A method and system for ultra-short-term wind power segmented prediction taking into account extreme cold wave weather samples

Abstract

The present invention discloses an ultra-short-term wind power segmented prediction method and system taking into account extreme cold wave weather samples, which relates to the technical field of wind power prediction, including using a sequence variational autoencoder (SeqVAE) algorithm to generate numerical weather forecast data and corresponding power samples. Then, a method based on a graph convolution (GCN) network and a bidirectional gated recurrent unit (BiGRU) is used to extract the cold wave power loss period in the entire period, and on this basis, a LightGBM prediction method is used for the normal weather period; a LightGBM‑Transformer prediction method is proposed for the power loss period to quantify the power loss generated. The present invention improves the accuracy of wind power prediction by performing combined power prediction separately for normal weather and cold wave weather.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power prediction, and more specifically to an ultra-short-term wind power segmented prediction method and system taking into account extreme cold wave weather samples. Background Art

[0002] At present, with the continuous construction of wind farms and the annual growth of wind power installed capacity, the reliable operation of wind turbines is becoming more and more important, which not only determines the economic benefits of wind farms, but also affects the safe and stable operation of the entire power grid. Wind turbines are installed outdoors, and extreme weather is one of the main reasons for interfering with their normal operation. Severe meteorological disasters may even cause failures.

[0003] Compared with the wind power forecast under normal climate, the wind power forecast under low temperature and cold wave weather is more difficult to predict due to its unique randomness and instability, and it also brings huge challenges to the reliable operation of wind farms. Cold wave weather is one of the most common extreme weather conditions, and wind power will fluctuate greatly in a short period of time, which greatly increases the difficulty of ultra-short-term wind power forecasting.

[0004] Therefore, how to improve the prediction accuracy of wind power is an urgent problem that technical personnel in this field need to solve. Summary of the invention

[0005] In view of this, the present invention provides an ultra-short-term wind power segmented prediction method and system taking into account extreme cold wave weather samples. In view of the characteristics of scarcity, extreme values ​​and high volatility of cold wave weather sample data, the Sequence Variational Autoencoder (SeqVAE) algorithm is used to generate numerical weather forecast data and corresponding power samples. At the same time, the prediction accuracy of wind power is improved by classifying power judgments for normal weather and cold wave weather.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] An ultra-short-term wind power segmented forecasting method taking into account extreme cold wave weather samples comprises:

[0008] Acquire historical wind farm NWP data and corresponding power data to generate a data set; perform data discrimination based on the cold wave time definition, and divide the data in the data set into normal weather samples and cold wave weather samples;

[0009] The normal weather samples are input into the LightGBM model, the LightGBM model is trained, and a normal weather power prediction model is obtained, and the wind power prediction value of normal weather is obtained through the normal weather power prediction model; the cold wave weather samples are input into the preset SeqVAE encoder for sample generation, and then the generated samples and the normal weather samples are assigned wind power loss states to obtain assigned samples with completed assignments;

[0010] The assigned samples are input into the GCN-BiGRU power loss period extraction model to capture the spatiotemporal characteristics of the assigned samples and extract the cold wave power loss period; the cold wave power loss period and the corresponding power data are input into the Transformer loss value prediction model for training to obtain a power loss prediction model, and the power loss is predicted based on the power loss prediction model; the ultra-short-term wind power prediction value is obtained based on the wind power and the power loss in normal weather.

[0011] Preferably, the training process of the SeqVAE encoder specifically includes: encoder q(m|x), decoder p θ (x|m), decoder p β (x|m), discriminator d δ (x|m), the encoder q(m|x) maps the cold wave weather sample to the latent space through a recurrent neural network, obtains the probability distribution of the latent vector, and extracts the latent vector from the probability distribution; the decoder p θ (x|m) and decoder p β (x|m) generates a reconstructed sequence according to each of the latent vectors, and inputs each reconstructed sequence into the discriminator d in combination with the cold wave weather samples in the data set δ (x|m), the discriminator d δ (x|m) feeds the output back to the decoder p θ (x|m) decoder p θ The parameters of (x|m) are updated and adjusted, and then the decoder p is adjusted. θ (x|m) sends the discriminator d δ The output of (x|m) is updated to adjust the decoder p β Parameters of (x|m).

[0012] Preferably, the parameters of the SeqVAE encoder are optimized by maximizing ELBO during training, as follows:

[0013] ELBO=E z～q(m|x) [logp θ (x|m)]-KL(q(m|x)||p(m))≤logp(x);

[0014] Among them, Ez～q(m|x) is the expected log-likelihood of the observed data x given the latent variable m, p(m) is the prior distribution of the latent variable m, usually the standard normal distribution N(0,l), and p(x) is the marginal likelihood of the observed data x.

[0015] Preferably, the obtained assigned sample specifically includes the generated sample and the normal weather sample including wind speed, wind direction, temperature, humidity, and air pressure data, and the time series in the sample is encoded, the power loss data in the cold wave weather is encoded as 0, and the normal weather data is encoded as 1, as follows:

[0016]

[0017] Among them, β is the percentage of climbing rate, C p is the operating capacity of the wind farm, is the power value of the normal weather sample using LightGBM, μ is the power loss percentage, P t is the power value at the current moment, and h is hours.

[0018] Preferably, the assigned sample is input into GCN to obtain a feature vector with spatial features, as follows:

[0019]

[0020] Where I is the identity matrix, is the self-circular correlation matrix, is the normalized adjacency matrix, D is the degree matrix, and D ii represents the diagonal elements of the degree matrix D, n k is the number of co-convolutional layers in the graph, σ k () is the activation function of the graph convolutional layer, is the weight matrix of the i-1th layer, X Feature is the feature matrix, represents the feature matrix of the i-th layer node, Indicates the degree of node i.

[0021] Preferably, the BiGRU consists of two independent GRU units running in opposite directions, one processing data forward according to the time series, and the other processing data backward. Through the bidirectional structure, the BiGRU model can capture both forward and backward information of the sequence data at the same time.

[0022] An ultra-short-term wind power segmented prediction system taking into account extreme cold wave weather samples, comprising:

[0023] The data discrimination module obtains historical wind farm NWP data and corresponding power data to generate a data set; performs data discrimination based on the cold wave time definition, and divides the data in the data set into normal weather samples and cold wave weather samples;

[0024] The sample assignment module inputs the normal weather sample into the LightGBM model, trains the LightGBM model, obtains the normal weather power prediction model, and obtains the wind power prediction value of normal weather through the normal weather power prediction model; inputs the cold wave weather sample into the preset SeqVAE encoder to generate samples, and then assigns the wind power loss state to the generated samples and the normal weather samples to obtain the assigned samples;

[0025] The power prediction module inputs the assigned samples into the GCN-BiGRU power loss period extraction model, captures the spatiotemporal characteristics of the assigned samples, and extracts the cold wave power loss period; inputs the cold wave power loss period and the corresponding power data into the Transformer loss value prediction model for training to obtain a power loss prediction model, and predicts the power loss according to the power loss prediction model; and obtains an ultra-short-term wind power prediction value according to the wind power and the power loss in normal weather.

[0026] It can be seen from the above technical solution that compared with the prior art, the present invention discloses a method and system for segmented prediction of ultra-short-term wind power taking into account extreme cold wave weather samples. In view of the characteristics of scarce cold wave weather sample data, extreme values ​​and high volatility, the Sequence Variational Autoencoder (SeqVAE) algorithm is used to generate numerical weather forecast data and corresponding power samples. At the same time, the prediction accuracy of wind power is improved by classifying power judgments for normal weather and cold wave weather. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0028] Figure 1 A schematic diagram of the method provided by the present invention;

[0029] Figure 2 A sequence variational autoencoder model diagram provided by the present invention;

[0030] Figure 3 The GCN model diagram provided by the present invention;

[0031] Figure 4 The BiGRU network structure diagram provided by the present invention;

[0032] Figure 5 The LightGBM histogram provided by the present invention;

[0033] Figure 6 A diagram of the Transformer power loss prediction framework provided by the present invention;

[0034] Figure 7 The real wind speed samples and the generated wind speed sample distribution diagram provided by the present invention;

[0035] Figure 8 The real humidity sample and the generated humidity sample distribution diagram provided by the present invention;

[0036] Fig. 9 A distribution diagram of the real power samples and generated power samples provided by the present invention;

[0037] Fig.10 A scatter plot of the actual temperature and power samples and the generated temperature and power samples provided by the present invention;

[0038] Fig.11 This is a comparison chart of the model effects provided by the present invention. DETAILED DESCRIPTION

[0039] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0040] Cold wave weather refers to the process of strong cold air activity over a large area. According to the standards given by the Central Meteorological Observatory, the definition of a cold wave is: the temperature drops by more than 8°C in 24 hours, or the temperature drops by more than 10°C in 48 hours. Based on this, this paper sets the start time of the cold wave event as the highest temperature threshold before the temperature drops sharply in a full period, and after one or more peaks and troughs, it finally returns to a certain temperature threshold. This threshold is set to 5°C in this invention.

[0041] The embodiment of the present invention discloses a super short-term wind power segmented prediction method taking into account extreme cold wave weather samples, such as Figure 1 As shown, including:

[0042] Obtain historical wind farm NWP data and corresponding power data to generate a data set; perform data discrimination based on the cold wave time definition and divide the data in the data set into normal weather samples and cold wave weather samples;

[0043] Input the normal weather samples into the LightGBM model, train the LightGBM model, obtain the normal weather power prediction model, and obtain the normal weather wind power prediction value through the normal weather power prediction model; input the cold wave weather samples into the preset SeqVAE encoder to generate samples, and then assign the wind power loss status to the generated samples and the normal weather samples to obtain the assigned samples;

[0044] The assigned samples are input into the GCN-BiGRU power loss period extraction model to capture the spatiotemporal characteristics of the assigned samples and extract the cold wave power loss period; the cold wave power loss period and the corresponding power data are input into the Transformer loss value prediction model for training to obtain the power loss prediction model, and the power loss prediction model is used to predict the power loss; the ultra-short-term wind power prediction value is obtained based on the wind power and power loss in normal weather.

[0045] In a specific embodiment, Figure 2 As shown in the figure, the training process of SeqVAE encoder specifically includes: encoder q(m|x), decoder p θ (x|m), decoder p β (x|m), discriminator d δ (x|m), the encoder q(m|x) maps the cold wave weather samples to the latent space through a recurrent neural network, obtains the probability distribution of the latent vector, and extracts the latent vector from the probability distribution; the decoder p θ (x|m) and decoder p β (x|m) generates a reconstructed sequence based on the latent vector, and inputs each reconstructed sequence into the discriminator d in combination with the cold wave weather samples in the dataset δ (x|m), discriminator d δ (x|m) feeds the output back to the decoder p θ (x|m) decoder p θ The parameters of (x|m) are updated and adjusted, and then the decoder p is adjusted. θ (x|m) sends the discriminator d δ The output of (x|m) is updated to adjust the decoder p β Parameters of (x|m).

[0046] where samples x and x′ are sampled from the training set. m is a potential vector. Sample f s and f′ are respectively obtained by the decoder p β and decoder p θ The feedback is generated by the discriminator d δ (x|m) is obtained from the input f′. Encoder p βis a conditional probability model that takes in latent variables m sampled from the latent space and generates a reconstructed sequence that is as similar as possible to the original input sequence x. The goal is to make the discriminator d δ (x|m) considers the generated sequence x to be real data, thereby maximizing the authenticity of the generated sequence.

[0047] Based on VAE, SeqVAE introduces VAE's policy gradient and adds an adversarial process similar to the GAN model in VAE training, so that it combines VAE's potential representation learning ability and GAN's generated data quality advantages to achieve better generated data effects in cold weather.

[0048] NWP data and corresponding power x under cold wave weather 1:T ={(X NWPt ,P t )|X NWPt ,P t ∈Ω ext ,t=1,…,T} to generate data,Ω ext is the real sample space of cold wave weather, X NWPt The input time series NWP data (including wind speed, wind direction, temperature, humidity, and air pressure) is as follows:

[0049] The encoder q(m|x) is used to convert the input sample x 1:T The probability distribution mapped to the latent vector m.

[0050] μ′(x) and σ′(x) are the parameters of the probability distribution of the latent vector m, representing the input sample x 1:T The central location in the latent space and the variance of each latent dimension.

[0051] Then the decoder p is initialized according to the latent vector m θ (x|m) and p β The initial state of (x|m) generates a reconstructed sequence sample f s and f′. In order to make the discriminator d δ (x|m) learns to distinguish the difference between the generated sequence and the real sequence, and puts f s Together with f′ and the real sample x′, it is input to the discriminator d δ (x|m). Finally, the discriminator d δ The output of (x|m) is fed back to the decoder p θ (x|m), for decoder p θ Update and adjust the (x|m) parameters to improve the quality and authenticity of generated samples.

[0052] RNN (Recurrent Neural Network) is a type of neural network that processes sequential data by repeatedly using the same neural network module over time steps.

[0053] The encoder uses RNN to process the input sequence X. The main purpose is to map the entire sequence to the mean vector μ and variance vector σ in the latent space, thereby encoding and learning the input sequence. This structure allows SeqVAE to effectively process sequence data of variable length and capture the temporal dependencies in the sequence for subsequent latent variable sampling and decoding reconstruction processes.

[0054] The processing process is:

[0055] Convert each element of the input sequence X into a processable vector representation;

[0056] Process the elements of the sequence one by one in time order. At each time step t, the RNN receives the current input x t and the hidden state h of the previous time step t-1 ;

[0057] Use the current input x t and the previous hidden state h t-1 Calculate the hidden state h at the current time step t , this hidden state h t Contains the input information of the current time step and the context information of the previous time step;

[0058] At each time step, the RNN generates an output y t ;

[0059] Repeat the above steps to process each time step in the sequence in turn until the entire sequence has been processed.

[0060] In a specific embodiment, the parameters of the SeqVAE encoder are optimized by maximizing ELBO during the training process. ELBO is a loss function in variational inference that is used to measure the difference between the likelihood of generated data and the prior distribution of the potential representation. By maximizing ELBO, the distribution of the generated data samples is similar to that of the training data, as follows:

[0061] ELBO=E z～q(m|x) [logp θ (x|m)]-KL(q(m|x)||p(m))≤logp(x);

[0062] Among them, E z～q(m|x)is the expected log-likelihood of the observed data x given the latent variable m, p(m) is the prior distribution of the latent variable m, usually the standard normal distribution N(0,l), and p(x) is the marginal likelihood of the observed data x, which is used to measure the ability of the model to generate data.

[0063] ELBO is divided into two parts: E z～q(m|x) [logp θ (x|m)] and KL(q(m|x)||p(m));

[0064] The first item describes the distance between the reconstructed sequence and the original sequence, usually called the reconstruction loss, and the second item describes the distance between the approximate posterior distribution and the prior distribution, usually called the kl divergence loss. SeqVAE simultaneously trains the encoder, decoder, and discriminator by maximizing ELBO to achieve the purpose of generating a distribution that is more consistent with the real data.

[0065] In a specific embodiment, the assignment sample obtained after the assignment is completed specifically includes the generated sample and the normal weather sample including five elements: wind speed, wind direction, temperature, humidity, and air pressure. The time series in the sample is encoded, and the power loss data during the cold wave period is encoded as 0, and the normal weather data is encoded as 1, as follows:

[0066]

[0067] Where β is the percentage of the ramp rate. In the binary classification problem, the setting of β can help identify the period of power loss in the cold wave. The value of β in the present invention is 0.06. p is the operating capacity of the wind farm, is the power value of the LightGBM normal weather sample, μ is the power loss percentage, which is used to describe the power loss that may occur in the wind farm under extreme weather conditions such as cold waves. The value of this invention is 0.04, P t is the power value at the current moment, and h is hours.

[0068] Then, the input is set as NWP time series data (assigned samples), the output is the corresponding encoded value, and training is performed in the GCN-BiGRU model.

[0069] In a specific embodiment, Figure 3As shown in the figure, GCN (Graph Convolutional Network) is a deep learning model specifically designed to process graph structured data. Traditional deep learning models such as CNN can only extract features of local spatial data, and cannot effectively capture meteorological sequences with complex topological information such as NWP (Numerical Weather Prediction), and cannot model the spatial dependencies between nodes well. In contrast, GCN can learn and utilize the relationship between nodes in the graph data structure to extract high-level features of the data. The figure shows the network structure of GCN. The figure shows the GCN model framework in which NWP data passes through multiple graph convolution layers and finally obtains a feature vector with a spatial structure.

[0070] An undirected graph is a basic concept in graph theory. It consists of a set of nodes and a set of edges. Each edge connects a pair of nodes and has no direction.

[0071] In this paper, we use GCN to construct the input NWP meteorological sequence into a graph structure:

[0072] In GCN, nodes are the basic units in the graph structure. Each node V = {v1, v2, ..., v n} represents the NWP meteorological data (wind speed, wind direction, temperature, air pressure, humidity, power) at a time point. i is the i-th node, and n is the total number of nodes. There is temporal continuity between adjacent time points, so edges represent the connection relationship between nodes. Finally, all nodes and their connection relationships are modeled into a graph.

[0073] The characteristics at time t in the graph can be expressed as: is the feature of the i-th node at time t.

[0074] Then construct the feature matrix:

[0075] Feature Matrix Represents the characteristics of node k in the past h time steps at time t.

[0076] Input the assigned sample into GCN to obtain the feature vector with spatial features, as follows:

[0077]

[0078] Where I is the identity matrix, is the self-circular correlation matrix, is the normalized adjacency matrix, D is the degree matrix, and D ii Denotes the diagonal elements of the degree matrix D, representing the degree of node i, n k is the number of co-convolutional layers in the graph, σ k () is the activation function of the graph convolutional layer, is the weight matrix of the i-1th layer, X Feature is the feature matrix, represents the feature matrix of the i-th layer node, It means calculating the degree of node i, which is the sum of all the connections of node i.

[0079] It means adding the identity matrix I to the original adjacency matrix C, which means adding a self-loop to each node. Adding a self-loop to each node can ensure that the feature does not lose its own information due to the normalization process.

[0080] In a specific embodiment, another key issue in the extraction of cold wave power loss periods is the time dependency of the model. Recurrent neural networks (RNNs) can establish connections between adjacent samples in a time series. However, when the input sequence is long, the problem of gradient vanishing occurs, so RNNs cannot learn their long-distance dependencies. Its variant, the long short-term memory (LSTM) model, as a variant of RNN, improves RNN by introducing selective memory and unit gates. LSTM solves the gradient vanishing problem that plagues RNNs and can learn long-term dependencies in sample data. Nevertheless, LSTM networks have a complex structure and a large number of parameters. This may lead to high computational complexity for training and reasoning. The GRU model is an improved variant of LSTM with an update gate and a reset gate. Compared with LSTM, GRU has fewer parameters and a simpler structure, which makes parameter convergence faster and reduces the risk of overfitting.

[0081] The present invention uses the BiGRU (Bidirectional Gated Recurrent Unit) model. BiGRU is bidirectional and considers the information of the sequence data in both the forward and reverse directions. This means that at each time step, BiGRU can use past and future contextual information, not just unidirectional historical information. This bidirectionality enables BiGRU to more comprehensively capture the long-term dependencies and dynamic changes in the time series, capture the temporal features of the time series from the last GCN layer, and use the NWP time series with spatial features obtained from GCN as the input of Bi-GRU. Figure 4As shown in the figure, we can see that BiGRU consists of two independent GRU units running in opposite directions, one processing data forward according to the time series, and the other processing data backward. Through this bidirectional structure, the BiGRU model can simultaneously capture the forward and backward information of the sequence data, thereby better understanding and predicting the sequence data.

[0082] First, analyze the structure of the GRU module. The update gate and reset gate in GRU are defined as follows:

[0083] z t =σ(W z *[x t ,h t-1 ]);

[0084] r t =σ(W r *[x t ,h t-1 ]);

[0085] Among them, x t is the input vector of the tth time step, which is the output of the last GCN layer. W z is the weight matrix of the update gate, W r is the reset weight matrix, h t-1 is the hidden state at the previous time step.

[0086] Calculate the reset gate r t With h t-1 The information obtained is:

[0087]

[0088] in: It means multiplying the corresponding positions in two vectors.

[0089] Then calculate the current cell state:

[0090]

[0091] Among them, tanh is the activation function and W is the weight.

[0092] The final output is:

[0093]

[0094] Among them, z t is the activation result of the update gate.

[0095] It can be seen that in the traditional GRU module, the information in the time series is passed forward without considering the subsequent time series information, so the GRU can only process unidirectional information. In the BiGRU model used in the present invention, two GRUs running in opposite directions are trained, and the output of the model is determined by these two GRUs. If the output of the forward GRU is The output of the reverse GRU is The output of BiGRU is as follows:

[0096]

[0097] In a specific embodiment, the present invention uses LightGBM to perform point prediction on the normal power period under the cold wave to obtain the basic power value.

[0098] like Figure 5 As shown in the figure, the LightGBM model mainly includes the histogram algorithm and the leaf-wise leaf growth strategy. The former can significantly reduce memory usage; the latter can grow a deeper decision tree with the same number of splits to obtain better prediction accuracy.

[0099] In a specific embodiment, Transformer is a more advanced algorithm in the field of time series prediction. Compared with long short-term memory (LSTM) and gated recurrent unit (GRU), the algorithm has the ability to model both long-term and short-term time series features. Transformer's multi-head attention mechanism reveals the importance of each time step in the historical power loss state. The more the number, the stronger the ability to capture long-distance dependent features.

[0100] Transformer model framework Figure 6 As shown, the number of layers of the encoder and decoder are Ne=Nd=4, the number of multi-head attention is n=6, and the feedforward network contains two fully connected layers, namely ReLU and linear activation layers.

[0101] Transformer calculates Q (query vector), K (keyword vector), and V (value vector) for the meteorological data at each timestamp, thereby calculating the feature vector Z(n) of the n-th multi-head attention, as shown in the following formula:

[0102]

[0103] Where: d k is the vector dimension; softmax(.) is the normalization function. It can realize multi-threaded synchronous calculation of multiple feature vectors, and convert the feature matrix Z of n self-attention calculations into (n)The final feature Z is concatenated as follows:

[0104] Z=C oncat (Z (1) ,...,Z (n) )W O ;

[0105] Where W O is the linear transformation weight matrix, C oncat (.) is the matrix concatenation function.

[0106] An ultra-short-term wind power segmented prediction system taking into account extreme cold wave weather samples, comprising:

[0107] The data discrimination module obtains historical wind farm NWP data and corresponding power data to generate a data set; it discriminates data based on the cold wave time definition and divides the data in the data set into normal weather samples and cold wave weather samples;

[0108] The sample assignment module inputs the normal weather samples into the LightGBM model, trains the LightGBM model, obtains the normal weather power prediction model, and obtains the normal weather wind power prediction value through the normal weather power prediction model; inputs the cold wave weather samples into the preset SeqVAE encoder to generate samples, and then assigns the wind power loss status to the generated samples and the normal weather samples to obtain the assigned samples;

[0109] The power prediction module inputs the assigned samples into the GCN-BiGRU power loss period extraction model, captures the spatiotemporal characteristics of the assigned samples, and extracts the cold wave power loss period; the cold wave power loss period and the corresponding power data are input into the Transformer loss value prediction model for training to obtain the power loss prediction model, and the power loss is predicted based on the power loss prediction model; the ultra-short-term wind power prediction value is obtained based on the wind power and power loss in normal weather.

[0110] In the present invention, the LightGBM model training: input temperature, humidity, wind speed, wind direction, air pressure, and the output is the corresponding power;

[0111] The LightGBM model uses: input weather sequence, output predicted power, that is, the power prediction value of normal weather;

[0112] Then the generated cold wave data is encoded (0 for power loss and 1 for normal);

[0113] Then train the GCN-BiGRU time segment extraction model:

[0114] Input temperature, humidity, wind speed, wind direction, air pressure and corresponding power, and output 0 or 1;

[0115] Use of the GCN-BiGRU model: input temperature, humidity, wind speed, wind direction, air pressure and corresponding power, and output predicted 0 or 1;

[0116] Then take out the power loss sequence encoded as 0 and put it into LightGBM for prediction. You can get a base value, and then subtract the true power value from this base value to get the loss value power.

[0117] Training of the Transformer loss value prediction model: input temperature, humidity, wind speed, wind direction, and air pressure, and output loss value power;

[0118] Use of the Transformer loss value prediction model: input temperature, humidity, wind speed, wind direction, and air pressure, and output the predicted loss value power;

[0119] Finally, the usage process of this model is: input a time period, put it into GCN-BiGRU for judgment, use lightGBM to predict the normal weather period and output it directly; if it is a power loss period: subtract the transformer loss value from the lightGBM prediction value to get the actual power. The sum of the two periods is the entire complete period.

[0120] The data set used in this invention is provided by Longyuan Electric Power, the world's largest wind power operator. The data used are the daily forecast values ​​of the main meteorological factors in the NWP of ten wind power plants from 2022 to 2023 and the measured values ​​of wind power, with a time resolution of 15 minutes. The data is divided into a training set, a validation set, and a test set. Among them, the data set obtained by small sample data enhancement of the No. 8 wind farm in 2022 is used as the training set, the real data of the cold wave weather in 2022 of the farm is used as the validation set, and the data of the No. 1 and No. 2 wind farms in 2022 are used as the test set to evaluate the effectiveness and generalization ability of the method of the invention.

[0121] The prediction performance evaluation indicators are as follows, where I NMAE and I NRMSE It is used to evaluate the accuracy of wind power forecasting, and QR is used to evaluate the accuracy of loss period prediction.

[0122] Mean absolute error:

[0123]

[0124] Root Mean Square Error:

[0125]

[0126] Accuracy:

[0127]

[0128] It indicates the ratio of the correct time points predicted during the loss period to the total time points during the entire period. The larger the value, the more effective the prediction.

[0129] In the above formula: N is the number of data; P i is the actual wind power value at time i; is the predicted wind power value at time i; is the predicted state encoding value at time t; o t is the actual state encoding value at the next time t;

[0130] The training and verification of the model of the present invention are implemented in the TensorFlow and PyTorch deep learning frameworks.

[0131] The present invention uses double Q graph and scatter plot to show the effect of generating small sample data of cold wave weather. Figure 7-10 As shown. The vertical and horizontal axes of the double Q graph represent the quantiles of the generated samples and the original samples, respectively. If the points on the graph are distributed near the diagonal line passing through the origin, the probability distributions of the two variables are more similar. The present invention uses the SeqVAE algorithm to generate and expand the weather and power data under the cold wave weather event of Wind Farm No. 8 in 2021. The original sample contains 6 cold wave weathers with a cumulative duration of 268 hours and a time resolution of 15 minutes, totaling 1067 samples, each of which contains 5 attributes: temperature, wind direction, wind speed, air pressure and power. The total number of samples generated by the expanded data reached 42254 samples. The double Q graphs of wind speed, temperature and power are shown in the figure respectively. It can be observed that the distribution of real meteorological samples and samples generated using SeqVAE model data is roughly the same, which shows that the SeqVAE algorithm used in the present invention has learned the meteorological and power characteristics under cold wave weather very well.

[0132] In order to verify the effectiveness of the method proposed in the present invention, taking a cold wave event from 5:00 on December 16, 2021 to 17:00 on December 18, 2021 at Wind Farm No. 1 as an example, the wind power in this period was predicted.

[0133] As shown in the figure, there was no power loss in the three periods from 5:00 on December 16, 2021 to 13:00 on December 16, 2021, from 8:30 on December 17, 2021 to 13:00 on December 17, 2021, and from 9:00 on December 18, 2021 to 13:00 on December 17, 2021, which are normal weather periods. The curve for power prediction using LightGBM is basically consistent with the true power curve. In the two periods from 13:00 on December 16, 2021 to 8:30 on December 17, 2021, and from 8:30 on December 17, 2021 to 9:00 on December 18, 2021, power loss occurred under the influence of cold wave weather. The LightGBM-Transformer combined model was used for power prediction, and the power prediction curve obtained was basically consistent with the true power curve without obvious errors.

[0134] At the same time, a single model LightGBM, a simple combined prediction model LightGBM+Transformer without loss period extraction, and S2S were used to predict the loss period, and XGBoost+Transformer was used for combined prediction, as well as the method used in the present invention for comparative experiments. Fig.11 It can be seen that compared with the single prediction method, the prediction accuracy of the simple combined prediction model without loss period extraction has been improved to a certain extent. Using GCN-BiGRU to extract the power loss period can not only capture the time series information, but also use graph convolution GCN to capture the spatial series information in the data, so the method of the present invention shows better prediction accuracy.

[0135] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0136] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined in the present invention may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown in the present invention, but will conform to the widest scope consistent with the principles and novel features disclosed in the present invention.

Claims

1. A super short-term wind power segmented forecasting method taking into account extreme cold wave weather samples, characterized in that: include: Obtain historical wind farm NWP data and corresponding power data to generate a data set; Based on the definition of cold wave events, data discrimination is performed to divide the data in the dataset into normal weather samples and cold wave weather samples; The normal weather samples are input into the LightGBM model, the LightGBM model is trained, and a normal weather power prediction model is obtained, and the wind power prediction value of normal weather is obtained through the normal weather power prediction model; the cold wave weather samples are input into the preset SeqVAE encoder for sample generation, and then the generated samples and the normal weather samples are assigned wind power loss states to obtain assigned samples with completed assignments; Input the assigned samples into the GCN-BiGRU power loss period extraction model, capture the spatiotemporal characteristics of the assigned samples, and extract the cold wave power loss period; input the cold wave power loss period and the corresponding power data into the Transformer loss value prediction model for training to obtain a power loss prediction model, and predict the power loss according to the power loss prediction model; obtain an ultra-short-term wind power prediction value according to the wind power and the power loss in normal weather; The parameters of the SeqVAE encoder are optimized by maximizing ELBO as follows: ELBO=E z～q(m|x) [logp θ (x|m)]-KL(q(m|x)||p(m))≤logp(x); Among them, E z～q(m|x) is the expected log-likelihood of the observed data x given the latent variable m, p(m) is the prior distribution of the latent variable m, usually the standard normal distribution N(0,l), and p(x) is the marginal likelihood of the observed data x.

2. The ultra-short-term wind power segmented forecasting method taking into account extreme cold wave weather samples according to claim 1 is characterized in that: The obtained assigned samples specifically include the generated samples and normal weather samples including wind speed, wind direction, temperature, humidity, and air pressure data. The time series in the samples are encoded, and the power loss data in cold wave weather is encoded as 0, and the normal weather data is encoded as 1, as follows: Among them, β is the percentage of climbing rate, C p is the operating capacity of the wind farm, is the power value of the normal weather sample using LightGBM, μ is the power loss percentage, P t is the power value at the current moment, and h is hours.

3. The ultra-short-term wind power segmented forecasting method taking into account extreme cold wave weather samples according to claim 1 is characterized in that: The assigned sample is input into GCN to obtain a feature vector with spatial features, as follows: Where I is the identity matrix, is the self-circular correlation matrix, is the normalized adjacency matrix, D is the degree matrix, and D ii represents the diagonal elements of the degree matrix D, n k is the number of co-convolutional layers in the graph, σ k () is the activation function of the graph convolutional layer, is the weight matrix of the i-1th layer, X Feature is the feature matrix, represents the feature matrix of the i-th layer node, Indicates the degree of node i.

4. The ultra-short-term wind power segmented forecasting method taking into account extreme cold wave weather samples according to claim 1 is characterized in that: BiGRU consists of two independent GRU units running in opposite directions, one processing data forward according to the time series, and the other processing data backward. Through the bidirectional structure, the BiGRU model can capture the forward and backward information of the sequence data at the same time.

5. An ultra-short-term wind power segmented prediction system taking into account extreme cold wave weather samples, applied to an ultra-short-term wind power segmented prediction method taking into account extreme cold wave weather samples as described in any one of claims 1 to 4, characterized in that: include: The data identification module obtains historical wind farm NWP data and corresponding power data to generate a data set; Based on the definition of cold wave time, data discrimination is performed to divide the data in the data set into normal weather samples and cold wave weather samples; A sample assignment module inputs the normal weather sample into the LightGBM model, trains the LightGBM model, obtains a normal weather power prediction model, and obtains a normal weather wind power prediction value through the normal weather power prediction model; inputs the cold wave weather sample into the preset SeqVAE encoder for sample generation, and then assigns the generated sample and the normal weather sample to the wind power loss state to obtain the assigned sample; a power prediction module inputs the assigned sample into the GCN-BiGRU loss power period extraction model, captures the spatiotemporal characteristics of the assigned sample, and extracts the cold wave loss power period; the cold wave loss power period and the corresponding power data are input into the Transformer loss value prediction model for training to obtain a loss power prediction model, and the loss power is predicted according to the loss power prediction model; and an ultra-short-term wind power prediction value is obtained according to the normal weather wind power and the loss power.

Images