WO2023070960A1

WO2023070960A1 - Wind power prediction method based on convolutional transformer architecture, and system and device

Info

Publication number: WO2023070960A1
Application number: PCT/CN2022/072596
Authority: WO
Inventors: 卢泽华; 李小翔; 任鑫; 曾谁飞; 杨永前; 王�华; 陈沐新; 张燧; 王青天; 冯帆; 王振荣
Original assignee: 中国华能集团清洁能源技术研究院有限公司
Priority date: 2021-10-29
Filing date: 2022-01-18
Publication date: 2023-05-04
Also published as: CN114021803A

Abstract

Disclosed in the present application are a wind power prediction method and apparatus based on a convolutional transformer architecture, and a storage medium. The method comprises: collecting meteorological data and operation data, and acquiring an embedding vector (S101); inputting the embedding vector into a power prediction network, wherein the power prediction network comprises an encoder and a decoder (S102); according to the encoder, acquiring a feature map which corresponds to the embedding vector (S103); and inputting the feature map into the decoder, so as to generate a predicted power (S104).

Description

A wind power prediction method, system and device based on convolution transformer architecture

Cross References to Related Applications

This application is based on a Chinese patent application with application number No. 202111274987.2 and a filing date of October 29, 2021, and claims the priority of this Chinese patent application. The entire content of this Chinese patent application is hereby incorporated by reference into this application.

technical field

This application relates to new energy power technology, and in particular to a wind power prediction method, system and equipment based on convolution transformer architecture.

Background technique

Wind power technology is becoming a major source of future electricity demand. A higher share of renewable energy technologies is critical to carbon-neutrally meeting the needs of future new power system grids, but also brings new grid operation challenges. Electric power companies need to predict the power generated by wind power in order to carry out power generation dispatching operations. Forecasting is a major enabler to ensure safe and economical integration of wind power, while creating links between many flexibility innovations at different levels of the power system to achieve synergies. Accurate wind power forecasting is an important and cost-effective element of energy management, which also facilitates effective and direct participation of wind power plants and aggregation systems in the electricity market and increases plant profitability through optimized supply planning.

In related technologies, the wind power generation power is predicted according to the recurrent neural network model, but the recurrent neural network has gradient disappearance and gradient explosion when the network is deepened, and the accuracy of power prediction is low.

Contents of the invention

The present application provides a wind power prediction method, system, and device based on a convolution transformer architecture.

According to the first aspect of the embodiments of the present application, a method for predicting wind power based on a convolutional transformer architecture is provided, including:

Collect meteorological data and operational data, and get embedding vectors.

The embedding vector is input into a power prediction network, which includes an encoder and a decoder.

A feature map corresponding to the embedding vector is obtained according to the encoder.

The feature maps are fed into a decoder to generate predicted powers.

In some embodiments, the time step of the weather data is t, and the weather data includes:

Power plant rated capacity, power generation unit model, number of power generation units and capacity expansion information.

The actual power of the factory station output table.

The height, speed and direction of the wind.

The wind speed at the height of the wind turbine hub and the wind direction at the height of the wind turbine hub.

Air temperature, air pressure, relative humidity.

In some embodiments, the time step of the operation data is t, and the operation data includes:

The name of the plant and station, the starting time of the report, and the forecast time.

Temperature, momentum flux, wind direction, wind speed, and relative humidity at various altitudes.

Sea level pressure, cloud cover, latent heat flux, sensible heat flux, shortwave radiation flux, longwave radiation flux, surface water pressure, total precipitation, large-scale precipitation, convective precipitation.

In some embodiments, the collection of meteorological data and operational data includes:

The collected meteorological data and operating data are normalized, and invalid data is cleaned.

In some embodiments, the obtaining the embedding vector includes:

Let the sliding window slide on the data, select the meteorological data and operating data in the sliding window, and generate an embedding vector.

In some embodiments, the encoder includes a self-attention layer and a feed-forward neural network, and obtaining the feature map corresponding to the embedding vector according to the encoder includes:

The embedding vectors are fed into the self-attention layer to generate query vector q, key vector k and value vector v.

Generate a vector score score according to the q and the k.

Generate a final score based on the score and normalization parameters.

The final score is normalized to generate a normalized score.

Calculate a weighted score vector based on v and the normalized score and calculate the sum of the weighted score vectors.

The sum of the weighted scoring vectors is input into the feedforward neural network, and the feature map is generated.

In some embodiments, the decoder includes a self-attention layer, an encoding-decoding attention layer, and a feed-forward neural network.

According to the second aspect of the embodiments of the present application, a power prediction network training method is provided, including:

Generate datasets from weather data and operational data.

Label the data set to generate a training data set.

The training data set is input into the power prediction network, and the training is carried out with the goal of minimizing the loss function.

In some embodiments, the labeling the data set to generate a training data set includes:

Mark the actual power corresponding to the meteorological data and operating data at each time point.

According to a third aspect of the embodiments of the present application, a wind power prediction device based on a convolution transformer architecture is provided, including:

The collection module is used to collect meteorological data and operational data, and obtain embedding vectors.

The input module is used to input the embedding vector into the power prediction network, and the power prediction network includes an encoder and a decoder.

A feature extraction module, configured to obtain a feature map corresponding to the embedding vector according to the encoder.

A prediction module, configured to input the feature map into a decoder to generate prediction power.

The actual power of the factory station output table.

The height, speed and direction of the wind.

Air temperature, air pressure, relative humidity.

In some embodiments, the collection module includes:

The data cleaning sub-module is used for normalizing the collected meteorological data and operating data, and cleaning invalid data.

In some embodiments, the collection module includes:

The first vector generation sub-module is used to make the sliding window slide on the data, select the meteorological data and operating data in the sliding window, and generate an embedding vector.

In some embodiments, the encoder includes a self-attention layer and a feed-forward neural network, and the feature extraction module includes:

The second vector generation sub-module is used to input the embedding vector into the self-attention layer to generate query vector q, key vector k and value vector v.

The first scoring submodule is configured to generate a vector scoring score according to the q and the k.

The second scoring submodule generates a final score according to the score and normalization parameters.

A third scoring submodule, normalizing the final score to generate a normalized score.

The fourth scoring submodule calculates a weighted scoring vector according to v and the normalized scoring and calculates the sum of the weighted scoring vectors.

The feature extraction submodule is used to input the sum of the weighted scoring vectors into the feedforward neural network and generate the feature map.

According to a fourth aspect of the embodiments of the present application, a power prediction network training device is provided, including:

The data acquisition module is used to generate data sets based on meteorological data and operational data.

The labeling module is used for labeling the data set to generate a training data set.

The training module is used to input the training data set into the power prediction network, and train with the goal of minimizing the loss function.

In some embodiments, the labeling module includes:

The marking sub-module is used to mark the actual power corresponding to the meteorological data and operating data at each time point.

According to a fifth aspect of the embodiments of the present application, a wind power prediction device based on a convolution transformer architecture is provided, including:

processor.

memory for storing said processor-executable instructions;

Wherein, the processor is configured to execute the instructions, so as to realize the wind power prediction method based on the convolution transformer architecture as described in any one of the above first aspects.

According to the sixth aspect of the embodiments of the present application, a non-transitory computer-readable storage medium is provided. When the instructions in the storage medium are executed by the processor of the wind power prediction device based on the convolution transformer architecture, the volume based The wind power prediction device of the convolutional transformer architecture can implement the wind power prediction method based on the convolutional transformer architecture as described in any one of the above first aspects.

According to the seventh aspect of the embodiments of the present application, a power prediction network training device is provided, including:

processor.

memory for storing said processor-executable instructions;

Wherein, the processor is configured to execute the instructions, so as to implement the power prediction network training method as described in the second aspect above.

According to an eighth aspect of the embodiments of the present application, there is provided a non-transitory computer-readable storage medium, when the instructions in the storage medium are executed by the processor of the power prediction network training device, the power prediction network training device can execute The power prediction network training method as described in the second aspect above.

According to a ninth aspect of the embodiments of the present application, there is provided a computer program product, the computer program product includes computer program code, when the computer program code is run on a computer, to execute the above-mentioned first aspect method.

According to a tenth aspect of the embodiments of the present application, there is provided a computer program product, the computer program product includes computer program code, when the computer program code is run on a computer, to execute the above-mentioned second aspect method.

According to an eleventh aspect of the embodiments of the present application, a computer program is provided, the computer program includes computer program code, and when the computer program code is run on a computer, the computer executes the above-mentioned first aspect. method.

According to a twelfth aspect of the embodiments of the present application, a computer program is provided, the computer program includes computer program code, and when the computer program code is run on a computer, the computer executes the above-mentioned second aspect. method.

The technical solutions provided by the embodiments of the present application have at least the following advantages.

By focusing on the data at multiple time points, the attention to local context information is enhanced, the influence of abnormal data on the prediction results is reduced, and the accuracy of power prediction is improved.

When calculating q and k, the convolution kernel is used to perform the convolution operation, so as to focus attention on the local context, so that more relevant features can be matched.

The improved power prediction network can fit faster, improve the prediction accuracy of the model in complex data sets, and achieve lower training loss.

Description of drawings

Fig. 1 is a flow chart of a wind power prediction method based on a convolutional transformer architecture according to an exemplary embodiment.

Fig. 2 is a flow chart of a wind power prediction method based on a convolution transformer architecture according to an exemplary embodiment.

Fig. 3 is a flowchart showing a method for training a power prediction network according to an exemplary embodiment.

Fig. 4 is a block diagram of a wind power prediction device based on a convolution transformer architecture according to an exemplary embodiment.

Fig. 5 is a block diagram of a wind power prediction device based on a convolution transformer architecture according to an exemplary embodiment.

Fig. 6 is a block diagram showing a power prediction network training device according to an exemplary embodiment.

Fig. 7 is a schematic diagram of a power prediction network prediction process according to an exemplary embodiment.

Fig. 8 is a schematic structural diagram of an encoder according to an exemplary embodiment.

Fig. 9 is a schematic structural diagram of a decoder according to an exemplary embodiment.

Fig. 10 is a block diagram of a wind power prediction device based on a convolution transformer architecture according to an exemplary embodiment.

Detailed ways

In order to enable ordinary persons in the art to better understand the technical solutions of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings.

The terms "first", "second" and the like in the specification and claims of the present application and the above drawings are used to distinguish similar objects, and are not necessarily used to describe a specific sequence or sequence. It is to be understood that the data so used are interchangeable under appropriate circumstances such that the embodiments of the application described herein can be practiced in sequences other than those illustrated or described herein. The implementations described in the following exemplary embodiments do not represent all implementations consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with aspects of the present application as recited in the appended claims.

Wind power technology is becoming a major source of future electricity demand. A higher share of renewable energy technologies is critical to carbon-neutrally meeting the needs of future new power system grids, but also brings new grid operation challenges. Electric power companies need to predict the power generated by wind power in order to carry out power generation dispatching operations. Forecasting is a major enabler to ensure safe and economical integration of wind power, while creating links between many flexibility innovations at different levels of the power system to achieve synergies. Accurate wind power forecasting is an important, cost-effective element of energy management, which also helps wind power plants and aggregation systems to participate effectively and directly in the electricity market, and to increase plant profitability through optimized supply planning.

Most wind power forecasting methods are based on time series analysis of wind turbine related data measured at fixed time intervals. In related technologies, a recurrent neural network model is used to analyze and predict time series sequences, but the recurrent neural network has gradient disappearance and gradient explosion when the network is deepened. Even LSTM networks are still incapable of capturing long-term dependencies. The Transformer architecture that emerged in the subsequent development has stronger long-term dependency modeling capabilities, and its effect on processing longer time series has been significantly improved. The method based on the recurrent neural network cannot completely eliminate gradient disappearance and gradient explosion when facing long sequences, while the Transformer architecture can perform better on long sequences, but the self-attention calculation method of the original Transformer architecture is not sensitive to local information. It makes the model vulnerable to outliers or abnormal data, which leads to biased predictions.

This application proposes a method, device and storage medium for wind power prediction based on convolutional transformer architecture.

Fig. 1 is a flow chart of a method for predicting wind power based on a convolution transformer architecture according to an exemplary embodiment. As shown in Fig. 1 , the method includes the following steps:

Step 101, collect meteorological data and operating data, and obtain embedding vectors.

In the embodiment of the present application, data needs to be collected to be input into the power prediction network. The power of wind power generation has two major factors: the operating state of the wind power generating set and the meteorological conditions around the wind power generating set. In this embodiment of the present application, the meteorological data and operating data are collected to predict the power of the wind energy generating set.

The meteorological data includes: the name of the wind power plant, the rated capacity, the model of the power generation unit, the quantity of the power generation unit, and capacity expansion information. The station output table in the station includes time and actual power. The meteorological data includes wind data, and the wind data includes: wind speed, wind direction, air temperature, air pressure and relative humidity at a specified altitude. The specified height can be adjusted by the implementer according to the actual situation, and this application does not limit the specified height. In a possible embodiment, the designated height is 10 meters, 30 meters, 50 meters, 70 meters and the height of the hub of the wind turbine. The running record includes the start time, end time and corresponding maximum output upper limit.

The operation data includes: the name of the factory station, the start time, the forecast time, the wind speed, wind direction, temperature, and relative humidity at the designated height. In a possible embodiment, the designated height is 10 meters, 30 meters meters, 70 meters, 100 meters. At the same time, it is also necessary to measure sea level pressure, cloud cover, latent heat flux, sensible heat flux, momentum flux, short-wave radiation flux, long-wave radiation flux, surface water pressure, total precipitation, large-scale precipitation, and convective precipitation.

It should be noted that the meteorological data and the operating data are collected periodically, and the meteorological data and the operating data are collected once every time step t. The specific value of t can be determined by the implementer according to the actual situation. The situation is adjusted, and this application does not limit t. In a possible embodiment, the time step t is 15 minutes.

Combining the weather data and operating data into time series data, this application implements the data collected at multiple time points to predict the wind power generation power at the next time point, and uses the sliding window to slide on the time series data to select several consecutive time points In order for the power prediction network to identify the time series sequence smoothly, the corresponding embedding vector is generated according to the data selected by the sliding window.

Step 102, input the embedding vector into a power prediction network, and the power prediction network includes an encoder and a decoder.

In the embodiment of the present application, the power prediction network is a neural network of a convolutional migration transformer architecture, and the power prediction network includes an encoder and a decoder.

Step 103, obtain the feature map corresponding to the embedding vector according to the encoder.

In the embodiment of the present application, the encoder includes a self-attention layer and a feed-forward neural network, the embedding vector is input into the self-attention layer and converted into a query vector q, a key vector k and a value vector v, and then the q, k, and v are input into the feedforward neural network to extract features to generate the feature map.

Step 104, input the feature map into a decoder to generate prediction power.

In the embodiment of the present application, the decoder includes a self-attention layer, an encoding-decoding attention layer and a feed-forward neural network, configured to reduce the dimensionality of the feature map to generate the prediction power.

The actual power of the factory station output table.

The height, speed and direction of the wind.

Air temperature, air pressure, relative humidity.

In the embodiment of the present application, the meteorological data includes: the name of the wind power plant, the rated capacity, the model of the power generation unit, the number of the power generation unit, and capacity expansion information. The station output table in the station includes time and actual power. The meteorological data includes wind data, and the wind data includes: wind speed, wind direction, air temperature, air pressure and relative humidity at a specified height. The specified height can be adjusted by the implementer according to the actual situation, and this application does not limit the specified height. In a possible embodiment, the designated height is 10 meters, 30 meters, 50 meters, 70 meters and the height of the hub of the wind turbine. The running record includes the start time, end time and corresponding maximum output upper limit.

In the embodiment of the present application, the operation data includes: the name of the plant station, the start time, the forecast time, the wind speed, wind direction, temperature, and relative humidity at the specified altitude. In a possible embodiment, the specified The height is 10 meters, 30 meters, 70 meters, 100 meters. At the same time, it is also necessary to measure sea level pressure, cloud cover, latent heat flux, sensible heat flux, momentum flux, short-wave radiation flux, long-wave radiation flux, surface water pressure, total precipitation, large-scale precipitation, and convective precipitation.

In the embodiment of the present application, in order to reduce the error of inputting the power prediction network data, it is necessary to clear invalid operating data and meteorological data. Perform data cleaning on the operation data and meteorological data, and delete abnormal data. In one possible embodiment, a threshold range is set to detect data that is significantly different from normal instances, or missing data and repeated measurements are detected by searching for null values. All detected errors and missing data are discarded from the initial dataset. At the same time, in order to prevent gradient explosion, the cleaned data needs to be normalized. In a possible embodiment, the normalized formula is:

Among them, x _norm is the normalized value, x is the original value, x _min is the minimum value in the original value, and x _max is the maximum value in the original value.

In some embodiments, the obtaining the embedding vector includes:

In the embodiment of this application, the meteorological data and operating data are combined into time-series data. This application implements the data collected at multiple time points to predict the wind power generation power at the next time point, and uses a sliding window to slide on the time-series data The data at several consecutive time points are selected, and the corresponding embedding vector is generated according to the data selected by the sliding window in order for the power prediction network to successfully identify the time series.

Fig. 2 is a flow chart of a wind power prediction method based on a convolutional transformer architecture shown according to an exemplary embodiment, the encoder includes a self-attention layer and a feed-forward neural network, as shown in Fig. 2 , the method Include the following steps:

Step 201, input the embedding vector into the self-attention layer to generate query vector q, key vector k and value vector v.

In the embodiment of the present application, the query vector q, key vector k, and value vector v corresponding to the embedding vector are obtained from the attention layer to perform subsequent score calculation and obtain the attention score.

Step 202, generating a vector score score according to the q and the k.

In the embodiment of the present application, the q and k are used to calculate the score of the embedding vector, and the calculation formula of the score is: score=|q×k|, and the score is obtained by multiplying q and k.

Step 203, generate a final score according to the score and normalization parameters.

In the embodiment of this application, in order to stabilize the gradient, it is necessary to normalize the score, that is, divide the score by the normalization parameter

In a possible embodiment, the d _k is the number of dimensions of the key vector k. In another possible embodiment, the score=112, the number of dimensions of k is 64, and the final score is

Step 204, normalize the final score to generate a normalized score.

In the embodiment of the present application, a normalization function is used to normalize the final score. In a possible embodiment, the normalization function is a softmax function, and the final score is input into the softmax function to generate the normalization score. The normalized score indicates the contribution of the embedded vector corresponding to the current time point to the predicted power. The higher the normalized score, the closer the relationship between the data corresponding to the embedded vector and the predicted power, and the greater the contribution to the predicted power . In a possible embodiment, the final score is 12, and a normalized score of 0.88 is output after being normalized by the softmax function, and the normalized score is used for subsequent z-weighting.

Step 205, calculate a weighted score vector according to v and the normalized score and calculate the sum of the weighted score vectors.

In the embodiment of the present application, the normalized score is multiplied by the v to obtain a weighted score vector, and each weighted score vector is added to a set to obtain the sum of the weighted score vectors.

Step 206, input the sum of the weighted scoring vectors into the feedforward neural network, and generate the feature map.

Then, the sum of the weighted scoring vectors is input into the feedforward neural network, and features are extracted to generate the feature map.

Fig. 8 is a schematic structural diagram of an encoder according to an exemplary embodiment. As shown in Figure 8, the encoder includes a self-attention layer and a feed-forward neural network.

In this embodiment of the present application, FIG. 9 is a schematic structural diagram of a decoder according to an exemplary embodiment. As shown in FIG. 9 , the decoder also has a self-attention layer and a feed-forward neural network of the encoder. Besides, there is an encoder-decoder layer (i.e., encoder-decoder attention layer) between these two layers to focus on relevant parts of the input embedding vector. The encoding-decoding attention layer is a fully connected network, wherein there are two layers of networks, the activation function of the first layer is ReLU, and the formulation of the ReLU activation function is expressed as

The sparse model achieved through ReLU can better mine relevant features and fit the training data; the second layer is a linear activation function. The whole encoding-decoding attention layer can be summarized as FFN(Z) function: FFN(Z)=max(0, ZW ₁ +b ₁ )W ₂ +b ₂ .

Fig. 7 is a schematic diagram of a power prediction network prediction process according to an exemplary embodiment. As shown in the figure, the meteorological data and operating data at four time points are selected through the sliding window, and the corresponding query vector q, key vector k, and value vector v are generated according to the convolution kernel, and input to the self-attention layer in the encoder. Computes the correlation score for attention and outputs the sum of weighted score vectors.

Fig. 3 is a flowchart of a method for training a power prediction network according to an exemplary embodiment. As shown in Fig. 3, the method includes the following steps:

Step 301, generating a data set according to meteorological data and operating data.

In the embodiment of the present application, after the meteorological data and operating data are collected by various sensors, a data set can be constructed to train the power prediction network. The data set is a time series data set, and the time step of the meteorological data and the operating data is t, and in a possible embodiment, the t=15 minutes. In a possible embodiment, different data segmentation methods are used to divide the data set, and the data set recorded within 2 years is divided into a training set and a test set. Extract 10 different training sets from the original time series, sequentially or randomly split the first-year evaluation dataset into 10% train, 30% train, 50% train, and 70% train set.

Step 302, label the data set to generate a training data set.

In the embodiment of the present application, the data in the data set are marked, and the meteorological data and operation data collected at each time point correspond to the actual power of wind power generation, so as to train the power prediction network.

Step 303, input the training data set into the power prediction network, and train with the goal of minimizing the loss function.

In the embodiment of the present application, the training data set is input into the power prediction network for iterative training, and the data at several consecutive time points are selected by sliding on the training data set with a sliding window and input into the power prediction network, The predicted power is output, and the predicted power is compared with the actual power to calculate a loss function. The parameters in the power prediction network are optimized with the goal of minimizing the loss function. After training, the recommended power prediction network can be obtained.

Fig. 4 is a block diagram of a wind power prediction device based on a convolution transformer architecture according to an exemplary embodiment. Referring to FIG. 4 , the device 400 includes an acquisition module 410 , an input module 420 , a feature extraction module 430 and a prediction module 440 .

The collection module 410 is used to collect meteorological data and operating data, and obtain embedded vectors.

The input module 420 is configured to input the embedding vector into a power prediction network, and the power prediction network includes an encoder and a decoder.

The feature extraction module 430 is configured to obtain a feature map corresponding to the embedding vector according to the encoder.

A prediction module 440, configured to input the feature map into a decoder to generate prediction power.

The actual power of the factory station output table.

The height, speed and direction of the wind.

Air temperature, air pressure, relative humidity.

In some embodiments, the collection module 410 includes:

The data cleaning sub-module 411 is used for normalizing the collected meteorological data and operating data, and cleaning invalid data.

In some embodiments, the collection module 410 includes:

The first vector generation sub-module 412 is used to make the sliding window slide on the data, select the meteorological data and operating data in the sliding window, and generate an embedding vector.

Fig. 5 is a block diagram of a wind power prediction device based on a convolution transformer architecture according to an exemplary embodiment. Referring to FIG. 5 , the apparatus 500 includes a second vector generation submodule 510 , a first scoring submodule 520 , a second scoring submodule 530 , a third scoring submodule 540 , a fourth scoring submodule 550 and a feature extraction submodule 560 .

The second vector generation sub-module 510 is configured to input the embedding vector into the self-attention layer to generate a query vector q, a key vector k and a value vector v.

The first scoring sub-module 520 is configured to generate a vector score score according to the q and the k.

The second scoring sub-module 530 generates a final score according to the score and normalization parameters.

The third scoring sub-module 540 normalizes the final score to generate a normalized score.

The fourth scoring sub-module 550 calculates a weighted scoring vector according to v and the normalized scoring and calculates the sum of the weighted scoring vectors.

The feature extraction sub-module 560 is configured to input the sum of the weighted scoring vectors into the feedforward neural network, and generate the feature map.

Fig. 6 is a block diagram showing a power prediction network training device according to an exemplary embodiment. Referring to FIG. 6 , the device 600 includes a data collection module 610 , a labeling module 620 and a training module 630 .

The data acquisition module 610 is configured to generate a data set according to meteorological data and operating data.

The labeling module 620 is configured to label the data set to generate a training data set.

A training module 630, configured to input the training data set into the power prediction network, and train with the goal of minimizing the loss function.

In some embodiments, the labeling module 620 includes:

The marking sub-module 621 is used to mark the actual power corresponding to the meteorological data and the operating data at each time point.

Regarding the apparatus in the foregoing embodiments, the specific manner in which each module executes operations has been described in detail in the embodiments related to the method, and will not be described in detail here.

Fig. 10 is a block diagram of an apparatus 1000 for realizing the wind power prediction method based on the convolution transformer architecture according to an exemplary embodiment.

In an exemplary embodiment, there is also provided a storage medium including instructions, such as a memory 1010 including instructions, and an interface 1030 , the instructions can be executed by the processor 1020 of the device 1000 to complete the above method. In some embodiments, the storage medium may be a non-transitory computer-readable storage medium, for example, the non-transitory computer-readable storage medium may be ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical disk. data storage devices, etc.

In an exemplary embodiment, a computer program product is also provided, the computer program product includes computer program code, and when the computer program code is run on a computer, the above method is executed.

In an exemplary embodiment, there is also provided a computer program, the computer program includes computer program code, and when the computer program code is run on a computer, it causes the computer to execute the above method.

All the embodiments of the present application may be implemented independently or in combination with other embodiments, and all of them shall be regarded as the scope of protection required by the present disclosure.

Claims

A method for predicting wind power based on convolution transformer architecture, characterized in that it includes:

Collect meteorological data and operational data, and obtain embedding vectors;

inputting the embedding vector into a power prediction network comprising an encoder and a decoder;

Obtaining a feature map corresponding to the embedding vector according to the encoder;

The feature maps are fed into a decoder to generate predicted powers.
The method according to claim 1, wherein the time step of the weather data is t, and the weather data includes:

Power station rated capacity, power generation unit model, number of power generation units and capacity expansion information;

The actual power of the plant output meter;

the height, speed and direction of the wind;

The wind speed at the height of the fan hub and the wind direction at the height of the fan hub;

Air temperature, air pressure, relative humidity.
The method according to claim 1 or 2, wherein the time step of the operating data is t, and the operating data includes:

The name of the plant and station, the starting time of reporting, and the forecasting time;

Temperature, momentum flux, wind direction, wind speed and relative humidity at each altitude;

Sea level pressure, cloud cover, latent heat flux, sensible heat flux, shortwave radiation flux, longwave radiation flux, surface water pressure, total precipitation, large-scale precipitation, convective precipitation.
The method according to any one of claims 1 to 3, wherein the collecting meteorological data and operating data includes:

The collected meteorological data and operating data are normalized, and invalid data is cleaned.
The method according to any one of claims 1 to 4, wherein said obtaining an embedding vector comprises:

Let the sliding window slide on the data, select the meteorological data and operating data in the sliding window, and generate an embedding vector.
The method according to any one of claims 1 to 5, wherein the encoder includes a self-attention layer and a feed-forward neural network, and the feature map corresponding to the embedding vector is obtained according to the encoder, include:

Input the embedding vector into the self-attention layer to generate query vector q, key vector k and value vector v;

Generate a vector scoring score according to the q and the k;

Generate a final score according to the score and normalization parameters;

normalizing the final score to generate a normalized score;

Calculate a weighted score vector according to v and the normalized score and calculate the sum of the weighted score vectors;

The sum of the weighted scoring vectors is input into the feedforward neural network, and the feature map is generated.
The method according to any one of claims 1 to 6, wherein the decoder comprises a self-attention layer, an encoding-decoding attention layer and a feed-forward neural network.
A power prediction network training method, characterized in that it is used to train the power prediction network described in any one of claims 1-7, comprising:

Generate datasets from meteorological data and operational data;

labeling the data set to generate a training data set;

The training data set is input into the power prediction network, and the training is carried out with the goal of minimizing the loss function.
The method according to claim 8, wherein said labeling said data set to generate a training data set comprises:

Mark the actual power corresponding to the meteorological data and operating data at each time point.
A wind power prediction device based on convolution transformer architecture, characterized in that it comprises:

The collection module is used to collect meteorological data and operational data, and obtain embedded vectors;

An input module, configured to input the embedding vector into a power prediction network, and the power prediction network includes an encoder and a decoder;

A feature extraction module, configured to obtain a feature map corresponding to the embedding vector according to the encoder;

A prediction module, configured to input the feature map into a decoder to generate prediction power.
A wind power prediction device based on convolution transformer architecture, characterized in that it comprises:

processor;

memory for storing said processor-executable instructions;

Wherein, the processor is configured to execute the instructions, so as to realize the wind power prediction method based on the convolution transformer architecture according to any one of claims 1-7.
A non-transitory computer-readable storage medium, characterized in that, when the instructions in the storage medium are executed by the processor of the wind power prediction device based on the convolution transformer architecture, the wind power prediction based on the convolution transformer architecture The device is capable of executing the wind power prediction method based on the convolution transformer architecture according to any one of claims 1 to 7.
A power prediction network training device, characterized in that it comprises:

processor;

memory for storing said processor-executable instructions;

Wherein, the processor is configured to execute the instructions, so as to realize the power prediction network training method as claimed in claim 8 or 9.
A non-transitory computer-readable storage medium, characterized in that, when the instructions in the storage medium are executed by the processor of the power prediction network training device, the power prediction network training device can execute the The power prediction network training method described above.
A computer program product, characterized in that the computer program product includes computer program code, and when the computer program code is run on a computer, the method according to any one of claims 1 to 7 is executed.
A computer program product, characterized in that the computer program product includes computer program code, and when the computer program code is run on a computer, the method as claimed in claim 8 or 9 is executed.
A computer program, characterized in that the computer program includes computer program code, and when the computer program code is run on a computer, the computer executes the method according to any one of claims 1 to 7.
A computer program, characterized in that the computer program includes computer program code, and when the computer program code is run on a computer, the computer is made to execute the method as claimed in claim 8 or 9.