A short-term wind power section probability prediction method

Abstract

This invention relates to the field of power system operation and planning technology, specifically to a short-term wind power segment probability prediction method. It utilizes deep learning to mine implicit information in data and nonlinear features in wind power sequences to generate prediction probability intervals. Simultaneously, a nonlinear weighting method is selected to improve the optimization performance of the particle swarm optimization algorithm (IPSO), addressing some problems of traditional algorithms and improving convergence speed. Furthermore, a hybrid artificial intelligence algorithm, CNN-LSTM, is selected to construct a prediction model based on a combination of SVM and quantile regression (IPSO-CNN-LSTM). After training, short-term wind power probability prediction is completed. The CNN network can extract latent features from sample data using convolutional kernels, while the Long Short-Term Memory (LSTM) network can capture long-term components, avoiding gradient vanishing and exploding phenomena found in some existing algorithms, thus improving the efficiency of wind power probability prediction.

Description

Technical Field

[0001] This invention relates to the field of power system operation and planning technology, specifically to a short-term wind power probability prediction method. Background Technology

[0002] With the development of society and technology, the role of energy is increasingly important. However, rapid development is accompanied by the continuous consumption of traditional fossil fuels, leading to the increasing attention paid to renewable energy by scholars. In the power industry, optimizing the traditional energy structure, increasing the proportion of renewable energy, and improving grid connection efficiency have become key research directions in global energy development. Wind energy, with its cleanliness and abundant reserves, has attracted considerable attention. However, the instability and uncertainty of wind speed cause significant fluctuations in the power generation efficiency of wind farms, making it difficult to schedule wind power in advance. Accurate power prediction can effectively mitigate the impact of wind power uncertainty and ensure the safe operation of the power system. Currently, the new trend in wind power prediction is artificial intelligence methods. However, traditional single-point prediction models cannot quantify the irregularities and uncertainties of wind power, and shallow learning models cannot fully extract the deep nonlinear characteristics in wind power sequences. Single prediction models are insufficient to capture the changing patterns in wind power sequences and achieve satisfactory prediction results. Therefore, new solutions are urgently needed to address this problem. Summary of the Invention

[0003] The purpose of this invention is to provide a short-term wind power probability prediction method, which aims to solve the technical problems that traditional point prediction methods are difficult to extract the implicit information in the subsequence of wind power data sets and that single prediction models are difficult to capture the changing patterns in wind power sequences.

[0004] To achieve the above objectives, the present invention provides a short-term wind power probability prediction method, comprising the following steps:

[0005] Step 1: Collect and process historical meteorological data to obtain the initial dataset for wind power probability prediction;

[0006] Step 2: Filter and supplement the initial dataset to obtain a wind power dataset, and then perform normalization processing;

[0007] Step 3: Divide the wind power dataset into training and testing sets, construct a prediction model based on the IPSO-CNN-LSTM algorithm combining SVM and quantile regression, and train and predict it.

[0008] Step 4: Adjust the model parameters based on the prediction results and errors until the results are close to the test set, and complete the short-term wind power probability prediction.

[0009] Optionally, the process of collecting and processing historical meteorological data to obtain the initial dataset for wind power probability prediction involves acquiring historical meteorological data for the region, processing the raw meteorological data and wind conditions by querying local logs, meteorological station records, and records from other climate monitoring systems, and obtaining the initial dataset for wind power probability prediction based on the wind speed fluctuations in historical predicted power and the numerical weather prediction (NWP) results of wind farms.

[0010] Optionally, the data for the numerical weather prediction (NWP) uses meteorological data from a selected wind farm throughout the year, including parameters such as wind speed, wind direction, temperature, and air pressure.

[0011] Optionally, the process of filtering and supplementing the initial dataset to obtain a wind power dataset and performing normalization processing involves filtering out the meteorological factors most strongly correlated with wind power, processing the abnormal data after filtering out the relevant data, filtering by clustering algorithms or manual methods, dividing the overall dataset into several conditional subsets, supplementing the abnormal data in each subset by methods such as cleaning or interpolation, and finally normalizing the obtained wind power dataset.

[0012] Optionally, the Pearson coefficient is used to filter meteorological factor data, and the meteorological data with the strongest correlation are retained for subsequent prediction; the clustering algorithm includes DBSCAN and K-means algorithm, and the manual method is to manually process the abnormal data.

[0013] Optionally, the execution process of step 3 involves dividing the obtained data into two groups: a training set and a test set. In the training set, a prediction model based on the IPSO-CNN-LSTM algorithm combining SVM and quantile regression is constructed using historical data. After training on the training set, predictions are made, and quantitative analysis is performed on the model output to obtain the upper and lower boundary values ​​of short-term wind power in quantile form at a given confidence level, as well as the prediction sequence for the day to be predicted. Finally, the prediction is compared with the pre-set test set.

[0014] Optionally, the construction process of the prediction model based on the IPSO-CNN-LSTM algorithm combining SVM and quantile regression includes the following steps:

[0015] The IPSO algorithm is selected to optimize the inertia weight ω;

[0016] By combining CNN and LSTM networks, the CNN network extracts latent features from sample data, while the LSTM network captures long-term components.

[0017] The QRNN model is used to reflect the nonlinearity of the data and to obtain the probability prediction interval at a given confidence level, and the obtained data is further analyzed.

[0018] Mean absolute error (MAE), root mean square error (RMSE), interval coverage (PICP), and interval average bandwidth (PINAW) were used as evaluation metrics for model accuracy.

[0019] This invention provides a short-term wind power probability prediction method. It utilizes deep learning to mine implicit information in data and nonlinear features in wind power sequences to generate prediction probability intervals. Simultaneously, a nonlinear weighting method is selected to improve the optimization performance of the particle swarm optimization algorithm, i.e., the IPSO algorithm, which solves some problems of traditional algorithms and improves convergence speed. Then, a hybrid artificial intelligence algorithm, the CNN-LSTM hybrid algorithm, is selected to construct a prediction model based on the combination of SVM and quantile regression. After training, the short-term wind power probability prediction is completed. The CNN network can extract latent features from the sample data by using convolutional kernels, while the Long Short-Term Memory (LSTM) network can capture long-term components, avoiding the gradient vanishing and exploding phenomena of some existing algorithms, thus improving the efficiency of wind power probability prediction. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of a convolutional neural network structure.

[0022] Figure 2 This is a schematic diagram of the memory cell structure of an LSTM network.

[0023] Figure 3 This is a schematic diagram of the information transmission process in an LSTM network.

[0024] Figure 4 This is a flowchart of the CNN-LSTM algorithm.

[0025] Figure 5 This is a schematic diagram illustrating the specific process of a short-term wind power probability prediction method according to the present invention. Detailed Implementation

[0026] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0027] This invention provides a short-term wind power probability prediction method, comprising the following steps:

[0028] S1: Collect and process historical meteorological data to obtain the initial dataset for wind power probability prediction;

[0029] S2: The initial dataset is filtered and supplemented to obtain a wind power dataset, and then normalized.

[0030] S3: Divide the wind power dataset into training and testing sets, construct a prediction model based on the IPSO-CNN-LSTM algorithm combining SVM and quantile regression, and perform training and prediction.

[0031] S4: Adjust the model parameters based on the prediction results and errors until the results are close to the test set, and complete the short-term wind power probability prediction.

[0032] The following explains the specific implementation steps:

[0033] In step S1, the specific execution process is as follows: obtain historical meteorological data of the region, process the raw meteorological data and wind conditions by querying local logs, meteorological station records, and other climate monitoring system records, and obtain the initial dataset for wind power probability prediction based on the wind speed fluctuation of historical predicted power and the numerical weather prediction (NWP) prediction results of wind farms.

[0034] The numerical weather prediction (NWP) data uses meteorological data from a wind farm throughout the year, including wind speed, wind direction, temperature, and air pressure.

[0035] The specific process of step S2 is as follows: Screen out the meteorological factors that are most strongly correlated with wind power, process the abnormal data after screening out the relevant data, and divide the overall dataset into several conditional subsets by screening through clustering algorithms or manual methods; supplement the abnormal data in each subset by cleaning or interpolation, and normalize the final new dataset.

[0036] Furthermore, in step S2 above, the Pearson coefficient is used to screen meteorological factor data, and the meteorological data with the strongest correlation are retained for subsequent prediction; the clustering algorithm mainly includes DBSCAN and K-means algorithm, and the manual method is to manually process abnormal data, clean or interpolate the abnormal data after screening to obtain wind power dataset, and then normalize the data.

[0037] For data selection, annual numerical weather prediction (NWP) data for a specific wind farm were used. The Pearson coefficient method was employed to select the meteorological factors with the strongest correlations as the dataset for subsequent predictions. Pearson analysis measures the correlation between two variables based on the distance between them; its specific functional formula is as follows:

[0038]

[0039] In the formula, ρxy is the Pearson correlation coefficient. When ρxy > 0, the two variables are positively correlated; otherwise, they are negatively correlated. xy The larger the value of |, the stronger the correlation between the two variables, where x and y are two n-dimensional vectors;

[0040] Step S3: Divide the obtained data into two groups: training set and test set. In the training set, construct a prediction model based on the IPSO-CNN-LSTM algorithm combining SVM and quantile regression using historical data. After training the model on the training set, make predictions and perform quantitative analysis on the model output to obtain the upper and lower boundary values ​​of short-term wind power in quantile form at a given confidence level and the prediction sequence for the day to be predicted. Finally, compare the prediction with the pre-set test set.

[0041] Specifically, after dividing the dataset into training and testing sets, a prediction model based on the IPSO-CNN-LSTM algorithm, which combines SVM and quantile regression, is constructed for training and prediction. First, the dataset is input into the IPSO-CNN-LSTM algorithm model, and the IPSO algorithm is used to improve the convergence speed and obtain the optimal inertia weight ω. This ω can be used as the weight of SVM for data output. Then, the dataset is put into CNN-LSTM to generate power prediction values. The prediction values ​​are then put into a model that combines SVM and quantile regression for quantitative analysis to generate a probability distribution, and further obtain the power interval prediction graph under a given confidence level.

[0042] Both the training and test sets are taken from the processed dataset. During training, only the training set data is used, and the resulting data is compared with the test set for verification. The prediction model structure of the IPSO-CNN-LSTM algorithm based on quantile regression is as follows:

[0043] Firstly, addressing the slow convergence speed of the model, we utilize the Particle Swarm Optimization (PSO) algorithm to optimize its hyperparameters. The hyperparameters to be optimized are mapped to particles, each sharing an individual extreme value and global comparison, continuously updating its position and iteratively optimizing. An inertial weight ω is introduced, which is positively correlated with the particle's global search ability and negatively correlated with its local search ability. However, in traditional PSO algorithms, ω is a fixed value, resulting in poor local and global search abilities and a tendency to get trapped in local optima. Therefore, we propose a nonlinear weighting method to improve the optimization performance of the PSO algorithm, namely the IPSO algorithm. The optimization formula for ω is:

[0044]

[0045] ωmax and ωmin are the maximum and minimum values ​​of the weights, respectively, with values ​​of 0.9 and 0.4; t is the current iteration number; Tmax is the maximum iteration number.

[0046] Furthermore, a convolutional neural network (CNN) is combined with a long short-term memory neural network (LSTM). CNN extracts latent features from the sample data, while LSTM captures long-term components. The processed dataset is used as input to the prediction model, and the power value is used as output. The model is represented as: P t =f(p t-n Given the wind power at time t, Pt represents the wind power at time t, pt-n represents the historical wind farm data, and k represents the preprocessed data. Finally, a deep neural network with a single hidden layer is used as the output layer, and the output result is the wind power at time t.

[0047] Convolutional Neural Networks (CNNs) consist of alternating convolutional and pooling layers. A ReLU activation function acts between each convolutional and pooling layer to accelerate model convergence. Data in the model is processed by the CNN, and all features are fused to obtain the CNN's feature description. This data is then passed to an LSTM (Laser-Based Memory) system. Typically, the input data at this point needs to be reshaped to the type that the LSTM can process. The LSTM, upon receiving the new input, determines what to retain and what to discard, using the sigmoid activation function to selectively forget and preserve important data. The data obtained from the input gate updates the state. Finally, the output gate determines the information carried over, transferring the new state and the hidden state to the next time step.

[0048] CNN networks can extract spatial relationships from multidimensional time-series data. They consist of convolutional and pooling layers, utilizing features such as weight sharing to accelerate training and improve generalization performance. The calculation of a one-dimensional convolution is as follows:

[0049]

[0050] In the formula Let f be the k-th convolutional mapping in layer l, f be the activation function, N be the number of input convolutional mappings, and * be the convolution operation. The bias of the k-th convolutional kernel corresponding to layer l is as follows: The principle of a CNN network is as follows: Figure 1 As shown;

[0051] The LSTM network is a type of recurrent neural network (RNN). It adds input gates, forget gates, and output gates to the hidden layers of a traditional RNN, and also adds units for storing memories. Its internal memory unit structure and information transfer process are as follows: Figure 2 , Figure 3 As shown in the figure, Ut is the preparatory value to be output to the hidden layer state, σ, l, and g are activation functions, where σ is the sigmoid function, l and g are the tanh functions, Ht is the hidden state at time t, ft, it, and ot are the forget gate, input gate, and output gate, respectively, xt is the input information at time t, and yt is the output information at time y. The mathematical calculation process of the LSTM network is as follows:

[0052] Forgotten Gate: f t =σ(W fx x t +W fx h t-1 +W fc C t-1 +b f );

[0053] Input gate: i t =σ(W ix x t +W ih h t-1 +W ic C t-1 +b i );U t =g(W cx x t +W ch h t-1 +b c );

[0054] C t =C t-1 f t +U t i t ;

[0055] Output gate: o t =σ(W ox x t +W oh h t-1 +W oc C t-1+b o );

[0056] The forget gate helps the LSTM determine which information will be deleted from the memory cell state; the input gate `it` determines the new information to be stored in the new cell state `Ct`; and the output gate is used to calculate the output value `Ht`. In the formula, `Ct-1ft` represents how much information is forgotten from `Ct-1`, `Utit` represents how much information is added to the new cell state `Ct`, `Wix`, `Wfx`, `Wox`, and `Wcx` are weight matrices connecting the input information `xt`; `Wic`, `Wfc`, and `Woc` are diagonal matrices connecting the neuron activation function output value `ct` and the gate function; `Wih`, `Wfh`, `Woh`, and `Wch` are weight matrices connecting the output signal `Ht`; and `bi`, `bf`, `bo`, and `bc` are the biases corresponding to the input gate, forget gate, output gate, and the pre-output value `Ut`. The algorithm flow after fitting is as follows: Figure 4 As shown.

[0057] Furthermore, the quantile regression (QRNN) model is used to reflect the nonlinearity of the data, and the data obtained from the CNN-LSTM algorithm model are further analyzed. Let the explanatory variables of the QRNN algorithm be X = [x1, x2, ..., xn], and the response variable be Y. The nonlinear response model of X to Y is:

[0058]

[0059]

[0060] In the formula, J is the number of hidden layer nodes, S is the number of input layer nodes, and g i (τ) represents the input result of the hidden layer, QY(τ|X) represents the conditional quantile of quantile τ, and f(1) represents the transformation function between the input layer and the hidden layer, taking the tanh function. For the weights and biases of the input layer, b (2) (τ) represents the weights and biases of the output layer, and f(2) represents the activation function between the hidden layer and the output layer. The weights and biases of the QRNN model are estimated by minimizing the loss function:

[0061]

[0062] In the formula, EYτ is the loss function, Yi is the explained variable of the i-th sample, and I(·) is the indicator function, whose range is:

[0063]

[0064] Suppose there are N hidden layer neurons in a CNN-LSTM network, then the output vector at time t is... By using this output vector as input to the fully connected layer, the output value can be obtained.

[0065]

[0066] In the formula, ωk and b are the weights and biases between the hidden layer output and the fully connected layer input, which are used to obtain the power prediction image under a given confidence level.

[0067] Furthermore, the mean absolute error (MAE) and root mean square error (RMSE) are used as evaluation metrics for model accuracy, and their expressions are as follows:

[0068]

[0069] In the formula, xt, Here, N represents the standardized true and predicted power values, and N is the number of tests and validations. The smaller the obtained MAE and RMSE values, the higher the model prediction accuracy and the better the prediction performance.

[0070] Furthermore, Predicted Interval Coverage (PICP) and Predicted Interval Average Bandwidth (PINAW) are used as evaluation metrics for interval prediction accuracy, and their expressions are as follows:

[0071]

[0072] In the formula, N is the number of evaluation samples. If the evaluation target value falls within the evaluation interval, then α i =1 otherwise = zero; β is the interval width, i.e., the difference between the upper and lower bounds of the interval. The larger the obtained PICP value, the higher the accuracy and reliability of the prediction; when the PICP value is constant, the smaller the PINAW value, the better the prediction effect.

[0073] Finally, the model parameters are adjusted based on the image data comparison until the predicted data is close to that of the test set, thus completing the short-term wind power probability prediction. The specific process is as follows: Figure 5 As shown.

[0074] In summary, the present invention has the following beneficial effects:

[0075] 1. The use of multivariate prediction models improves computational efficiency and accuracy, and is also conducive to the stable operation of the power system, promotes wind curtailment and consumption, and improves energy utilization.

[0076] 2. It helps to improve the competitiveness of wind power in the power industry;

[0077] 3. It helps power-related departments determine reasonable maintenance and repair times for wind turbines, thereby improving the power generation efficiency of wind turbines.

[0078] The above description discloses only one preferred embodiment of the present invention, and should not be construed as limiting the scope of the present invention. Those skilled in the art will understand that all or part of the processes of the above embodiments can be implemented, and equivalent changes made in accordance with the claims of the present invention are still within the scope of the invention.

Claims

1. A short-term wind power zone probabilistic prediction method, characterized in that, Includes the following steps: Step 1: Collect and process historical meteorological data to obtain the initial dataset for wind power probability prediction; Step 2: Filter and supplement the initial dataset to obtain a wind power dataset, and then perform normalization processing; Step 3: Divide the wind power dataset into training and testing sets, construct a prediction model based on SVM and quantile regression using the IPSO-CNN-LSTM algorithm, and train and predict the data. The execution process of step 3 is as follows: the obtained data is divided into two groups: training set and test set. In the training set, a prediction model based on the IPSO-CNN-LSTM algorithm combining SVM and quantile regression is constructed using historical data. After training on the training set, prediction is performed. Quantitative analysis is performed on the model output to obtain the upper and lower boundary values ​​of short-term wind power in quantile form at a given confidence level and the prediction sequence of the day to be predicted. Finally, it is compared with the pre-set test set. The construction process of the prediction model based on the IPSO-CNN-LSTM algorithm combining SVM and quantile regression includes the following steps: The IPSO algorithm is selected to optimize the inertia weight ω; By combining CNN and LSTM networks, the CNN network extracts latent features from sample data, while the LSTM network captures long-term components. The QRNN model is used to reflect the nonlinearity of the data and the prediction probability intervals at different confidence levels are obtained. The obtained data is then further analyzed. Mean absolute error (MAE) and root mean square error (RMSE), interval coverage (PICP) and interval average bandwidth (PINAW) are used as evaluation metrics for model accuracy. Step 4: Adjust the model parameters based on the prediction results and errors until the results are close to the test set, and complete the short-term wind power probability prediction.

2. The short-term wind power segment probability prediction method as described in claim 1, characterized in that, The process of collecting and processing historical meteorological data to obtain the initial dataset for wind power probability prediction involves acquiring historical meteorological data for the region, processing the raw meteorological data and wind conditions by querying local logs, meteorological station records, and records from other climate monitoring systems, and obtaining the initial dataset for wind power probability prediction based on the wind speed fluctuations in historical predicted power and the numerical weather prediction (NWP) results of wind farms.

3. The short-term wind power zone probability prediction method as described in claim 2, characterized in that, The numerical weather prediction (NWP) data uses meteorological data from a selected wind farm throughout the year, with parameters including wind speed, wind direction, temperature, and air pressure.

4. The short-term wind power zone probability prediction method as described in claim 3, characterized in that, The process of filtering and supplementing the initial dataset to obtain a wind power dataset and then normalizing it involves filtering out the meteorological factors most strongly correlated with wind power, processing the abnormal data after filtering out the relevant data, and dividing the overall dataset into several conditional subsets by filtering through clustering algorithms or manual methods; supplementing the abnormal data in each subset by methods such as cleaning or interpolation, and finally normalizing the obtained wind power dataset.

5. The short-term wind power segment probability prediction method as described in claim 4, characterized in that, The Pearson coefficient was used to screen meteorological factor data, and the meteorological data with the strongest correlation were retained for subsequent prediction. The clustering algorithm included DBSCAN and K-means algorithm, and the manual method was to manually process the outlier data.

Images