Wind power prediction method and system in extreme weather

Abstract

The invention discloses a wind power prediction method in extreme weather, which comprises the following steps of: constructing a pre-training model, and training the pre-training model by taking non-extreme weather meteorological prediction data as input and wind power prediction power as output until a loss function of the pre-training model is converged; assigning hyper-parameters and parameters of the pre-training model to the meta-training model until a loss function of a source domain query set model in the meta-training model is converged; assigning parameters of the source domain query set model to the target domain model, and finely adjusting the target domain model through samples in extreme weather; and inputting the meteorological prediction data in the extreme weather into the target domain model to predict the wind power. In addition, a cross-task risk loss minimum function is designed in pre-training and meta-training to serve as a loss function of a prediction model, and the generalization ability of the prediction model in cross-task iterative training is promoted. Through the wind power prediction method, the problems of insufficient sample size and insufficient prediction precision in extreme weather are solved.

Description

Technical Field

[0001] The present invention belongs to the field of power system planning and operation, and specifically relates to a method and system for predicting wind power under extreme weather conditions. Background Art

[0002] Accurate wind power forecasting helps ensure electricity supply and promote wind energy consumption. However, wind turbine power generation is affected by weather, especially extreme weather. Wind power generation in extreme weather conditions can impact the safe and stable operation of the power system. Frequent extreme weather events and the increase in wind power installed capacity further exacerbate this risk. Therefore, accurate wind power forecasting in extreme weather conditions is particularly important.

[0003] Deep learning methods are widely used due to their powerful data mining capabilities. However, data samples of wind power under extreme weather conditions are relatively few, making it difficult to meet the data sample requirements of deep learning models. Current research mainly addresses this issue through methods such as sample augmentation and transfer learning. However, sample augmentation can introduce additional bias, and training of generative models also faces the problem of insufficient sample size. While transfer learning can theoretically solve this small sample modeling problem, the input information (weather conditions), weather-power mapping relationship, and target power characteristics under extreme weather conditions all undergo significant changes, which can lead to domain shift in deep learning. Therefore, in extreme weather wind power prediction modeling, both the insufficient sample size and domain shift issues should be considered. Furthermore, the prediction model should be suitable for wind power prediction modeling under multiple types of extreme weather conditions.

[0004] Based on the above defects, a new small sample learning framework is provided to address the problems of insufficient sample size and domain offset in wind power prediction under various types of extreme weather conditions, so as to improve the accuracy of wind power prediction under various types of extreme weather conditions. Summary of the Invention

[0005] In response to the shortcomings of the existing technology and the need for improvement, the present invention provides a method for predicting wind power under extreme weather conditions. The purpose of the present invention is to address the problem of insufficient sample size under extreme weather conditions by providing a small sample learning framework to improve the prediction accuracy of wind power under extreme weather conditions, including:

[0006] S1: Construct a neural network model including a feature extractor and a base learner, wherein the feature extractor serves as a shared parameter, the base learner serves as a specific parameter for a specific task, and each nonlinear layer in the neural network model is preceded and followed by a lightweight parameter layer;

[0007] S2: Divide the source domain data into k tasks and use a clustering method to cluster the weather forecast data as input. Each task includes a source domain support set and a source domain query set.

[0008] S3: Training the neural network model using source domain data as a pre-training model for predicting wind power under extreme weather conditions, wherein the neural network model training uses actual wind power of the wind farm as target data and meteorological forecast data as input of the pre-training model, and updating the shared parameters and the specific parameters so that the loss function of the pre-training model converges, thereby obtaining a trained pre-training model;

[0009] S4: Construct neural network models for the source domain support set, source domain query set, and target domain respectively to obtain the source domain support set model, source domain query set model, and target domain model, and initialize the parameters of the source domain support set model, source domain query set model, and target domain model. The hyperparameters and parameters of the models are the same as those of the trained pre-trained model.

[0010] S5: Train the source domain support set model, sample k* tasks from k tasks, and sample from each corresponding task samples as the support set, and update the parameter η in the source domain support set model through the first sampling sample SS , where η SS are the parameters of the base learner in the source domain support set model;

[0011] S6: Train the source domain query set model, and use the k samples from step S5 * Resampling in tasks samples, and according to the second sampling sample and [ξ SS ,η SS ]Calculate the gradient and update the parameters of the source domain query set model based on the gradient [SS {1,2} ,η SQ ], where [ξ SS ,η SS ] represents the parameters of the feature extractor and base learner in the source domain support set model, [SS {1,2} ,η SQ ] represents the scaling parameters and offset parameters of the lightweight parameter layer in the pre-trained model and the parameters of the base learner in the source domain query set model;

[0012] S7: Repeat S5 and S6 until the loss function of the source domain query set model converges;

[0013] S8: Assign the parameters of the source domain query set model to the target domain model, and fine-tune the target domain model parameters based on samples under extreme weather conditions [SS {1,2} ,η T ], to obtain the target domain model for predicting wind power under extreme weather conditions;

[0014] S9: Input the meteorological forecast data into the target domain model in step S8 to predict wind power.

[0015] The present invention also discloses a wind power prediction system under extreme weather conditions, comprising:

[0016] Pre-training module: A neural network is used for model training as a pre-training model for wind power prediction under non-extreme weather conditions. The pre-training model uses actual wind power data under non-extreme weather conditions as target data and meteorological forecast data as input parameters of the pre-training model. The pre-training model includes a feature extractor and a basic learner, wherein the feature extractor serves as a shared parameter and the basic learner serves as a specific parameter for a specific task. In addition, each nonlinear layer in the pre-training model is preceded and followed by a lightweight parameter layer.

[0017] Meta-training module: divide the tasks for non-extreme weather and use the method of steps S5-S7 above to train the model parameters across tasks;

[0018] Fine-tuning module: This module uses meteorological forecast data under extreme weather conditions to fine-tune the parameters within the neural network in the meta-training module to obtain a target domain model for predicting wind power under extreme weather conditions.

[0019] Prediction module: It includes the target domain model, which inputs meteorological forecast data into the target domain model to predict wind power.

[0020] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects:

[0021] 1. A model training method was established for wind power prediction under extreme weather conditions. This method combines transfer learning with a meta-training model to adapt the prediction model to learning from a small number of samples. The meta-training model further adapts the prediction model to multi-task processing and cross-task generalization through inter-task iteration, thereby alleviating the domain shift problem under various extreme weather conditions.

[0022] 2. A lightweight parameter layer is embedded in the pre-trained and meta-trained models to align the shifted domain distribution. High-performance deep learning models are difficult to apply to small-sample learning due to the risk of overfitting. This lightweight parameter layer enables deep learning models to be applied to wind power forecasting modeling in extreme weather conditions.

[0023] 3. We designed a cross-task risk loss minimization function as the loss function for the training model. We added a second-order gradient term to the traditional MSE (Mean Squared Error) loss function, and made this second-order gradient term approach 0 during training. This allows the model to achieve simultaneous optimization in different tasks, thereby guiding the model to extract general knowledge and promoting its generalization ability in cross-task iterative training. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 Flowchart of a method for predicting wind power under extreme weather conditions according to an embodiment of the present invention;

[0025] Figure 2 Schematic diagram of parameter update in a method for predicting wind power under extreme weather conditions according to an embodiment of the present invention;

[0026] Figure 3 4 is a module diagram of a wind power prediction system under extreme weather conditions according to an embodiment of the present invention. DETAILED DESCRIPTION

[0027] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0028] In the present invention, the terms "first", "second", etc. (if any) in the present invention and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0029] Example 1:

[0030] The present invention discloses a method for predicting wind power under extreme weather conditions, comprising the following steps:

[0031] S1: Construct a neural network model including a feature extractor ξ and a base learner η, where the feature extractor is used as a shared parameter and the base learner is used as a specific parameter for a specific task. Each nonlinear layer in the neural network model is preceded and followed by a lightweight parameter layer, including SS {1,2} The two trainable parameters are called scaling parameter SS1 and offset parameter SS2. The relationship between scaling parameter SS1 and offset parameter SS2 is as follows:

[0032]

[0033] Among them, x * and y * are the input and output of the lightweight parameter layer respectively.

[0034] S2: Divide the source domain data into k tasks and cluster them using a clustering method with weather forecast data as input. Each task includes a source domain support set and a source domain query set. Specifically, the source domain data here is sample data under normal weather conditions, with weather data as input and wind power as output, that is, weather data corresponds to wind power data. When clustering weather forecast data, the main focus is on clustering source domain data with the same type of weather characteristics.

[0035] S3: Train the neural network model based on the source domain data as a pre-trained model for wind power prediction under extreme weather conditions F PRE , the neural network model training uses the actual wind power of the wind farm as the target data and the meteorological forecast data as the pre-training model F PRE Input, update shared parameters and specific parameters; the shared parameters refer to the pre-trained model F PRE The parameters of the feature extractor in , the specific parameters refer to the pre-trained model F PRE Parameters of the base learner in . Specifically, the pre-trained model F PRE The update process of shared parameters and specific parameters is as follows:

[0036] [ξ0,η0]→[ξ0,η0]-σ0▽L CTRM ([ξ0,η0])

[0037] Where [ξ0,η0] represents the parameters of the feature extractor and base learner in the pre-trained model, σ0 is the learning rate, and the loss function L CTRM The expression is:

[0038]

[0039] Among them, L task·i is the empirical loss (MSE) of the i-th task, and λ is the weight of the latter term. m and n represent two sub-datasets. Specifically, the mini-batch dataset used to calculate the loss function is divided into m and n sub-datasets for the calculation of the latter term. The two sub-datasets m and n do not overlap.

[0040] S4: Construct neural network models for the source domain support set, source domain query set, and target domain respectively to obtain the source domain support set model F SS , source domain query set model F SQ and the target domain model F T , initialize the source domain support set model F SS , source domain query set model F SQ and the target domain model F T The parameters of the model are the same as those of the pre-trained model F PRE Same, pre-trained model F PREThe parameters in the source domain support set model F are transferred to the above source domain support set model F SS , source domain query set model F SQ and the target domain model F T In the embodiment of the present invention, the source domain support set model F SS , source domain query set model F SQ and the target domain model F T With the pre-trained model F PRE The architecture of the source domain support set model F SS , source domain query set model F SQ With the pre-trained model F PRE The expression of the loss function is also the same, and the source domain support set model F SS and the source domain query set model F SQ The combination is called a meta-trained model.

[0041] S5: Source domain support set model F SS Perform training and update F SS Parameters of , sample k* tasks from k tasks, and sample from each corresponding task samples as the support set, and update the source domain support set model F through the first sampling sample SS The parameter η in SS , where η SS is the source domain support set model F SS The parameters of the basic learner in the k* task. The first sample here refers to the sampled Specifically, for the source domain support set model F SS Update the parameter η during training SS The expression is as follows:

[0042]

[0043] where σ SS It's F SS The learning rate, [ξ SS ,η SS ] indicates F SS The above expression means to update η with the value on the right of the arrow. SS , the parameter update process in the following expressions is similar.

[0044] S6: Query the source domain model F SQ Perform training and update F SQ The parameters of k sampled from step S5 * Resampling in tasks samples, and according to the second sampling sample and [ξ SS ,η SS]Calculate the gradient, where the second sampling sample refers to k * Sampled from the task samples, and update the source domain query set model F according to the gradient SQ Parameters [SS {1,2} ,η SQ ], where [ξ SS ,η SS ] represents the source domain support set model F SQ Parameters of the feature extractor and base learner in [SS {1,2} ,η SQ ] represents the pre-trained model F PRE The scaling parameters, offset parameters, and source domain query set model F of the lightweight parameter layer SQ Specifically, update the source domain query set model F SQ The parameter process is as follows:

[0045]

[0046] where σ SQ It's F SQ The learning rate, L CTRM is the loss function of the source domain query set model.

[0047] S7: Repeat S5 and S6 until the source domain query set model F SQ In an optional embodiment of the present invention, the process of repeatedly executing S5 and S6 is as follows: first, repeat S5 for a preset number of iterations to obtain the final parameter η SS Then execute step S6, for example, according to the preset number of iterations, where the number of iterations is set based on experience, after the iteration is completed, the source domain query set model F SQ The loss function converges, so that the difference between the predicted power and the actual power can be minimized. At this time, the source domain query set model F SQ The above is just an example of repeatedly executing S5 and S6. In fact, during the training process, the source domain query set model F SQ The specific way of convergence of the loss function is not limited to this.

[0048] S8: The source domain query set model F SQ The parameters of the target domain model F are assigned to T , and fine-tune the target domain model F based on samples under extreme weather conditions T Parameters [SS {1,2} ,η T ], obtain the target domain model F for predicting wind power under extreme weather conditions T Specifically, it refers to the source domain query set model F trained in step S7 SQThe parameters of the model F are transferred to the target domain T , and then fine-tune the target domain model F based on samples under extreme weather conditions T The parameters of the lightweight parameter layer SS in {1,2} and the parameters η of the base learner T Specifically, the sample update model parameters under a certain type of extreme weather [SS {1,2} ,η T ], the update process is as follows:

[0049]

[0050] Among them, σ t is the target domain model F T The learning rate, [SS {1,2} ,η T ] is the target domain model F T The parameter L represents the MSE loss function of the target domain model.

[0051] S9: Input the weather forecast data into the target domain model F in step S8 T In this paper, wind power is predicted. The target domain model F T It is used to predict wind power based on meteorological forecast data. In the embodiment of the present invention, the meteorological forecast data here is weather forecast data under extreme weather conditions, including but not limited to strong winds, freezing, extreme cold, heavy rain and other weather data. By constructing a pre-trained model F PRE And the meta-training model (source domain support set model F SS and the source domain query set model F SQ ), transfer the parameters in the meta-training model to the target domain model F T , adapting to learning with a small number of samples, and at the same time building a lightweight parameter layer in each model corresponding to the offset domain distribution, solving the problem that high-performance deep learning models are difficult to apply to small-sample learning due to the risk of overfitting. Figure 1 This is a simplified process for predicting wind power under extreme weather conditions disclosed in the present invention. PRE , source domain support set model F SS and the source domain query set model F SQ , and fine-tuning, the parameter update process is as follows Figure 2 shown.

[0052] Specifically, such as Figure 2 As shown in the neural network model constructed in the present invention, a lightweight parameter layer is set before and after each two nonlinear modules. When there is only a linear layer between two lightweight parameter layers, one lightweight parameter layer can be ignored. For the specific training process, refer to Figure 2As shown in the figure, 1) Pre-training process: All parameters of the pre-trained model are trained using the entire dataset. The updated parameters in the pre-trained model are then used to train the source domain support set model, the source domain query set model, and the target domain model obtained after fine-tuning. 2) Meta-training model training process: This process is divided into source domain support set model training and source domain query set model training. The parameters of the pre-trained model are assigned to the source domain support set model and the source domain query set model as initial parameters. During the source domain support set model training process, the tasks after conventional weather classification are sampled, and some samples within the tasks are sampled. Only the parameters of the base learner are updated, and the parameters of the base learner are used for gradient calculation. The parameters of the temporal convolutional layer and the lightweight parameter layer in the source domain support set model are consistent with those in the pre-trained model. During the source domain query set model training process, non-overlapping samples in the support set sampling task are sampled. Based on the source domain support set model with updated parameters, the gradients of the lightweight parameter layer and the base learner are calculated, and the corresponding parameters of the source domain query set model are updated. 3) Fine-tuning process: After the loss function of the meta-training model process converges, specifically the loss function of the source domain query set model converges, the target domain model is fine-tuned based on samples under different extreme weather conditions to train the lightweight parameter layer and the basic learner layer parameters. For example, when the target domain model is to predict wind power under strong winds, the meteorological data under strong winds is used to fine-tune the target domain model. At this time, the target domain model is used to predict wind power under strong winds. For wind power prediction under other extreme conditions, the target domain model is obtained in the same way.

[0053] Example 2:

[0054] like Figure 3 The figure shows a module diagram of a wind power prediction system under extreme weather conditions. The wind power prediction system includes: a pre-training module, a meta-training module, a fine-tuning module, and a prediction module. The pre-training module uses a neural network for model training as a pre-training model for predicting wind power under non-extreme weather conditions. The pre-training model training uses actual wind power data under non-extreme weather conditions as target data and meteorological forecast data as input parameters of the pre-training model. The pre-training model includes a feature extractor and a basic learner, wherein the feature extractor serves as a shared parameter and the basic learner serves as a specific parameter for a specific task. Lightweight parameter layers are provided before and after each nonlinear layer in the pre-training model. The meta-training module is used to divide tasks for non-extreme weather conditions and to train model parameters across tasks using the method of steps S5-S7 in Example 1. The fine-tuning module selects meteorological forecast data under extreme weather conditions and fine-tunes the parameters within the neural network of the meta-training module to obtain a target domain model for predicting wind power under extreme weather conditions. The prediction module includes a trained target domain model, and meteorological forecast data is input into the trained target domain model to predict wind power.

[0055] It should be noted that the meta-training module includes the source domain support set model and the source domain query set model. The process of updating parameters during the training of the pre-training model, the source domain support set model and the source domain query set model is described in Example 1 and will not be described in detail to avoid redundancy.

[0056] In order to verify the wind power prediction method and system under extreme weather conditions disclosed in the present invention, data from a Chinese wind farm A was collected, where the total capacity of the wind farm is 400.5MW. Specifically, the wind power data comes from actual data with a time resolution of 1h, and the time length is two years. For example, the wind power data of 2018 is used as the source domain data for training, and the wind power data of 2019 is used to verify the effect of the model, that is, as a test set. The input meteorological forecast data includes: 100m wind speed, 100m wind direction, surface air pressure, surface temperature, and surface humidity. First, a neural network model is constructed according to the method in Example 1. The specific neural network model selects the TCN architecture. The feature extractor in the architecture consists of six residual modules, each module contains two TCN layers, and the TCN layer follows the ReLU activation function and the dropout layer. The convolution kernel is set to 2, the packet loss rate is 0.2, and the number of channels = [128, 96, 64, 48, 32, 16, 8]. The Adam optimizer was selected, with a learning rate of 0.0002 and a beta coefficient of [0.5, 0.999]. The number of tasks was set to 10, meaning that after inputting the weather forecast data, the number of clusters was 10. During pre-training, the number of iterations was set to 80,000. The model parameters obtained after the pre-training model were used to initialize the parameters of the source domain support set model, the source domain query set model, and the target domain model. When training the source domain support set model, 5 tasks were sampled from the 10 tasks, and 10 samples were sampled from each corresponding task to train the source domain support set model, and the parameters of the source domain support set model were updated. Furthermore, when training the source domain query set model, 10 samples were resampled from the 5 tasks sampled above to train the source domain query set model, and the parameters of the source domain query set model were updated. The source domain support set model and the source domain query set model were iteratively trained 70,000 times, based on empirical values. When the iterative training reached 70,000 times, the source domain query set model converged, meaning its loss function converged. Furthermore, after the source domain query set model converged, the parameters of the source domain query set model were assigned to the target domain model. Sample sets were collected for high winds, cold waves, frost, and high temperatures, and the parameters of the target domain model were fine-tuned to obtain the corresponding target domain prediction model for extreme weather conditions. In other words, when the input sample set is meteorological data for high winds, the fine-tuned target domain model is used to predict wind power in high winds, and so on.

[0057] In the above implementation, the normalized root mean square error (RMSE) and mean absolute error (MAE) are used as evaluation indicators:

[0058]

[0059] Where Cap is the capacity of the node, N e is the number of events evaluated, and each event contains N = 12 time points.

[0060] The framework GTL (Generalizable Transfer-learning) that migrates the meta-training model parameters superimposed by the pre-training model constructed in the present invention to the target domain model is compared with mature deep learning architectures, and the results are shown in Table 1. Compared with other small sample learning methods, the results are shown in Table 2. The RMSE and MAE of GTL are lower than those of other comparison methods. Compared with several mature deep learning architectures, the RMSE of the prediction method in the present invention is reduced by 5.97%-14.28%, and the MAE is reduced by 2.86%-13.76%. Compared with several small sample learning architectures, the RMSE of the prediction method in the present invention is reduced by 3.31%-43.55%, and the MAE is reduced by 3.03%-43.39%.

[0061] Table 1. Error statistics of prediction results of GTL framework and other mature deep learning architectures

[0062]

[0063] Table 2. Error statistics of prediction results of GTL framework and other small sample learning architectures

[0064]

[0065] It can be seen from this that the wind power prediction method under extreme weather conditions disclosed in the present invention significantly improves the accuracy and efficiency of prediction.

[0066] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for predicting wind power under extreme weather conditions, characterized in that: include: S1: Construct a neural network model including a feature extractor and a base learner, wherein the feature extractor serves as a shared parameter, the base learner serves as a specific parameter for a specific task, and each nonlinear layer in the neural network model is preceded and followed by a lightweight parameter layer; S2: Divide the source domain data into k tasks and use a clustering method to cluster the weather forecast data as input. Each task includes a source domain support set and a source domain query set. S3: Training the neural network model using source domain data as a pre-training model for predicting wind power under extreme weather conditions, wherein the neural network model training uses actual wind power of the wind farm as target data and meteorological forecast data as input of the pre-training model, and updating the shared parameters and the specific parameters so that the loss function of the pre-training model converges, thereby obtaining a trained pre-training model; S4: Construct neural network models for the source domain support set, source domain query set, and target domain respectively to obtain the source domain support set model, source domain query set model, and target domain model, and initialize the parameters of the source domain support set model, source domain query set model, and target domain model. The hyperparameters and parameters of the models are the same as those of the trained pre-trained model. S5: Train the source domain support set model, sample k* tasks from k tasks, and sample from each corresponding task samples as the support set, and update the parameter η in the source domain support set model through the first sampling sample SS , where η SS are the parameters of the base learner in the source domain support set model; S6: Train the source domain query set model, and use the k samples from step S5 * Resampling in tasks samples, and according to the second sampling sample and [ξ SS ,η SS ]Calculate the gradient and update the parameters of the source domain query set model based on the gradient [SS {1,2} ,η SQ ], where [ξ SS ,η SS ] represents the parameters of the feature extractor and base learner in the source domain support set model, [SS {1,2} ,η SQ ] represents the scaling parameters and offset parameters of the lightweight parameter layer in the pre-trained model and the parameters of the base learner in the source domain query set model; S7: Repeat S5 and S6 until the loss function of the source domain query set model converges; S8: Assign the parameters of the source domain query set model to the target domain model, and fine-tune the target domain model parameters based on samples under extreme weather conditions [SS {1,2} ,η T ], to obtain the target domain model for predicting wind power under extreme weather conditions; S9: Input the meteorological forecast data into the target domain model in step S8 to predict wind power.

2. The method for predicting wind power under extreme weather conditions according to claim 1, characterized in that: The lightweight parameter layer in the pre-trained model includes the scaling parameter SS1 and the offset parameter SS2, which satisfy the following expression: Among them, x * and y * are the input and output of the lightweight parameter layer respectively.

3. The method for predicting wind power under extreme weather conditions according to claim 1, characterized in that: Loss function L of pre-trained model, source domain support set model, and source domain query set model CTRM The expression is as follows: Among them, k represents the number of tasks, L task.i is the MSE loss function for the i-th task, and λ is the weight of the latter term in the loss function. m and n are two sub-datasets, which are divided from the mini-batch dataset used to calculate the loss function and used for the calculation of the latter term. The two sub-datasets m and n do not overlap.

4. The method for predicting wind power under extreme weather conditions according to claim 3, characterized in that: The pre-trained model is based on the risk loss function L CTRM The process of minimizing and optimizing the shared parameters and the specific parameters is as follows: Where [ξ0,η0] represents the parameters of the feature extractor and base learner in the pre-trained model, and σ0 is the learning rate.

5. The method for predicting wind power under extreme weather conditions according to claim 3, characterized in that: The process of updating specific parameters of the source domain support set model is as follows: where σ SS is the learning rate of the source domain support set model, [ξ SS ,η SS ] represents the parameters of the feature extractor and base learner in the source domain support set model.

6. The method for predicting wind power under extreme weather conditions according to claim 3, characterized in that: Source domain query set model update [SS {1,2} ,η SQ The process of ] is as follows: where σ sq is the learning rate of the source domain query set model, SS1 and SS2 are the scaling parameter and offset parameter of the lightweight parameter layer in the pre-trained model, respectively.

7. The method for predicting wind power in extreme weather conditions according to claim 3, characterized in that: Target domain model update [SS {1,2} ,η T The process of ] is as follows: Among them, σ t is the learning rate of the target domain model, [SS {1,2} ,η T ] are the parameters of the target domain model, and L represents the MSE loss function of the target domain model.

8. The method for predicting wind power in extreme weather conditions according to claim 1, characterized in that: The weather forecast data includes wind speed, wind direction, temperature, humidity, and pressure.

9. A wind power prediction system under extreme weather conditions, characterized in that: include: Pre-training module: A neural network is used for model training as a pre-training model for wind power prediction under non-extreme weather conditions. The pre-training model uses actual wind power data under non-extreme weather conditions as target data and meteorological forecast data as input parameters of the pre-training model. The pre-training model includes a feature extractor and a basic learner, wherein the feature extractor serves as a shared parameter and the basic learner serves as a specific parameter for a specific task. In addition, each nonlinear layer in the pre-training model is preceded and followed by a lightweight parameter layer. Meta-training module: divides tasks for non-extreme weather and uses the method of steps S5-S7 in claim 1 to train model parameters across tasks; Fine-tuning module: This module uses meteorological forecast data under extreme weather conditions to fine-tune the parameters within the neural network in the meta-training module to obtain a target domain model for predicting wind power under extreme weather conditions. Prediction module: It includes the target domain model, which inputs meteorological forecast data into the target domain model to predict wind power.

10. The extreme weather wind power prediction system according to claim 9, characterized in that: The meta-training module includes a source domain support set model and a source domain query set model. The loss function L of the pre-training model, the source domain support set model, and the source domain query set model is CTRM The same expression is as follows: Among them, L task·i is the empirical loss MSE of the i-th task, and λ is the weight of the latter term. m and n are two sub-datasets, which are divided from the mini-batch dataset used to calculate the loss function and used for the calculation of the latter term. The two sub-datasets m and n do not overlap.

Images