then filled using the cubic spline interpolation method. When the missing values ​​were less than 20%, the mean or median was used to fill in the missing values ​​for continuous variables, while categorical variables were either left unfilled or treated as a single category, or the mode was used for filling in the missing values. When the missing values ​​were between 20% and 80%, the mean or median was used to fill in the missing values ​​for continuous variables, and an indicator dummy variable was created for each variable containing missing values. When the missing values ​​were greater than 80%, an indicator dummy variable was created for each variable containing missing values ​​and participated in subsequent modeling, while the original variables were no longer used.

[0093] The Prophet algorithm is used to predict temperature change trends, and combined with cold wave warnings, the time periods of cold wave weather are extracted.

[0094] Specifically, trend functions, periodic functions, and special node accidental factor functions are established respectively. The input historical time series is decomposed using the established 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 separately, and additive regression models are established to predict the temperature change trend. Based on the rate of change of the trend change point and combined with the occurrence time of historical cold wave events, the rate of change is defined as α. When the condition |α|≥10 is met, it indicates that a short-term strong temperature change has occurred, and the complete cold wave weather period in which the short-term strong change occurred is extracted.

[0095] The sample data for cold wave periods were expanded using the DoppelGANger network, and the network parameters were set.

[0096] For details, please 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. A GAN model is constructed based on the generator G and the discriminator D. 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. Both 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. When training the generative adversarial network, the generator and the discriminator are connected into a network, and the DoppelGANger network is used to comprehensively consider the hidden relationship between different sequences and the temporal dependency within the sequence. At the same time, each group of data is normalized separately. The generator and the discriminator continuously and alternately update based on the processed data until the training ends.

[0097] It should be further explained that the specific expression for L(G,D) is as follows:

[0098]

[0099] In the formula, as well as G(·) and D(·) represent the expected values ​​of the distributions 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.

[0100] The MLP generator is used to decouple meteorological data and power data to generate subsequences. At each time step, meteorological data is fed into the time series generator, and the power data is preprocessed based on the generated meteorological data.

[0101] Specifically, the bottom layer of the MLP generator is used as the input layer, receiving multiple n-dimensional vectors. Then, the hidden layer is fully connected to the input layer. The input vector X is input into the hidden layer, and the output representation f(W1X+b1) is obtained, where W1 represents the weights, b1 represents the biases, and f represents the tanh function tanh(a) = (e^(-b1 / a)). a -e -a ) / (e a +e -a ), where a=W1X+b1, after the output layer receives the output of the hidden layer, it generates softmax(W2X1+b2) for the output layer through softmax regression to obtain the subsequences of each group after the decoupling of meteorological data and power data, where X1 represents the output of the hidden layer f(W1X+b1), where W2 represents the weight of the output layer, and b2 represents the bias of the output layer to obtain the subsequences of each group after the decoupling of meteorological data and power data, where X1 represents the output of the hidden layer f(W1X+b1).

[0102] Each data feature is normalized, and the maximum and minimum values ​​of each time series are used as random variables to generate samples.

[0103] It should be further explained that the data is normalized using global min and max values ​​and stored as global constants. Training is then performed on this normalized data. For example, wind power is a measured value, denoted as f. i Meteorological influencing factors are metadata, denoted as A. i (e.g., wind speed / direction, temperature, precipitation, etc.) Model collapse can occur in both measurement data and metadata. Optimizing the standard normalization method involves normalizing each time series sample separately and storing the min / max values ​​as "false" metadata. (This differs from the original data.) iA i Training on ) is The GAN then learns to generate these two pseudo-metadata sets, defining the min / max values ​​for each time series, and then rescales the measurements to the actual range.

[0104] The generated subsequences and pseudo-attributes are used as inputs, and the target sequence is generated through RNN batch processing.

[0105] It should be further explained that the basic structure of an RNN includes an input layer, hidden layers, and an output layer. The state of the hidden layer is updated with each time step and serves as one of the inputs for the next time step.

[0106] Furthermore, it should be noted that the hidden state update formula for RNN is:

[0107] a <t>< / t> =tanh(W ax x <t>< / t> +W aa a <t-1>< / t-1> +b a )

[0108] In the formula: x <t>< / t> It is the input at the current time step, a <t>< / t> It is the hidden state at the current time step, a <t-1>< / t-1> It is the hidden state of the previous time step, W ax and W aa It is the weight matrix, b a It is the bias term, and tanh() is the activation function.

[0109] The output formula for an RNN is:

[0110]

[0111] In the formula: It is the output of the current time step, W ya It is the weight matrix, b y It is a bias term;

[0112] Compared to the single-sample generation of traditional generative adversarial networks, a batch generation method is proposed to reduce the number of RNN passes. In each pass of the RNN, S samples are generated (e.g., historical wind power data with a data sampling point every 15 minutes, and S consecutive sampling points), where S is an adjustable parameter. This effectively reduces the number of RNN passes.

[0113] A short-term wind power prediction model based on an improved SF-TSN-TFT neural network was established to make separate predictions for normal weather and cold wave periods, and the prediction results were combined to form a preliminary prediction result.

[0114] 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;

[0115] The TFT network model specifically includes a gating mechanism, a feature selection network, a static covariate encoder, and a timing processing module;

[0116] As a further aspect of the present invention, the specific calculation formula for the gating mechanism is as follows:

[0117] GRN ω (a,c)=LayerNorm(a+GLU ω (η1))

[0118] η1=W 1,ω η2+b 1,ω

[0119] η2=ELU(W 2,ω a+W 3,ω a+b 2,ω )

[0120] In the formula, η1 and η2 both represent intermediate layers; LayerNorm represents the normalization of the standard layer; ω represents a weight sharing index; ELU represents the exponential linear unit activation function, when W 2,ω a+W 3,ω a+b 2,ω When >> 0, the ELU activation function will be used as the identity function, when W 2,ω a+W 3,ω a+b 2,ω When << 0, the ELU activation function produces a constant output, performing linear processing; GLU represents the gated linear unit, and its specific calculation formula is as follows:

[0121] GLU ω (γ)=σ(W 4,ω γ+b 4,ω )⊙(W 5,ω γ+b 5,ω )

[0122] In the formula, σ(·) represents the activation function; W (·) With b (·) ⊙ represents the weight and bias respectively; ⊙ represents the Hadamard product of the elements.

[0123] The feature selection network selects relevant input variables at each time step. Different feature inputs are selected using individual variables with different weights. There are three types of input variables: static features, dynamic time-varying features, and dynamic time-invariant features. Assuming ξ represents the transformed input of the j-th feature at time t, then:

[0124]

[0125] In the formula, Ξ t This represents a flat vector representing all historical input feature vectors at time t; feature selection weights are provided to the external context vectors through GRN and Softmax layers, respectively, and the weights are calculated as follows:

[0126]

[0127] In the formula, v χt c represents a feature selection weight vector; s The context vector c represents the information obtained from the static covariate encoder. For static variables, since it has already accessed static information, it is the context vector c. s Omitted; at each time step, by giving each ξ t Input your own GRN to use an additional non-linear processing layer:

[0128]

[0129] In the formula, ξ t The feature vector represents the processed variables; the weights are shared across all time points t, and then weights are selected based on the processed features. The weighted result is as follows:

[0130]

[0131] In the formula, Represents the feature vector v χt The j-th element;

[0132] Static covariate encoders integrate static features into the network and modulate temporal dynamics by encoding context vectors;

[0133] The time series processing module learns the long-term and short-term time relationships from historical and measured time series variable inputs;

[0134] Multi-head attention mechanisms multiply and then sum the importance coefficients of each data point. Full attention mechanisms use a request-primary key-value model for calculation, and the expression for establishing the attention mechanism is as follows:

[0135] Attention(Q,K,V)=A(Q,K)V

[0136]

[0137] In the formula, A() represents the normalization function. Based on the current attention mechanism, different heads are used for different representation subspaces through a multi-head attention mechanism. Its specific form is as follows:

[0138]

[0139] In the formula, as well as W represents the weights of specific heads Q, K, and V, respectively; H Represents all head H of the linear combination h Serial output;

[0140] The multi-head attention mechanism is adjusted to achieve value sharing among all heads, and additive aggregation of all heads is adopted, with the specific calculation formula for additive aggregation as follows:

[0141] InterpretableMultiHead(Q,K,V)=HW H

[0142]

[0143] In the formula, W V W represents the weight value shared by all heads. H Used for the final linear mapping;

[0144] The spatiotemporal normalization module specifically includes a time normalization module and a spatial normalization module, which are used to refine high-frequency and local components, respectively.

[0145] The time normalization module is used to refine the high-frequency components from the input mixed signal, and the high-frequency components and low-frequency components are represented by the following expressions:

[0146]

[0147] In the formula, and These represent the high-frequency and low-frequency components of the i-th time series at time t, respectively. and These represent the local high-frequency and low-frequency components of the i-th time series at time t, respectively. and These represent the global high-frequency components at time t;

[0148] If the rate of change of the low-frequency components is much smaller than that of the high-frequency components, that is, each low-frequency component is approximately equal to a constant over a period of time, then the high-frequency components are represented as:

[0149]

[0150] In the formula, ε represents a relatively small constant that maintains numerical stability; C i,t It is observable; and Representing C under low-frequency influence on the i-th time series respectively i,t The mean and standard deviation; and Each represents a pair of learnable vectors;

[0151] The spatial normalization module is used to refine the local components, which consist of local high-frequency components and local low-frequency components. The summarized local and global components are represented by the following expressions:

[0152]

[0153] In the formula, and Let represent the global and local components at time t, respectively; substituting the estimates of the four unobservable variables into the formula, we obtain the representation of the local components as follows:

[0154] By employing a multiple error correction strategy, different types of residuals are predicted and multiple corrections are performed. The correction is iterated multiple times, and the error correction iteration process is terminated when the residuals become unpredictable or the maximum number of iterations is reached.

[0155] For details, please refer to Figure 5 It can be seen that, assuming the prediction residual and prediction result of the v-th training dataset in the c-th correction are... and There are M training sets in total. The prediction residuals and prediction results of the test data in the c-th correction are: and Based on the different main sources of error for different error types, this study explores the main influencing factors of different errors and conducts correlation tests between error sequences and meteorological variables to detect the predictability of residuals. If predictable, the ARIMA model uses these residuals for training; if unpredictable, the error correction process is stopped, and the trained ARIMA model is used for prediction to obtain corrected training set and test set prediction results. In the new round of correction, a correction shrinkage rate *r* is introduced to control the progress of error correction, and the formula is used during the iterative correction process. Calculate the residuals, where M represents the total number of elements in the training dataset. When K = 1, 2, ..., M, it indicates that the model is corrected using the training set. T represents the total number of elements in the test dataset. When K = 1, 2, ..., T, it indicates that the final prediction is made on the test set. When the residuals become unpredictable or the maximum number of iterations is reached, the error correction iteration process is terminated, and the prediction results after multiple corrections are output. Among them, forecasting SF-TSN-TFT This represents the uncorrected result after the initial prediction of the model.

[0156] In addition, it should be noted that the specific calculation formula for the correlation test is as follows:

[0157]

[0158] In the formula, x i Represents sequence attributes, i.e., relevant influencing factors; y i This represents the target prediction sequence, i.e., the error sequence; as well as represents the mean of variables x and y, respectively; n represents the total number of data points; r represents the Pearson correlation coefficient.

[0159] 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 weather conditions, characterized in that, The prediction method comprises the following specific steps: I. Select the historical wind power data and meteorological data of the wind farm, and preprocess the input wind farm meteorological data and power data; II. Use the Prophet algorithm to predict the temperature change trend, combine the cold wave warning, and extract the cold wave weather occurrence period; III. Expand the cold wave weather period sample data through the DoppelGANger network, and set the network parameters; IV. Use the MLP generator to decouple the meteorological data and power data to generate sub-sequences, feed the meteorological data to the time series generator at each time step, and preprocess the power data according to the generated meteorological data; V. Normalize each data feature, and use the maximum and minimum values of each time series as a random variable for sample generation; VI. Use the generated sub-sequences and pseudo attributes as inputs to generate the target sequence through RNN batch processing; VII. Establish a short-term wind power prediction model based on the improved SF-TSN-TFT neural network, separately predict the normal weather and cold wave weather periods, and combine the prediction results to form the preliminary prediction results; VIII. Use multiple error correction strategies to predict and correct different types of residuals, and perform multiple iterations of correction. When the residual is unpredictable or the maximum number of iterations is reached, terminate the iteration process of error correction; IX. Add the preliminary prediction results to the corrected residual sequence to obtain the final wind power prediction results; The specific steps of using the MLP generator to decouple the meteorological data and power data to generate sub-sequences in step IV are as follows: S4.1: The bottom layer of the MLP generator is taken as the input layer, and a plurality of n-dimensional vectors are input. Then the hidden layer is fully connected with 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, wherein 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 ), wherein a=W1X+b1; S4.2: After the output layer receives the hidden layer output, the output layer output softmax(W2X1+b2) is generated through softmax regression, where W2 represents the weight of the output layer, b2 represents the bias of the output layer, and X1 represents the output f(W1X+b1) of the hidden layer.

2. The short-term wind power prediction method considering the cold wave weather condition 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: Use the change point grouping algorithm to identify the shutdown data and some scattered abnormal data in the wind power curve data set, and then use the quartile algorithm to divide a group of data in ascending order into 4 parts, each accounting for 25% of the entire data set. The lower quartile is denoted as Q1, the median is denoted as Q2, and the upper quartile is denoted as Q3; S1.2: Calculate the difference between the upper quartile and the lower quartile, denoted as the interquartile range I. QR According to Q1-1.5I QR and Q3+1.5I QR Obtain the inner limit of outliers [F] L F U ], and remove data that is outside the 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 the 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 classification variables are not filled in or are treated as a single category for processing, 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 a dummy variable is created for each variable containing missing values. When the missing values are greater than 80%, a dummy variable is created for each variable containing missing values, and is used in subsequent modeling, while the original variable is not used.

3. The short-term wind power prediction method considering the cold wave weather condition according to claim 2, characterized in that, The temperature change trend prediction using the Prophet algorithm in step II, combined with the cold wave warning, extracts the specific steps of the cold weather occurrence period as follows: S2.1: A trend function, a periodic function and a special node accidental factor function are respectively established, and the input historical time series is decomposed through the established functions to decompose the sequence into trend items, periodic items, special time node accidental factor items and random fluctuation items; S2.2: The trend items, periodic items, special time node accidental factor items and random fluctuation items obtained by decomposition are respectively modeled, and an additive regression model is established to predict the temperature change trend; S2.3: According to the rate of change of the trend change point, and combined with the historical cold wave event occurrence time, the rate of change is defined as α, when the condition |α|≥10 is met, it indicates that the temperature has a short-time strong change, and the complete cold weather period of the short-time strong change is extracted.

4. The short-term wind power prediction method considering the cold wave weather condition according to claim 1, characterized in that, The specific steps of step III for expanding the cold weather period sample data through the DoppelGANger network are as follows: S3.1: The generating model capturing the data distribution is taken as the generator G, and the discriminant model estimating the probability from the training data sample is taken as the discriminator D, and the GAN model is constructed based on the generator G and the discriminator D; S3.2: Random noise data z is input into the generator G, and corresponding wind power data G(z) is generated, real power data x and generated data G(z) are input into the discriminator for discrimination, and then the discrimination results of the discriminator are fed back to the generator and the discriminator, and the two update their parameters according to the feedback results L(G,D) to optimize, wherein the feedback result L(G,D) is the loss parameter of the model, and its specific expression is as follows: wherein and respectively denote the expected values of the real power data x and the noise data z distribution; G(·) and D(·) represent the data output by the generator and discriminator, respectively; S3.3: 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 dependence within the sequence are considered through the DoppelGANger network, and each group of data is normalized separately, and the generator and the discriminator are continuously and alternately updated based on the processed data until the training is completed.

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

6. The method for short-term wind power prediction considering cold wave meteorological conditions according to claim 1, characterized in that, The specific steps of step VIII for predicting different types of residuals and completing multiple corrections through multiple error correction strategies are as follows: S5.1: Assume the prediction residual and the prediction result of the vth training dataset in the cth correction are and M training sets, the prediction residual and the prediction result of the test data in the cth correction are and S5.2: 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 variable is performed, and the correlation test calculation formula is as follows: where x i represents the sequence property, i.e. the relevant influencing factor; y i represents the target prediction sequence, i.e. the error sequence; and respectively represent the mean of the variables x and y; n represents the total number of data points; r represents the Pearson correlation coefficient; S5.3: Detect the predictable state of the residual, if it is predictable, use the ARIMA model to train the residual, if it is not predictable, stop the error correction process, and use the trained ARIMA model for prediction to obtain the corrected training set prediction result and test set prediction result; S5.4: In the new round of correction, the correction shrinkage rate r1 is introduced to control the progress of error correction, and the formula Calculate the residual error, where M represents the total number of training data sets, and when K = 1, 2,..., M, it means that the model is corrected using the training set; T represents the total number of test data sets, and when K = 1, 2,..., T, it means the final prediction of the test set, when the residual error is unpredictable or reaches the maximum number of iterations, terminate the iteration process of error correction, and output the prediction results after multiple corrections where Forecasting SF-TSN-TFT Represents the uncorrected result after the preliminary prediction of the model.

Images