Method and System for Wind Power Forecasting Correction Based on Cold Wave and Icing Shutdown Capacity

Abstract

A method and system for predicting and correcting wind power capacity under cold waves, taking into account icing shutdown capacity, includes: acquiring historical wind power and meteorological data of a wind farm; fitting historical wind speed and wind power into a wind speed-power curve; inputting wind speed into the curve to obtain ideal wind power; extracting icing power based on the deviation rate between ideal wind power and historical wind power, and meteorological data, with the remainder used as conventional power; estimating shutdown capacity and obtaining the final icing dataset through interpolation and dataset expansion; using conventional power and corresponding meteorological data as the conventional dataset to train a wind power prediction model and a wind turbine icing shutdown capacity prediction model; inputting real-time collected meteorological data into the two models to predict wind power and shutdown capacity; if the current meteorological data meets the meteorological conditions, correcting the predicted wind power. This invention effectively considers the impact of icing shutdown capacity on wind turbine handling, thereby improving prediction accuracy.

Description

Technical Field

[0001] This invention belongs to the field of wind power prediction technology, and more specifically, relates to a method and system for correcting wind power prediction based on the capacity of wind turbines shut down due to icing during cold waves. Background Technology

[0002] The highly random and fluctuating nature of wind power output contradicts the rigid requirements of real-time supply-demand balance and frequency and voltage stability in the power system, a contradiction that becomes increasingly prominent with the increase in installed wind power capacity. Therefore, research on wind power forecasting methods has become a key challenge in current wind power forecasting. Cold waves often accompany icing of wind turbine blades, causing wind turbines to experience reduced output due to the additional load. In such cases, the turbine output is not strongly correlated with conventional meteorological factors for power forecasting, such as wind speed. Considering only wind speed and direction data will introduce significant errors, threatening grid security.

[0003] Currently, there are three main problems with wind power prediction during cold waves: 1) It is difficult to determine whether icing will occur during a cold wave, and it is difficult to extract samples of icing moments from historical wind turbine output data; 2) There are few samples of wind turbines experiencing icing during cold waves, and sample expansion needs to be considered; 3) When humidity, temperature, and wind speed reach certain conditions during a cold wave, icing may occur, at which point wind turbine operating efficiency will be low or even shut down, which no longer conforms to the conventional power prediction rules, and current research results lack specificity in considering shutdown capacity. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for predicting and correcting wind power capacity under cold waves and considering icing shutdown capacity.

[0005] The present invention adopts the following technical solution.

[0006] The first aspect of this invention proposes a method for correcting wind power prediction based on wind power capacity during cold waves and icing shutdowns, including: The historical wind power and corresponding meteorological data for the first cycle set by a certain wind farm are obtained. The meteorological data includes temperature, humidity and wind speed, and the historical wind power is preprocessed. The historical wind speed and corresponding wind power are fitted to a set piecewise cubic polynomial, and the piecewise cubic polynomial is used as the wind speed-power curve. Historical wind speeds are input into the wind speed-power curve to obtain ideal wind power; based on the deviation rate between the ideal wind power and the corresponding historical wind power, and whether the meteorological data meets the set meteorological conditions, icing power is extracted from all historical wind power, and other historical wind power besides icing power is taken as conventional power. Based on the ratio of ideal wind power to historical wind power, the outage capacity is estimated. The meteorological data and the corresponding outage capacity corresponding to the extracted icing power are interpolated to generate continuous time-series meteorological data and outage capacity data as the original icing dataset. The original icing dataset is sample-expanded to obtain the final icing dataset. The normal power and the corresponding meteorological data are used as the normal dataset. A wind power prediction model was trained based on a regular dataset, and a wind turbine icing shutdown capacity prediction model was trained based on the final icing dataset. The meteorological data collected in real time within the second cycle prior to the current moment are input into the wind power prediction model and the wind turbine icing shutdown capacity prediction model, respectively. The predicted wind power and shutdown capacity are output. If the meteorological data at the current moment meets the set meteorological conditions, the predicted wind power is corrected by the difference between the predicted shutdown capacity and the installed capacity.

[0007] Preferably, the step of fitting historical wind speeds and corresponding wind power into a set piecewise cubic polynomial specifically involves: Multiply the historical wind speed by the linear coefficient and add the bias term to obtain the intermediate coefficient; When the intermediate coefficient is located in the first set interval, the wind power is the first cubic polynomial of the intermediate coefficient; When the intermediate coefficient is located in the second set interval, the wind power is the second cubic polynomial of the intermediate coefficient; The coefficients of the cubic term, the coefficients of the quadratic term, and the constant term are all different in the first and second cubic polynomials.

[0008] Preferably, the step of extracting icing power from all historical wind power specifically involves: If the deviation rate between the ideal wind power and the corresponding historical wind power exceeds the set deviation threshold and the corresponding meteorological data meets the set meteorological conditions, then the corresponding historical wind power is the icing power. The deviation rate is the absolute value of the difference between the ideal wind power and the corresponding historical wind power divided by the ideal wind power. The set meteorological conditions are that the temperature is below a set temperature threshold, the humidity is above a set humidity threshold, the wind speed exceeds a set wind speed threshold, and the rate of change of the wind speed is less than a set change threshold.

[0009] Preferably, the outage capacity is estimated based on the ratio of ideal wind power to historical wind power, specifically: Subtract the ratio of ideal wind power to historical wind power from 1, and multiply the result by the corresponding installed capacity to obtain the estimated outage capacity.

[0010] Preferably, the original icing dataset is augmented to obtain the final icing dataset, specifically as follows: The original icing dataset is input into the TimeVAE model embedded in the GRU layer. The original icing dataset consists of multiple icing vectors, each of which is a vector composed of meteorological data and corresponding downtime capacity for a time step. The GRU layer is embedded in the encoder of the TimeVAE model. The GRU layer processes the entire original icing dataset step by step to obtain the hidden state for each time step. The hidden state of the final time step is input into two parallel fully connected layers using the LeakyReLU activation function to output the mean and log-variance of the latent variables. The standard deviation is calculated based on the log-variance. The Hadamard product of the mean, the standard deviation, and a random noise vector that follows a standard normal distribution is calculated as the latent variable. The latent variables are input into the decoder of the TimeVAE model to generate the final icing dataset.

[0011] Preferably, the model parameters in the improved TimeVAE model with embedded GRU layers are obtained by training on the original icing dataset, and the loss function is the reconstruction loss minus the set KL divergence weight coefficient multiplied by the KL divergence loss.

[0012] Preferably, a wind power prediction model is trained based on a conventional dataset, and a wind turbine icing shutdown capacity prediction model is trained based on the final icing dataset, specifically as follows: The wind power prediction model is an improved iTransformer model, which is improved by adding a TCN model for convolution and unifying the activation function to LeakyReLU; the wind power prediction model takes meteorological data as input and outputs the predicted wind power. The wind turbine icing shutdown capacity prediction model is an x-LSTM model, with meteorological data as input and predicted shutdown capacity as output.

[0013] Preferably, the predicted wind power output is corrected using the difference between the predicted outage capacity and the installed capacity, specifically as follows: The corrected wind power is the predicted wind power multiplied by the difference between the corresponding installed capacity and the predicted outage capacity. The predicted wind power is the predicted normalized wind power, which is equal to the actual wind power divided by the corresponding installed capacity.

[0014] The second aspect of this invention proposes a wind power prediction and correction system based on the method described in the first aspect of this invention, which includes a data acquisition module, a wind speed-power curve fitting module, an icing power extraction module, a dataset construction module, a model training module, and a wind power prediction and correction module, specifically: Data acquisition module: used to acquire the historical wind power and corresponding meteorological data within the first cycle set by a certain wind farm. The meteorological data includes temperature, humidity and wind speed, and preprocesses the historical wind power. Wind speed-power curve fitting module: used to fit historical wind speed and corresponding wind power into a set piecewise cubic polynomial, and the piecewise cubic polynomial is used as the wind speed-power curve. Icing power extraction module: It is used to input historical wind speed into the wind speed-power curve to obtain ideal wind power; based on the deviation rate between the ideal wind power and the corresponding historical wind power and whether the meteorological data meets the set meteorological conditions, it extracts icing power from all historical wind power, and other historical wind power other than icing power is used as regular power. Dataset construction module: It is used to estimate the outage capacity based on the ratio of ideal wind power to historical wind power. The extracted meteorological data corresponding to the icing power and the corresponding outage capacity are used as the original icing dataset. The original icing dataset is expanded to obtain the final icing dataset. The regular power and corresponding meteorological data are used as the regular dataset. Model training module: Train a wind power prediction model based on a regular dataset, and train a wind turbine icing shutdown capacity prediction model based on the final icing dataset; Wind power prediction and correction module: The real-time meteorological data collected before the current time within the second cycle is input into the wind power prediction model and the wind turbine icing shutdown capacity prediction model, respectively. The predicted wind power and shutdown capacity are output, and the predicted wind power is corrected by the difference between the predicted shutdown capacity and the installed capacity.

[0015] A third aspect of the invention provides an apparatus comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, the processor performing steps of the wind power prediction correction method for considering ice-induced shutdown capacity during cold waves as described in the first aspect of the invention.

[0016] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, uses the steps of the wind power prediction correction method for considering ice-induced shutdown capacity during cold waves as described in the first aspect of the present invention.

[0017] The beneficial effects of the present invention are that, compared with the prior art, (1) the present invention fully considers the sample quality, considers the cold wave icing meteorological conditions to screen the icing dataset, divides the data into two categories, which facilitates the subsequent classification modeling process, thereby improving the accuracy and reliability of the prediction results; (2) This invention constructs a calculation formula for wind turbine icing shutdown capacity, expands the icing dataset, establishes a mapping relationship between meteorological data and icing shutdown capacity, realizes data on shutdown capacity under different meteorological conditions, and provides high-quality data for subsequent shutdown capacity prediction; (3) The present invention constructs a wind power prediction model based on a conventional dataset and a shutdown capacity prediction model based on an icing dataset, respectively, and outputs power prediction values ​​and shutdown capacity prediction values. Based on the prediction results of the two models, a correction formula under icing conditions is constructed to effectively take into account the impact of icing shutdown capacity on wind turbine handling, thereby improving the prediction accuracy and ensuring the stable operation of the system when there is icing under cold wave. Attached Figure Description

[0018] Figure 1 This is a scatter plot of wind speed and power. Figure 2 Here is a flowchart for extracting icing power; Figure 3 For time variational autoencoder architecture; Figure 4 To improve the iTransformer architecture; Figure 5 This is a flowchart of the present invention; Figure 6 Double Q plots before and after sample augmentation; Figure 7 A comparison chart of wind power predictions from different models. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0020] like Figure 5 As shown, Embodiment 1 of the present invention proposes a method for correcting wind power prediction based on icing shutdown capacity during cold waves, comprising: S1. Obtain the historical wind power and corresponding meteorological data for the first cycle set by a certain wind farm station. The meteorological data includes temperature, humidity and wind speed, and preprocess the historical wind power. It should be noted that the preprocessing of historical wind power is to divide the wind power by the installed capacity of the corresponding wind turbine to obtain the normalized wind power. All wind power mentioned thereafter refers to the normalized wind power.

[0021] Specifically, the first period is set to a year or several years. In this embodiment, meteorological factors (humidity, temperature, wind speed) and wind power of a 64MW wind farm in a certain province from November 2022 to January 2024 were collected. This includes a small amount of historical data during cold waves. The sampling period is 15 minutes, with a total of 35,040 data points. When collecting raw meteorological data and historical power data, it is necessary to ensure that the time span and sampling frequency of the data are consistent and cover a sufficiently long time period to include different weather conditions and seasonal changes.

[0022] S2, such as Figure 1 As shown, the historical wind speed and corresponding wind power are fitted to a set piecewise cubic polynomial, and the piecewise cubic polynomial is used as the wind speed-power curve. In this preferred embodiment, the step of fitting the historical wind speed and corresponding wind power into a set piecewise cubic polynomial specifically involves: Multiply the historical wind speed by the linear coefficient and add the bias term to obtain the intermediate coefficient; When the intermediate coefficient is located in the first set interval, the wind power is the first cubic polynomial of the intermediate coefficient; When the intermediate coefficient is located in the second set interval, the wind power is the second cubic polynomial of the intermediate coefficient; The coefficients of the cubic term, the coefficients of the quadratic term, and the constant term are all different in the first and second cubic polynomials.

[0023] Specifically, the formula is:

[0024] in, Indicates ideal wind power output. As an intermediate coefficient, For wind speed, The coefficients are linear. For bias terms, , , These are the coefficients of the cubic term, the coefficients of the quadratic term, and the constant term of the first type of cubic polynomial; , , These are the coefficients of the cubic term, the coefficients of the quadratic term, and the constant term in the second type of cubic polynomial; For the first defined interval, The upper limit of the first interval is set. The lower limit of the first interval is set; other cases (i.e.) () represents the second defined interval. , , , , , , and All parameters are obtained through fitting.

[0025] S3. Input historical wind speeds into the wind speed-power curve to obtain ideal wind power; such as Figure 2 As shown, based on the deviation rate between the ideal wind power and the corresponding historical wind power, and whether the meteorological data meets the set meteorological conditions, the icing power is extracted from all historical wind power, and the other historical wind power besides the icing power is taken as the conventional power. In this preferred embodiment, the step of extracting icing power from all historical wind power specifically involves: If the deviation rate between the ideal wind power and the corresponding historical wind power exceeds the set deviation threshold and the corresponding meteorological data meets the set meteorological conditions, then the corresponding historical wind power is the icing power. The deviation rate is the absolute value of the difference between the ideal wind power and the corresponding historical wind power, divided by the ideal wind power; deviation rate The calculation formula is:

[0026] in, Let be the i-th actual wind power output.

[0027] The set meteorological conditions are that the temperature is below a set temperature threshold, the humidity is above a set humidity threshold, the wind speed exceeds a set wind speed threshold, and the rate of change of the wind speed is less than a set change threshold.

[0028] It should be noted that there is usually a small deviation between the actual and theoretical wind power output. In rare cases, cold waves can cause icing on wind turbine blades. In such situations, the wind turbines may experience reduced output or even shut down due to freezing due to the additional load. In these cases, the turbine output is not significantly correlated with conventional meteorological factors such as wind speed, resulting in a large deviation between the actual and theoretical power output, generally exceeding 50% of the theoretical value. Therefore, the deviation rate threshold is set at 50%.

[0029] S4. Estimate the outage capacity based on the ratio of ideal wind power to historical wind power. Interpolate the meteorological data and outage capacity corresponding to the extracted icing power to generate continuous time-series meteorological data and outage capacity data as the original icing dataset. Expand the original icing dataset to obtain the final icing dataset. Use the regular power and corresponding meteorological data as the regular dataset. In this preferred embodiment, the outage capacity is estimated based on the ratio of ideal wind power to historical wind power, specifically as follows: Subtract the ratio of ideal wind power to historical wind power from 1, and multiply the result by the corresponding installed capacity to obtain the estimated outage capacity.

[0030] In this preferred embodiment, the original icing dataset is augmented to obtain the final icing dataset, specifically as follows: The original icing dataset is input into the TimeVAE (Time Variational Autoencoder) model, which is embedded in a GRU layer. The TimeVAE model architecture is as follows. Figure 3 As shown, the original icing dataset consists of multiple icing vectors. Each icing vector is a vector composed of meteorological data at a time step and the corresponding downtime capacity. The GRU layer is embedded in the encoder of the TimeVAE model. The GRU layer processes the entire original icing dataset step by step to obtain the hidden state corresponding to each time step. The hidden state of the final time step is input into two parallel fully connected layers using the LeakyReLU activation function to output the mean and log-variance of the latent variables respectively. The standard deviation is calculated based on the log-variance. The mean plus the standard deviation and the Hadamard product of the random noise vector that follows the standard normal distribution are calculated as the latent variable. The latent variables are input into the decoder of the TimeVAE model to generate the final icing dataset.

[0031] It should be noted that the TimeVAE model is an improved variational autoencoder specifically designed for time series optimization. It introduces a time-aware structure on top of VAE, enabling it to better capture trends, cycles, and fluctuations over time, making it more suitable for scenarios with strong temporal sequences, strong coupling, and non-stationarity, such as wind turbine operation data. However, the basic CNN in TimeVAE struggles to capture long-range temporal dependencies and multi-scale patterns, and still has limitations in capturing long-range dependencies in wind turbine icing-related time series data. The Gate Recurrent Unit (GRU), with its simple gating mechanism, can effectively alleviate the gradient vanishing problem of traditional recurrent neural networks and accurately capture the dynamic changes accumulated over time in time series. Therefore, using GRU to improve the TimeVAE architecture can significantly improve the model's performance in modeling wind turbine icing time series data. To improve model training stability and avoid uneven gradient propagation caused by different activation functions, the activation function of the fully connected layers in the TimeVAE encoder and decoder is unified to the LeakyReLU activation function, such as... Figure 3 The diagram shows the structure of the improved variational autoencoder.

[0032] Specifically, a GRU layer is embedded before the fully connected layer at the input of a traditional TimeVAE encoder, thereby more accurately inferring the latent variables at the current time based on historical information. The sequence in the icing dataset input to the encoder is as follows:

[0033] In the formula: W t For meteorological data, To indicate the dimension of its superscript, meteorological data dimensions D W It is three-dimensional, including wind speed. v t ,humidity l t ,temperature q t ; C t Shutdown capacity; W ext The allowable set (i.e., constraint space) of meteorological data and wind power for the joint sample space under cold wave icing conditions. t For time step; T The length of the time series; X Input data for the model (raw icing dataset); D Let be the total dimension of the input feature vector, and D = D W +1=4.

[0034] Before input, the sequence is processed by the Min-Max normalization module to normalize the sequence. X Standardization processing is performed to eliminate the dimensional differences between meteorological data and downtime capacity. The standardized sequence is as follows. for:

[0035]

[0036] in, After standardization .

[0037] Then, the GRU layer processes the input... Time-step processing is the core of the improved variational autoencoder, capturing long-range dependencies in the cold wave icing time series data while discarding invalid short-term fluctuations. The final output is the hidden state, which integrates the dynamic features of the entire time series. h T To provide high-quality temporal features for subsequent potential spatial mapping, the following formula represents the calculation process:

[0038] In the formula: r t To reset the door, To reset the gate input weights, , All are fixed values. To reset the weight matrix of the hidden state of the door, To reset the door's bias, The sigmoid activation function satisfies , Indicates the independent variable; z t To update the door, To update the gate input weights, To update the hidden state weight matrix of the gate, To update the gate's bias term; In the candidate hidden state, Input weights for the candidate hidden states. Let be the hidden state weight matrix of the candidate hidden states. The bias term for the candidate hidden state. It represents the Hadamardi (or Hadama) stack; It is the hyperbolic tangent function; h t The current hidden state, 1- z t Indicates the percentage of historical hidden states that are retained. z t This represents the update ratio of the current candidate hidden state. The weighted sum of the two ensures the transmission of long-term time series information and timely updates of new features at the current moment, adapting to the dynamic characteristics of cold wave icing time series data.

[0039] Next, the data enters the fully connected layer mapping stage, setting the final hidden state dimension of the GRU layer output. H =32, dimensionality compression is performed through a fully connected layer, mapping to a low-dimensional latent space while outputting the mean and log-variance of the latent variables. This step uses two independent fully connected layers to map the mean and log-variance respectively, and the activation function is uniformly improved to the LeakyReLU activation function to avoid the dead neuron problem of ReLU and improve gradient propagation efficiency.

[0040] In the formula: , These are the weight matrix and bias term of the mean-connected layer, respectively. , Here are the weight matrix and bias terms for the log-variance fully connected layer; KAs a potential spatial dimension, and K < H To achieve dimensional compression; For the activation function, satisfying a = 0.01 is a negative slope.

[0041] Then, latent variable sampling is performed. In variational inference, latent variables... z Need to be from the posterior distribution The latent variables are obtained through sampling, but the direct sampling process is non-differentiable, making it impossible to train the model via backpropagation. Therefore, a reparameterization technique is employed to transform the sampling process into a differentiable operation, ensuring that the model can be trained normally while preserving the Gaussian distribution characteristics of the latent variables.

[0042] In the formula: The standard deviation of the latent variable, The variance is logarithmic. The mean of the latent variables; Let be a random noise vector that follows a standard normal distribution, where The identity matrix is ​​used, and the introduction of noise can enhance the generalization ability of the model and avoid overfitting.

[0043] In this embodiment, the model parameters in the improved TimeVAE model with embedded GRU layer are obtained by training on the original icing dataset, and the loss function is the reconstruction loss minus the set KL divergence weight coefficient multiplied by the KL divergence loss.

[0044] Specifically, the loss function It consists of two parts: reconstruction loss and KL divergence, expressed as:

[0045] In the formula: For all learnable parameters of the encoder; represents the learnable parameters of the GRU layer, ensuring that the parameters of the GRU layer and the fully connected layer are updated synchronously during training; F represents the learnable parameters of the decoder, which are trained independently with the encoder parameters to jointly minimize the overall loss function. For reconstruction loss, mean squared error loss is used. Given the potential encoding z, The conditional probability distribution of occurrence; Let KL divergence be the KL divergence. The prior distribution of the latent code z is used to measure the difference in the posterior distribution of the encoder output, and plays a regularization role to avoid model overfitting. The KL divergence weight coefficient is used to balance the reconstruction loss and the regularization term, adapting to the characteristics of cold wave icing data.

[0046] The encoder and decoder are trained jointly. Through backpropagation, the parameters of the encoder's GRU layer, fully connected layer, and decoder are updated synchronously to minimize the overall loss function. Ultimately, the closed loop is achieved by the GRU layer accurately extracting temporal dependent features, the fully connected layer reasonably mapping them to the latent space, and the decoder accurately reconstructing the input sequence.

[0047] The quality of sample expansion is evaluated by generating a double-Q plot before and after expansion, as well as a heat map showing the correlation between wind speed, humidity, temperature, and shutdown capacity. Figure 6 To expand the double-Q plot before and after the sample expansion, a comparison revealed that the distribution of the sample and the correlation between various meteorological factors and the shutdown capacity did not change significantly before and after the expansion.

[0048] S5. Train a wind power prediction model based on a regular dataset, and train a wind turbine icing shutdown capacity prediction model based on the final icing dataset. In this preferred embodiment, the wind power prediction model is trained based on a conventional dataset, and the wind turbine icing shutdown capacity prediction model is trained based on the final icing dataset. Specifically: like Figure 4 As shown, the wind power prediction model is an improved iTransformer model, which is improved by adding a TCN model for convolution and unifying the activation function to LeakyReLU; the wind power prediction model takes meteorological data as input and outputs the predicted wind power.

[0049] It should be noted that this model, while retaining the core structure (Softmax in the attention mechanism and gated Sigmoid in TCN), uniformly replaces all optional nonlinear activation layers, including those after TCN convolution, the Transformer feedforward network, and the output head, with LeakyReLU. This avoids the neuron death problem caused by ReLU and is consistent with the extended model. It is particularly suitable for low wind speed segments (sparse gradients) and abrupt change segments (requiring negative responses) commonly found in wind power data, such as... Figure 4 The diagram shows the structure of the improved power prediction model. The following equation illustrates the principle of the improved conventional power prediction model:

[0050]

[0051]

[0052]

[0053] In the formula: for tThe original input features at any given time include normalized multi-source data such as wind speed, temperature, humidity, and historical power. W embed This is the learnable weight matrix of the embedding layer, responsible for mapping low-dimensional original features to a high-dimensional feature space; b embed The bias term of the embedding layer; Embed is the embedded feature; W couple This is the weight matrix of the coupled coding layer; b couple This corresponds to the bias term; the enhanced coupling timing feature of Enc0 is the initial input of the encoder; f scal For adaptive multi-scale convolutional layer output; s The kernel size is {3, 5, 7}. For convolution kernel; These are the weight coefficients corresponding to the convolution kernel; The input is the temporal feature after initial encoding, and it is also the output of the iTansformer transpose layer. b conv,s The bias term corresponds to the convolution kernel; Conv(·) is a one-dimensional temporal convolution operation, the purpose of which is to extract local temporal variation features; f corr These are time-series autocorrelation features, and long-distance periodic dependence features extracted through the autocorrelation mechanism. W 1 ，W 2 represents the linear transformation weights of the two features, used to align the dimensions of the two features and achieve feature fusion; b fusion For the bias term of the fusion layer; MHA out To invest more attention in output; b 1, b 2 represents the linear layer bias term; FFN out This is the final output of the feedforward network in iTransformer.

[0054] The wind turbine icing shutdown capacity prediction model is an x-LSTM model, with meteorological data as input and predicted shutdown capacity as output.

[0055] x-LSTM includes sLSTM and mLSTM structures, with a clear division of labor. It can preserve global features for a long time while balancing efficiency and accuracy.

[0056] 1) Compared to LSTM, the sLSTM model uses normalized states, hidden states for recurrent connections, input gates to control the extent to which new information is added to the model, forget gates to control the extent to which previous unit states are retained, and output gates to control the flow of information from unit states to hidden states. The forget gate is combined with the input gate to form another state to provide numerical stability of the exponential gate within the model. 2) Compared with traditional LSTM, mLSTM's scalar cell states can store more complex structured information. The input gate controls the merging of new information into memory, and the forget gate determines how much previous memory is retained.

[0057] S6. Input the meteorological data collected in real time within the second cycle before the current time into the wind power prediction model and the wind turbine icing shutdown capacity prediction model respectively, and output the predicted wind power and shutdown capacity. If the meteorological data at the current time meets the set meteorological conditions, the predicted wind power is corrected by the difference between the predicted shutdown capacity and the installed capacity.

[0058] In this preferred embodiment, the predicted wind power output is corrected using the difference between the predicted outage capacity and the installed capacity, specifically as follows: The corrected wind power is the predicted wind power multiplied by the difference between the corresponding installed capacity and the predicted outage capacity. The predicted wind power is the predicted normalized wind power, which is equal to the actual wind power divided by the corresponding installed capacity.

[0059] The formula is:

[0060] in, For the corresponding installed capacity, For the predicted wind power, For the predicted outage capacity, This represents the corrected wind power output.

[0061] Figure 7 A comparison chart of day-ahead wind power prediction using various models is provided. M0 represents the prediction model of this invention, M1 is the TCN-Transformer prediction model, M2 is the iTransformer prediction model, and M3 is the Transformer prediction model. Table 1 shows the calculated mean absolute error, root mean square error, and coefficient of determination for the power prediction results of the four models. Based on the following results, the mean absolute error of this model is approximately 10%, the root mean square error is within 13%, and R0... 2 Closest to 1.

[0062] Table 1 Power Prediction Evaluation Results

[0063] To further verify the predictive performance of this invention, mean absolute error, root mean square error, and coefficient of determination are still used as evaluation indicators.

[0064] The shutdown capacity of wind turbines for the next 5 days is predicted. Table 2 shows the calculated mean absolute error, root mean square error, and coefficient of determination of the power prediction results from three prediction models. Based on the following results, the mean absolute error of this model is within 5%, and the root mean square error is within 15%.

[0065] Table 2 Comparison of prediction performance before and after correction

[0066] Table 3 shows the relevant evaluation indicators. It can be seen that the prediction error is effectively reduced after correction, and the coefficient of determination is closer to 1. The wind power prediction correction model considering the icing shutdown capacity has a good prediction effect. The mean absolute error and root mean square error both decreased by about 15% before and after correction. The final mean absolute error is within 5% and the root mean square error is within 10%, indicating a good prediction effect. The prediction evaluation before and after correction is shown in Table 3.

[0067] Table 3 Prediction and Evaluation Before and After Correction

[0068] Embodiment 2 of the present invention proposes a wind power prediction and correction system based on the method described in Embodiment 1 of the present invention, which includes a data acquisition module, a wind speed-power curve fitting module, an icing power extraction module, a dataset construction module, a model training module, and a wind power prediction and correction module, specifically as follows: Data acquisition module: used to acquire the historical wind power and corresponding meteorological data within the first cycle set by a certain wind farm. The meteorological data includes temperature, humidity and wind speed, and preprocesses the historical wind power. Wind speed-power curve fitting module: used to fit historical wind speed and corresponding wind power into a set piecewise cubic polynomial, and the piecewise cubic polynomial is used as the wind speed-power curve. Icing power extraction module: It is used to input historical wind speed into the wind speed-power curve to obtain ideal wind power; based on the deviation rate between the ideal wind power and the corresponding historical wind power and whether the meteorological data meets the set meteorological conditions, it extracts icing power from all historical wind power, and other historical wind power other than icing power is used as regular power. Dataset construction module: It is used to estimate the outage capacity based on the ratio of ideal wind power to historical wind power. The extracted meteorological data corresponding to the icing power and the corresponding outage capacity are used as the original icing dataset. The original icing dataset is expanded to obtain the final icing dataset. The regular power and corresponding meteorological data are used as the regular dataset. Model training module: Train a wind power prediction model based on a regular dataset, and train a wind turbine icing shutdown capacity prediction model based on the final icing dataset; Wind power prediction and correction module: The real-time meteorological data collected before the current time within the second cycle is input into the wind power prediction model and the wind turbine icing shutdown capacity prediction model, respectively. The predicted wind power and shutdown capacity are output, and the predicted wind power is corrected by the difference between the predicted shutdown capacity and the installed capacity.

[0069] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0070] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0071] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0072] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0073] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for modifying wind power prediction under cold wave considering icing shutdown capacity, characterized in that, include: The historical wind power and corresponding meteorological data for the first cycle set by a certain wind farm are obtained. The meteorological data includes temperature, humidity and wind speed, and the historical wind power is preprocessed. The historical wind speed and corresponding wind power are fitted to a set piecewise cubic polynomial, and the piecewise cubic polynomial is used as the wind speed-power curve. Historical wind speeds are input into the wind speed-power curve to obtain ideal wind power; based on the deviation rate between the ideal wind power and the corresponding historical wind power, and whether the meteorological data meets the set meteorological conditions, icing power is extracted from all historical wind power, and other historical wind power besides icing power is taken as conventional power. Based on the ratio of ideal wind power to historical wind power, the outage capacity is estimated. The meteorological data and the corresponding outage capacity corresponding to the extracted icing power are interpolated to generate continuous time-series meteorological data and outage capacity data as the original icing dataset. The original icing dataset is sample-expanded to obtain the final icing dataset. The normal power and the corresponding meteorological data are used as the normal dataset. A wind power prediction model was trained based on a regular dataset, and a wind turbine icing shutdown capacity prediction model was trained based on the final icing dataset. The meteorological data collected in real time within the second cycle prior to the current moment are input into the wind power prediction model and the wind turbine icing shutdown capacity prediction model, respectively. The predicted wind power and shutdown capacity are output. If the meteorological data at the current moment meets the set meteorological conditions, the predicted wind power is corrected by the difference between the predicted shutdown capacity and the installed capacity.

2. The wind power prediction correction method based on icing shutdown capacity during cold waves according to claim 1, characterized in that: The process of fitting historical wind speeds and corresponding wind power into a predefined piecewise cubic polynomial is as follows: Multiply the historical wind speed by the linear coefficient and add the bias term to obtain the intermediate coefficient; When the intermediate coefficient is located in the first set interval, the wind power is the first cubic polynomial of the intermediate coefficient; When the intermediate coefficient is located in the second set interval, the wind power is the second cubic polynomial of the intermediate coefficient; The coefficients of the cubic term, the coefficients of the quadratic term, and the constant term are all different in the first and second cubic polynomials.

3. The wind power prediction correction method based on icing shutdown capacity during cold waves according to claim 1, characterized in that: The extraction of icing power from all historical wind power output specifically involves: If the deviation rate between the ideal wind power and the corresponding historical wind power exceeds the set deviation threshold and the corresponding meteorological data meets the set meteorological conditions, then the corresponding historical wind power is the icing power. The deviation rate is the absolute value of the difference between the ideal wind power and the corresponding historical wind power divided by the ideal wind power. The set meteorological conditions are that the temperature is below a set temperature threshold, the humidity is above a set humidity threshold, the wind speed exceeds a set wind speed threshold, and the rate of change of the wind speed is less than a set change threshold.

4. The wind power prediction correction method based on icing shutdown capacity during cold waves according to claim 1, characterized in that: The estimated outage capacity is based on the ratio of ideal wind power to historical wind power, specifically: Subtract the ratio of ideal wind power to historical wind power from 1, and multiply the result by the corresponding installed capacity to obtain the estimated outage capacity.

5. The wind power prediction correction method based on icing shutdown capacity during cold waves according to claim 1, characterized in that: The original icing dataset was augmented to obtain the final icing dataset, specifically: The original icing dataset is input into the TimeVAE model embedded in the GRU layer. The original icing dataset consists of multiple icing vectors, each of which is a vector composed of meteorological data and corresponding downtime capacity for a time step. The GRU layer is embedded in the encoder of the TimeVAE model. The GRU layer processes the entire original icing dataset step by step to obtain the hidden state for each time step. The hidden state of the final time step is input into two parallel fully connected layers using the LeakyReLU activation function to output the mean and log-variance of the latent variables. The standard deviation is calculated based on the log-variance. The Hadamard product of the mean, the standard deviation, and a random noise vector that follows a standard normal distribution is calculated as the latent variable. The latent variables are input into the decoder of the TimeVAE model to generate the final icing dataset.

6. The wind power prediction correction method based on icing shutdown capacity during cold waves according to claim 5, characterized in that: The model parameters in the improved TimeVAE model with embedded GRU layers are obtained by training on the original icing dataset, and the loss function is the reconstruction loss minus the set KL divergence weight coefficient multiplied by the KL divergence loss.

7. The wind power prediction correction method based on icing shutdown capacity during cold waves according to claim 1, characterized in that: A wind power prediction model was trained based on a regular dataset, and a wind turbine icing shutdown capacity prediction model was trained based on the final icing dataset. Specifically: The wind power prediction model is an improved iTransformer model, which is improved by adding a TCN model for convolution and unifying the activation function to LeakyReLU; the wind power prediction model takes meteorological data as input and outputs the predicted wind power. The wind turbine icing shutdown capacity prediction model is an x-LSTM model, with meteorological data as input and predicted shutdown capacity as output.

8. The wind power prediction correction method based on icing shutdown capacity during cold waves according to claim 1, characterized in that: The predicted wind power output is corrected using the difference between the predicted outage capacity and the installed capacity, specifically as follows: The corrected wind power is the predicted wind power multiplied by the difference between the corresponding installed capacity and the predicted outage capacity. The predicted wind power is the predicted normalized wind power, which is equal to the actual wind power divided by the corresponding installed capacity.

9. A wind power prediction and correction system based on the method of any one of claims 1-8, considering the shutdown capacity due to icing during cold waves, comprising a data acquisition module, a wind speed-power curve fitting module, an icing power extraction module, a dataset construction module, a model training module, and a wind power prediction and correction module, characterized in that: Data acquisition module: used to acquire the historical wind power and corresponding meteorological data within the first cycle set by a certain wind farm. The meteorological data includes temperature, humidity and wind speed, and preprocesses the historical wind power. Wind speed-power curve fitting module: used to fit historical wind speed and corresponding wind power into a set piecewise cubic polynomial, and the piecewise cubic polynomial is used as the wind speed-power curve. Icing power extraction module: It is used to input historical wind speed into the wind speed-power curve to obtain ideal wind power; based on the deviation rate between the ideal wind power and the corresponding historical wind power and whether the meteorological data meets the set meteorological conditions, it extracts icing power from all historical wind power, and other historical wind power other than icing power is used as regular power. Dataset construction module: It is used to estimate the outage capacity based on the ratio of ideal wind power to historical wind power. The extracted meteorological data corresponding to the icing power and the corresponding outage capacity are used as the original icing dataset. The original icing dataset is expanded to obtain the final icing dataset. The regular power and corresponding meteorological data are used as the regular dataset. Model training module: Train a wind power prediction model based on a regular dataset, and train a wind turbine icing shutdown capacity prediction model based on the final icing dataset; Wind power prediction and correction module: The real-time meteorological data collected before the current time within the second cycle is input into the wind power prediction model and the wind turbine icing shutdown capacity prediction model, respectively. The predicted wind power and shutdown capacity are output, and the predicted wind power is corrected by the difference between the predicted shutdown capacity and the installed capacity.

10. An apparatus comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, the processor performing steps of the wind power prediction correction method for cold wave and icing shutdown capacity as described in any one of claims 1 to 8.

11. A computer-readable storage medium storing a computer program that, when executed by a processor, uses the steps of the wind power prediction correction method for considering icing shutdown capacity during cold waves as described in any one of claims 1 to 8.

Images