CN115169544A - A short-term photovoltaic power generation power prediction method and system - Google Patents

Abstract

The invention discloses a short-term photovoltaic power generation power prediction method and system. First, the characteristic data of the target domain photovoltaic power station and the source domain photovoltaic power station are obtained, the data is preprocessed and divided into training data and test data, and then the GRU‑DANN adversarial transfer learning model is constructed and trained to obtain the trained GRU-DANN confrontation transfer learning model, and finally input the test data into the trained GRU-DANN confrontation transfer learning model to obtain the target power of the photovoltaic power station in the target domain, and generate the power time series corresponding to the target power according to the time series. The present invention can automatically extract the features required to establish a power prediction model of a few-sample photovoltaic power station from the data of the multi-sample photovoltaic power station, realize the effective migration of the multi-sample photovoltaic power station to the few-sample photovoltaic power station, and improve the efficiency of the few-sample photovoltaic power station. Power prediction accuracy for power plants.

Description

A short-term photovoltaic power generation power prediction method and system

technical field

The invention relates to the technical field of photovoltaic power stations, in particular to a short-term photovoltaic power generation power prediction method and system.

Background technique

As a new energy source, the large-scale grid connection of solar energy brings challenges to the economical, safe and stable operation of the power system. Therefore, accurate photovoltaic power generation power prediction is of great significance to the power system.

Photovoltaic power generation has the characteristics of randomness, intermittency and volatility. The prediction model of photovoltaic power generation requires a large amount of sample data for training and simulation. However, due to the lack of raw data for newly built photovoltaic power stations, the prediction accuracy of photovoltaic power is low. Effective use of multi-data photovoltaic power stations to help low-data photovoltaic power stations to establish a photovoltaic power prediction model will further improve the accuracy of wind farm power prediction. So far, the existing low-data photovoltaic power generation prediction methods are all based on simple model migration or parameter migration to establish the prediction model. These simple methods cannot automatically extract from the multi-data photovoltaic power station data. The features required for photovoltaic power prediction models are prone to negative migration.

Therefore, how to realize the effective migration of multi-data photovoltaic power stations and then establish a small-data photovoltaic power prediction model is an urgent problem to be solved.

SUMMARY OF THE INVENTION

The invention provides a short-term photovoltaic power generation power prediction method and system. The source domain photovoltaic power station represents a multi-sample photovoltaic power station, and the target domain photovoltaic power station represents a few-sample photovoltaic power station. Automatically extract the features required to establish a power prediction model of a few-sample photovoltaic power station, and realize the effective migration of a multi-sample photovoltaic power station to a few-sample photovoltaic power station, so as to improve the power prediction accuracy of the few-sample photovoltaic power station.

The technical scheme of the present invention is as follows:

A short-term photovoltaic power generation power prediction method, comprising the following steps:

S1. Obtain the characteristic data of the photovoltaic power station in the target domain and the photovoltaic power station in the source domain, and perform data preprocessing;

S2. Divide the preprocessed feature data of the target domain photovoltaic power station into two parts, one part and the preprocessed source domain photovoltaic power station feature data are used as training data, and the other part is used as test data;

S3. Use the GRU feature extractor, regression predictor and domain classifier to build a GRU-DANN adversarial transfer learning model, and input training data to train the GRU-DANN adversarial transfer learning model. The process is as follows:

S31, using the GRU feature extractor to extract initial time features from the training data;

S32. Input the extracted initial time feature into the regression predictor to obtain the predicted value of photovoltaic power generation of the initial time feature, and calculate the regression loss function by using the predicted value of photovoltaic power generation and the measured value of photovoltaic power generation. When the regression loss function converges, Take the initial time feature as the target time feature;

S33, input the target time feature into the domain classifier, determine the data domain source of the target time feature through the domain classifier, and calculate the domain loss between the data domain source and the real domain source through the binary cross-entropy formula;

S34. According to the adversarial domain adaptation between the GRU feature extractor and the domain classifier, the parameters of the GRU feature extractor and the domain classifier are continuously updated. When the domain loss converges, the GRU feature extractor obtains from the source domain and the target domain. Domain invariant features, the GRU-DANN adversarial transfer learning model training is completed;

S4. Input the test data into the trained GRU-DANN adversarial transfer learning model, obtain the target power of the photovoltaic power station in the target domain, and generate a power time series corresponding to the target power according to the time series.

The short-term photovoltaic power generation power prediction method based on the gated recurrent neural network and the domain confrontation neural network of the present invention can effectively extract the source photovoltaic power station and the target photovoltaic power station data through the feature extractor of the gated recurrent neural network (GRU). The domain adversarial neural network (DANN) can find the domain invariant features between the source domain and the target domain that can effectively help the target domain photovoltaic power station to establish a prediction model, and the present invention can effectively improve the prediction accuracy of photovoltaic power generation.

Further, in step S1, the characteristic data includes data of power, temperature, humidity, direct solar radiation intensity, solar scattering intensity and wind speed;

The process of preprocessing feature data is as follows:

Perform min-max normalization on the power, temperature, humidity, direct solar radiation intensity, solar scattering intensity and wind speed data to obtain the processed power sequence P, temperature sequence T, humidity sequence H, and direct solar radiation intensity sequence D, The solar scattering intensity sequence S and the wind speed sequence W.

Further, in step S2, the preprocessed power, temperature, humidity, direct solar radiation intensity, solar scattering intensity and wind speed data of the source domain photovoltaic power station are all used as training data, while the preprocessed power, temperature and wind speed data of the target domain photovoltaic power station are used as training data. , humidity, direct solar radiation intensity, solar scattering intensity and wind speed data, the data of three different meteorological days of sunny, cloudy and rainy days are selected as test data, and the rest are used as training data.

It should be noted that the sequence data of the six parameters of power, temperature, humidity, solar direct radiation intensity, solar scattering intensity and wind speed form a data string at a certain moment, that is, at different moments, multiple data strings can be formed. , and the multiple data strings form a data set, which is the form of training data or test data set.

Further, in step S3, the construction process of the GRU-DANN adversarial transfer learning model is as follows:

The GRU-DANN adversarial transfer learning model uses the GRU feature extractor, and is connected to the regression predictor and the domain classifier after the GRU feature extractor, where the GRU feature extractor and the domain classifier are connected through a gradient reversal layer;

The time features are obtained by inputting data to the GRU feature extractor, and then the time features are input to the regression predictor and the domain classifier respectively to obtain the corresponding photovoltaic power generation power prediction data and domain label prediction data.

Further, in step S3, the GRU feature extractor includes two GRU layers and an activation function Tanh, and the two GRU layers include 6 neurons and 64 neurons respectively;

The regression predictor includes three fully connected layers, and the three fully connected layers include 100 neurons, 100 neurons and 1 neuron respectively;

The domain classifier consists of two fully connected layers with 100 neurons and 1 neuron, respectively.

Further, in step S32, the process of calculating the regression loss function through the predicted value of photovoltaic power generation and the measured value of photovoltaic power generation is as follows:

的公式如下：The regression loss of power generation prediction is defined as the mean square error, that is, the regression loss function

The formula is as follows:

表示训练数据的样本数量，

和

分别表示实测值和预测值。In the formula,

represents the number of samples in the training data,

and

represent the measured and predicted values, respectively.

Further, in step S33, the process of calculating the domain loss between the data domain source and the real domain source through the binary cross-entropy formula is as follows:

Domain loss is defined as binary cross entropy, and the binary cross entropy formula is as follows:

表示域损失，

和

分别表示实际域标签和预测域标签，其中，源域的域标签为0，目标域的域标签为1。In the formula,

represents the domain loss,

and

represent the actual domain label and the predicted domain label, respectively, where the domain label of the source domain is 0 and the domain label of the target domain is 1.

Further, in step S34, the process of updating the parameters of the GRU feature extractor and the domain classifier is as follows:

By training the GRU feature extractor, the features of the source domain and the target domain are extracted, and then the extracted features are input into the domain classifier. By continuously training the GRU feature extractor to extract the domain-invariant features between the source domain and the target domain, the domain classifier cannot correctly identify the domain labels, that is, the extracted features cannot be distinguished from the source domain or the target domain. At this time, the domain loss Convergence, the GRU feature extractor can successfully extract the domain invariant features between the source domain and the target domain, then the parameter update of the GRU feature extractor and the domain classifier is completed, indicating that the training of the GRU-DANN adversarial transfer learning model is completed.

来表示，下式表示其正向和反向传播过程：Further, since the GRU feature extractor and the domain classifier have opposite effects on the domain loss during the training process of the GRU-DANN adversarial transfer learning model, the purpose of the feature extractor is to make the domain classifier unable to distinguish the source of the extracted features, even if the The domain loss is maximized, and the purpose of the domain classifier is to accurately distinguish the source of the features extracted by the GRU feature extractor, even if the domain loss is minimized, this min-max operation cannot simultaneously pass through the neural network backpropagation process. The gradient update is directly implemented, so a gradient reversal layer (GRL) is added between the GRU feature extractor and the domain classifier. The function of the gradient reversal layer is to multiply the gradient passed into the gradient reversal layer by a negative number, so that in the The training goals of the network before and after the gradient inversion layer are opposite, and the gradient inversion layer uses a pseudo function

to express, the following formula expresses its forward and backward propagation process:

代表单位矩阵，

是用于实现回归损失和域损失之间权衡的超参数，

代表当前批次数，

代表当前迭代数，

代表迭代总次数，

表示源域和目标域的最小总批次数，

为迭代进程相对值，即当前迭代次数与总迭代次数的比率，

为常数10。In the formula,

represents the identity matrix,

are the hyperparameters used to achieve the trade-off between regression loss and domain loss,

represents the current batch number,

represents the current iteration number,

represents the total number of iterations,

represents the minimum total number of batches for the source and target domains,

is the relative value of the iteration process, that is, the ratio of the current number of iterations to the total number of iterations,

is the constant 10.

The invention also provides a short-term photovoltaic power generation power prediction system, which includes a data processing module, a model building module and an output module respectively connected to the control center in communication;

The data processing module is used to obtain the characteristic data of the photovoltaic power station in the target domain and the photovoltaic power station in the source domain, perform data preprocessing, and then divide the preprocessed characteristic data of the photovoltaic power station in the target domain into two parts. The preprocessed feature data of the source-domain photovoltaic power station is used as training data, and the other part is used as test data, and the training data and test data are sent to the control center;

The model building module uses GRU feature extractor, regression predictor and domain classifier to build a GRU-DANN adversarial transfer learning model, and obtains training data from the control center, and inputs the training data to carry out the GRU-DANN adversarial transfer learning model. The training process is as follows:

Extract initial temporal features from training data using GRU feature extractor;

Input the extracted initial time feature into the regression predictor to obtain the predicted value of photovoltaic power generation of the initial time feature, and calculate the regression loss function through the predicted value of photovoltaic power generation power and the measured value of photovoltaic power generation power. When the regression loss function converges, the initial Temporal features as target temporal features;

Input the target time feature into the domain classifier, determine the data domain source of the target time feature through the domain classifier, and calculate the domain loss between the data domain source and the real domain source through the binary cross entropy formula;

According to the adversarial domain adaptation between the GRU feature extractor and the domain classifier, the parameters of the GRU feature extractor and the domain classifier are continuously updated. When the domain loss converges, the GRU feature extractor obtains the domain difference between the source domain and the target domain. If the feature is changed, the GRU-DANN adversarial transfer learning model training is completed;

The model building module outputs the trained GRU-DANN adversarial transfer learning model to the control center;

The control center inputs the test data into the trained GRU-DANN adversarial transfer learning model, outputs the target power of the photovoltaic power station in the target domain, and generates a power time series corresponding to the target power according to the time sequence;

The output module is used for outputting and displaying the predicted power time series.

Compared with the prior art, the present invention has the following beneficial effects:

(1) The short-term photovoltaic power generation power prediction method proposed by the present invention combines deep learning and adversarial domain adaptation methods for photovoltaic power generation power prediction for the first time. The proposed GRU-DANN adversarial transfer learning model can significantly improve photovoltaic power generation. Predictive power generation performance of power stations.

(2) The GRU feature extractor of the present invention is used to automatically extract temporal features across the source domain and the target domain. DANN finds the domain between the source domain and the target domain through the adversarial domain adaptation of the GRU feature extractor and the domain classifier. Invariant features, thus completing the training of the GRU-DANN adversarial transfer learning model. The invention realizes the effective migration of the multi-sample photovoltaic power station data to the few-sample photovoltaic power station data, and the trained model can be directly applied to help predict the power generation of the target photovoltaic power station without causing the prediction performance to decline due to domain transfer. , which can effectively improve the power prediction accuracy of few-sample photovoltaic power stations, and has certain practical significance for short-term photovoltaic power generation power prediction.

Description of drawings

FIG. 1 is a flow chart of the short-term photovoltaic power generation power prediction method of the present invention.

FIG. 2 is a frame diagram of the GRU-DANN adversarial transfer learning model of the present invention.

FIG. 3 is a frame diagram of the short-term photovoltaic power generation power prediction system of the present invention.

FIG. 4 is an effect diagram of forecasting the forecast data on sunny days by the short-term photovoltaic power generation power forecasting method of the present invention.

FIG. 5 is an effect diagram of forecasting the forecast data in cloudy days by the short-term photovoltaic power generation power forecasting method of the present invention.

FIG. 6 is an effect diagram of predicting the prediction data in rainy days by the short-term photovoltaic power generation power prediction method of the present invention.

Detailed ways

The accompanying drawings are for illustrative purposes only, and should not be construed as limitations on this patent; in order to better illustrate the present embodiment, some parts of the accompanying drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product; for those skilled in the art It is understandable to the artisan that certain well-known structures and descriptions thereof may be omitted from the drawings. The positional relationships described in the drawings are only for exemplary illustration, and should not be construed as a limitation on the present patent.

Example 1:

As shown in FIG. 1 and FIG. 2 , this embodiment provides a short-term photovoltaic power generation power prediction method, which includes the following steps:

S1. Obtain the characteristic data of the photovoltaic power station in the target domain and the photovoltaic power station in the source domain, and perform data preprocessing;

S2. Divide the preprocessed feature data of the target domain photovoltaic power station into two parts, one part and the preprocessed source domain photovoltaic power station feature data are used as training data, and the other part is used as test data;

S3. Use the GRU feature extractor, regression predictor and domain classifier to build a GRU-DANN adversarial transfer learning model, and input training data to train the GRU-DANN adversarial transfer learning model. The process is as follows:

S31, using the GRU feature extractor to extract initial time features from the training data;

S32. Input the extracted initial time feature into the regression predictor to obtain the predicted value of photovoltaic power generation of the initial time feature, and calculate the regression loss function by using the predicted value of photovoltaic power generation and the measured value of photovoltaic power generation. When the regression loss function converges, Take the initial time feature as the target time feature;

S33, input the target time feature into the domain classifier, determine the data domain source of the target time feature through the domain classifier, and calculate the domain loss between the data domain source and the real domain source through the binary cross-entropy formula;

S34. According to the adversarial domain adaptation between the GRU feature extractor and the domain classifier, the parameters of the GRU feature extractor and the domain classifier are continuously updated. When the domain loss converges, the GRU feature extractor obtains from the source domain and the target domain. Domain invariant features, the GRU-DANN adversarial transfer learning model training is completed;

S4. Input the test data into the trained GRU-DANN adversarial transfer learning model, obtain the target power of the photovoltaic power station in the target domain, and generate a power time series corresponding to the target power according to the time series.

The short-term photovoltaic power generation power prediction method based on the gated recurrent neural network and the domain confrontation neural network of the present invention can effectively extract the source photovoltaic power station and the target photovoltaic power station data through the feature extractor of the gated recurrent neural network (GRU). The domain adversarial neural network (DANN) can find the domain invariant features between the source domain and the target domain that can effectively help the target domain photovoltaic power station to establish a prediction model, and the present invention can effectively improve the prediction accuracy of photovoltaic power generation.

In step S1 of this embodiment, the characteristic data includes data of power, temperature, humidity, direct solar radiation intensity, solar scattering intensity and wind speed;

The process of preprocessing feature data is as follows:

Perform min-max normalization on the power, temperature, humidity, direct solar radiation intensity, solar scattering intensity and wind speed data to obtain the processed power sequence P, temperature sequence T, humidity sequence H, and direct solar radiation intensity sequence D, The solar scattering intensity sequence S and the wind speed sequence W.

In step S2 of this embodiment, the power, temperature, humidity, direct solar radiation intensity, solar scattering intensity and wind speed data preprocessed by the photovoltaic power station in the source domain are all used as training data, while the preprocessed data from the photovoltaic power station in the target domain are all used as training data. For power, temperature, humidity, direct solar radiation intensity, solar scattering intensity and wind speed data, the data of three different meteorological days of sunny, cloudy and rainy days are selected as test data, and the rest are used as training data.

It should be noted that the sequence data of the six parameters of power, temperature, humidity, solar direct radiation intensity, solar scattering intensity and wind speed form a data string at a certain moment, that is, at different moments, multiple data strings can be formed. , and the multiple data strings form a data set, which is the form of training data or test data set.

Among them, the training data is the data used to train the GRU-DANN adversarial transfer learning model, and the GRU-DANN adversarial transfer learning model is composed of the gate recurrent unit (GRU) feature extractor, regression predictor and domain classification The model constructed by the three devices is used to automatically extract the common features of the multi-data photovoltaic power station and the less-data photovoltaic power station, and help the less-data photovoltaic power station to predict the photovoltaic power generation.

In step S3 of this embodiment, the construction process of the GRU-DANN adversarial transfer learning model is as follows:

The GRU-DANN adversarial transfer learning model uses the GRU feature extractor, and is connected to the regression predictor and the domain classifier after the GRU feature extractor, where the GRU feature extractor and the domain classifier are connected through a gradient reversal layer;

The time features are obtained by inputting data to the GRU feature extractor, and then the time features are input to the regression predictor and the domain classifier respectively to obtain the corresponding photovoltaic power generation power prediction data and domain label prediction data.

In step S3 of this embodiment, the GRU feature extractor includes two GRU layers and an activation function Tanh, and the two GRU layers include 6 neurons and 64 neurons respectively;

The formula of the activation function Tanh is as follows:

和

为重置门和更新门，

为隐含层的状态，

和

为输入与输出，

为上一个输出，

、

、

、

、

、

为权重参数矩阵，

、

、

为偏置参数矩阵，

为矩阵乘法，

为Sigmod函数；In the formula,

and

To reset gates and update gates,

is the state of the hidden layer,

and

for input and output,

for the previous output,

,

,

,

,

,

is the weight parameter matrix,

,

,

is the bias parameter matrix,

for matrix multiplication,

is the Sigmod function;

The regression predictor includes three fully connected layers, and the three fully connected layers include 100 neurons, 100 neurons and 1 neuron respectively;

The domain classifier consists of two fully connected layers with 100 neurons and 1 neuron, respectively.

In step S31 of this embodiment, the GRU feature extractor extracts the initial time features of the source domain and the target domain respectively from the training data; the initial time features are the power sequence P, the temperature sequence T, the humidity sequence H, and the direct solar radiation intensity. The implicit data information in the sequence D, the scattering intensity sequence S and the wind speed sequence W, the implicit data information refers to the information related to power.

In step S32 of this embodiment, the process of calculating the regression loss function through the predicted value of photovoltaic power generation and the measured value of photovoltaic power generation is as follows:

的公式如下：The regression loss of power generation prediction is defined as the mean square error, that is, the regression loss function

The formula is as follows:

表示训练数据的样本数量，

和

分别表示实测值和预测值。In the formula,

represents the number of samples in the training data,

and

represent the measured and predicted values, respectively.

In step S33 of the present embodiment, the process of calculating the domain loss between the data domain source and the real domain source through the binary cross-entropy formula is as follows:

Domain loss is defined as binary cross entropy, and the binary cross entropy formula is as follows:

表示域损失，

和

分别表示实际域标签和预测域标签，其中，源域的域标签为0，目标域的域标签为1。In the formula,

represents the domain loss,

and

represent the actual domain label and the predicted domain label, respectively, where the domain label of the source domain is 0 and the domain label of the target domain is 1.

In step S34 of this embodiment, the process of updating the parameters of the GRU feature extractor and the domain classifier is as follows:

By training the GRU feature extractor, the features of the source domain and the target domain are extracted, and then the extracted features are input into the domain classifier. By continuously training the GRU feature extractor to extract the domain-invariant features between the source domain and the target domain, the domain classifier cannot correctly identify the domain labels, that is, the extracted features cannot be distinguished from the source domain or the target domain. At this time, the domain loss Convergence, the GRU feature extractor can successfully extract the domain invariant features between the source domain and the target domain, then the parameter update of the GRU feature extractor and the domain classifier is completed, indicating that the training of the GRU-DANN adversarial transfer learning model is completed.

来表示，下式表示其正向和反向传播过程：Among them, since the GRU feature extractor and the domain classifier have opposite effects on the domain loss during the training process of the GRU-DANN adversarial transfer learning model, the purpose of the feature extractor is to make the domain classifier unable to distinguish the source of the extracted features, even if the The domain loss is maximized, and the purpose of the domain classifier is to accurately distinguish the source of the features extracted by the GRU feature extractor, even if the domain loss is minimized, this min-max operation cannot simultaneously pass through the neural network backpropagation process. The gradient update is directly implemented, so a gradient reversal layer (GRL) is added between the GRU feature extractor and the domain classifier. The function of the gradient reversal layer is to multiply the gradient passed into the gradient reversal layer by a negative number, so that in the The training goals of the network before and after the gradient inversion layer are opposite, and the gradient inversion layer uses a pseudo function

to express, the following formula expresses its forward and backward propagation process:

代表单位矩阵，

是用于实现回归损失和域损失之间权衡的超参数，

代表当前批次数，

代表当前迭代数，

代表迭代总次数，

表示源域和目标域的最小总批次数，

为迭代进程相对值，即当前迭代次数与总迭代次数的比率，

为常数10。In the formula,

represents the identity matrix,

are the hyperparameters used to achieve the trade-off between regression loss and domain loss,

represents the current batch number,

represents the current iteration number,

represents the total number of iterations,

represents the minimum total number of batches for the source and target domains,

is the relative value of the iteration process, that is, the ratio of the current number of iterations to the total number of iterations,

is the constant 10.

The short-term photovoltaic power generation power prediction method proposed by the present invention combines deep learning and adversarial domain adaptation methods for photovoltaic power generation power prediction for the first time. The proposed GRU-DANN confrontation transfer learning model can significantly improve the power generation power of photovoltaic power stations. Predictive performance.

Example 2:

As shown in FIG. 3 , this embodiment also provides a short-term photovoltaic power generation power prediction system, which is used to realize the short-term photovoltaic power generation power prediction method in the above-mentioned embodiment 1. Model module, output module;

The data processing module is used to obtain the characteristic data of the photovoltaic power station in the target domain and the photovoltaic power station in the source domain, perform data preprocessing, and then divide the preprocessed characteristic data of the photovoltaic power station in the target domain into two parts. The preprocessed feature data of the source-domain photovoltaic power station is used as training data, and the other part is used as test data, and the training data and test data are sent to the control center;

The model building module uses GRU feature extractor, regression predictor and domain classifier to build a GRU-DANN adversarial transfer learning model, and obtains training data from the control center, and inputs the training data to carry out the GRU-DANN adversarial transfer learning model. The training process is as follows:

Extract initial temporal features from training data using GRU feature extractor;

Input the extracted initial time feature into the regression predictor to obtain the predicted value of photovoltaic power generation of the initial time feature, and calculate the regression loss function through the predicted value of photovoltaic power generation power and the measured value of photovoltaic power generation power. When the regression loss function converges, the initial Temporal features as target temporal features;

Input the target time feature into the domain classifier, determine the data domain source of the target time feature through the domain classifier, and calculate the domain loss between the data domain source and the real domain source through the binary cross entropy formula;

According to the adversarial domain adaptation between the GRU feature extractor and the domain classifier, the parameters of the GRU feature extractor and the domain classifier are continuously updated. When the domain loss converges, the GRU feature extractor obtains the domain difference between the source domain and the target domain. If the feature is changed, the GRU-DANN adversarial transfer learning model training is completed;

The model building module outputs the trained GRU-DANN adversarial transfer learning model to the control center;

The control center inputs the test data into the trained GRU-DANN adversarial transfer learning model, outputs the target power of the photovoltaic power station in the target domain, and generates a power time series corresponding to the target power according to the time sequence;

The output module is used for outputting and displaying the predicted power time series.

Example 3:

This embodiment uses specific data to verify the validity of the short-term photovoltaic power generation power prediction method in the above-mentioned embodiment 1, and the specific process is as follows:

In step S1, the power, temperature, humidity, direct solar radiation intensity, solar scattering intensity and wind speed data of the three photovoltaic power stations in Australia from 2018/01/01/0:00 to 2018/12/29/23:40 are obtained As the characteristic data of the photovoltaic power station in the source domain, obtain the power, temperature, humidity, direct solar radiation intensity, solar scattering intensity and wind speed data of another photovoltaic power station in the same Australian time period as the characteristic data of the photovoltaic power station in the target domain;

In step S2, the data is processed according to the method in Embodiment 1, and training data and test data are divided;

In step S3, the GRU-DANN confrontation transfer learning model is trained according to the training data obtained above, and the trained GRU-DANN confrontation transfer learning model is obtained;

In step S4, the test data is input into the trained GRU-DANN adversarial transfer learning model, the target power of the photovoltaic power station in the target domain is obtained, and the power time series corresponding to the target power is generated according to the time sequence;

Among them, the output target power is a power prediction point every 20 minutes, and a 72×1 tensor photovoltaic power time series is generated according to 72 power prediction points in a day.

As shown in FIG. 4-FIG. 6, in this embodiment, the power prediction effects of the photovoltaic power station in the output target domain in sunny days, cloudy days and rainy days are obtained respectively. It can be seen that the present invention effectively improves the power prediction accuracy of a few-sample photovoltaic power station.

To sum up, the GRU feature extractor of the present invention is used to automatically extract temporal features across the source domain and the target domain, and DANN finds between the source domain and the target domain through the adversarial domain adaptation of the GRU feature extractor and the domain classifier. Domain invariant features, thus completing the training of the GRU-DANN adversarial transfer learning model.

The invention realizes the effective migration of the multi-sample photovoltaic power station data to the few-sample photovoltaic power station data, and the trained model can be directly applied to help predict the power generation of the target photovoltaic power station without causing the prediction performance to decline due to domain transfer. , which can effectively improve the power prediction accuracy of few-sample photovoltaic power stations, and has certain practical significance for short-term photovoltaic power generation power prediction.

Obviously, the above-mentioned embodiments of the present invention are only examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those of ordinary skill in the art, changes or modifications in other different forms can also be made on the basis of the above description. There is no need and cannot be exhaustive of all implementations here. Any modifications, equivalent replacements and improvements made within the spirit and principle of the present invention shall be included within the protection scope of the claims of the present invention.

Claims (10)

1. A short-term photovoltaic power generation power prediction method, comprising the following steps: S1. Obtain the characteristic data of the photovoltaic power station in the target domain and the photovoltaic power station in the source domain, and perform data preprocessing; S2. Divide the preprocessed feature data of the target domain photovoltaic power station into two parts, one part and the preprocessed source domain photovoltaic power station feature data are used as training data, and the other part is used as test data; S3. Use the GRU feature extractor, regression predictor and domain classifier to build a GRU-DANN adversarial transfer learning model, and input training data to train the GRU-DANN adversarial transfer learning model. The process is as follows: S31, using the GRU feature extractor to extract initial time features from the training data; S32. Input the extracted initial time feature into the regression predictor to obtain the predicted value of photovoltaic power generation of the initial time feature, and calculate the regression loss function by using the predicted value of photovoltaic power generation and the measured value of photovoltaic power generation. When the regression loss function converges, Take the initial time feature as the target time feature; S33, input the target time feature into the domain classifier, determine the data domain source of the target time feature through the domain classifier, and calculate the domain loss between the data domain source and the real domain source through the binary cross-entropy formula; S34. According to the adversarial domain adaptation between the GRU feature extractor and the domain classifier, the parameters of the GRU feature extractor and the domain classifier are continuously updated. When the domain loss converges, the GRU feature extractor obtains from the source domain and the target domain. Domain invariant features, the GRU-DANN adversarial transfer learning model training is completed; S4. Input the test data into the trained GRU-DANN adversarial transfer learning model, obtain the target power of the photovoltaic power station in the target domain, and generate a power time series corresponding to the target power according to the time series. 2. A short-term photovoltaic power generation power prediction method according to claim 1, wherein in step S1, the characteristic data includes data of power, temperature, humidity, direct solar radiation intensity, solar scattering intensity and wind speed; The process of preprocessing feature data is as follows: Perform min-max normalization on the power, temperature, humidity, direct solar radiation intensity, solar scattering intensity and wind speed data to obtain the processed power sequence P, temperature sequence T, humidity sequence H, and direct solar radiation intensity sequence D, The solar scattering intensity sequence S and the wind speed sequence W. 3. A short-term photovoltaic power generation power prediction method according to claim 2, characterized in that, in step S2, the preprocessed power, temperature, humidity, direct solar radiation intensity, solar scattering intensity and The wind speed data are all used as training data, and the preprocessed power, temperature, humidity, solar direct radiation intensity, solar scattering intensity and wind speed data of the target domain photovoltaic power station are selected from three different meteorological days: sunny, cloudy and rainy days. as test data and the rest as training data. 4. A short-term photovoltaic power generation power prediction method according to claim 1, wherein in step S3, the construction process of the GRU-DANN confrontation transfer learning model is as follows: The GRU-DANN adversarial transfer learning model uses the GRU feature extractor, and is connected to the regression predictor and the domain classifier after the GRU feature extractor, where the GRU feature extractor and the domain classifier are connected through a gradient reversal layer; The time features are obtained by inputting data to the GRU feature extractor, and then the time features are input to the regression predictor and the domain classifier respectively to obtain the corresponding photovoltaic power generation power prediction data and domain label prediction data. 5. A short-term photovoltaic power generation power prediction method according to claim 1, wherein in step S3, the GRU feature extractor comprises two GRU layers and an activation function Tanh, and the two GRU layers respectively comprise 6 neurons and 64 neurons; The regression predictor includes three fully connected layers, and the three fully connected layers include 100 neurons, 100 neurons and 1 neuron respectively; The domain classifier consists of two fully connected layers with 100 neurons and 1 neuron, respectively. 6. A short-term photovoltaic power generation power prediction method according to claim 1, wherein in step S32, the process of calculating the regression loss function by the photovoltaic power generation power predicted value and the photovoltaic power generation power measured value is as follows: The regression loss of power generation prediction is defined as the mean square error, that is, the regression loss function

The formula is as follows:

In the formula,

represents the number of samples in the training data,

and

represent the measured and predicted values, respectively. 7 . The short-term photovoltaic power generation power prediction method according to claim 1 , wherein, in step S33 , the process of calculating the domain loss between the data domain source and the real domain source through the binary cross-entropy formula is as follows: 8 . Domain loss is defined as binary cross entropy, and the binary cross entropy formula is as follows:

In the formula,

represents the domain loss,

and

represent the actual domain label and the predicted domain label, respectively, where the domain label of the source domain is 0 and the domain label of the target domain is 1. 8. The short-term photovoltaic power generation power prediction method according to claim 1, wherein in step S34, the process of updating the parameters of the GRU feature extractor and the domain classifier is as follows: By training the GRU feature extractor, the features of the source domain and the target domain are extracted, and then the extracted features are input into the domain classifier. By continuously training the GRU feature extractor to extract the domain-invariant features between the source domain and the target domain, the domain classifier cannot correctly identify the domain labels, that is, the extracted features cannot be distinguished from the source domain or the target domain. At this time, the domain loss Convergence, the GRU feature extractor can successfully extract the domain invariant features between the source domain and the target domain, then the parameter update of the GRU feature extractor and the domain classifier is completed, indicating that the training of the GRU-DANN adversarial transfer learning model is completed. 9. A short-term photovoltaic power generation power prediction method according to claim 8, wherein, since the GRU feature extractor and the domain classifier have opposite effects on the domain loss during the training process of the GRU-DANN adversarial transfer learning model, The purpose of the feature extractor is to make the domain classifier unable to distinguish the source of the extracted features, that is, to maximize the domain loss, while the purpose of the domain classifier is to accurately distinguish the source of the features extracted by the GRU feature extractor, that is, to obtain the domain loss. Minimization, this minimum-maximum operation cannot be directly implemented by the gradient update in the neural network backpropagation process at the same time, so a gradient inversion layer is added between the GRU feature extractor and the domain classifier. The function of the gradient inversion layer is Multiply the gradient passed to the gradient inversion layer by a negative number, so that the training goals of the network before and after the gradient inversion layer are opposite, and the gradient inversion layer uses a pseudo function

to express, the following formula expresses its forward and backward propagation process:

In the formula,

represents the identity matrix,

are the hyperparameters used to achieve the trade-off between regression loss and domain loss,

represents the current batch number,

represents the current iteration number,

represents the total number of iterations,

represents the minimum total number of batches for the source and target domains,

is the relative value of the iteration process, that is, the ratio of the current number of iterations to the total number of iterations,

is the constant 10. 10. A short-term photovoltaic power generation power prediction system, characterized in that it comprises a data processing module, a model building module and an output module respectively connected to a control center in communication; The data processing module is used to obtain the characteristic data of the photovoltaic power station in the target domain and the photovoltaic power station in the source domain, perform data preprocessing, and then divide the preprocessed characteristic data of the photovoltaic power station in the target domain into two parts. The preprocessed feature data of the source-domain photovoltaic power station is used as training data, and the other part is used as test data, and the training data and test data are sent to the control center; The model building module uses GRU feature extractor, regression predictor and domain classifier to build a GRU-DANN adversarial transfer learning model, and obtains training data from the control center, and inputs the training data to carry out the GRU-DANN adversarial transfer learning model. The training process is as follows: Extract initial temporal features from training data using GRU feature extractor; Input the extracted initial time feature into the regression predictor to obtain the predicted value of photovoltaic power generation of the initial time feature, and calculate the regression loss function through the predicted value of photovoltaic power generation power and the measured value of photovoltaic power generation power. When the regression loss function converges, the initial Temporal features as target temporal features; Input the target time feature into the domain classifier, determine the data domain source of the target time feature through the domain classifier, and calculate the domain loss between the data domain source and the real domain source through the binary cross entropy formula; According to the adversarial domain adaptation between the GRU feature extractor and the domain classifier, the parameters of the GRU feature extractor and the domain classifier are continuously updated. When the domain loss converges, the GRU feature extractor obtains the domain difference between the source domain and the target domain. If the feature is changed, the GRU-DANN adversarial transfer learning model training is completed; The model building module outputs the trained GRU-DANN adversarial transfer learning model to the control center; The control center inputs the test data into the trained GRU-DANN adversarial transfer learning model, outputs the target power of the photovoltaic power station in the target domain, and generates a power time series corresponding to the target power according to the time sequence; The output module is used for outputting and displaying the predicted power time series.

Images