Wind power plant power prediction method of IRBMO-TCN-RVM model

Abstract

The invention provides a wind power plant power prediction method based on an IRBMO-TCN-RVM model, and the method comprises the steps: collecting historical wind power data and corresponding meteorological data, carrying out the cleaning and normalization processing, and dividing a training set, a verification set and a test set; determining a topological structure of the TCN-RVM according to the sample data, preprocessing the sample data, inputting the sample data into the TCN-RVM, and calculating a global error; cubic chaotic mapping, a dynamic search strategy and an early warning mechanism are added into a standard RBMO algorithm to obtain an IRBMO algorithm, global errors serve as a fitness function, and the position of a red mouth blue magpie is updated; and taking the optimal red beak blue magpie as the optimal weight and threshold of the TCN-RVM to construct a wind power plant power prediction model, and performing analysis and prediction on new wind power plant power data by using the trained wind power plant power prediction model. The method comprises the following steps: cleaning and normalizing wind power plant power data; cubic chaotic mapping is introduced, the diversity of populations is enhanced, the global search capability is improved, local optimum is avoided, and the prediction efficiency is improved.

Description

Technical Field

[0001] The invention relates to the technical field of wind farm power prediction, and in particular to a wind farm power prediction method based on an IRBMO-TCN-RVM model. Background Art

[0002] In recent years, countries around the world have paid more and more attention to environmental issues. People are increasingly looking for alternative energy sources to fossil fuels [2]. Wind energy has key advantages in terms of technical maturity, greenhouse gas emissions, synergy with electric heat exchange and heat storage technology, and far-reaching economic benefits. Wind power generation is a very promising technology for the development and utilization of wind energy. It has the advantages of low carbon emissions, environmental protection, and large storage capacity. Wind power forecasting is of great significance for promoting the development of renewable energy and ensuring the safety of the power grid.

[0003] Accurately predicting wind power generation has become a hot topic, and the literature surrounding this issue is increasing. Power generation prediction refers to predicting the actual output power of power generation equipment in the future by analyzing historical data and environmental factors. Temperature, humidity, air pressure, wind direction, wind speed, etc. have become important parameters affecting wind power generation, and wind power data is a time series[8]. Therefore, the development of high-precision models plays an important role in rationally allocating wind farm resources and improving power generation efficiency.

[0004] To address the above challenges and model power generation forecasting, a wide range of strategies have been proposed in the past. Machine learning (ML) and deep learning (DL) are the most important models for wind power forecasting. Machine learning (ML) is one of the commonly used models for wind power forecasting. Although the above models have made considerable progress, it is still difficult to obtain ideal results using deep learning for wind power forecasting due to the non-stationary nature of meteorological factors. Deep learning (DL) models can be very effective in predicting wind power after training. Most existing model methods involve the decomposition and optimization of data parameters. Optimization algorithms aim to quickly determine the optimal parameters of the prediction model. Swarm intelligence optimization algorithms inspired by biological populations have been one of the common methods to improve model accuracy.

[0005] RBMO is a new type of swarm intelligence optimization algorithm. TCN is a deep learning model for processing sequence data. TCN mainly captures long-distance dependencies in sequences through a series of convolutional layers and dilated convolutions. RVM is a sparse probabilistic model for classification and regression tasks. Summary of the invention

[0006] In view of the technical problems of low efficiency and complex calculation of existing wind power generation prediction methods, the present invention proposes a wind farm power prediction method of IRBMO-TCN-RVM model, which is composed of an improved red-billed blue magpie optimization algorithm (IRBMO), a temporal convolutional network (TCN) and a relevance vector machine (RVM). IRBMO integrates Cubic chaotic mapping, dynamic search strategy and early warning mechanism with the red-billed blue magpie optimization algorithm. TCN automatically extracts time series features in wind farm power. The relevance vector machine RVM accurately predicts the features extracted by TCN. In order to achieve the above purpose, the technical solution of the present invention is implemented as follows:

[0007] The present invention collects historical wind power data and corresponding meteorological data, performs cleaning and normalization processing, and divides the data into a training set, a validation set and a test set; selects features that have a significant impact on wind power according to data analysis; uses a TCN model to process time series data and extracts features that represent complex time series relationships of data; inputs the features extracted by TCN into an RVM for training, and the RVM predicts wind power according to the features; determines the values ​​of an input layer, a hidden layer and an output layer of a TCN-RVM neural network according to the processed data, and initializes a red-billed blue magpie population, and uses the prediction error of the TCN-RVM neural network initialized by an individual as the fitness value of the individual, thereby determining a fitness function; sorts the obtained fitness values ​​and selects the optimal red-billed blue magpie; compares the optimal individual obtained in each iteration, and replaces the optimal individual if the optimal individual obtained in each iteration has a better fitness value than the optimal red-billed blue magpie; IRBMO evaluates the model performance according to the validation set, and adjusts the hyperparameters of TCN and RVM to improve the prediction accuracy; and uses a test set to evaluate the generalization ability of the model to ensure that the model makes accurate predictions on wind power data.

[0008] The main idea of ​​the improved algorithm (IRBMO) of the present invention is to replace the original random generation method with Cubic chaotic mapping, so as to prevent the initial value of the population from being concentrated in a certain local area, increase the diversity of the population, ensure the algorithm's ability to explore certain areas in the initial stage, and integrate the dynamic search strategy so that the red-billed blue magpie uses multiple paths to adjust its own position. The dynamic search strategy is dynamically adjusted according to the number of iterations, which increases the adaptive ability, increases the early warning mechanism of the red-billed blue magpie, and improves the convergence speed. When the red-billed blue magpie is aware of the danger, the red-billed blue magpie at the edge of the group will quickly move to a safe area to obtain a better position, and construct the IRBMO-TCN-RVM model to predict the power forecast of the wind farm. And the present invention:

[0009] (1) An improved algorithm IRBMO is proposed and compared with other optimization algorithms to prove its suitability for model parameter optimization.

[0010] (2) An IRBMO-TCN-RVM hybrid model is proposed. IRBMO is used to optimize the hyperparameters of TCN and RVM, TCN is used to extract time series features, and RVM is used to accurately classify the time features of the data.

[0011] (3) Compare the IRBMO-TCN-RVM model with other machine learning, deep learning models and their combination models. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] 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 some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0013] Figure 1 It is a schematic diagram of the overall process of the present invention.

[0014] Figure 2 The figure is a specific implementation flow chart of the present invention.

[0015] Figure 3 This is the convergence diagram of the F1-F7 function.

[0016] Figure 4 This is the convergence diagram of the F8-F13 function.

[0017] Figure 5 This is the convergence diagram of the F14-F23 function. DETAILED DESCRIPTION

[0018] 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.

[0019] like Figure 1 As shown, the embodiment of the present invention provides a wind farm power prediction method based on the IRBMO-TCN-RVM model, and the steps are as follows:

[0020] Step 1: Collect historical wind power data and corresponding meteorological data, clean and normalize them, and divide them into training set, validation set and test set. Based on data analysis, select features that have a significant impact on wind power.

[0021] Each group of the wind farm power data includes 13 kinds of meteorological index information; the 13 kinds of meteorological index information are wind speed at 10m of the wind tower, wind speed at 30m of the wind tower, wind speed at 50m of the wind tower, wind speed at 70m of the wind tower, wind speed at hub height, wind direction at 10m of the wind tower, wind direction at 30m of the wind tower, wind direction at 50m of the wind tower, wind direction at 70m of the wind tower, wind direction at hub height, temperature, air pressure and humidity.

[0022] Step 2: Determine the topological structure of the TCN-RVM neural network based on the sample data, preprocess the sample data and input it into the TCN-RVM neural network, and calculate the global error of the TCN-RVM neural network.

[0023] The method of clustering wind farm power using TCN-RVM neural network is:

[0024] S31. Normalize using the "mapminmax" function:

[0025]

[0026] Among them, x d,min is the minimum value, x d,max is the maximum value, x d is the dth information, y d is the normalized value of the d-th indicator information.

[0027] S32, using random numbers to randomly set the vector w of the initial connection weight value between the hidden layer and the input layer ih and the connection weight w between the hidden layer and the output layer ho , set the maximum number of cycles t max .

[0028] S33, input wind power as an input vector to the input layer, input vector p = (p1, p2, p3 ..., p m ), where the training sample K a =(K a,1 ,K a,2 ,K a,3 ,…,K a,501 ).

[0029] S34, calculate the Euclidean distance between the weight vector of the hidden layer and the input vector: select a set of training samples x i =(x i,1 ,x i,2 ,x i,3 ,…,x i,501 ), each time the training data is input, the neuron that is closest to the input vector is found as the winning neuron h through the size of the connection weight, and the error function E1 is defined as the distance between the input vector and the connection weight vector:

[0030]

[0031] The role of the error function E1 is to calculate the initial weights and thresholds. q represents the number of input vectors in the input layer, d o (1) is the initial expected input vector, y o (1) The actual initial input vector.

[0032] S35, label the winning neurons corresponding to the wind power, and return to step S34 until all inputs are completed;

[0033] S36, judging whether the global error meets the requirements: when the error reaches the preset accuracy or the number of learning times is greater than the set maximum number of times t max , the algorithm ends and the global error value is used as the fitness function of the IRBMO algorithm. Otherwise, the next learning sample and the corresponding expected output are selected, and steps S33 to S36 are executed cyclically. The global error is:

[0034]

[0035] Among them, m is the number of input vectors, d is the number of o (K) is the expected input vector, y o (K) The actual input vector

[0036] Step 3: Use the global error of step 2 as the fitness function of the IRBMO algorithm, add Cubic chaotic mapping, dynamic search strategy and early warning mechanism to the standard RBMO algorithm to obtain the IRBMO algorithm, save and update the position of the optimal red-billed blue magpie.

[0037] like Figure 2 As shown in the figure, the initial parameters are set, the features that have a significant impact on wind power are selected, the fitness function is the called benchmark function, and the search is performed through Cubic chaos mapping, dynamic search strategy and early warning mechanism. max Indicates the maximum number of iterations.

[0038] The IRBMO-TCN-RVM wind power prediction influencing factors and data are used as the thresholds and weights of the initial TCN-RVM neural network. The individual fitness is calculated through the fitness function. The individual corresponding to the optimal fitness value is found by improving the red-billed blue magpie optimization algorithm, that is, the optimal threshold and weight are found, and the optimal solution is found.

[0039] Red-billed blue magpie optimizer (RBMO) is an optimization algorithm proposed by Shengwei Fu and Ke Li in 2024, inspired by the hunting patterns and biological characteristics of red-billed blue magpies. RBMO imitates the three different stages of the hunting process of red-billed blue magpies: searching for food, attacking prey, and storing food. In a d-dimensional optimization problem, red-billed blue magpie is a 1×d-dimensional array representing the problem.

[0040] SandCat i =[X1,X2,…,X d ] (1)

[0041] Each variable value (X1, X2, ..., X d ) are all floating point numbers. Each X i Must be between the upper and lower bounds

[0042] In the initialization phase of RBMO, the random generation of the red-billed blue magpie population may not fully cover the entire search space in some cases. This leads to insufficient diversity, which makes the algorithm lose the ability to explore certain areas in the early stages. The sequence generated by the Cubic chaotic map has global traversal and can cover the entire search space. This helps prevent the initial value of the population from being concentrated in a certain local area, increases the diversity of the population, and ensures the algorithm's ability to explore certain areas in the early stages.

[0043] Therefore, the Cubic chaotic map is introduced to replace the original random generation method. The mathematical model is as follows:

[0044]

[0045] Where ρ is the control parameter, and the Cubic chaotic map sequence is between 0 and 1. When ρ = 2.595, the chaotic variable has better ergodicity.

[0046] In the process of searching for food, red-billed blue magpies usually act in small groups (2 to 5 individuals) or clusters (more than 10 individuals) to improve search efficiency. They use various techniques, such as jumping, walking, or searching trees on the ground to obtain food resources. This adaptability and flexibility enables red-billed blue magpies to adopt different hunting strategies according to environmental conditions and available resources to ensure sufficient food supply. When the group explores for food, equation (3) is used, and equation (4) is used for cluster search. The above model is defined as follows:

[0047]

[0048] In the formula, t is the current iteration number, X i(t+1) represents the position of the i-th new search agent, p represents the number of red-billed blue magpies hunted by a small group of 2 to 5 randomly selected from all searched individuals, and X m represents the randomly selected mth individual, X i represents the i-th individual, and X rs (t) represents the search agent randomly selected in the current iteration.

[0049]

[0050] Where q represents the number of search agents when the cluster explores food, which is between 10 and n. Again, it is randomly selected from the entire population.

[0051] The above formula lacks a dynamic adjustment mechanism. As iterations proceed, it is difficult to adapt to the exploration needs of different stages, which affects the global search capability of the algorithm. To address this problem, a dynamic search strategy is proposed. The dynamic search strategy enables the red-billed blue magpie to use multiple paths to adjust its position. The dynamic search strategy is dynamically adjusted according to the number of iterations to increase the adaptive capability. This strategy can effectively improve the global search performance of the algorithm. The dynamic search factor is:

[0052] Z=ek*cos(π*(1-t / t Max )) (5)

[0053] The improved formula is as follows:

[0054]

[0055] Where l is a random value in [0,1], l∈[0,1].

[0056] Red-billed blue magpies display a high level of hunting proficiency and cooperation when chasing prey. The strategies they adopt include rapid pecking, jumping to catch prey, or flying to catch insects. When acting in small groups, the main targets are usually small prey or plants. The corresponding mathematical model is shown in equation (7). When acting in groups, red-billed blue magpies are able to jointly target large prey such as large insects or small vertebrates. The mathematical representation of this behavior is outlined in equation (8). This predatory hunting behavior highlights the variety of strategies and skills possessed by the red-billed blue magpie, making it a versatile predator that is good at successfully obtaining food in a variety of situations. This stage is calculated by the following formula:

[0057]

[0058] In the formula, X food (t) indicates the location of the food, Randn represents random numbers used to generate a standard normal distribution (mean 0, standard deviation 1).

[0059] After the second stage, the red-billed blue magpie warning mechanism is added to increase the speed of the red-billed blue magpie to avoid natural enemies. In the corresponding algorithm, the convergence speed is improved. When the red-billed blue magpie realizes the danger, the red-billed blue magpie at the edge of the group will quickly move to a safe area to get a better position. The red-billed blue magpie in the middle of the group will move randomly to get closer to other red-billed blue magpies. When f i >f g When f i >f g This indicates that the red-billed blue magpie in the middle of the population is aware of the danger and needs to get closer to other red-billed blue magpies to avoid being preyed on. K represents the direction in which the sparrow moves and is also the step length control parameter.

[0060] The above mathematical expression is as follows:

[0061]

[0062] Where K is a random number that follows a normal distribution with a mean of 0 and a variance of 1, K∈[-1,1]. i is the fitness value of the current red-billed blue magpie individual. g and f w are the current global best and worst fitness values ​​respectively. worst (t+1) is the worst position. ε is a constant to avoid zero in the denominator.

[0063] In addition to finding and attacking food, the red-billed blue magpie also stores excess food in tree holes or other hidden places for future consumption, thereby ensuring a stable food supply in times of shortage. This process preserves solution information, helping individuals find the global optimum. Its mathematical model is shown below:

[0064]

[0065] In the formula, and They respectively represent the fitness values ​​of the i-th red-billed blue magpie before and after the position is updated.

[0066] Initialize the red-billed blue magpie population, use the red-billed blue magpie individuals to represent the initial weights and thresholds of the TCN-RVM neural network, use the prediction error of the TCN-RVM neural network initialized by the individual as the fitness value of the individual, and determine the threshold and number of weights of the wind power generation prediction model.

[0067] Update the population of red-billed blue magpies: sort by the fitness values ​​obtained, select the best red-billed blue magpie individuals; compare the best red-billed blue magpie individuals obtained in each iteration, and replace them if the fitness values ​​obtained in each iteration are better. Save and update the positions of red-billed blue magpies through Cubic chaos mapping, dynamic search strategy and early warning mechanism.

[0068] Determine whether the IRBMO algorithm has reached the maximum number of iterations t set initially Max , if the maximum number of iterations t has been reached Max , then the optimal red-billed blue magpie output is the optimal weight and threshold corresponding to the TCN-RVM neural network. If it is not reached, the number of iterations is increased by 1, and the IRBMO algorithm is continuously calculated until the maximum number of iterations t is reached. Max .

[0069] Step 4: Use the optimal red-billed blue magpie individual obtained in step 3 as the optimal weight and threshold of the TCN-RVM neural network to build a wind power prediction model, and use the sample data to train the wind power plant power prediction model, and use the trained wind power plant power prediction model to analyze and predict new data.

[0070] The optimal red-billed blue magpie individual obtained by the IRBMO algorithm is used as the optimal weight and threshold of the TCN-RVM neural network. When the output error is less than the set error, the calculation ends. Otherwise, the above process is calculated repeatedly until the error output by the TCN-RVM neural network is less than the initial set error, and the calculation is terminated. The wind power plant power prediction model is analyzed and predicted by the TCN-RVM neural network model that has been determined.

[0071] As the initial weights and thresholds to build a TCN-RVM neural network, the data after washing, normalization and feature selection are substituted into the TCN-RVM neural network for training, and the global error is used for the fitness function of the IRBMO algorithm. The basic idea of ​​the present invention is to use the initial weights and thresholds of the TCN-RVM neural network represented by individuals, and the prediction error of the TCN-RVM neural network initialized by individual values ​​as the fitness value of the individual, and find the optimal solution through training, that is, the optimal individual is used as the initial weight of the optimal TCN-RVM neural network.

[0072] The improved red-billed blue magpie algorithm is used to optimize the TCN-RVM neural network weights and global error, and sigmoid is selected as the kernel function to train the sample data. The implementation method of establishing the wind power plant power prediction model is as follows:

[0073] S1: Initialize the population and generate a random population.

[0074] S2: Set the maximum number of iterations n, collect historical wind power data and corresponding meteorological data, clean and normalize them, and divide them into training sets, validation sets, and test sets. According to data analysis, select features that have a significant impact on wind power, train them, and calculate the fitness of individuals in the population through the fitness function.

[0075] S3: When the optimal individual of the population reaches the termination iteration, the optimization process is exited and the optimal solution is obtained, and step 5) is executed; otherwise, step S4 is executed.

[0076] S4: Perform fitness value comparison on the contemporary surviving population, compare the best red-billed blue magpie individual with the individual obtained in each iteration, if the individual obtained in each iteration has a larger fitness value than the best red-billed blue magpie individual, replace it and execute step S3.

[0077] S5: Get the optimal weight and the optimal global error, and bring the optimal weight and the optimal global error into the TCN-RVM neural network. Through the determined TCN-RVM neural network model, the wind power plant power prediction model is constructed, and thus the model is proposed:

[0078]

[0079] in, is the wind farm power prediction model; h L Output after L layers of convolution;

[0080] * represents convolution operation, and the convolution kernel size is k;

[0081] is the activation function; W l and b l l are the weight matrix and bias term of the layer respectively; is the RVM kernel function, h i L is the i-th correlation vector; σ is the parameter of the Gaussian kernel function.

[0082] TCN is a new network architecture based on CNN, adding dilated convolution, causal convolution and residual blocks. The size of the effective window of dilated convolution grows exponentially with the number of layers. Dilated convolution increases the receptive field by introducing "holes" between convolution kernels, thereby obtaining temporal dependencies without increasing the number of layers. The expression is as follows:

[0083]

[0084] Where s represents the time point index in the output sequence F, and the expansion rate d = [1,…,2 L], dilation rate d = [1, 2, 4]. f is the convolution kernel. f(i) is the filter function. X s-d·i Enter the value of the sliding window for the previous time step.

[0085] The expression strength of a neural network is inextricably linked to the depth of the network. However, as the depth of the network increases, the neural network often suffers from the degradation of network capabilities such as gradient disappearance, so TCN adds skip connections to solve this problem. Residual connections are used to solve the problem of gradient disappearance in deep network training. It allows gradients to flow directly through some layers by adding skip connections. Its operation expression is:

[0086] X h =δ(G(X h-1 )+X h-1 ) (13)

[0087] Where δ(·) is the activation function. h-1 is the error of the previous layer. G(·) is part of X h-1 Through the transformation of dilated causal convolution, weight normalization, activation function and random deactivation. h For jumpered output.

[0088] RVM is a linear regression model based on the Bayesian framework. Compared with the support vector machine (SVM), which assigns deterministic values, RVM processes the model weights probabilistically under the Bayesian framework. This means that RVM can take into account the uncertainty of weights. In the RVM multi-input single-output regression prediction model, it is assumed that the input variables are fixed. Output variables are set to fixed Where N is the total number of samples. The relationship between input and output needs to satisfy the following formula:

[0089]

[0090] In the formula, k(x,x i ) is the kernel function. w is the weight vector. θ i The mean is 0 and the variance is σ 2 Gaussian noise. When w and σ 2 When is known, the output probability distribution can be obtained as follows:

[0091]

[0092] Where w=[w1,w2,…,w N ], φ is the kernel matrix defined as:

[0093] φ=[φ(x1),φ(x2),…,φ(x N )] T (16)

[0094] φ(x i )=[1,k(x i ,x1),k(x i ,x2),…,k(x i ,x N )] (17)

[0095] In order to avoid over-learning, i.e. over-fitting, in the solution process, w is defined as a Gaussian prior distribution with a mean of zero, which is defined as:

[0096]

[0097] Where α is an N+1-dimensional hyperparameter vector, defined as α=(α0,α1,…,α N ).

[0098] When the output y is given, the posterior probability distribution formula of other parameters is:

[0099] p(w,α,σ 2 |y)=p(w|y,α,σ 2 )p(α,σ 2 |y) (19)

[0100] According to the above formula, combined with Bayes' theorem, the posterior probability distribution of w is derived as follows:

[0101]

[0102] Where Σ and u are the mean square error and mean of w respectively, that is, Σ=(σ -2 Φ T Φ+A) -1 and u=σ -2 ∑Φ T t. Where A is the hyperparameter from α0 to α N A diagonal matrix composed of .

[0103] The process of RVM model training includes calculating the posterior probability And when the maximum value is obtained, the hyperparameter α and noise variance σ are obtained 2 The solution is obtained through continuous iterative operations, completing the training of the RVM regression prediction model.

[0104] Variable definition:

[0105] Input vector: p=(p1,p2,p3…,p m ) (twenty one)

[0106] Hidden layer input vector: hi = (hi1, hi2, ..., hi c ) (twenty two) Hidden layer output vector: ho = (ho1,ho2,…,ho c ) (twenty three) Output layer input vector: yi = (yi1, yi2, ..., yi q ) (twenty four) Output layer output vector: yo=(yo1,yo2,…,yo q ) (25) Expected output vector: d o =(d1,d2,…,d q ) (26) Number of sample data: K i =(K i,1 ,K i,2 ,K i,3 ,…,K i,501 ,) (27)

[0107] Among them, x o is the sample data after dimensionality reduction, the hidden layer input vector = the input vectors related to it * the connection weights between the input layer and the hidden layer, the output vector of the hidden layer is used as the input vector of the output, and the output layer output vector = the output layer input vector related to it * the connection weights between the hidden layer and the output layer.

[0108] Parameter initialization of TCN-RVM neural network: Set the number of nodes in the input layer, hidden layer, and output layer of the TCN-RVM neural network to n, c, and q.

[0109] Assign a random number in the interval (-1,1) to each connection weight, give the error function E1, the calculation accuracy value and the maximum number of learning times t Max , the error function is:

[0110]

[0111] Among them, d o (K) is the expected output function, yo(K) is the actual output function

[0112] Randomly select the Kth sample data and the corresponding expected output. The input sample data and the corresponding expected output are:

[0113] p(K)=(p1(K),p2(K),…,p m (K)) (29) d o (K)=(d1(K),d2(K),…,d q (K)) (30)

[0114] Calculate the input and output of each neuron in the hidden layer

[0115] ho h (K) = f(hi h (K)) h=1,2,…,p (32) yo o (K) = f(yi o (K)) o=1,2,…q (34)

[0116] Among them, w ih is the connection weight between the input layer and the hidden layer, w ho is the connection weight between the hidden layer and the output layer, b h is the threshold of each neuron in the hidden layer, b o is the threshold of each neuron in the output layer, and f() is the activation function sigmoid function.

[0117] Using the expected output and actual output of the network, calculate the partial derivative of the error function with respect to each neuron in the output layer (find the lowest point in the error surface), and calculate the partial derivative formula of the error function with respect to each neuron in the output layer:

[0118]

[0119] Using the connection weights from the hidden layer to the output layer, the δ o (k) and the output of the hidden layer to calculate the partial derivative δ of the error function with respect to each neuron in the hidden layer h (k):

[0120]

[0121] Using the δ of each neuron in the output layer o (k) and the output of each neuron in the hidden layer to modify the link weight w ho (k):

[0122]

[0123] Using the δ of each neuron in the hidden layer h (k) and the input-corrected connection weights of each neuron in the input layer.

[0124]

[0125] Calculate the global error

[0126]

[0127] When the partial derivative of the error with respect to the weight is greater than zero, the weight adjustment is negative, the actual output is greater than the expected output, and the weight is adjusted in the direction of reduction, so that the difference between the actual output and the expected output is reduced.

[0128]

[0129] When the partial derivative of the error with respect to the weight is less than 0, the weight adjustment amount is positive, the actual output is less than the expected output, and the weight is adjusted in the direction of increase so that the difference between the actual output and the expected output is reduced.

[0130] Determine whether the network error meets the requirements. When the error reaches the preset accuracy or the number of learning times is greater than the set maximum number, the algorithm ends and the calculated global error value is used for the fitness function of the IRBMO algorithm. Otherwise, select the next learning sample and the corresponding expected output to enter the next round of learning.

[0131] The weights and thresholds are updated by calculating the error between the network and the expected output. The test data is substituted into the trained TCN-RVM neural network model to predict the wind power generation.

[0132] Function test experiment and result analysis:

[0133] (1) Parameter settings:

[0134] The present invention selects the red-billed blue magpie optimization algorithm (RBMO), particle swarm optimization algorithm (PSO), gray wolf optimization algorithm (GWO), sparrow optimization algorithm (SCSSA) with sine-cosine and Cauchy mutation, and whale optimization algorithm (GSWOA) with global search strategy. In order to make the algorithm fair, the population and number of iterations of the compared algorithms are set to be unified. The population is set to 50 and the number of iterations is set to 500.

[0135] (2) Test function:

[0136] The present invention selects 23 benchmark functions to evaluate the 6 algorithms. 1-8 are single-peak benchmark functions, 9-14 are multi-peak functions, and 14-23 are mixed functions.

[0137] Table 1 Benchmark function test

[0138]

[0139]

[0140]

[0141] (3) Experimental results and analysis

[0142] The red-billed blue magpie optimization algorithm (RBMO), particle swarm optimization algorithm (PSO), gray wolf optimization algorithm (GWO), sparrow optimization algorithm integrating sine-cosine and Cauchy mutation (SCSSA) and whale optimization algorithm with global search strategy (GSWOA) were selected to compare the results of function optimization with the improved red-billed blue magpie optimization algorithm (IRBMO) of the present invention to verify the performance of the IRBMO algorithm. In order to ensure fairness during comparison, the mean value and standard deviation of the fitness of the benchmark function are calculated respectively. The mean value can reflect the optimization accuracy of each algorithm, and the standard deviation can intuitively reflect the robustness and stability of each algorithm. According to the experimental results, it can be seen that for all test functions F1-F23, the mean value and standard deviation of the IRBMO algorithm of the present invention are all significantly better than those of RBMO, PSO, GWO, SCSSA and GSWOA algorithms. Figure 3 , Figure 4 , Figure 5 As shown in Table 2, except for some functions whose convergence graphs are larger than those of other algorithms, the remaining convergence curves of the IRBMO algorithm are better than those of other algorithms, and it shows that the IRBMO algorithm has fast optimization speed and high stability.

[0143] Table 2 Results of test functions

[0144]

[0145]

[0146]

[0147] The present invention uses PCA to reduce the dimension of wind power data; initializes relevant parameters, processes sample data, and determines the fitness function of the IRBMO algorithm by initializing the initial weights and thresholds of the TCN-RVM neural network under the premise of determining the topological structure of the TCN-RVM neural network; initializes the population of red-billed blue magpies; sorts by fitness; saves and updates the elite red-billed blue magpies using cross-mutation and adaptive weights; when the maximum number of iterations set initially is reached, the optimal weight corresponding to the elite red-billed blue magpies is output. If not reached, the number of iterations is increased by 1, and the cycle is returned to continue until the maximum number of iterations is reached; the TCN-RVM neural network is constructed using the optimal weights and thresholds; the network and expected output errors are calculated, and when the output error is less than the set error, the calculation is terminated, and the trained TCN-RVM neural network is used for power prediction of wind farms.

[0148] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A wind farm power prediction method based on the IRBMO-TCN-RVM model, characterized in that: The steps are as follows: Step 1: Collect historical wind power data and corresponding wind farm power data, use PCA method to reduce the dimension of the collected wind farm power data, clean and normalize it, and divide it into training set, validation set and test set; Step 2: Based on data analysis, select the features that have a significant impact on wind power. Step 3: Determine the topological structure of the TCN-RVM neural network based on the sample data, preprocess the sample data and input it into the TCN-RVM neural network, and calculate the global error of the TCN-RVM neural network; Step 4: Add Cubic chaotic mapping, dynamic search strategy and early warning mechanism to the standard RBMO algorithm to obtain the IRBMO algorithm. Use the global error of step 3 as the fitness function of the IRBMO algorithm to save and update the position of the elite red-billed blue magpie. Step 5: Use the elite red-billed blue magpie individuals obtained in step 4 as the optimal weights and thresholds of the TCN-RVM neural network to construct a wind farm power prediction model, and use sample data to train the wind farm power prediction model. Use the trained wind farm power prediction model to analyze and predict new wind farm power data.

2. The wind farm power prediction method based on the IRBMO-TCN-RVM model according to claim 1 is characterized in that: Historical wind power data and corresponding meteorological data are collected, cleaned and normalized, and divided into training sets, validation sets and test sets. Based on data analysis, features that have a significant impact on wind power are selected.

3. The wind farm power prediction method based on the IRBMO-TCN-RVM model according to claim 2 is characterized in that: The method for obtaining the sample data is as follows: the wind farm data is formed into a matrix X, and the covariance matrix of the matrix X is obtained; the eigenvalue and eigenvector of the covariance matrix are obtained by using a linear algebra method; Calculate the variance contribution rate and cumulative variance contribution rate corresponding to each eigenvalue; arrange the variance contribution and cumulative variance contribution rate into a matrix from top to bottom according to the corresponding size, and the cumulative variance contribution rate T i ≥0.85, then select the data of the rows corresponding to the first p eigenvalues ​​to form the matrix L; The data after dimensionality reduction is Among them, l pp represents the value of the pth index of the pth variable, x p Indicates the p-th row of the wind farm power data matrix P.

4. The wind farm power prediction method based on the IRBMO-TCN-RVM model according to any one of claims 1 to 3, characterized in that: Each group of the wind farm power data includes 13 kinds of meteorological index information; the 13 kinds of meteorological index information are respectively wind speed at 10m of the wind tower, wind speed at 30m of the wind tower, wind speed at 50m of the wind tower, wind speed at 70m of the wind tower, wind speed at hub height, wind direction at 10m of the wind tower, wind direction at 30m of the wind tower, wind direction at 50m of the wind tower, wind direction at 70m of the wind tower, wind direction at hub height, temperature, air pressure and humidity.

5. The wind farm power prediction method based on the IRBMO-TCN-RVM model according to claim 4 is characterized in that: The input is 13 sets of data; the structure of the TCN-RVM neural network is a 5-10-1 structure, that is, the input layer has 5 neurons, the hidden layer has 10 neurons, and the output layer is a 1-layer neural network.

6. The wind farm power prediction method of the IRBMO-TCN-RVM model according to claim 5 is characterized in that: The method for calculating the global error of the TCN-RVM neural network in step 2 is: S1. Normalize all meteorological index information; S2, use random numbers to randomly set the initial connection weight vector w between the hidden layer and the input layer of the TCN-RVM neural network ih , the connection weight w between the hidden layer and the output layer ho , set the maximum number of cycles t max ; S3, input each meteorological index information as an input vector into the input layer of the TCN-RVM neural network; S4, calculating the Euclidean distance between the connection weight vector of the hidden layer and the input vector, and finding the neuron that is most similar to the input vector as the winning neuron h according to the size of the connection weight; S5, label the winning neurons corresponding to the meteorological indicators until all inputs are completed; S6. Determine whether the global error meets the requirements: When the global error reaches the preset accuracy or the number of iterations is greater than the set maximum number t max , the global error is used as the fitness function of the IRBMO algorithm; otherwise, the next sample data and the corresponding expected output are selected, and steps S3 to S6 are executed cyclically.

7. The wind farm power prediction method of the IRBMO-TCN-RVM model according to claim 6 is characterized in that: The global error is: Among them, q is the number of input vectors in the neural network input layer, d o (K) is the expected input vector, m is the number of input vectors, y o (K) The actual input vector.

8. The wind farm power prediction method of the IRBMO-TCN-RVM model according to claim 1 or 6, characterized in that: The ISSCO algorithm's wandering around the elite red-billed blue magpie and the dynamic search strategy's wandering around the red-billed blue magpie are parallel searches; The sequence generated by the Cubic chaotic map has global traversal and can cover the entire search space. It helps to prevent the initial value of the population from being concentrated in a local area, increases the diversity of the population, and ensures the algorithm's ability to explore certain areas in the initial stage. That is, ρ is the control parameter, and the Cubic chaotic mapping sequence is between 0 and 1. When ρ = 2.595, the chaotic variable has better ergodicity; The method of the dynamic search strategy is: the dynamic search factor is Among them, t is the current iteration number, t MAX is the maximum number of iterations; The red-billed blue magpie uses multiple paths to adjust its position. The dynamic search strategy is dynamically adjusted according to the number of iterations to increase the adaptive ability. This strategy can effectively improve the global search performance of the algorithm: X(t+1)=e z*l ·cos(2πl)·r·(X bc (t)-rand(0,1)·X bc (t)); r=r G ×rand(0,1) Among them, X(t+1) is the next iteration position; X bc (t) is the optimal solution elite red-billed blue magpie; X(t) represents the current position of the red-billed blue magpie; rand represents a random number between (0,1); r G Used to guide the parameter R to achieve the transfer of control between these stages; S M Simulates the auditory characteristics of the red-billed blue magpie, initially set to 2. The implementation method of the early warning mechanism is: Where: ω_I is a normally distributed random number with a mean of 0 and a variance of 1, K∈[-1,1]; ω max is the fitness value of the current red-billed blue magpie individual; f g and f w are the current global best and worst fitness values ​​respectively; X worst (t) is the worst position. ε is a constant to avoid zero in the denominator.

9. The wind farm power prediction method of the IRBMO-TCN-RVM model according to claim 8, characterized in that: In the IRBMO algorithm, the red-billed blue magpie is initialized, and the red-billed blue magpie individual is used to represent the initial weights and thresholds of the TCN-RVM neural network. The prediction error of the TCN-RVM neural network initialized by the individual is used as the fitness value of the individual to determine the threshold and number of weights of the wind farm power; Sort the obtained fitness values ​​by size and select the best red-billed blue magpie individual; Compare the best red-billed blue magpie individuals obtained in each iteration; Determine whether the IRBMO algorithm has reached the maximum number of iterations initially set. If so, the optimal red-billed blue magpie individual output is the optimal weight and threshold corresponding to the TCN-RVM neural network; if not, the number of iterations is increased by 1, and the IRBMO algorithm is continued to cycle until the maximum number of iterations t is reached. MAX ; In the training of the wind farm power prediction model, when the output error is less than the set error, the calculation ends; otherwise, the calculation is repeated until the global error output by the TCN-RVM neural network is less than the initial set error.

10. The wind farm power prediction method of the IRBMO-TCN-RVM model according to claim 1, 6, 7 or 9, characterized in that: The wind farm power prediction model is: in, is the wind farm power prediction model; h L Output after L layers of convolution; h L =φ(W L *φ(W L *Lφ(W1*X+b1)+b L-1 )+b L ); * represents the convolution operation, the convolution kernel size is k; φ(·) is the activation function; W l and b l l are the weight matrix and bias term of the layer respectively; is the RVM kernel function, h i L is the i-th correlation vector; σ is the parameter of the Gaussian kernel function.

Images