A multi-network combined multi-region time-frequency-space wind power output prediction method

Abstract

The application provides a multi-network combined multi-region time-frequency-space wind power output prediction method. The method first uses wind farm original data for training of an Inception V4 network, a time convolution network, a bidirectional gate recurrent neural network and a Universal Transformer network. The output results of the Inception V4 network and the time-space convolution network are input into the Universal Transformer network, and the output result of the Universal Transformer network is added to the output result of the bidirectional gate recurrent neural network to obtain a wind power output prediction result of the region. The data processing methods of the remaining regions are the same as the above method. Finally, the prediction results of the regions are processed by a three-layer full connection layer and added to obtain a final wind power output prediction. The method can improve the accurate prediction of wind farm output data.

Description

Technical Field

[0001] This invention belongs to the field of new energy and artificial intelligence, and involves a new generation of artificial intelligence methods applicable to power output prediction of new energy wind farms. Background Technology

[0002] Existing shallow learning models for wind power output prediction are not good at capturing data features, resulting in low prediction accuracy and difficulty in meeting the needs of practical applications. On the other hand, traditional deep learning models use fixed network structures and methods, which are difficult to adapt to different wind energy prediction problems and environmental conditions. They also lack the ability to dynamically update network parameters, resulting in a lack of flexibility and adaptability.

[0003] Furthermore, existing models for wind power output prediction cannot be matched with real-world application conditions that differ significantly from the model training conditions. They have weak generalization ability, limited application scope, and cannot be applied on a large scale.

[0004] Therefore, to address the aforementioned problems, a multi-network, multi-regional time-frequency-space wind power output prediction method is proposed. This method enables the model to have self-optimization capabilities, allowing it to adjust hyperparameters based on changes in real-time data. Simultaneously, the method improves the versatility of the wind power output prediction model, enabling it to adapt to different operating conditions and provide wind power output predictions. Summary of the Invention

[0005] This invention proposes a multi-network combined multi-regional time-frequency-space wind power output prediction method. This method applies discrete wavelet transform and complete ensemble empirical mode decomposition methods in data processing, and comprehensively utilizes various network structures to achieve more accurate wind farm output prediction. The method has the ability to adjust hyperparameters based on real-time data changes, improving the model's generalizability, generalization ability, and prediction accuracy. This facilitates more precise planning of wind farm output, improves wind energy utilization, reduces power system operating costs and risks, enhances power system reliability and stability, improves the economic efficiency and reliability of wind farms, saves energy, reduces carbon emissions, and achieves sustainable development and environmental protection. The steps in the usage process are as follows:

[0006] Step (1): First, collect 96×7 sets of raw wind farm data every 15 minutes for 7 consecutive days. Each set of data includes the raw power output data of the wind farm, wind speed, wind direction, ambient temperature, air density, air pressure, eddies, wind turbine height, average pitch torque, average pitch angle, number of wind turbine blades, rotor speed, generator active power, generator reactive power, precipitation, and air pollution data.

[0007] Step (2): First, input the original dataset of the wind farm, including the original power output data of the wind farm, wind speed, wind direction, ambient temperature, air density, air pressure, eddies, wind turbine height, average pitch torque, average pitch angle, number of wind turbine blades, rotor speed, generator active power, generator reactive power, precipitation, and air pollution data.

[0008] The input raw wind farm data is cleaned, and then the raw wind farm output data is randomly divided into training set, validation set and test set in a ratio of 7:2:1.

[0009] Step (3): Extract the historical wind power output data sequence x n Perform complete set empirical mode decomposition:

[0010] For x n Adding Gaussian white noise with a mean of 0 once yields the noisy data x′. n for:

[0011]

[0012] in, ei represents the Gaussian white noise that follows a normal distribution added in the i-th experiment, where i = 1, 2, ..., I; e0 is the deweighting coefficient of the Gaussian white noise;

[0013] Performing an I-th complete set of empirical mode decomposition on it and taking the average yields the first mode component IMF1(n):

[0014]

[0015] in, The i-th component of the complete set empirical mode decomposition;

[0016] Calculate the first residual component r 1n for:

[0017] r 1n =x 1n -IMF1(n) (3)

[0018] For the first residual component r obtained 1n Noise is added to obtain the residual component r′ after noise addition. 1n for:

[0019]

[0020] Where e1 is the Gaussian white noise deweighting coefficient; E1() is the first modal component of the complete set empirical mode decomposition; The Gaussian white noise that follows a normal distribution is added in the i-th experiment; r 1nThis is the first residual component;

[0021] For r′ 1n The second modal component IMF2(n) obtained by performing complete set empirical mode decomposition is:

[0022]

[0023] Where E1() is the first modal component of the complete set empirical mode decomposition; r 1n e1 represents the first residual component; e1 is the Gaussian white noise deweighting coefficient. The Gaussian white noise that follows a normal distribution is added in the i-th experiment;

[0024] Let k = 2, 3, ..., K, calculate the k-th residual component r. kn :

[0025] r kn =r (k-1)n -IMF k (n) (6)

[0026] Where r (k-1)n For the (k-1)th residual component, IMF k (n) represents the k-th modal component, K = I, which is the total number of complete set empirical mode decompositions performed;

[0027] For the k-th residual component r kn Adding noise yields:

[0028]

[0029] Among them, e k E represents the weighted coefficients of Gaussian white noise. k () represents the k-th modal component of the complete set empirical mode decomposition; The Gaussian white noise that follows a normal distribution is added in the i-th experiment; i = 1, 2, ..., I; r kn This is the k-th residual component;

[0030] For r′ kn Perform complete set empirical mode decomposition to obtain the (k+1)th mode component (IMF). k+1 (n) is:

[0031]

[0032] Where E k () represents the k-th mode component of the complete set empirical mode decomposition, r kn For the k-th residual component, e k The weighted coefficients are Gaussian white noise; The Gaussian white noise that follows a normal distribution is added in the i-th experiment;

[0033] The decomposition ends when the residual component is monotonic and the number of its extreme points is less than 3, and the final residual component r is obtained. n for:

[0034]

[0035] Among them, IMF k This is the k-th modal component;

[0036] Step (4): Input the historical wind power output data sequence after complete set empirical mode decomposition into the bidirectional gated recurrent neural network after Bayesian optimization to obtain the processed data after processing by the bidirectional gated recurrent neural network;

[0037] The state of the forward hidden layer at time t of the bidirectional gated recurrent unit for:

[0038]

[0039] The state of the backward hidden layer at time t of the bidirectional gated recurrent unit for:

[0040]

[0041] The state h of the hidden layer at time t of the bidirectional gated recurrent unit t for:

[0042]

[0043] The GRU() function is a non-linear transformation of the input time data, x t For input, The output of the forward hidden layer state at time t. The output is fed back to the hidden layer at time t. The output is the forward hidden layer state at time t-1. The output to the hidden layer after time t-1; w t Represents the state of the forward hidden layer at time t. The corresponding weight, v t Let t be the state of the backward hidden layer. The corresponding weight, b t This represents the bias corresponding to the hidden layer state of the bidirectional gated recurrent neural network at time t;

[0044] Step (5): Transfer historical wind speed and direction data y t Perform discrete wavelet transform processing:

[0045] Input signal y t With a set of discrete wavelet DW a,b The integral of (t) is DWY(y) t ) a,b for:

[0046]

[0047] Where 'a' is the scaling factor of the wavelet function, and 'b' is the time localization factor. Z is an integer. Let k be the Daubechies wavelet function. z Let represent the translation parameter of the wavelet function on the time axis, and j represent the scaling parameter of the wavelet function. This represents a time-time stretching and translation of the wavelet function;

[0048] The discrete wavelet transform can be formalized into a scaling function (14) and a discrete wavelet function (15).

[0049] Scaling Function for:

[0050]

[0051] Discrete wavelet function for:

[0052]

[0053] Approximate part for:

[0054]

[0055] Detailed information section for:

[0056]

[0057] in for The complex conjugate function; the coefficients of the discrete wavelet transform are derived from the approximate part. and detailed information section Composed of two parts, y can be obtained through equations (16) and (17). t The discrete wavelet transform coefficients; the approximation part is a low-pass filter, and the detailed information part is a high-pass filter; the original signal is decomposed three times to obtain multi-level neighborhood features;

[0058] Each decomposition has two branches corresponding to low-frequency features and high-frequency features; in the first stage of the discrete wavelet transform, the original signal is decomposed into a low-frequency part "approximate component 1" and a high-frequency part "detail component 1"; then, the low-frequency part "approximate component 1" is continuously decomposed into "approximate component 2" and "detail component 2".

[0059] After performing the above operations three times, the multi-resolution frequency domain features "approximate component", "detail component 1", "detail component 2" and "detail component 3" will be obtained to reflect the common frequency characteristics of the line signal.

[0060] Step (6): Input the processed data into the spatiotemporal convolutional network, and pass it through multiple extended causal convolutional layers, weight normalization, ReLU activation, spatial dropout operation, and sum with the data that only performs one-dimensional convolution operation to obtain the processed data;

[0061] Step (7): Collect historical data on ambient temperature, air density, air pressure, eddies, wind turbine height, average pitch torque, average pitch angle, number of wind turbine blades, turbine speed, generator active power, generator reactive power, precipitation, and air pollution data. n Normalization is performed, and the normalized data S n for:

[0062]

[0063] in, -z represents e n The power, e is the natural constant, z n Historical data representing ambient temperature, air density, air pressure, eddies, wind turbine height, average pitch torque, average pitch angle, number of wind turbine blades, rotor speed, generator active power, generator reactive power, precipitation, and air pollution data.

[0064] The normalized data S(n)′ for generating the RGB image is:

[0065] S′ n =S n ×255 (19)

[0066] The data is normalized to the range of 0-255 and rounded down, and then an RGB image is generated based on the normalized integer data.

[0067] Step (8): Input the parameters of the Inception V4 memory convolutional network, including the weights W. network and bias b network The policy parameters θ are a set:

[0068] θ=[W networkb network ]

[0069] Step (9): Input the RGB image into the Inception V4 memory convolutional network, and obtain the processed data through convolutional layers, pooling layers, and fully connected layers;

[0070] Step (10): Input the data processed by the Inception V4 network and the data processed by the spatiotemporal convolutional network into the Universal Transformer network, and sum the results with the results processed by the Bayesian optimized bidirectional gated recurrent neural network to obtain the wind power output prediction of region I.

[0071] Step (11): Compare the obtained wind power output prediction results with the verification set, adjust the network structure parameters according to the comparison results, further optimize the network structure parameters, and improve the accuracy of the prediction results;

[0072] Step (12): Compare and analyze the wind power output prediction results obtained from the optimized network processing with the test set to obtain the prediction accuracy.

[0073] Step (13): For regions II to N, collect 24×7 sets of data every hour for 7 consecutive days; the data processing method and network structure are the same as those for region I.

[0074] Step (14): The wind power output prediction results obtained from each region are processed through three fully connected layers to obtain the final wind power output prediction results.

[0075] The present invention has the following advantages and effects compared with the prior art:

[0076] (1) This invention proposes a multi-network combined multi-regional time-frequency-space wind power output prediction method. This method employs multiple processing methods and network models, including discrete wavelet transform, complete ensemble empirical mode decomposition and Inception V4 network, spatiotemporal convolutional network, Bayesian optimized bidirectional gated recurrent neural network, and Universal Transformer network. By selectively choosing different processing methods and models, the accuracy of wind farm output prediction results can be improved, and it has good generalization ability.

[0077] (2) This invention decomposes a prediction problem into a prediction problem in multiple regions, a prediction problem in multiple time scales, and a prediction problem in multiple frequencies; and adopts a deep fully connected layer to combine the advantages of multiple networks, avoiding the limitations of a single network; and finally can greatly improve the accuracy of prediction.

[0078] (3) The method of the present invention is conducive to more accurate prediction of wind farm output. The high-precision prediction results of this method can provide more reliable data support for power system dispatch and energy planning, which is conducive to improving the utilization efficiency of wind farms, further increasing the proportion of renewable energy, and reducing dependence on traditional energy.

[0079] (4) The method of the present invention is conducive to saving energy, reducing carbon emissions, effectively avoiding energy waste caused by excessive reserve capacity, and reducing power shortage and energy shortage problems caused by insufficient power generation.

[0080] (5) This invention can effectively reduce the instability and unpredictability of wind farm operation, and reduce the negative impacts and carbon emissions of wind farms. Therefore, the method of this study helps to achieve energy conservation and emission reduction goals and promote sustainable development. Attached Figure Description

[0081] Figure 1 This is a diagram illustrating the overall framework of the prediction method of this invention.

[0082] Figure 2 This is a spatiotemporal convolutional network diagram of the method of this invention.

[0083] Figure 3 This is a diagram of the bidirectional gated recurrent neural network structure of the method of this invention.

[0084] Figure 4 This is a diagram of the Inception V4 network structure of the method of this invention.

[0085] Figure 5 This is a diagram of the Universal Transformer network structure of the method of this invention. Detailed Implementation

[0086] This invention proposes a multi-network combined multi-regional time-frequency-space wind power output prediction method, which is described in detail below with reference to the accompanying drawings:

[0087] Figure 1 This is a diagram illustrating the overall framework of the prediction method of this invention. The specific steps for predicting wind power output in a wind farm are as follows:

[0088] First, input the original dataset of the wind farm, including the original power output data of the wind farm, wind speed, wind direction, ambient temperature, air density, air pressure, eddies, wind turbine height, average pitch torque, average pitch angle, number of wind turbine blades, rotor speed, generator active power, generator reactive power, precipitation, and air pollution data.

[0089] The input raw wind farm data is cleaned, and then the raw wind farm output data is randomly divided into training set, validation set and test set in a ratio of 7:2:1.

[0090] Next, the historical wind speed and direction data are subjected to discrete wavelet transform, and the transformed data is input into a temporal convolutional network to obtain processed data 2; at the same time, the historical data of ambient temperature, air density, air pressure, eddy current, wind turbine height, average pitch torque, average pitch angle, number of wind turbine blades, rotor speed, generator active power, generator reactive power, precipitation, and air pollution data are normalized and multiplied by 255 and rounded down to the integer range of 0 to 255 to generate RGB images;

[0091] Next, the RGB image is input into the Inception V4 network to obtain processed data 3; data 2 and data 3 are input together into the Universal Transformer network, and then processed data 4 is added to processed data 1 to obtain the wind power output prediction result for region I; it is determined whether the network structure parameters have been optimized. If not, the obtained prediction result is compared with the historical wind power output data of the validation set to optimize the network structure parameters; if the network structure parameters have been optimized, proceed to the next step.

[0092] Finally, all steps are repeated to obtain the wind power output prediction results for Region I. The data processing method and network structure for other regions are the same as for Region I. The wind power output prediction results for each region are processed through three fully connected layers to obtain the final wind power output prediction results for the wind farm.

[0093] Figure 2 This is a diagram of the spatiotemporal convolutional network structure of the method of this invention.

[0094] Wind speed and direction data are input into a spatiotemporal convolutional network. After passing through several spatiotemporal convolutional network residual modules, global average pooling operations, and fully connected layers, the processed output data is finally obtained.

[0095] Figure 3 This is a diagram of the bidirectional gated recurrent neural network structure of the method of this invention.

[0096] A bidirectional gated recurrent neural network consists of an input layer, a forward hidden layer, a backward hidden layer, and an output layer.

[0097] The input layer contains input data, which is passed to the forward hidden layer and the backward hidden layer at each time step. That is, the data flows to two gated recurrent units in opposite directions at the same time. The forward hidden layer processes the timing information of the input sequence from left to right, and the backward hidden layer processes the timing information from right to left. The output sequence of the output layer is determined by these two gated recurrent units.

[0098] Figure 4 This is a diagram of the Inception V4 network structure of the method of this invention.

[0099] First, the input RGB image passes through the Stem layer. Then, it sequentially passes through four Inception-A modules, a Reduction-A module, seven Inception-B modules, a Reduction-B module, three Inception-C modules, a global average pooling layer, a Dropout operation, a fully connected layer, a Softmax layer, and an output layer.

[0100] Figure 5 This is a diagram of the Universal Transformer network structure of the method of this invention.

[0101] The Universal Transformer network consists of two parts: a loop encoder and a loop decoder.

[0102] First, the input sequence is symbolically embedded and fed into the recurrent encoder. In the recurrent encoder, time steps and positions are embedded first, then the sequence is fed into the multi-head attention mechanism, followed by a Dropout operation. This result is then summed with the input sequence without the multi-head attention mechanism and Dropout operation. Next, the summed sequence is subjected to layer normalization and processed using a transformation function, then Dropout is performed again and summed with the sequence that has only undergone layer normalization. The result is then subjected to layer normalization. This process is repeated T times, and the final result is fed into the multi-head attention mechanism in the recurrent decoder.

[0103] First, the input sequence is embedded and shifted right by one bit to obtain the target sequence. The target sequence symbol is then embedded and input into the recurrent decoder. In the recurrent decoder, the input sequence is first embedded with time steps and positions, and a multi-head self-attention mechanism is input. Then, a dropout operation is performed, and the result is summed with the sequence without multi-head self-attention and dropout. The summed result is then layer-normalized, input into the multi-head attention mechanism, and dropout is performed on this sequence and the sequence input from the recurrent encoder. This result is then summed with the sequence that has only undergone layer normalization. The summed result is then layer-normalized, processed using a transformation function, and then dropout is performed. This result is then summed with the sequence that has only undergone layer normalization. The resulting result is layer-normalized. This process is repeated T times. Finally, the result is processed using Softmax to obtain the output sequence.

[0104] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A multi-network combined multi-regional time-frequency-space wind power output prediction method, characterized in that, This method employs discrete wavelet transform and complete ensemble empirical mode decomposition in data processing, and comprehensively utilizes various network structures to achieve more accurate predictions of wind farm output. It possesses the ability to adjust hyperparameters based on real-time data changes, improving the model's generalizability, generalization ability, and prediction accuracy. This facilitates more precise planning of wind farm output, increases wind energy utilization, reduces power system operating costs and risks, enhances power system reliability and stability, improves the economic efficiency and reliability of wind farms, saves energy, reduces carbon emissions, and achieves sustainable development and environmental protection. The steps in its application are as follows: Step (1): First, collect 96×7 sets of raw wind farm data every 15 minutes for 7 consecutive days. Each set of data includes the raw power output data of the wind farm, wind speed, wind direction, ambient temperature, air density, air pressure, eddies, wind turbine height, average pitch torque, average pitch angle, number of wind turbine blades, rotor speed, generator active power, generator reactive power, precipitation, and air pollution data. Step (2): First, input the original dataset of the wind farm, including the original power output data of the wind farm, wind speed, wind direction, ambient temperature, air density, air pressure, eddies, wind turbine height, average pitch torque, average pitch angle, number of wind turbine blades, rotor speed, generator active power, generator reactive power, precipitation, and air pollution data. The input raw wind farm data is cleaned, and then the raw wind farm output data is randomly divided into training set, validation set and test set in a ratio of 7:2:

1. Step (3): Extract the historical wind power output data sequence x n Perform complete set empirical mode decomposition: For x n Adding Gaussian white noise with a mean of 0 once yields the noisy data x′. n for: in, ei represents the Gaussian white noise that follows a normal distribution added in the i-th experiment, where i = 1, 2, ..., I; e0 is the deweighting coefficient of the Gaussian white noise; Performing an I-th complete set of empirical mode decomposition on it and taking the average yields the first mode component IMF1(n): in, The i-th component of the complete set empirical mode decomposition; Calculate the first residual component r 1n for: r 1n =x 1n -IMF1(n) (3) For the first residual component r obtained 1n Noise is added to obtain the residual component r′ after noise addition. 1n for: Where e1 is the Gaussian white noise deweighting coefficient; E1() is the first modal component of the complete set empirical mode decomposition; The Gaussian white noise that follows a normal distribution is added in the i-th experiment; r 1n This is the first residual component; For r′ 1n The second modal component IMF2(n) obtained by performing complete set empirical mode decomposition is: Where E1() is the first modal component of the complete set empirical mode decomposition; r 1n e1 represents the first residual component; e1 is the Gaussian white noise deweighting coefficient. The Gaussian white noise that follows a normal distribution is added in the i-th experiment; Let k = 2, 3, ..., K, calculate the k-th residual component r. kn : r kn =r (k-1)n -IMF k (n) (6) Where r (k-1)n For the (k-1)th residual component, IMF k (n) represents the k-th modal component, K = I, which is the total number of complete set empirical mode decompositions performed; For the k-th residual component r kn Adding noise yields: Among them, e k E represents the weighted coefficients of Gaussian white noise. k () represents the k-th modal component of the complete set empirical mode decomposition; The Gaussian white noise that follows a normal distribution is added in the i-th experiment; i = 1, 2, ..., I; r kn This is the k-th residual component; For r′ kn Perform complete set empirical mode decomposition to obtain the (k+1)th mode component (IMF). k+1 (n) is: Where E k () represents the k-th mode component of the complete set empirical mode decomposition, r kn For the k-th residual component, e k The weighted coefficients are Gaussian white noise; The Gaussian white noise that follows a normal distribution is added in the i-th experiment; The decomposition ends when the residual component is monotonic and the number of its extreme points is less than 3, and the final residual component r is obtained. n for: Among them, IMF k This is the k-th modal component; Step (4): Input the historical wind power output data sequence after complete set empirical mode decomposition into the bidirectional gated recurrent neural network after Bayesian optimization to obtain the processed data after processing by the bidirectional gated recurrent neural network; The state of the forward hidden layer at time t of the bidirectional gated recurrent unit for: The state of the backward hidden layer at time t of the bidirectional gated recurrent unit for: The state h of the hidden layer at time t of the bidirectional gated recurrent unit t for: The GRU() function is a non-linear transformation of the input time data, x t For input, The output of the forward hidden layer state at time t. The output is fed back to the hidden layer at time t. The output is the forward hidden layer state at time t-1. The output to the hidden layer after time t-1; w t Represents the state of the forward hidden layer at time t. The corresponding weight, v t Let t be the state of the backward hidden layer. The corresponding weight, b t This represents the bias corresponding to the hidden layer state of the bidirectional gated recurrent neural network at time t; Step (5): Transfer historical wind speed and direction data y t Perform discrete wavelet transform processing: Input signal y t With a set of discrete wavelet DW a,b The integral of (t) DWT(y) t ) a,b for: Where a is the scaling factor of the wavelet function, and b is the time localization factor. Z is an integer. Let k be the Daubechies wavelet function. z Let represent the translation parameter of the wavelet function on the time axis, and j represent the scaling parameter of the wavelet function. This represents a time-time stretching and translation of the wavelet function; The discrete wavelet transform can be formalized into a scaling function (14) and a discrete wavelet function (15). Scaling Function for: Discrete wavelet function for: Approximate part for: Detailed information section for: in for The complex conjugate function; the coefficients of the discrete wavelet transform are derived from the approximate part. and detailed information section Composed of two parts, y can be obtained through equations (16) and (17). t The discrete wavelet transform coefficients; the approximation part is a low-pass filter, and the detailed information part is a high-pass filter; the original signal is decomposed three times to obtain multi-level neighborhood features; Each decomposition has two branches corresponding to low-frequency features and high-frequency features; in the first stage of the discrete wavelet transform, the original signal is decomposed into a low-frequency part "approximate component 1" and a high-frequency part "detail component 1"; then, the low-frequency part "approximate component 1" is continuously decomposed into "approximate component 2" and "detail component 2". After performing the above operations three times, the multi-resolution frequency domain features "approximate component", "detail component 1", "detail component 2" and "detail component 3" will be obtained to reflect the common frequency characteristics of the line signal. Step (6): Input the processed data into the spatiotemporal convolutional network, and pass it through multiple extended causal convolutional layers, weight normalization, ReLU activation, spatial dropout operation, and sum with the data that only performs one-dimensional convolution operation to obtain the processed data; Step (7): Collect historical data on ambient temperature, air density, air pressure, eddies, rotor height, average pitch torque, average pitch angle, number of rotor blades, rotor speed, generator active power, generator reactive power, precipitation, and air pollution data. n Normalization is performed, and the normalized data S n for: in, -z represents e n The power, e is the natural constant, z n Historical data representing ambient temperature, air density, air pressure, eddies, wind turbine height, average pitch torque, average pitch angle, number of wind turbine blades, rotor speed, generator active power, generator reactive power, precipitation, and air pollution data. The normalized data S(n)′ for generating the RGB image is: S′ n =S n ×255 (19) The data is normalized to the range of 0-255 and rounded down, and then an RGB image is generated based on the normalized integer data. Step (8): Input the parameters of the Inception V4 memory convolutional network, including the weights W. network and bias b network The policy parameters θ are a set: θ=[W network ,b network ] Step (9): Input the RGB image into the Inception V4 memory convolutional network, and obtain the processed data through convolutional layers, pooling layers, and fully connected layers; Step (10): Input the data processed by the Inception V4 network and the data processed by the spatiotemporal convolutional network into the Universal Transformer network, and sum the results with the results processed by the Bayesian optimized bidirectional gated recurrent neural network to obtain the wind power output prediction of region I. Step (11): Compare the obtained wind power output prediction results with the verification set, adjust the network structure parameters according to the comparison results, further optimize the network structure parameters, and improve the accuracy of the prediction results; Step (12): Compare and analyze the wind power output prediction results obtained from the optimized network processing with the test set to obtain the prediction accuracy. Step (13): For regions II to N, collect 24×7 sets of data every hour for 7 consecutive days; the data processing method and network structure are the same as those for region I. Step (14): The wind power output prediction results obtained from each region are processed through three fully connected layers to obtain the final wind power output prediction results.

Images