CN117200200A - Training method of photovoltaic power prediction model - Google Patents

Abstract

The application relates to a training method of a photovoltaic power prediction model, belongs to the technical field of photovoltaic power generation, and aims to solve the problems that the existing photovoltaic power prediction model has low prediction precision and the existing deep learning network model has low calculation efficiency when applied to photovoltaic power prediction. The method of the application comprises the following steps: classifying the daily weather of the region where the photovoltaic electric field to be predicted is located in a historical period into stationary weather and turning weather; classifying weather in each history period in a stationary weather day and a turning weather day into a plurality of sub-weather types based on generalized weather types; constructing a training data set of each sub-weather type; training the pre-established photovoltaic power prediction model through training data sets of all sub-weather types respectively to obtain sub-photovoltaic power prediction models corresponding to all the sub-weather types. The photovoltaic power prediction model trained by the method has higher prediction precision.

Description

A training method for photovoltaic power prediction model

Technical field

The invention relates to the technical field of photovoltaic power generation, and in particular to a training method for a photovoltaic power prediction model.

Background technique

Photovoltaic power generation is a renewable, clean, and flexible distributed energy source that plays an important role in meeting the growing demand for clean energy worldwide. As the integration of photovoltaic power generation has brought significant economic and environmental benefits, the penetration rate of photovoltaic power generation has gradually increased, but its high penetration rate has also brought many new problems to the operation of the existing power grid system. In particular, photovoltaic output is volatile and intermittent. When a high proportion of photovoltaic power stations are connected to the power grid, it will have an impact on the power system. In order to solve the above problems, the demand for photovoltaic output forecasting continues to increase. Among them, distributed photovoltaic power station output forecasting based on refined weather classification identification in micro-meteorological environments is an important area of photovoltaic output forecasting. Through effective classification of weather types, it can Improve the accuracy of photovoltaic power prediction.

In the existing technology, most of the weather classification technologies used for photovoltaic output prediction are subdivided for all weather types. For example, weather types are divided into three categories through clarity index, or cross subdivision is introduced by introducing total cloud cover. Then the sunny day types are divided into three categories. However, in engineering applications, due to the uncertainty of meteorological environmental factors, turning weather with large fluctuations in a day (that is, weather with drastic changes in meteorology) often has a greater impact on photovoltaic output and the stability and dispatch of the power grid. This impact is usually considered when designing photovoltaic grid-connected energy storage. However, the previous classification of weather types cannot well identify the turning weather days when large power fluctuations occur. Therefore, based on the existing weather classification method Effective classification of model training samples cannot be achieved, resulting in low prediction accuracy of the prediction model, and therefore more accurate photovoltaic power prediction results cannot be obtained.

Most of the existing photovoltaic power prediction models are shallow models, but shallow models have limitations in feature selection, generalization capabilities, and processing of complex samples. There is a problem of low accuracy when using shallow models to predict photovoltaic power. Deep learning models, such as convolutional neural networks, can extract deeper features of data compared to traditional shallow models, and have greatly improved accuracy. However, in the field of photovoltaic power prediction, in order to ensure prediction accuracy, the amount of training data for photovoltaic power prediction models is huge, and deep learning models such as traditional neural networks have complex structures and suffer from low computational efficiency.

Contents of the invention

In view of the above analysis, embodiments of the present invention aim to provide a training method for a photovoltaic power prediction model to solve the problem of low prediction accuracy of existing photovoltaic power prediction models and the problem when the existing deep learning network model is applied to photovoltaic power prediction. There is a problem of low computational efficiency.

An embodiment of the present invention provides a training method for a photovoltaic power prediction model. The method includes the following steps:

Based on meteorological changes and photovoltaic output fluctuations, the daily weather in the area where the photovoltaic farm is to be predicted is classified into stable weather and transitional weather during the historical period;

Based on generalized weather types, the weather in each historical period of stable weather days and turning weather days is classified into multiple sub-weather types;

Preprocess the historical weather data and historical photovoltaic data for the historical period corresponding to each sub-weather type respectively, and construct a training data set for each sub-weather type;

The pre-established photovoltaic power prediction model is trained through the training data sets of each sub-weather type respectively to obtain the sub-photovoltaic power prediction model corresponding to each sub-weather type.

Based on the further improvement of the above method, the variance cost function and the cross-entropy cost function are used to construct the loss function of the photovoltaic power prediction model,

The variance cost function is: loss_1=(yy ₀ ) ² /2;

The cross entropy cost function is: loss_2=-[yln(y ₀ )+(1-y ₀ )ln(1-y ₀ )];

The loss function is: Loss=q ₁ loss_1+q ₂ loss_2;

In the formula, y is the output value of the photovoltaic power prediction model, y ₀ is the actual measured value of photovoltaic power, Loss is the total loss, loss_1 is the variance loss, loss_2 is the cross entropy loss, q ₁ and q ₂ are the weight coefficients.

Based on further improvement of the above method, the photovoltaic power prediction model is a lightweight convolutional neural network model, and the lightweight convolutional neural network model includes:

The first convolutional layer is used to extract shallow features of the input data and output shallow feature data;

Maximum pooling layer, used to reduce the dimensionality of shallow feature data;

Multiple shuffleNet modules are used to extract deep semantic features from the reduced-dimensional shallow feature data using point-wise group convolution and channel shuffling;

The second convolutional layer is used to reduce the channel dimension of deep semantic features and output prediction results.

Based on further improvement of the above method, in the training data set, the historical weather data is the input data of the photovoltaic power prediction model, and the historical photovoltaic data is the output correction data of the photovoltaic power prediction model.

Based on the further improvement of the above method, the historical weather data and historical photovoltaic data of the historical period corresponding to each sub-weather type are preprocessed respectively to construct a training data set for each sub-weather type, including:

Perform correlation analysis on historical weather data and historical photovoltaic power data for the historical period corresponding to the same sub-weather type, and select at least two types of meteorological characteristic data with high correlation from the historical weather data to construct the input data sequence of the sub-weather type. .

Based on further improvements of the above method, the method also includes:

Establish a first weather identification model, train the first weather identification model through historical weather data of stable weather days and turning weather days, and obtain the trained first weather identification model, which is used to identify whether the weather type on the predicted day is turning weather or turning weather. Smooth weather;

Based on further improvements of the above method, the method also includes:

Establish a first sub-weather identification model, train the first sub-weather identification model through historical weather data of historical periods corresponding to each sub-weather type on stable weather days, and obtain the trained first sub-weather identification model for identifying weather The sub-weather types of each forecast period for forecast days with stationary weather type;

Establish a second sub-weather identification model, train the second sub-weather identification model through the historical weather data of the historical period corresponding to each sub-weather type on the turning weather day, and obtain the trained second sub-weather identification model for identifying weather The sub-weather types of each forecast period for forecast days with type turning weather.

Based on the further improvement of the above method, based on meteorological changes and photovoltaic output fluctuations, the daily weather in the historical period where the photovoltaic farm is to be predicted is classified into stable weather and transitional weather, including:

Obtain power change rate data based on historical photovoltaic power data of the photovoltaic electric field to be predicted;

The power change rate data is clustered and divided into at least two categories based on the SOM network, and the category with the largest power fluctuation is selected and used as a potential turning weather sample;

Calculate the clarity index of the corresponding period of the potential turning weather sample based on historical solar radiation data;

The weather type on the day of the potential turning weather sample with a clarity index less than the clarity threshold is identified as turning weather, and the weather type on other days in the historical period is identified as stationary weather.

Based on the further improvement of the above method, the power change rate is the difference between the hourly power, calculated according to the following formula:

ΔP=P _i+1 -P _i ;

In the formula, ΔP is the hourly power change rate, Pi ₊₁ is the power value at time i+1 every day, and _Pi is the power value at time i every day.

Based on the further improvement of the above method, based on generalized weather types, the K-means clustering method is used to classify the weather in each historical period of stable weather days and turning weather days, and divide it into multiple sub-weather types.

Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

1. In the present invention, considering that turning weather with large changes in the weather during the day will cause large fluctuations in photovoltaic power, which will have a greater impact on photovoltaic output, firstly, based on weather changes and photovoltaic output fluctuations, the daily data in the historical period will be calculated. The weather is divided into two categories: stable weather and transitional weather. Then, the weather in each historical period of the day is subdivided based on the generalized weather type, and multiple sub-weather types under stable weather days and transitional weather days are obtained to realize the weather analysis. Fine division of types, and then the photovoltaic power prediction model is trained through the preprocessed historical period data corresponding to each sub-weather type, so that the trained photovoltaic power prediction model has high prediction accuracy, and can then obtain More accurate photovoltaic power prediction results.

2. In the present invention, the sub-photovoltaic power prediction model is a lightweight convolutional neural network model. By improving the model structure, the model can reduce the amount of calculation and improve the prediction efficiency while ensuring the prediction accuracy. Among them, a self-attention module is added to Extract more accurate features and improve the shuffleNet unit to ensure that there is little loss in accuracy while effectively reducing the computational load of the network.

3. In the present invention, the power change rate is used to represent the historical photovoltaic power data fluctuations, and then the power change rate data is clustered for primary screening, and then combined with the astronomical and meteorological factors-clarity index for secondary screening, so as to more accurately classify the occurrence of photovoltaic power data. The day of turning weather with larger power fluctuations can be identified to achieve effective classification of weather types and ensure the accuracy of photovoltaic power prediction results.

4. In the present invention, it is considered that the weather will change at each time period of the day, especially the weather changes greatly at different time periods during the turning weather day, which will have a greater impact on photovoltaic output. Therefore, in the present invention, based on the generalized weather type , the K-means clustering method is used to classify the weather in each historical period of stable weather days and turning weather days into multiple sub-weather types to ensure that the weather types in each prediction period of the prediction day can be accurately identified.

In the present invention, the above technical solutions can also be combined with each other to achieve more preferred combination solutions. Additional features and advantages of the invention will be set forth in the description which follows, and in part, some advantages will be apparent from the description, or may be learned by practice of the invention. The objectives and other advantages of the invention may be realized and obtained by the disclosure particularly pointed out in the description and drawings.

Description of the drawings

The drawings are for the purpose of illustrating specific embodiments only and are not to be construed as limitations of the invention. Throughout the drawings, the same reference characters represent the same components.

Figure 1 is a flow chart of a model training method for photovoltaic power prediction according to an embodiment of the present invention;

Figure 2 is a flow chart for classifying daily weather in a historical period in step 1 of the embodiment of the present invention.

Detailed ways

The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings. The drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

The present invention provides a training method for a photovoltaic power prediction model, as shown in Figure 1. The method includes the following steps:

Step 1: Based on meteorological changes and photovoltaic output fluctuations, classify the daily weather in the historical period where the photovoltaic farm is to be forecast, and divide it into stable weather and transitional weather;

Step 2: Based on generalized weather types, classify the weather in each historical period of stable weather days and turning weather days, and divide it into multiple sub-weather types;

Step 3: Preprocess the historical weather data and historical photovoltaic data for the historical period corresponding to each sub-weather type respectively, and construct a training data set for each sub-weather type;

Step 4: Train the pre-established photovoltaic power prediction model through the training data sets of each sub-weather type respectively to obtain the sub-photovoltaic power prediction model corresponding to each sub-weather type.

During implementation, the corresponding trained sub-photovoltaic power prediction model is selected according to the sub-weather type of each prediction period on the prediction day to obtain the photovoltaic predicted power.

Compared with the existing technology, in the present invention, considering that turning weather with large changes in the weather during the day will cause large fluctuations in photovoltaic power, which will have a greater impact on photovoltaic output, therefore, first, based on weather changes and photovoltaic output fluctuations, the The daily weather in the historical period is divided into two categories: stable weather and transitional weather. Then the weather in each historical period of the day is subdivided based on the generalized weather type to obtain multiple sub-weather types for stable weather days and transitional weather days. , achieves a refined division of weather types, and then trains the photovoltaic power prediction model through the preprocessed historical period data corresponding to each sub-weather type, so that the trained photovoltaic power prediction model has a higher prediction Accuracy, thereby enabling more accurate photovoltaic power prediction results.

It should be noted that in the embodiment of the present invention, turning weather refers to weather in which the weather changes and the weather changes can cause large fluctuations in photovoltaic power, excluding weather in which the weather changes but does not cause large fluctuations in photovoltaic power.

The above-mentioned historical period is generally selected over a longer period of time, such as one year or more, so as to obtain more training samples of changing weather, which is beneficial to improving prediction accuracy.

In one embodiment, as shown in Figure 2, in step 1, based on meteorological changes and photovoltaic output fluctuations, the daily weather in the historical period in the area where the photovoltaic farm is to be predicted is classified into stable weather and transitional weather. Weather, including:

Step 11: Obtain power change rate data based on historical photovoltaic power data of the photovoltaic electric field to be predicted;

Step 12: Cluster the power change rate data based on the SOM network and divide it into at least two categories, and filter out the category with the largest power fluctuation and use it as a potential turning weather sample;

Step 13: Calculate the clarity index of the corresponding period of the potential turning weather sample based on historical solar radiation data;

Step 14: Identify the weather type on the day of the potential turning weather sample whose clarity index is less than the clarity threshold as turning weather, and the weather type on other dates in the historical period as stationary weather.

The weather type classification method of the embodiment of the present invention is based on historical photovoltaic power data and historical solar radiation data of the same period. During implementation, considering that drastic changes in meteorological conditions are usually accompanied by large fluctuations in photovoltaic output, the power changes are first used The rate meets the fluctuation identification amplitude numerically, but non-weather factors can also cause power fluctuations. Especially at sunrise or sunset on a sunny day, the photovoltaic power will fluctuate greatly due to the influence of the sun's altitude angle. Therefore, when identifying turning weather, it is not only necessary to pay attention to the power The characteristic of fluctuations also requires the introduction of physical quantities that can reflect the fluctuations in illumination radiation caused by weather changes. Among them, the clarity index can reflect the transparency of the atmosphere and is closely related to weather conditions and solar radiation. Therefore, in the embodiment of the present invention, The clarity index is used for secondary screening to filter out the sunny data in the potential turning weather samples clustered according to the power change rate, and only retain the data of the real turning weather that will cause large power fluctuations, thereby realizing the identification of turning weather. . That is, if the photovoltaic output on a certain day can meet the fluctuation amplitude requirements of the power change rate, and at the same time the corresponding atmospheric conditions can be met on that day, the weather type on that day can be identified as turning weather with large fluctuation amplitude.

Compared with the existing technology, in the embodiment of the present invention, the power change rate is used to represent historical photovoltaic power data fluctuations, and then the power change rate data is clustered for primary screening, and then combined with astronomical and meteorological factors-clarity index for secondary screening , thereby more accurately identifying turning weather with larger power fluctuations, achieving effective classification of weather types, and ensuring the accuracy of photovoltaic power prediction results.

In addition, the historical photovoltaic power data and historical solar radiation data used in the embodiment of the present invention are data obtained from the photovoltaic electric field to be predicted and subjected to data quality inspection.

Preferably, step 11, obtaining power change rate data based on historical photovoltaic power data of the photovoltaic electric field to be predicted, includes the following steps:

Step 1101: Calculate the solar altitude angle of the target area, and use the period when the solar altitude angle is greater than the preset angle as the photovoltaic output statistical time period;

Step 1102: Calculate the power change rate of the photovoltaic output statistical period according to the historical photovoltaic output data to obtain power change rate data.

Among them, the solar altitude angle refers to the angle between the sun's rays and the normal line of the ground plane. During implementation, considering that photovoltaic output has significant time period characteristics, the initial historical photovoltaic power data is screened by calculating the solar altitude angle in the area where the photovoltaic electric field to be predicted is located, so that effective data can be better utilized.

The solar altitude angle is calculated according to the following formula:

ω=(t-12)×15°;

In the formula, α _s is the solar altitude angle, θ _z is the zenith angle, is the latitude of the target area, δ is the declination angle, ω is the hour angle, n is the serial number of the date in the year, and t is the hour.

For example, on January 1 of every year, n=1; on December 31 of a normal year, n=365; on December 31 of a leap year, n=366.

The hour angle is calculated from the Q point on the meridian circle (from the sun’s noon). The clockwise direction is positive and the counterclockwise direction is negative, that is, it is negative in the morning and positive in the afternoon. Its value is equal to the time from noon ( hours) times 15°.

In step 1101, the preset angle ranges from 5° to 15°, preferably 10°. That is, it is preferable to use the period when the solar altitude angle is greater than 10° as the photovoltaic output statistical period.

Further specifically, the power change rate is the hourly power difference, calculated according to the following formula: ΔP=P _i+1 -P _i ; where ΔP hourly power change rate, P _i+1 is daily i+ The power value at time 1, _Pi is the power value at time i every day.

During implementation, the historical photovoltaic power data are first arranged according to the time series hourly values, and then the hourly power change rate is obtained.

Step 12: Cluster the power change rate data based on the SOM network and divide it into at least two categories, and filter out the category with the largest power fluctuation as a potential turning weather sample.

Since the current standards are difficult to define a value for the size of power fluctuations, in order to screen out a type of data samples with sufficiently large negative fluctuations in the power change rate, in the embodiment of the present invention, the SOM network (Self-Organizing Mapping Neural Network) is used. Network) conducts unsupervised self-organized clustering on the power change rate data and divides it into multiple categories, thereby screening out the data sample with the largest power fluctuation and using it as a potential turning weather sample.

Among them, the SOM network is a tutor-less learning network. It automatically searches for the inherent laws and essential attributes in data samples, self-organizes and adaptively changes the network parameters and structure, and can adapt to the complex patterns of photovoltaic power data, accurately and effectively. Classification.

Further, in step 12, the category with the largest negative power fluctuation amplitude is screened out and used as a potential turning weather sample. The category with the largest negative power fluctuation amplitude is the category with a negative power change rate and the largest fluctuation amplitude.

Among them, the category with the largest negative power fluctuations is used as a potential turning weather sample because the positive power fluctuations with large amplitudes mostly occur when the weather changes from cloudy or rainy and snowy weather to sunny days, while clarity is used. When the index identifies turning weather, sunny data will be filtered out. Therefore, during clustering and screening, the category with the largest negative power fluctuations is directly selected as a potential turning weather sample, which is beneficial to improving processing efficiency.

At the same time, the photovoltaic power decline caused by the instability of meteorological factors during photovoltaic output fluctuations has the greatest impact on the stability and dispatch of the power grid. Moreover, photovoltaic power has a significant time periodicity with the change of the sun's altitude angle throughout the day, and the amplitude appears larger during the day. Large positive power fluctuations will also be accompanied by large negative power fluctuations. Therefore, during the clustering screening, the category with the largest positive power fluctuations was not selected and used as a potential turning weather sample, which would not be used to predict turning weather. The identification results and photovoltaic predicted power results have a greater impact.

Specifically, the number of clusters of the SOM network is set to 5. In this way, in step 12, the power change rate data is clustered based on the SOM network and divided into five categories. Among them, the category with the largest negative power fluctuations is separated from the other four categories. The other four categories mainly include sunny days, rainy days, power Types such as sunny to overcast with little fluctuation, and the type with the largest negative power fluctuations also include sunny days and sudden changes in weather.

Step 13: Calculate the clarity index of the day of the potential turning weather sample based on the historical solar radiation data of the same period.

Specifically, the clarity index on the day of the potential turning weather sample is calculated according to the following formula: In the formula, k _T is the clarity index, I is the total radiation, and I ₀ is the total solar radiation on the extraterrestrial horizontal plane.

The total solar radiation on the extraterrestrial horizontal plane is calculated according to the following formula:

In the formula, E _sc is the solar constant, γ is the correction value of the solar radiation flux at the upper boundary of the atmosphere caused by changes in the distance between the sun and the earth, and ω _s is the hour angle from sunrise to sunset.

Step 14: Identify the weather type on the day of the potential turning weather sample whose clarity index is less than the clarity threshold as turning weather.

In step 14, the weather type at the time point where the potential turning weather sample is located is identified through the clarity index. Wherein, the definition threshold ranges from 0.1 to 0.3, preferably 0.2.

Generally speaking, the corresponding weather when k _T <0.2 is light rain, showers, light snow, light fog, haze, moderate rain and above, moderate snow and above, etc.

In one embodiment, in step 2, based on generalized weather types, the K-means clustering method is used to classify the weather in each historical period of stable weather days and turning weather days, and divide it into multiple sub-weather types.

Wherein, the generalized weather types include: sunny, cloudy, cloudy, rain and snow.

In the embodiment of the present invention, it is considered that the weather will change at various times of the day, especially the weather changes greatly at different times of the day during turning weather, which will have a greater impact on photovoltaic output. Therefore, in the present invention, based on the generalized weather type , the K-means clustering method is used to classify the weather in each historical period of stable weather days and turning weather days into multiple sub-weather types to ensure that the weather types in each prediction period of the prediction day can be accurately identified.

During implementation, photovoltaic output is cyclically related to solar irradiation. Therefore, the statistical period for photovoltaic output and corresponding weather data is generally from 8:00 to 18:00 every day. Divide the daily statistical period of the historical period into multiple historical periods according to preset time intervals, for example, each historical period is 1 hour. Then, obtain the historical weather data for each historical period of the stable weather day, and divide the above data into four categories through the K-means clustering method. The corresponding four sub-weather types are stable sunny weather, stable cloudy weather, stable cloudy weather, Smooth rain and snow weather. Similarly, obtain the historical weather data for each historical period of the turning weather day, and divide the above data into four categories through the K-means clustering method. The corresponding four sub-weather types are turning sunny weather, turning cloudy weather, and turning cloudy weather. , turning rain and snow weather.

In one embodiment, step 3, preprocessing historical weather data and historical photovoltaic data for the historical period corresponding to each sub-weather type, and constructing a training data set for each sub-weather type, includes: Correlation analysis is performed on the historical weather data and historical photovoltaic power data of the corresponding historical period, and at least two types of meteorological characteristic data with high correlation are selected from the historical weather data to construct the input data sequence of the sub-weather type.

In this embodiment, after the weather types of each historical period of each day in the historical period are divided into multiple sub-weather types, correlation analysis is performed on the historical weather data and historical photovoltaic power data of the historical period corresponding to the same sub-weather type, and the filtered out Meteorological characteristic data that are highly correlated with photovoltaic output under each sub-weather type is used for model training, which can not only reduce the sample size but also ensure prediction accuracy.

Specifically, the input data sequences of the four sub-weather types corresponding to the stable weather days include five types of meteorological characteristic data, including irradiance, humidity, temperature, cloud cover and visibility, in order from high to low in correlation;

The input data sequence of the four sub-weather types corresponding to the turning weather day includes five types of meteorological characteristic data, including irradiance, humidity, visibility, cloud cover and temperature, in order from high to low in correlation.

It should be noted that the historical weather data and historical photovoltaic power data of each historical period in the historical period are used as a training sample, that is, the training data set of each sub-weather type includes one or more historical periods corresponding to the sub-weather type. Historical weather data and historical photovoltaic power data. In addition, in each training sample, weather data is the input data, and photovoltaic power data is the output correction data (label).

In one embodiment, the photovoltaic power prediction model in step 4 is a lightweight convolutional neural network model, and the lightweight convolutional neural network model includes:

The first convolutional layer is used to extract shallow features of the input data and output shallow feature data;

Maximum pooling layer, used to reduce the dimensionality of shallow feature data;

Multiple shuffleNet modules are used to extract deep semantic features from the reduced-dimensional shallow feature data using point-wise group convolution and channel shuffling;

The second convolutional layer is used to reduce the channel dimension of deep semantic features and output prediction results.

In the present invention, in order to improve training efficiency, a lightweight convolutional neural network model is constructed as a photovoltaic power prediction model. Wherein, the lightweight convolutional neural network model is an improved ShuffleNet model.

Specifically, the first convolutional layer is used to extract shallow features of the input data, reduce the data dimension, and increase the number of feature channels.

The dimensionally reduced feature data is passed through multiple ShuffleNet modules to extract deep semantic features. Each module continues to reduce the dimensionality of feature data while increasing the number of feature channels, and gradually extracts deep semantic features of edge points. ShuffleNet uses point-wise group convolution and channel shuffling to effectively reduce the computational complexity of the network while ensuring little loss in accuracy, greatly reducing computational costs.

Specifically, in order to enhance the feature extraction capability of the network, each ShuffleNet module includes a ShuffleNet unit and a channel attention module connected in sequence.

A channel attention module is added after each ShuffleNet unit structure to further extract more accurate features. The channel attention module consists of global average pooling, 2 fully connected layers, a ReLU layer and a Sigmoid activation layer. It generates corresponding weights for each channel, allowing the network to make independent selections based on the importance of each feature channel. The features output by the ShuffleNet unit are multiplied by the weights calculated by the channel attention module to obtain features that increase the attention weight.

The improved ShuffleNet model of this application changes the global average pooling and fully connected layer at the end of the existing ShuffleNet model into one convolution layer, that is, the second convolution layer. The convolution kernel size of this convolution layer is 1×1 , the number is 2, and the step size is 1.

Specifically, the ShuffleNet unit includes a convolution layer, a channel shuffle layer (ChannelShuffle), a depth convolution layer, a convolution layer, and a fusion layer connected in sequence. In the ShuffleNet unit, the two convolutional layers are both 1×1 convolutional layers, and the depth convolutional layer is a 3×3 depth convolutional layer.

By constructing the above lightweight neural network structure, the model can reduce the computational complexity and improve detection efficiency while ensuring prediction accuracy.

After building a lightweight convolutional neural network model, the model is trained based on the classified training sample data. During the network training process, the gradient descent method is used to backpropagate the parameters to update and optimize the model parameters, thereby obtaining the corresponding eight sub-weather types. The trained sub-sub photovoltaic power prediction model.

Specifically, the variance cost function and the cross-entropy cost function are used to construct the loss function Loss of the model. When the model reaches the required accuracy or reaches the predetermined number of iterations, the training is stopped and the trained sub-photovoltaic power prediction model is obtained.

The variance cost function is:

The cross entropy cost function is: loss_2=-[yln(y ₀ )+(1-y ₀ )ln(1-y ₀ )];

The loss function is: Loss=q ₁ loss_1+q ₂ loss_2;

In the formula, y is the output value of the photovoltaic power prediction model, y ₀ is the actual measured value (target value) of photovoltaic power; in the formula, q ₁ and q ₂ are weight coefficients, which are set according to the actual situation.

The method of the embodiment of the present invention also includes: after obtaining the trained photovoltaic power prediction model, using the root mean square error and the absolute value error to evaluate the accuracy of the model prediction results.

Specifically, the calculation formula of the root mean square error e _RMSE is:

The calculation formula of absolute value error e _MAE is:

In the formula: P _{t, pre} is the predicted value of photovoltaic power at time t; P _t is the actual measured value of photovoltaic power at time t; N is the prediction period; C _ap is the rated power of photovoltaic output of the power station.

After training the above photovoltaic power prediction model, the model can be used to predict photovoltaic power. When predicting photovoltaic power, first determine whether the weather type on the forecast day is stable weather or turning weather, and then further determine the weather type in each forecast period on the forecast day.

Specifically, the following method is used to determine the sub-weather type corresponding to each prediction period on the prediction day:

Step 501: Establish a first weather identification model, and train the first weather identification model through historical weather data of stable weather days and turning weather days;

Step 502: Establish a first sub-weather identification model, and train the first sub-weather identification model through historical weather data of historical periods corresponding to each sub-weather type on stable weather days;

Step 503: Establish a second sub-weather identification model, and train the second sub-weather identification model through the historical weather data of the historical period corresponding to each sub-weather type of the turning weather day;

Step 504: Enter the weather forecast data on the forecast day into the trained first weather identification model to obtain the weather type on the forecast day, and then select the corresponding trained sub-weather identification model based on the weather type on the forecast day, and add the weather forecast data on the forecast day. The weather type of each prediction period is input into the selected trained sub-weather identification model, and the weather type of each prediction period of the prediction day is obtained.

In the embodiment of the present invention, the first weather identification model used to identify the weather type on the predicted day and the first sub-weather identification model and the second sub-weather model used to identify each prediction period of the predicted day are all accurately classified. It is trained with historical weather data, so that it can effectively identify whether the forecast day is a turning weather day with large power fluctuations, and more accurately identify the weather types in each forecast period of the forecast day, which is conducive to improving Accuracy of photovoltaic power prediction results.

Specifically, the first weather identification model takes the weather data of a certain day as input and uses turning weather or stable weather as output. The first sub-weather identification model takes weather data for a certain period of time in a stable weather day as input, and uses four sub-weather types as stable sunny weather, stable cloudy weather, stable cloudy weather, and stable rainy and snowy weather as output; the second The sub-weather identification model takes the weather data of a certain period of time in a turning weather day as input, and uses four sub-weather types as turning sunny weather, turning cloudy weather, turning cloudy weather, and turning rainy and snowy weather as output.

The above-mentioned first weather recognition model, first sub-weather recognition model and second sub-weather recognition model are all deep learning models, such as kohonen model, neural network model, etc.

Those skilled in the art can understand that all or part of the process of implementing the method of the above embodiments can be completed by instructing relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. Wherein, the computer-readable storage medium is a magnetic disk, an optical disk, a read-only memory or a random access memory, etc.

The above are only preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person familiar with the technical field can easily think of changes or modifications within the technical scope disclosed in the present invention. All substitutions are within the scope of the present invention.

Claims (10)

1. A training method for a photovoltaic power prediction model, characterized in that the method includes the following steps: Based on meteorological changes and photovoltaic output fluctuations, the daily weather in the area where the photovoltaic farm is to be predicted is classified into stable weather and transitional weather during the historical period; Based on generalized weather types, the weather in each historical period of stable weather days and turning weather days is classified into multiple sub-weather types; Preprocess the historical weather data and historical photovoltaic data for the historical period corresponding to each sub-weather type respectively, and construct a training data set for each sub-weather type; The pre-established photovoltaic power prediction model is trained through the training data sets of each sub-weather type respectively to obtain the sub-photovoltaic power prediction model corresponding to each sub-weather type. 2. The method according to claim 1, characterized in that a variance cost function and a cross-entropy cost function are used to construct the loss function of the photovoltaic power prediction model, The variance cost function is: loss_1=(yy ₀ ) ² /2; The cross entropy cost function is: loss_2=-[yln(y ₀ )+(1-y ₀ )ln(1-y ₀ )]; The loss function is: Loss=q ₁ loss_1+q ₂ loss_2; In the formula, y is the output value of the photovoltaic power prediction model, y ₀ is the actual measured value of photovoltaic power, Loss is the total loss, loss_1 is the variance loss, loss_2 is the cross entropy loss, q ₁ and q ₂ are the weight coefficients. 3. The method according to claim 1 or 2, characterized in that the photovoltaic power prediction model is a lightweight convolutional neural network model, and the lightweight convolutional neural network model includes: The first convolutional layer is used to extract shallow features of the input data and output shallow feature data; Maximum pooling layer, used to reduce the dimensionality of shallow feature data; Multiple shuffleNet modules are used to extract deep semantic features from the reduced-dimensional shallow feature data using point-wise group convolution and channel shuffling; The second convolutional layer is used to reduce the channel dimension of deep semantic features and output prediction results. 4. The method according to claim 3, characterized in that the ShuffleNet module includes a ShuffleNet unit and a channel attention module connected in sequence. 5. The method according to claim 4, wherein the ShuffleNet unit includes a convolution layer, a channel shuffling layer, a depth convolution layer, a convolution layer and a fusion layer connected in sequence. 6. The method according to claim 1 or 2, characterized in that, in the training data set, the historical weather data is the input data of the photovoltaic power prediction model, and the historical photovoltaic data is the output correction data of the photovoltaic power prediction model. 7. The method according to claim 1 or 2, characterized in that the historical weather data and historical photovoltaic data of the historical period corresponding to each sub-weather type are preprocessed respectively to construct a training data set for each sub-weather type. ,include: Perform correlation analysis on historical weather data and historical photovoltaic power data for the historical period corresponding to the same sub-weather type, and select at least two types of meteorological characteristic data with high correlation from the historical weather data to construct the input data sequence of the sub-weather type. . 8. The method according to claim 1 or 2, characterized in that based on meteorological changes and photovoltaic output fluctuations, the daily weather in the historical period of the area where the photovoltaic field to be predicted is located is classified into stable weather and Turning weather, including: Obtain power change rate data based on historical photovoltaic power data of the photovoltaic electric field to be predicted; The power change rate data is clustered and divided into at least two categories based on the SOM network, and the category with the largest power fluctuation is screened out and used as a potential turning weather sample; Calculate the clarity index of the corresponding period of the potential turning weather sample based on historical solar radiation data; The weather type on the day of the potential turning weather sample with a clarity index less than the clarity threshold is identified as turning weather, and the weather type on other days in the historical period is identified as stationary weather. 9. The method according to claim 8, characterized in that the power change rate is the difference between the hourly power, calculated according to the following formula: ΔP=P _i+1 -P _i ; In the formula, ΔP is the hourly power change rate, Pi ₊₁ is the power value at time i+1 every day, and _Pi is the power value at time i every day. 10. The method according to claim 1 or 2, characterized in that, based on the generalized weather type, the K-means clustering method is used to classify the weather in each historical period of the stable weather day and the turning weather day, and divide it into a plurality of sub-sections. Weather type.