A wind power output prediction method based on convolutional attention fusion and confidence region enhancement transfer.

Abstract

This invention proposes a wind power output prediction method based on convolutional attention fusion and confidence domain-enhanced transfer learning. The method comprises the following steps: First, data preprocessing is performed, splitting the dataset into dataset 1 and dataset 2. Second, a memory-based convolutional network using dataset 1 and an Inception module is trained, and the network parameters are tuned using a confidence domain policy optimization method. Finally, based on the transfer operation, the trained memory-based convolutional network using dataset 2 is used to predict the wind farm output. This method enables accurate prediction of wind farm output data, improving energy utilization and reducing carbon emissions.

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 lack the ability to capture data features, resulting in low prediction accuracy and difficulty in meeting the needs of practical applications. While traditional deep learning models have sufficient depth and breadth to capture the features contained in the data, they are trained on historical data and do not have the ability to dynamically update parameters. This leads to performance degradation issues in practical applications.

[0003] In addition, existing models for wind power output prediction cannot adapt to real-world application conditions that differ greatly from the model training conditions, resulting in poor versatility and making them unsuitable for large-scale application.

[0004] Therefore, a wind power output prediction method based on convolutional attention fusion, confidence domain reinforcement, and transfer learning is proposed. This method combines reinforcement learning with scaled attention convolutional networks, enabling the model to self-optimize and adjust hyperparameters according to changes in real-time data. Simultaneously, the method includes transfer operations to improve the model's generality, allowing it to adapt to different operating conditions. Summary of the Invention

[0005] This invention proposes a wind power output prediction method based on convolutional attention fusion, confidence domain reinforcement, and transfer learning. This method combines reinforcement learning, transfer learning, and attention-based convolutional networks for wind farm output prediction. The method has the ability to adjust hyperparameters based on real-time data changes, improving the model's generality and the accuracy of wind farm output prediction results. This facilitates more precise wind farm output planning, improves wind energy utilization, and thus saves energy and reduces carbon emissions. The steps in its application are as follows:

[0006] Step (1): Input the original data of the wind farm, including the original power output data, historical operation data, basic data of the wind farm, meteorological forecast data and actual meteorological data;

[0007] Step (2): Clean the input wind farm raw data, and then randomly divide it according to the ratio ω:1. The divided datasets are denoted as dataset 1 and dataset 2 respectively.

[0008] Step (3): Input the parameters of the memory convolutional network based on the scaling attention mechanism and the Inception module, including the weights W. network and bias b networkLet the loop number i = 1, and the input confidence region policy optimization method parameters, including the initial policy parameters θ0 and the relative entropy constraint δ. The policy parameters θ are a set:

[0009] θ=[W network ,b network (1)

[0010] Step (4): The memory convolutional network based on the scaling attention mechanism and the Inception module is trained using dataset 1. The data processing flow in the memory convolutional network based on the scaling attention mechanism and the Inception module is as follows: First, the data is simultaneously input into the scaling attention mechanism architecture and the Inception V3 module for processing. The Inception V3 module is the Inception V3 neural network. Then, the two processing results are simultaneously input into the gated recurrent unit and the fully connected layer for processing. Finally, the prediction result is output. The data processing process of the fully connected layer is as follows:

[0011] M fc =W fc ·N fc +b fc (2)

[0012] Where, N fc M is the input matrix; fc For the output matrix; W fc b is the weight matrix; fc It is the bias vector;

[0013] The architecture of the scaling attention mechanism is as follows: a stem module and an Inception-A module are connected sequentially to form an SA structure. The four SA structures simultaneously receive and process the data input from the memory convolutional network based on the scaling attention mechanism and the Inception module. The outputs of the first and second SA structures are multiplied and processed by a softmax operation, then multiplied together with the output of the third SA structure. After passing through a convolutional layer, they are fed together with the output of the fourth SA structure and then sequentially fed into the scaling processing layer and the fully connected layer for further processing.

[0014] The stem module and the Inception-A module involve convolution modules and threshold operations. The convolution module includes a convolution operation, a batch normalization operation and a threshold operation. The batch normalization operation processes the data as shown in equations (3) and (4).

[0015]

[0016]

[0017] Where ε is a constant; μ B The average value of the input; σ 2 B represents the variance of the input elements; B represents the B-th dimension of the current layer, i.e., the B-th neuron node; x Batch，o For any input element; For x Batch，o Normalized corresponding elements; o is the index; β is the offset; ρ is the factor; y Batch，o for The corresponding element after processing;

[0018] The data processing procedure of the threshold operation is shown in equation (5);

[0019]

[0020] Where q represents the input of the threshold operation, and z(q) represents the output of the threshold operation;

[0021] The softmax processing procedure is as follows:

[0022]

[0023] Where x softmax，u x softmax，v These are the u-th and v-th input elements, respectively; S u For x softmax，u The softmax value; exp(x softmax，v x is a natural number e softmax，v The power; j is the total number of input elements; u and v are both numbers;

[0024] The scaling processing layer's processing procedure is as follows:

[0025] y scaling，p =l scaling x scaling，p +c (7)

[0026] Among them l scaling c is the scaling factor; c is the bias factor; x scaling，p For the p-th input element, y scaling，p This is the p-th output element, where p is the index;

[0027] Step (5): Calculate the root mean square error (RMSE) between the predicted and true values ​​of the memory convolutional network based on the scaling attention mechanism and the Inception module. Determine whether the RMSE value is less than a set value. If the condition is met, the memory convolutional network based on the scaling attention mechanism and the Inception module is considered to have completed training, and proceed to step (13); if the condition is not met, the memory convolutional network based on the scaling attention mechanism and the Inception module is considered not to have completed training, and proceed to step (6).

[0028] Step (6): Set the reward R and state s of the confidence region policy method as shown in equations (8) and (9). Action a represents whether to use the confidence region policy method on the weight W. network and bias b network Update; use the confidence region strategy method to adjust the weights W. network and bias b network Perform strategy learning and parameter tuning, and record the number of optimization attempts as k=1;

[0029] R = -RMSE (8)

[0030]

[0031] Where θ k θ k-1 θ1 represents the strategy parameters in the k-th, k-1, and 1-th optimizations, respectively; ||||2 is the 2-norm;

[0032] Step (7): Execute policy π in the environment k =π(θ) k ), collect trajectory D k ={τ t}, where π is not a strategy. k Let τ be the strategy in the k-th optimization. t The specific trajectory includes actions and states, and t is the time step;

[0033] Step (8): Calculate the current reward and dominance function

[0034]

[0035]

[0036]

[0037] in Let the action value function be... State value function, a t and a t+1These represent the actions at time step t and time step t+1, respectively. t+1 The state at time step t+1 Let γ be the reward at time step t+1, γ be the reward discount factor, and α be the exponent of γ, where α∈N. * ; Indicates from s t+1 and a t+1 Start, until s ∞ and a ∞ ,right Seeking expectations; Indicates from s t+1 and a t Start, until s ∞ and a ∞ ,right Expectation; s ∞ and a ∞ These represent the state and action at time step ∞;

[0038] Step (9): Calculate the policy gradient

[0039]

[0040] π k (a t |s t ) = P(a = a t |s=s t (14)

[0041] Where s t The state at time step t; For the gradient operator; logπ k (a t |s t ) represents π with base 10 k (a t |s t The logarithm of ). This means that at θ = θ k Gradient operator operations are performed under the condition that P(a=a) t |s=s t ) is s = s t Under the condition a=a t The probability of;

[0042] Step (10): Calculate step size constraints using conjugate gradients

[0043]

[0044] in The Hessian matrix is ​​the relative entropy, and the general formula for calculating the Hessian matrix H is:

[0045]

[0046] Where f is an arbitrary multivariate function, x H，1 x H，2 , ..., Let f be the independent variable. The number of independent variables;

[0047] Step (11): Update policy θ k+1 As shown in equations (17) and (18);

[0048] θ k+1 =θ k +Δθ (17)

[0049]

[0050] Where Δθ is the corrected step size; for Transpose of;

[0051] Step (12): Determine θ k+1 If the constraints are not met, let k = k + 1, update the state s according to equation (9), and execute action a as "do not use the confidence region strategy method on weight W". network and bias b network Perform the update and return to step (7) to continue execution; if the constraints are met, update the loop number i = i + 1 and the strategy parameter θ. i =θ k+1 Action a is to "use the confidence region strategy method to adjust the weights W". network and bias b network Perform the update and return to step (4) to continue execution;

[0052] Step (13): Based on the transfer operation, use the pre-trained memory convolutional network based on scaling attention mechanism and Inception module to make predictions based on dataset 2 and output the prediction results. The specific steps of the prediction are the same as those in step (4).

[0053] The present invention has the following advantages and effects compared with the prior art:

[0054] (1) Combining reinforcement learning, transfer learning, and scaled attention convolutional networks enables the model to adjust hyperparameters based on changes in real-time data, thereby improving the model's generalizability and the accuracy of wind farm output prediction. This approach combines three major branches of artificial intelligence methods—reinforcement learning, transfer learning, and image convolutional neural networks—to improve prediction accuracy.

[0055] (2) It is conducive to more precise planning of wind farm output and improve wind energy utilization.

[0056] (3) It is conducive to saving energy, reducing carbon emissions, and helping to achieve "carbon peaking" and "carbon neutrality". Attached Figure Description

[0057] Figure 1 This is a framework diagram of wind farm output prediction using the method of the present invention.

[0058] Figure 2 This is a diagram of the memory convolutional network framework based on the scaling attention mechanism and the Inception module of the present invention.

[0059] Figure 3 This is a flowchart of the confidence region strategy optimization method of the present invention. Detailed Implementation

[0060] The wind power output prediction method proposed in this invention, which uses convolutional attention fusion and confidence domain-enhanced transfer learning, is described in detail below with reference to the accompanying drawings:

[0061] Figure 1 This is a framework diagram for wind farm power output prediction using the method of this invention. The process for wind farm power output prediction is as follows:

[0062] First, input the raw data of the wind farm, including the raw power output data, historical operation data, basic data of the wind farm, meteorological forecast data and actual meteorological data;

[0063] Then, the input data is cleaned and then randomly divided according to the ratio ω∶1. The resulting datasets are denoted as dataset 1 and dataset 2, respectively.

[0064] Secondly, the input parameters are the parameters of the memory convolutional network based on the scaling attention mechanism and the Inception module, and the parameters of the confidence region policy optimization method.

[0065] Subsequently, the memory-based convolutional network based on the scaling attention mechanism and the Inception module was trained using dataset 1, and the root mean square error (RMSE) between the predicted and true values ​​was calculated. The network was then judged to be trained successfully if the RMSE was less than a set value; otherwise, the network was considered not yet trained, and the confidence region strategy was used to fine-tune the network's weights and biases.

[0066] Finally, based on the transfer operation, the pre-trained memory convolutional network based on the scaling attention mechanism and the Inception module is used to make predictions on dataset 2, and the prediction results are output.

[0067] Figure 2This is a diagram of the memory convolutional network framework based on the scaling attention mechanism and the Inception module of the present invention.

[0068] The memory-based convolutional network structure, based on the scaling attention mechanism and the Inception module, is divided into an initialization part and a reinforcement part. The initialization part includes a stem module, an Inception-A module, and an Inception V3 module, where the stem module and the Inception-A module are connected by the scaling attention mechanism. The InceptionV3 module is the InceptionV3 neural network. The reinforcement part consists of a gated recurrent unit and two fully connected layers.

[0069] The data processing flow in the memory convolutional network based on the scaling attention mechanism and the Inception module is as follows: First, the data is simultaneously input into the scaling attention mechanism architecture and the Inception V3 module for processing. Then, the two processing results are simultaneously input into the gated recurrent unit and the fully connected layer for processing. Finally, the prediction result is output. The architecture of the scaling attention mechanism is as follows: a stem module and an Inception-A module are connected sequentially to form an SA structure. Four SA structures simultaneously receive and process the data input to the memory convolutional network based on the scaling attention mechanism and the Inception module. The outputs of the first and second SA structures undergo a multiplication operation and a softmax process, and then are multiplied together with the output of the third SA structure. After processing through a convolutional layer, they are sequentially input into the scaling processing layer and the fully connected layer for processing along with the output of the fourth SA structure.

[0070] Figure 3 This is a flowchart of the confidence region policy optimization process of the method of this invention. The confidence region policy optimization process is as follows:

[0071] First, read the initial policy parameters θ0.

[0072] Secondly, execute policy π in the environment. k Collect trajectory D k ;

[0073] Then, calculate the rewards in sequence. Advantage function Policy gradient

[0074] Subsequently, the conjugate gradient was calculated.

[0075] Finally, update the policy θ k+1 =θ k +Δθ, determine θ k+1Check if the constraints are met. If not, set k = k + 1 and return to "Execute policy π in the environment". k Collect trajectory D k Continue execution; if the constraints are met, the loop ends.

[0076] 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 wind power output prediction method of convolution attention fusion confidence domain reinforcement transfer, characterized in that, Combining reinforcement learning, transfer learning and attention mechanism-based convolutional network for wind farm output prediction; The steps in use are as follows: Step (1): inputting wind farm original data, including wind farm original output data, historical operation data, station basic data, meteorological prediction data and meteorological measured data; Step (2): The inputted wind farm original data is cleaned and then scaled The segmented data sets are respectively recorded as data set 1 and data set 2; Step (3): input the parameters of the memory convolutional network based on the scaled attention mechanism and the Inception module, including weights and biases ; record the number of cycles , input the parameters of the confidence domain policy optimization method, including the initial policy parameters and the relative entropy constraint , the policy parameters are a set: (1) Step (4): training the memory convolutional network based on the scaling attention mechanism and the Inception module using dataset 1, wherein the data processing process in the memory convolutional network based on the scaling attention mechanism and the Inception module is as follows: firstly, the data is simultaneously input into the scaling attention mechanism architecture and the Inception V3 module for processing, and the Inception V3 module is an Inception V3 neural network; then, the two processing results are sequentially input into the gate recurrent unit and the full connection layer for processing, and finally, the prediction result is output, wherein the processing process of the full connection layer on the data is as follows: (2) wherein, is an input matrix; is an output matrix; is a weight matrix; is a bias vector; The architecture of the scaling attention mechanism is as follows: a stem module and an Inception-A module are sequentially connected to form an S-A structure, four S-A structures simultaneously receive the input data of the memory convolutional network based on the scaling attention mechanism and the Inception module and process the data; the outputs of the first S-A structure and the second S-A structure are subjected to a multiplication operation and a softmax processing, and then are subjected to a multiplication operation with the output of the third S-A structure, and after being processed by a convolution layer, are sequentially input into the scaling processing layer and the full connection layer together with the output of the fourth S-A structure for processing; In the stem module and the Inception-A module, a convolution module and a threshold operation are involved, wherein the convolution module includes a convolution operation, a batch normalization operation and a threshold operation, and the processing process of the batch normalization operation on the data is as shown in formula (3) and formula (4); (3) (4) in It is a constant; This is the average value of the input; The variance of the input elements; Represents the current layer's first The dimension, i.e., the first dimension One neuron node; For any input element; for Normalized corresponding elements; For numbering; This is the offset; As a factor; for The corresponding element after processing; The processing process of the threshold operation on the data is as shown in formula (5); (5) wherein, an input representative of the threshold operation, an output representative of the threshold operation; The process of the softmax processing is as follows: (6) wherein , are the first and the second input element, respectively; is the softmax value of ; is the th power of the natural number e; is the total number of input elements; and are both numbers. The processing process of the scaling processing layer is as follows: (7) wherein is a scaling factor; is a bias factor; is the input element, output element, output element, is a number; Step (5): calculating the root mean square error (RMSE) of the prediction value and the true value of the memory convolutional network based on the scaling attention mechanism and the Inception module, and determining whether the value of the root mean square error is less than a set value, if the condition is met, it is considered that the training of the memory convolutional network based on the scaling attention mechanism and the Inception module is completed, and step (13) is entered; if the condition is not met, it is considered that the training of the memory convolutional network based on the scaling attention mechanism and the Inception module is not completed, and step (6) is entered; Step (6): Set the reward R and state s of the confidence domain policy method as shown in equation (8) and equation (9), and the action representing whether the weight and bias are updated using the confidence domain policy method; the weight and bias are policy learned and parameter optimized using the confidence domain policy method, and the optimization number is numbered ; (8) (9) wherein , , are the first , strategy parameters in the first optimization; is the 2-norm; Step (7): executing the policy in the environment , collecting trajectories where is a policy, is the policy in the th optimization, is a specific trajectory, including actions and states, is a time step; Step (8): Compute the current reward and advantage function ; (10) (11) (12) wherein is the action value function, is the state value function, and are respectively the action at time step the state at time step is the reward at time step is the reward discount factor is the exponential of ; ; ; denotes the expectation over and starting from and , over ; denotes the expectation over and starting from and , over ; and are respectively the state at time step the action at time step Step (9): Compute policy gradient ; (13) (14) wherein is the state at time step is the gradient operator; denotes the logarithm to the base 10; denotes the logarithm to the base 10; denotes the operation of the gradient operator in the case of denotes the operation of the gradient operator in the case of is the probability under the condition is the probability under the condition is the probability under the condition Step (10): Compute step size constraint using conjugate gradient ; (15) wherein is the relative entropy of the matrix H, the calculation formula of the matrix H is: (16) where f is any multivariate function, is the argument of f, is the number of arguments; Step (11): Updating the policy As shown in Equation (17) and Equation (18); (17) (18) wherein is the modified step size; is the transpose of Step (12): judging whether the constraint condition is satisfied or not If the constraint condition is not satisfied, let update the state according to equation (9), execute the action "not updating the weight and bias using the confidence region strategy method" and return to step (7) to continue the execution; if the constraint condition is satisfied, update the cycle number , the strategy parameter , execute the action "updating the weight and bias using the confidence region strategy method" and return to step (4) to continue the execution. Step (13): based on the transfer operation, using the trained memory convolutional network based on the scaling attention mechanism and the Inception module to predict and output the prediction result based on dataset 2, and the specific steps of the prediction are the same as those in step (4).

Images