A method for short-term wind power prediction in newly built wind farms with limited data
By combining evolutionary generative adversarial networks and bidirectional gated cyclic unit models with cross-cutting optimization algorithms, the problem of insufficient data in wind power prediction for newly built wind farms was solved, improving the accuracy of wind power prediction and achieving efficient data augmentation and model optimization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GUANGDONG UNIV OF TECH
- Filing Date
- 2022-08-19
- Publication Date
- 2026-07-03
Smart Images
Figure CN115640868B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of wind power prediction, and more specifically, to a method for short-term prediction of wind power in newly built wind farms with limited data. Background Technology
[0002] Due to the large exploitable capacity and low investment cost of wind energy resources, wind power generation has become the dominant focus of renewable energy development. The large-scale and rapid deployment of wind power means a large number of newly built wind farms will be put into operation. However, such large-scale wind power grid connection will pose a significant challenge to the safe operation of the power system. Therefore, to ensure the safety of wind power grid connection, exploring a new technology to improve the accuracy of wind power prediction for newly built wind farms with limited data is of great significance.
[0003] To date, research on wind power prediction by scholars both domestically and internationally has mainly categorized into physical models, statistical models, and artificial intelligence models. Among these, artificial intelligence methods are widely used in wind power prediction due to their powerful nonlinear fitting capabilities. However, since most artificial intelligence methods are data-driven, a lack of sufficient wind power data can easily lead to overfitting, increasing the difficulty of model training and resulting in lower short-term wind power prediction accuracy. Some studies have attempted to use transfer learning methods to improve prediction algorithm performance in situations with insufficient samples; however, most renewable energy data owners are unwilling to share data with other operators in practical applications due to data privacy concerns. Therefore, although various prediction methods are publicly available, improving the accuracy of wind power prediction for newly built wind farms remains a pressing problem that needs to be solved. Summary of the Invention
[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for short-term prediction of wind power in newly built wind farms with limited data, thereby achieving high-quality data enhancement of limited wind power data and improving the prediction accuracy of wind power.
[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:
[0006] A method for short-term wind power prediction in newly built wind farms with limited data is provided, comprising the following steps:
[0007] S10. Obtain the raw wind data collected by the original wind farm sensor, and use the Pearson correlation coefficient method to select features of all wind factors in the raw wind data, wherein the wind factors include wind power, wind speed, wind direction, atmospheric pressure, temperature and humidity;
[0008] S20. Normalize the feature data selected in step S10 to the [0,1] value range to construct the input matrix X of the small data time series. input =[Xpower X speed X direction ], where X power X speed X direction These represent power sequence, wind speed sequence, and wind direction sequence, respectively.
[0009] S30. Construct an evolutionary generative adversarial network (EGAN), and transmit the original small data input matrix X to the EGAN, where X is expressed as:
[0010]
[0011] In the formula: These represent the values of power, wind speed, and wind direction of the wind farm at time t, respectively, and T represents the time length of the input data.
[0012] S40. Evolutionary Generative Adversarial Networks learn by studying the marginal distribution of the original, limited data, and then generate new data X with modal diversity and the same statistical similarity. generated ;
[0013] S50. Generate the new data X generated The augmented data X is formed by concatenating the original small data input matrix X with the original small data input matrix X. new And will enhance data X new The data is transmitted to a bidirectional gated cyclic unit (BiGRU) model, and the BiGRU model is used to process the input data X. new Time-related information is extracted bidirectionally;
[0014] S60. To address the problem that the hyperparameters of the Dense layer in the BiGRU network are prone to getting trapped in local optima, the CSO algorithm is introduced to optimize the hyperparameters of the Dense layer and obtain the predicted wind power sequence.
[0015] Preferably, in step S10, the feature selection using the Pearson correlation coefficient method is implemented as follows:
[0016] S11. Calculate the Pearson coefficient score r of other characteristics and wind power in the original wind power data. The calculation process is as follows:
[0017]
[0018] In the formula: n is the length of the feature data; X represents the other 5 meteorological features, including wind speed, wind direction, atmospheric pressure, temperature, and humidity; Y represents the average value of the corresponding feature; Y represents wind power. This represents the average wind power output.
[0019] S12: Based on the Pearson coefficient score r of different features and wind power, features with negative scores are discarded and features with positive scores are retained.
[0020] Preferably, in step S30, the EGAN model is constructed according to the following steps:
[0021] S31: The EGAN model consists of a generator and a discriminator;
[0022] S32: Transmit several sets of noise data z that follow a uniform or normal distribution to the generator to produce a new random variable G(z), i.e., virtual data;
[0023] S33: The virtual data G(z) and real data are fed together into the discriminator for discrimination. The discriminator identifies the label of the input sample, i.e., real or fake. During training, the generator and discriminator update their network parameters using a specific loss function and gradient descent method. The generator's loss function L... G The loss function L of the discriminator D The definitions are shown in equations (3) and (4) respectively:
[0024]
[0025]
[0026] In equations (3) to (4), z represents noise; G(z) is a new random number generated by the generator; D(x) represents the discriminator's judgment output on the real data x; D(G(x)) represents the discriminator's judgment output on G(z); This represents the mathematical expectation of the distribution of noisy data and generated data; This represents the mathematical expectation of the real data and its distribution; log(.) is the logarithmic function.
[0027] S34: Set different mutation operators, i.e., different objective functions, to guide the initial generator to evolve and produce a generator population, narrowing the distance between the generated data and the distribution of real data from different perspectives, and the population iteratively evolves in a given environment, i.e., a discriminator; mutation operators include the Minimax mutation operator. Heuristic mutation operator and Least-Squares mutation operator The corresponding functions are shown in equations (5) to (7):
[0028]
[0029]
[0030]
[0031] S35: Introducing a quality fitness evaluation function during evolution. and diversity fitness value evaluation function The performance of the individual generator is evaluated, and both are calculated as shown in equations (8) and (9):
[0032]
[0033]
[0034] In equations (8) to (9), E x E represents the mathematical expectation of real data. z This represents the mathematical expectation of the noise data;
[0035] Combining equations (8) and (9) to set the individual performance global fitness value function As in equation (10):
[0036]
[0037] In equation (10), γ is the balance factor for balancing the two fitness value functions;
[0038] S36: Generator offspring individuals obtained after mutation and evaluation, ranked by individual fitness. Sort the generators from highest to lowest fitness, retain the generator with the highest fitness value as the parent for the next iteration, and output the generator with the highest fitness value when the preset number of iterations is reached.
[0039] Preferably, in step S40, the BiGRU model is constructed according to the following steps:
[0040] S41: The BiGRU model includes two-directional predictions, each containing a gated recurrent unit (GRU) network.
[0041] S42: After BiGRU performs bidirectional mining of the temporal correlation of the data in the time series, it outputs the wind power sequence for future time moments.
[0042] Preferably, step S42 is performed according to the following steps:
[0043] Update Gate Z t and reset door r t These are used to control the amount of old time-series memory retained and the combination of old and new information, and to determine the network output, ensuring that important time information in the long-term time series is not cleared during the iteration process; the network calculation formulas are shown in equations (11) to (15):
[0044] r t =σ(W r·[h t-1 ,x t ]+b) (11)
[0045] z t =σ(W z ·[h t-1 ,x t (12)
[0046]
[0047]
[0048] y t =σ(W o ·h t (15)
[0049] In equations (11) to (15): W r To reset the gate weight matrix; W z To update the gate weight matrix; W is the weight matrix for state updates. o The output weight matrix; x t b is the input at time t; b is the bias vector; h t It is the output of the GRU unit at time t; To retain the memory of the previous state; y t is the output at time t; the operator e indicates element-wise multiplication of two matrices; σ and tanh represent the Sigmoid and tanh activation functions, respectively.
[0050] Preferably, in step S60, the detailed steps of the cross-sectional algorithm are as follows:
[0051] S61: Initialize the population P; set the population size to N, and the dimension of each particle in the population to D, then the population P can be expressed as P = {P1, P2, ..., P...} n}, particle P r =[p r1 ,p r2 ,...,p rd ] T ;
[0052] S62: After the initial population is established, the fitness of the initial population is calculated according to a specific fitness function, and the corresponding fitness value is assigned to each particle.
[0053] S63: Perform a horizontal crossover operation. Horizontal crossover searches for offspring within half the population size of the hypercube. This is achieved by randomly selecting two individuals without repetition from the population to perform the crossover operation, thus exchanging information between different individuals. Randomly select any two particles P from the parent generation. i and Pj Then, the horizontal offspring P after performing the horizontal crossover operation can be obtained through equations (16) and (17). i,hc and P j,hc :
[0054] P i,hc =r1P i +(1-r1)P j +c1(P i -P j (16)
[0055] P j,hc =r2P j +(1-r2)P i +c2(P j -P i (17)
[0056] In equations (16) to (17), r1, r2, c1, and c2 are all random numbers with values in the range [0, 1].
[0057] S64: Perform a vertical crossover operation; vertical crossover searches different dimensions within the same individual to exchange information across different dimensions within the same individual, thereby helping the population escape local optima from fixed dimensions; for the r-th particle P in the population... r The vertical offspring P after the vertical cross operation in dimensions d1 and d2. rd1,hc It can be obtained from equation (18):
[0058]
[0059] In the formula, r is a random number with a value in the range [0,1]; p rd1 p rd2 P represents the r-th particle in the population. r The d1 and d2 dimensions;
[0060] S65: After performing a complete crossover, retain the population with better fitness values as the parents for the next iteration.
[0061] Preferably, in step S62, the specific fitness function is expressed as:
[0062]
[0063] Among them, y i (t) represents the actual wind power value at time t in the test set. Let n be the power prediction value of the prediction model at time t, and n represent the test set length.
[0064] Compared with the prior art, the beneficial effects of the present invention are:
[0065] The present invention provides a method for short-term prediction of wind power in newly built wind farms with limited data based on evolutionary generative adversarial networks and bidirectional gated recurrent units. The method uses evolutionary computation to optimize the generative adversarial network, enabling the generative model to efficiently learn the marginal distribution of the original limited data and generate new data with modal diversity and similar marginal distributions. This method compensates for the limitations of the original small-scale data and provides practical help in improving the accuracy of wind power prediction in newly built wind farms with limited data.
[0066] The present invention provides a method for short-term wind power prediction of newly built wind farms with limited data, based on evolutionary generative adversarial networks and bidirectional gated recurrent units. This method uses a cross-cutting optimization algorithm to optimize the weights and bias terms of the Dense layer in the BiGRU network, which can effectively avoid the model getting trapped in local optima and help it find the global optimum. This method has a significant effect on improving the accuracy of wind power prediction for newly built wind farms with limited data. Attached Figure Description
[0067] Figure 1 A flowchart illustrating a method for short-term wind power forecasting with limited data for newly constructed wind farms;
[0068] Figure 2 A schematic diagram illustrating the steps involved in constructing an evolutionary generative adversarial network;
[0069] Figure 3 A rendering of a method for short-term wind power prediction with limited data for newly built wind farms; Detailed Implementation
[0070] The present invention will be further described below with reference to specific embodiments. The accompanying drawings are for illustrative purposes only, representing schematic diagrams rather than actual physical objects, and should not be construed as limiting the scope of this patent. To better illustrate the embodiments of the present invention, some components in the drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions of the product. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.
[0071] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present patent. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.
[0072] Example 1
[0073] like Figure 1 The illustration shows an embodiment of the short-term wind power prediction method for newly built wind farms with limited data according to the present invention, which includes the following steps:
[0074] S10. Obtain the raw wind data collected by the original wind farm sensor, and use the Pearson correlation coefficient method to select features of all wind factors in the raw wind data, wherein the wind factors include wind power, wind speed, wind direction, atmospheric pressure, temperature and humidity;
[0075] S20. Normalize the feature data selected in step S10 to the [0,1] value range to construct the input matrix X of the small data time series. input =[X power X speed X direction ], where X power X speed X direction These represent power sequence, wind speed sequence, and wind direction sequence, respectively.
[0076] S30. Construct an evolutionary generative adversarial network (EGAN), and transmit the original small data input matrix X to the EGAN, where X is expressed as:
[0077]
[0078] In the formula: These represent the power, wind speed, and wind direction values of the wind farm at time t, respectively, and T represents the time length of the input data.
[0079] S40. Evolutionary Generative Adversarial Networks learn by studying the marginal distribution of the original, limited data, and then generate new data X with modal diversity and the same statistical similarity. generated To achieve data augmentation;
[0080] S50. Generate the new data X generated The augmented data X is formed by concatenating the original small data input matrix X with the original small data input matrix X. new And will enhance data X new The data is transmitted to a bidirectional gated cyclic unit (BiGRU) model, and the BiGRU model is used to process the input data X. new Time-related information is extracted bidirectionally;
[0081] S60. To address the problem that the hyperparameters of the Dense layer in the BiGRU network are prone to getting trapped in local optima, the CSO algorithm is introduced to optimize the hyperparameters of the Dense layer and obtain the predicted wind power sequence.
[0082] In step S10, the feature selection using the Pearson correlation coefficient method is implemented as follows:
[0083] S11. Calculate the Pearson coefficient score r of other characteristics and wind power in the original wind power data. The calculation process is as follows:
[0084]
[0085] In the formula: n is the length of the feature data; X represents the other 5 meteorological features, including wind speed, wind direction, atmospheric pressure, temperature, and humidity; Y represents the average value of the corresponding feature; Y represents wind power. This represents the average wind power output.
[0086] S12: Based on the Pearson coefficient score r of different features and wind power, features with negative scores are discarded and features with positive scores are retained. In this embodiment, wind speed and wind direction are ultimately selected as the two features.
[0087] Example 2
[0088] This embodiment is similar to Embodiment 1, except that:
[0089] In step S30, the EGAN model is constructed according to the following steps, such as Figure 2 As shown:
[0090] S31: The EGAN model consists of a generator and a discriminator;
[0091] S32: Transmit several sets of noise data z that follow a uniform or normal distribution to the generator to produce a new random variable G(z), i.e., virtual data;
[0092] S33: The virtual data G(z) and real data are fed together into the discriminator for discrimination. The discriminator identifies the label of the input sample, i.e., real or fake. During training, the generator and discriminator update their network parameters using a specific loss function and gradient descent method. The generator's loss function L... G The loss function L of the discriminator D The definitions are shown in equations (3) and (4) respectively:
[0093]
[0094]
[0095] In equations (3) to (4), z represents noise; G(z) is a new random number generated by the generator; D(x) represents the discriminator's judgment output on the real data x; D(G(x)) represents the discriminator's judgment output on G(z); This represents the mathematical expectation of the distribution of noisy data and generated data; It represents the mathematical expectation of the real data and the distribution of the real data; log(.) is the logarithmic function.
[0096] S34: Set different mutation operators, i.e., different objective functions, to guide the initial generator to evolve and produce a generator population, narrowing the distance between the generated data and the distribution of real data from different perspectives, and the population iteratively evolves in a given environment, i.e., a discriminator; mutation operators include the Minimax mutation operator. Heuristic mutation operator and Least-Squares mutation operator The corresponding functions are shown in equations (5) to (7):
[0097]
[0098]
[0099]
[0100] S35: Introducing a quality fitness evaluation function during evolution. and diversity fitness value evaluation function The performance of the individual generator is evaluated, and both are calculated as shown in equations (8) and (9):
[0101]
[0102]
[0103] In equations (8) to (9), E x E represents the mathematical expectation of real data. z The mathematical expectation of the noise data.
[0104] Combining equations (8) and (9) to set the individual performance global fitness value function As in equation (10):
[0105]
[0106] In equation (10), γ is the balance factor for balancing the two fitness value functions.
[0107] S36: Generator offspring individuals obtained after mutation and evaluation, ranked by individual fitness. Sort the generators from highest to lowest fitness, retain the generator with the highest fitness value as the parent for the next iteration, and output the generator with the highest fitness value when the preset number of iterations is reached.
[0108] In this embodiment, both the generator and the discriminator are constructed from four fully connected neural networks; the generator has 100, 128, 128, and 72 neurons, and the discriminator has 72, 128, 128, and 1 neuron, respectively.
[0109] Example 3
[0110] This embodiment is similar to Embodiment 1 or Embodiment 2, except that:
[0111] In step S40, the BiGRU model is constructed according to the following steps:
[0112] S41: The BiGRU model includes two-directional predictions, each containing a gated recurrent unit (GRU) network.
[0113] S42: After bidirectional mining of the temporal correlation of the data in the time series using the BiGRU model, the wind power series for future moments is output.
[0114] In step S41, the GRU network consists of three neural network layers with 4, 8 and 16 neurons respectively; the output layer of the BiGRU consists of two Dense layers with 50 and 1 neurons respectively.
[0115] Step S42 is performed as follows:
[0116] Update Gate Z t and reset door r t These are used to control the amount of old time-series memory retained and the combination of old and new information, and to determine the network output, ensuring that important time information in the long-term time series is not cleared during the iteration process; the network calculation formulas are shown in equations (11) to (15).
[0117] r t =σ(W r ·[h t-1 ,x t ]+b) (11)
[0118] z t =σ(W z ·[h t-1 ,x t (12)
[0119]
[0120]
[0121] y t =σ(W o ·h t (15)
[0122] In equations (11) to (15): W r To reset the gate weight matrix; W z To update the gate weight matrix; W is the weight matrix for state updates. o The output weight matrix; x t b is the input at time t; b is the bias vector; h t It is the output of the GRU unit at time t; To retain the memory of the previous state; y t is the output at time t; the operator e indicates element-wise multiplication of two matrices; σ and tanh represent the Sigmoid and tanh activation functions, respectively.
[0123] In step S60, the detailed steps of the cross-sectional algorithm are as follows:
[0124] S61: Initialize the population P; set the population size to N, and the dimension of each particle in the population to D, then the population P can be expressed as P = {P1, P2, ..., P...} n}, particle P r =[p r1 ,p r2 ,...,p rd ] T ;
[0125] S62: After the initial population is established, the fitness of the initial population is calculated according to a specific fitness function, and the corresponding fitness value is assigned to each particle.
[0126] S63: Perform a horizontal crossover operation. Horizontal crossover searches for offspring within half the population size of the hypercube. This is achieved by randomly selecting two individuals without repetition from the population to perform the crossover operation, thus exchanging information between different individuals. Randomly select any two particles P from the parent generation. i and P j Then, the horizontal offspring P after performing the horizontal crossover operation can be obtained through equations (16) and (17). i,hc and P j,hc :
[0127] P i,hc =r1P i +(1-r1)P j +c1(P i -P j (16)
[0128] P j,hc =r2P j +(1-r2)P i +c2(P j -P i (17)
[0129] In equations (16) to (17), r1, r2, c1, and c2 are all random numbers with values in the range [0, 1].
[0130] S64: Perform a vertical crossover operation; vertical crossover searches different dimensions within the same individual to exchange information across different dimensions within the same individual, thereby helping the population escape local optima from fixed dimensions; for the r-th particle P in the population... r The vertical offspring P after the vertical cross operation in dimensions d1 and d2. rd1,hc It can be obtained from equation (18):
[0131]
[0132] In the formula, r is a random number with a value in the range [0,1]; p rd1 ,p rd2 P represents the r-th particle in the population. r The d1 and d2 dimensions.
[0133] S65: After performing a complete crossover, retain the population with better fitness values as the parents for the next iteration.
[0134] In step S62, the specific fitness function is expressed as:
[0135]
[0136] Among them, y i (t) represents the actual wind power value at time t in the test set. Let n be the power prediction value of the prediction model at time t, and n represent the test set length.
[0137] In step S60, the prediction performance of the BiGRU-CSO prediction model is evaluated using the mean absolute error (MAE), root mean square error (RMSE), and mean absolute error percentage (MAPE).
[0138]
[0139]
[0140]
[0141] Where y represents the actual value of the wind power output in the test cluster. To test the average value of the actual concentrated wind power, This represents the power prediction value from the prediction model.
[0142] The prediction model of this invention is denoted as the EGAN-BiGRU-CSO model. This application plots the prediction curves of the EGAN-BiGRU-CSO model and the EGAN-BiGRU model, as shown below. Figure 3 As shown, the predicted curves of both models are compared with the actual wind power variation curves over time. Analysis reveals that both the EGAN-BiGRU-CSO model and the EGAN-BiGRU model have high prediction accuracy, but:
[0143] The EGAN-BiGRU-CSO model exhibits higher prediction accuracy compared to the EGAN-BiGRU model. This is because: The EGAN-BiGRU-CSO model employs a cross-optimization algorithm to optimize the weights and biases of the Dense layer in the BiGRU network, effectively avoiding local optima and helping the model find the global optimum. This significantly improves the prediction accuracy of wind power for newly built wind farms with limited data. Furthermore, the EGAN-BiGRU-CSO model utilizes evolutionary computation to optimize the generative adversarial network, enabling the generative model to efficiently learn the marginal distribution of the original limited data and generate new data with modal diversity and similar marginal distributions. This compensates for the limitations of the original small-scale data, practically contributing to improving the prediction accuracy of wind power for newly built wind farms with limited data.
[0144] In the specific implementation of the above embodiments, the technical features can be combined in any non-contradictory way. For the sake of brevity, not all possible combinations of the above technical features are described. However, as long as the combination of these technical features is not contradictory, it should be considered to be within the scope of this specification.
Claims
1. A method for short-term prediction of wind power in newly built wind farms with limited data, characterized in that, Includes the following steps: S10. Obtain the raw wind data collected by the original wind farm sensor, and use the Pearson correlation coefficient method to select features of all wind factors in the raw wind data, wherein the wind factors include power, wind speed, wind direction, atmospheric pressure, temperature, and humidity; S20. Normalize the feature data selected in step S10 to the [0,1] value range to construct the input matrix X of the small data time series. input =[X power X speed X direction ], where X power X speed X direction These represent the power sequence, wind speed sequence, and wind direction sequence, respectively. S30. Construct an evolutionary generative adversarial network (EGAN), and transmit the original small amount of data input matrix X to the EGAN, where X is expressed as: In the formula: These represent the values of power, wind speed, and wind direction of the wind farm at time t, respectively, and T represents the time length of the input data. S40. Evolutionary Generative Adversarial Networks (EGANs) learn by studying the marginal distribution of raw, limited data, thereby generating new data X with modal diversity and similar statistical similarity. generated To achieve data augmentation; S50. Generate the new data X generated The augmented data X is formed by concatenating the original small data input matrix X with the original small data input matrix X. new And will enhance data X new The data is transmitted to a bidirectional gated cyclic unit (BiGRU) model, and the BiGRU model is used to process the input data X. new Time-related information is extracted bidirectionally; S60. To address the problem that the hyperparameters of the Dense layer in the BiGRU network are prone to getting trapped in local optima, the CSO algorithm is introduced to optimize the hyperparameters of the Dense layer and obtain the predicted wind power sequence.
2. The short-term wind power prediction method for newly built wind farms with limited data according to claim 1, characterized in that, In step S10, the feature selection using the Pearson correlation coefficient method is implemented as follows: S11. Calculate the Pearson coefficient score r of other characteristics and wind power in the original wind power data. The calculation process is as follows: In the formula: n is the length of the feature data; X represents the other 5 meteorological features, including wind speed, wind direction, atmospheric pressure, temperature, and humidity; Y represents the average value of the corresponding feature; Y represents wind power. This represents the average wind power output. S12: Based on the Pearson coefficient score r of different features and wind power, features with negative scores are discarded and features with positive scores are retained.
3. The short-term wind power prediction method for newly built wind farms with limited data according to claim 1, characterized in that, In step S30, the EGAN model is constructed according to the following steps: S31: The EGAN model consists of a generator and a discriminator; S32: Transmit several sets of noise data z that follow a uniform or normal distribution to the generator to produce a new random variable G(z), i.e., virtual data; S33: The virtual data G(z) and real data are fed together into the discriminator for discrimination. The discriminator identifies the label of the input sample, i.e., real or fake. During training, the generator and discriminator update their network parameters using a specific loss function and gradient descent method. The generator's loss function L... G The loss function L of the discriminator D The definitions are shown in equations (3) and (4) respectively: In equations (3) to (4), z represents noise; G(z) is a new random number generated by the generator; D(x) represents the discriminator's judgment output on the real data x; D(G(x)) represents the discriminator's judgment output on G(z); This represents the mathematical expectation of the distribution of noisy data and generated data; This represents the mathematical expectation of the real data and its distribution; log(.) is the logarithmic function. S34: Set different mutation operators, i.e., different objective functions, to guide the initial generator to evolve and produce a generator population, narrowing the distance between the generated data and the distribution of real data from different perspectives, and the population iteratively evolves in a given environment, i.e., a discriminator; mutation operators include the Minimax mutation operator. Heuristic mutation operator and Least-Squares mutation operator The corresponding functions are shown in equations (5) to (7): S35: Introduce a quality fitness evaluation function during evolution. and diversity fitness value evaluation function The performance of the individual generator is evaluated, and both are calculated as shown in equations (8) and (9): In equations (8) to (9), E x E represents the mathematical expectation of real data. z This represents the mathematical expectation of the noise data; Combining equations (8) and (9) to set the individual performance global fitness value function As in equation (10): In equation (10), γ is the balance factor for balancing the two fitness value functions; S36: Generator offspring individuals obtained after mutation and evaluation, ranked by individual fitness. Sort the generators from highest to lowest fitness, retain the generator with the highest fitness value as the parent for the next iteration, and output the generator with the highest fitness value when the preset number of iterations is reached.
4. The short-term wind power prediction method for newly built wind farms with limited data according to claim 3, characterized in that, Both the generator and the discriminator are constructed from four fully connected layers of neural networks; the generator has 100, 128, 128, and 72 neurons, respectively, and the discriminator has 72, 128, 128, and 1 neurons, respectively.
5. The method for short-term wind power prediction in newly built wind farms with limited data, as described in claim 1, is characterized in that... In step S40, the BiGRU model is constructed according to the following steps: S41: The BiGRU model includes two-directional predictions, each containing a gated recurrent unit (GRU) network. S42: After BiGRU performs bidirectional mining of the temporal correlation of the data in the time series, it outputs the wind power sequence for future time moments.
6. The method for short-term wind power prediction in newly built wind farms with limited data, as described in claim 1, is characterized in that... In step S41, the GRU network consists of three neural network layers with 4, 8 and 16 neurons respectively; the output layer of the BiGRU consists of two Dense layers with 50 and 1 neurons respectively.
7. The short-term wind power prediction method for newly built wind farms with limited data according to claim 5, characterized in that, Step S42 is performed as follows: Update Gate Z t and reset door r t These are used to control the amount of old time-series memory retained and the combination of old and new information, and to determine the network output, ensuring that important time information in the long-term time series is not cleared during the iteration process; the network calculation formulas are shown in equations (11) to (15): r t =σ(W r ·[h t-1 ,x t ]+b) (11) z t =σ(W z ·[h t-1 ,x t ]) (12) y t =σ(W o ·h t ) (15) In equations (11) to (15): W r To reset the gate weight matrix; W z To update the gate weight matrix; W is the weight matrix for state updates. o The output weight matrix; x t b is the input at time t; b is the bias vector; h t It is the output of the GRU unit at time t; To retain the memory of the previous state; y t is the output at time t; the operator e indicates element-wise multiplication of two matrices; σ and tanh represent the Sigmoid and tanh activation functions, respectively.
8. The method for short-term wind power prediction in newly built wind farms with limited data, as described in claim 1, is characterized in that... In step S60, the detailed steps of the cross-sectional algorithm are as follows: S61: Initialize the population P; set the population size to N, and the dimension of each particle in the population to D, then the population P can be expressed as P = {P1, P2, ..., P...} n }, particle P r =[p r1 ,p r2 ,...,p rd ] T ; S62: After the initial population is established, the fitness of the initial population is calculated according to a specific fitness function, and the corresponding fitness value is assigned to each particle. S63: Perform a horizontal crossover operation. Horizontal crossover searches for offspring within half the population size of the hypercube. This is achieved by randomly selecting two individuals without repetition from the population to perform the crossover operation, thus exchanging information between different individuals. Randomly select any two particles P from the parent generation. i and P j Then, the horizontal offspring P after performing the horizontal crossover operation can be obtained through equations (16) and (17). i,hc and P j,hc : P i,hc =r1P i +(1-r1)P j +c1(P i -P j ) (16) P j,hc =r2P j +(1-r2)P i +c2(P j -P i ) (17) In equations (16) to (17), r1, r2, c1, and c2 are all random numbers with values in the range [0, 1]. S64: Perform a vertical crossover operation; vertical crossover searches different dimensions within the same individual to exchange information across different dimensions within the same individual, thereby helping the population escape local optima from fixed dimensions; for the r-th particle P in the population... r The vertical offspring after the vertical cross operation in dimensions d1 and d2. It can be obtained from equation (18): In the formula, r is a random number with a value in the range [0,1]; p rd1 p rd2 P represents the r-th particle in the population. r The d1 and d2 dimensions; S65: After performing a complete crossover, retain the population with better fitness values as the parents for the next iteration.
9. The method for short-term wind power prediction in newly built wind farms with limited data, as described in claim 8, is characterized in that... In step S62, the specific fitness function is expressed as: Among them, y i (t) represents the actual wind power value at time t in the test set. Let n be the power prediction value of the prediction model at time t, and n represent the test set length.
10. The method for short-term wind power prediction in newly built wind farms with limited data, according to any one of claims 1 to 9, is characterized in that, In step S60, the prediction performance of the BiGRU-CSO prediction model is evaluated using the mean absolute error (MAE), root mean square error (RMSE), and mean absolute error percentage (MAPE). Where y represents the actual value of the wind power output in the test cluster. To test the average value of the actual concentrated wind power, This represents the power prediction value from the prediction model.