CN116187498A - A photovoltaic power prediction method based on frequency domain decomposition - Google Patents

Abstract

The invention discloses a photovoltaic power generation power prediction method based on frequency domain decomposition, which relates to the technical field of photovoltaic power generation, and comprises the following steps: collecting historical data of a photovoltaic power plant, constructing a multi-dimensional time sequence data sample set, cleaning the multi-dimensional time sequence data sample set by using a quarter-point intra-distance algorithm, and separating the multi-dimensional time sequence data sample set into a sunlight intensity data set and a power sequence data set; constructing a neural network model based on the frequency domain decomposition, the neural network model comprising: the system comprises an illumination prediction module, a power prediction module and a fusion module; using a sunlight intensity data set and a power sequence data set as a training neural network model, wherein the sunlight intensity data set is input into a light prediction module, the power sequence data set is input into a power prediction module, and output results of the light prediction module and the power prediction module are input into a fusion module together for processing; and the data is processed by the fusion module and then the result is output as the predicted photovoltaic power generation power of the photovoltaic power plant.

Description

A Photovoltaic Power Prediction Method Based on Frequency Domain Decomposition

technical field

The present invention relates to the technical field of photovoltaic power generation, in particular to a method for predicting photovoltaic power generation power based on frequency domain decomposition.

Background technique

Photovoltaic power generation is a power generation method that uses solar panels to receive solar radiation and convert it into electrical energy. Because it needs to rely on sunlight to obtain energy, the power generation is affected by factors such as seasons, cloudy and sunny, and day and night. The output power of photovoltaic power generation systems Exhibits randomness and intermittency. In addition, due to the difficulty of energy storage, when most of the electricity in the grid comes from photovoltaic power generation, the stability of the power system will be affected, which limits the development of the photovoltaic industry. Therefore, forecasting the power of photovoltaic power generation has become a concern of people.

There are two main ideas for the prediction of photovoltaic power generation data: the use of statistical methods and the use of physical methods. Among them, the statistical method is to conduct statistical analysis on historical data to find out its internal laws for prediction; the physical method is to use meteorological parameters as input values and solve the predicted power through a physical model. Both have their own limitations: the statistical method only emphasizes the regularity of the data, and cannot respond to sudden meteorological changes in a timely manner; while the physical method cannot take into account both the cycle and the law, and the prediction results are often not accurate enough.

With the rapid development of artificial intelligence methods, deep learning-based models have been developed and applied in many fields. Deep learning is a new branch of machine learning methods. Using deep learning-based models, such as networks such as RNN or LSTM, to predict photovoltaic power generation has received attention. Compared with traditional physical and statistical methods, the deep learning model can mine deep features from photovoltaic power sequences and obtain more accurate prediction results. However, due to the inherent defects of the RNN network, it cannot handle long-term/power sequences. LSTM can only predict the historical law of power, and cannot take into account sudden meteorological conditions.

Contents of the invention

The purpose of the present invention is to propose a method for forecasting photovoltaic power generation that can take into account both historical data and weather conditions.

The technical solution of the present invention is to provide a method for predicting photovoltaic power generation based on frequency domain decomposition, the method comprising:

S1. Collect the historical data of photovoltaic power plants to construct a multi-dimensional time-series data sample set, use the interquartile point distance algorithm to clean the multi-dimensional time-series data sample set, and separate it into a sunshine intensity data set and a power sequence data set;

S2. Construct a neural network model based on frequency domain decomposition. The neural network model includes: an illumination prediction module, a power prediction module, and a fusion module; use the sunshine intensity data set and the power sequence data set as training sets to train the neural network model, wherein the sunshine intensity data set Input the illumination prediction module, the power sequence data set is input to the power prediction module, and the output results of the illumination prediction module and the power prediction module are input into the fusion module for processing;

S3. After the data is processed by the fusion module, the result is output as the predicted photovoltaic power generation power of the photovoltaic power plant;

Among them, the structure of the illumination prediction module and the power prediction module in step S2 are the same, both are two-layer encoder and one-layer decoder, the data is input to the decoder after two rounds of encoding, and the initialization sequence input into the decoder is processed into the final The results are output to the fusion module.

In any one of the above technical solutions, further, the encoder includes: a frequency learning module, a cycle-trend decomposition module and a forward propagation module, the residual connections between each module, after the data is input into the encoder, it is sequentially processed by the above modules and then output .

In any one of the above technical solutions, further, the encoder processing process includes:

The data first enters the frequency learning module, and the time series data is obtained after processing. The time series data and the unprocessed data are connected by residuals, and then sent to the cycle-trend decomposition module. The cycle-trend decomposition module decomposes the input data and obtains the time series The periodic component and trend component above, discarding the trend component and then outputting the periodic component, also make a residual link with the signal input to this module, and send it to the forward propagation module; after the data output by the forward propagation module is fused with the input data, Then enter a period-trend decomposition module, discard the trend component, and finally output the period component as the result of the encoder.

In any of the above technical solutions, further, the two-layer encoder is divided into a first-layer encoder and a second-layer encoder, both of which have exactly the same structure, and the result processed by the two-layer encoder is decomposed into value and key input decoding device.

In any one of the above technical solutions, further, the decoder includes: a frequency learning module, a cycle-trend decomposition module, a frequency domain attention module and a forward propagation module, and the residual connection between each module, after the data is input into the decoder, sequentially Output after processing by the above modules.

In any one of the above technical solutions, further, the decoder processing process includes:

First initialize a periodic component and a trend component, input the initialized periodic component into the frequency learning module, make a residual link between the results before and after the transformation, and then send the connection result to the cycle-trend decomposition module, and the trend component output by the cycle-trend decomposition module Make a residual link with the initialization trend, and the result is recorded as A; the periodic component output by the period-trend decomposition module is sent to the frequency domain attention module together with the sequence output by the encoder;

The frequency-domain attention module links the processed results with the output of the cycle-trend decomposition module as a residual, and sends them to the next cycle-trend decomposition module; the cycle-trend decomposition module decomposes it into period volume and trend volume, The period quantity enters the forward propagation module, the trend quantity and the result A are connected by residual, and the result is recorded as B;

The results of the input and output of the forward propagation module are connected by residuals and then sent to the next cycle-trend decomposition module. The cycle-trend decomposition module decomposes it into periodic quantities and trend quantities, and connects the trend quantities with B as residuals. The result is recorded as C;

Finally, the periodic quantity decomposed by the cycle-trend decomposition module is fused with the trend quantity C, and the result is output to the fusion module.

In any of the above technical solutions, further, the Gaussian window processing is specifically:

Among them, n is the serial number of discrete sequence data points, and M is the width of the window.

In any of the above technical solutions, further, the weighted average formula is:

Among them, L(n) is the output value, that is, the value put into the fully connected layer, and l(n) is the original sequence data point.

The beneficial effects of the present invention are:

The technical solution in the present invention aims at the periodicity and instability of photovoltaic power generation. Two identical network modules are designed to predict sunshine and power respectively, and then integrate the prediction results to obtain the prediction result of photovoltaic power generation. This method not only It can extract the timing characteristics of photovoltaic power generation, and can also take into account the impact of meteorological conditions on photovoltaic power generation, making up for the shortcomings of statistical methods and physical methods, thereby improving the accuracy of forecast values;

By introducing a frequency-domain attention module in the decoder, the time-domain data is transferred to the frequency domain. For time-domain data with obvious periodicity such as illumination information and photovoltaic power generation, the frequency domain often contains more intuitive information. The learning of frequency domain information by the model can predict the results more accurately.

Description of drawings

The advantages of the above and additional aspects of the present invention will become apparent and readily understood from the following description of embodiments when taken in conjunction with the accompanying drawings, in which:

Fig. 1 is a schematic flow chart of a method for predicting photovoltaic power generation based on frequency domain decomposition according to an embodiment of the present invention;

Fig. 2 is a structural diagram of an illumination prediction module or a power prediction module of a photovoltaic power prediction method based on frequency domain decomposition according to an embodiment of the present invention;

Fig. 3 is the coder structural diagram of the photovoltaic generation power prediction method based on frequency domain decomposition according to an embodiment of the present invention;

Fig. 4 is the structural diagram of the frequency learning module of the coder of the photovoltaic power generation power prediction method based on frequency domain decomposition according to an embodiment of the present invention;

FIG. 5 is a structural diagram of a decoder for a photovoltaic power prediction method based on frequency domain decomposition according to an embodiment of the present invention;

Fig. 6 is a structural diagram of a frequency domain attention module of a decoder of a photovoltaic power generation power prediction method based on frequency domain decomposition according to an embodiment of the present invention;

Fig. 7 is a fusion module processing flow chart of a photovoltaic power prediction method based on frequency domain decomposition according to an embodiment of the present invention;

Fig. 8 is a comparison curve between predicted power generation and actual power generation of a photovoltaic power prediction method based on frequency domain decomposition according to an embodiment of the present invention.

Detailed ways

In order to understand the above-mentioned purpose, features and advantages of the present invention more clearly, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the case of no conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other.

In the following description, many specific details are set forth in order to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described here. Therefore, the protection scope of the present invention is not limited by the following disclosure. Limitations of specific embodiments.

As shown in Figure 1, this embodiment provides a method for predicting photovoltaic power generation based on frequency domain decomposition, which includes:

S1. Collect the historical data of photovoltaic power plants to construct a multi-dimensional time-series data sample set, use the interquartile distance algorithm to clean the multi-dimensional time-series data sample set, and separate it into a sunshine intensity data set and a power sequence data set.

Specifically, the collected historical data of the photovoltaic power plant includes time, illumination amplitude, temperature, air pressure and actual power, and the above data at the same moment are saved as a multi-dimensional time series data sample as a local file. The frequency of sampling has accumulated 5,000 multi-dimensional time-series data samples, which cover photovoltaic power generation data affected by various scene factors, which is more comprehensive.

The interquartile distance algorithm is used to detect abnormal data, and after cleaning the data set, the data is reconstructed and standardized, and separated into sunshine intensity data set and power sequence data set according to the type.

S2. Construct a neural network model based on frequency domain decomposition. The neural network model includes: an illumination prediction module, a power prediction module, and a fusion module; use the sunshine intensity data set and the power sequence data set as training sets to train the neural network model, wherein the sunshine intensity data set Input the illumination prediction module, the power sequence data set is input to the power prediction module, and the output results of the illumination prediction module and the power prediction module are input into the fusion module for processing.

As shown in Figure 2, the illumination prediction module and the power prediction module have the same structure, both of which are two-layer encoder and one-layer decoder. After two rounds of encoding, the data is input to the decoder, and the initialization sequence will be additionally input into the decoder. The final result is processed and output to the fusion module.

As shown in Figure 3, the encoder includes: a frequency learning module, a cycle-trend decomposition module, and a forward propagation module. The residual connection between each module prevents the gradient from exploding or disappearing. After the data is input into the encoder, it is processed by the above modules in turn and then output .

再将其与参数矩阵R做线性运算，得到的结果定义为/>

接着对/>

进行频率补全后，做傅里叶逆变换得到处理完成的时序数据y并输出。As shown in Figure 4, after the data enters the frequency learning module, it is linearly transformed first, and the result is denoted as q, and the discrete Fourier transform is performed on q, and the result is denoted as Q, and the result obtained after downsampling Q is denoted as

Then perform a linear operation with the parameter matrix R, and the obtained result is defined as />

Next to />

After frequency completion, inverse Fourier transform is performed to obtain the processed time series data y and output it.

Then, the data processed by the frequency learning module and the unprocessed data are residually connected and sent to the cycle-trend decomposition module. The cycle-trend decomposition module decomposes the input data and obtains the periodic components and trends in time series. Component, the periodic component is output after discarding the trend component, which is also connected with the signal input to this module by residual link, and sent to the forward propagation module.

The forward propagation module is a fully connected network whose parameters are used as training parameters. After the data output by the forward propagation module is fused with the input data, it enters a cycle-trend decomposition module. After discarding the trend component, the cycle component is finally input to the second-layer encoder as a result, and the second-layer encoder and the first layer The structure of one-layer encoder is exactly the same, and the result processed by two-layer encoder is decomposed into value and key input decoder.

As shown in Figure 5, the decoder consists of: frequency learning module, cycle-trend decomposition module, frequency domain attention module and forward propagation module.

In the decoder, first initialize a period component and a trend component, input the initialized period component into the frequency learning module, make the residual link of the results before and after the transformation, and then send the connection result to the period-trend decomposition module, period-trend decomposition The trend component output by the module is residually linked with the initialization trend, and the result is denoted as A; the period component output by the period-trend decomposition module is sent to the frequency domain attention module together with the sequence output by the encoder.

The frequency-domain attention module links the processed results with the output of the cycle-trend decomposition module as a residual, and sends them to the next cycle-trend decomposition module; the cycle-trend decomposition module decomposes it into period volume and trend volume, The period quantity enters the forward propagation module, the trend quantity and the result A are connected by residuals, and the result is recorded as B.

The forward propagation module uses the same fully connected layer as the encoder, and the results of the input and output are residually connected and then sent to the next cycle-trend decomposition module. The cycle-trend decomposition module decomposes it into cycle quantities and trend quantities, and Make a residual connection between the trend quantity and B, and record the result as C. Finally, the periodic quantity decomposed by the cycle-trend decomposition module is fused with the trend quantity C, and the result is output to the fusion module.

As shown in Figure 6, the value v and the key k output by the encoder are first input into the frequency domain attention module. Specifically, the input from the encoder to the frequency domain attention module of the decoder is divided into two ways, and the value v and the key k are input respectively. k; secondly, record the result of the network layer in front of the decoder as q input, perform a linear operation on q and the key k, and then perform a linear operation on the result and the value v to obtain the attention result, and then perform frequency completion and summing on the attention result Inverse Fourier transform to obtain time series data, and this result will be passed to the next layer for learning.

S3. After the data is processed by the fusion module, the result is output as the predicted photovoltaic power of the photovoltaic power plant.

As shown in Figure 7, the data output by the illumination prediction module and the power prediction module are respectively processed by a Gaussian window with a step size of 1 and a β value of 1.2 and then weighted and averaged, and the output results of both are input into the fusion module. The forward propagation module finally outputs the result data after being processed by the fully connected layer.

The Gaussian window formula is:

Among them, n is the serial number of discrete sequence data points, and M is the width of the window.

The weighted average formula is:

Among them, L(n) is the output value, that is, the value put into the fully connected layer, and l(n) is the original sequence data point.

In another embodiment of the present invention, the photovoltaic power generation trend of a village is predicted, as shown in Figure 8, the results predicted by the photovoltaic power generation prediction method provided by the present invention fit well with the actual power generation results , with high accuracy.

In summary, the present invention proposes a method for predicting photovoltaic power generation based on frequency domain decomposition, including:

S1. Collect the historical data of photovoltaic power plants to construct a multi-dimensional time-series data sample set, use the interquartile distance algorithm to clean the multi-dimensional time-series data sample set, and separate it into a sunshine intensity data set and a power sequence data set.

S2. Construct a neural network model based on frequency domain decomposition. The neural network model includes: an illumination prediction module, a power prediction module, and a fusion module; use the sunshine intensity data set and the power sequence data set as training sets to train the neural network model, wherein the sunshine intensity data set Input the illumination prediction module, the power sequence data set is input to the power prediction module, and the output results of the illumination prediction module and the power prediction module are input into the fusion module for processing.

S3. After the data is processed by the fusion module, the result is output as the predicted photovoltaic power generation power of the photovoltaic power plant.

Among them, the structure of the illumination prediction module and the power prediction module in step S2 are the same, both are two-layer encoder and one-layer decoder, the data is input to the decoder after two rounds of encoding, and the initialization sequence input into the decoder is processed into the final The results are output to the fusion module.

The steps in the present invention can be adjusted, combined and deleted according to actual needs.

Although the present invention has been disclosed in detail with reference to the accompanying drawings, it should be understood that these descriptions are illustrative only and are not intended to limit the application of the present invention. The protection scope of the present invention is defined by the appended claims, and may include various changes, modifications and equivalent solutions for the invention without departing from the protection scope and spirit of the present invention.

Claims (9)

1. A method for forecasting photovoltaic power generation based on frequency domain decomposition, characterized in that the method comprises: S1. Collect historical data of photovoltaic power plants to construct a multi-dimensional time-series data sample set, use the interquartile distance algorithm to clean the multi-dimensional time-series data sample set, and separate it into a sunshine intensity data set and a power sequence data set; S2. Constructing a neural network model based on frequency domain decomposition, the neural network model comprising: an illumination prediction module, a power prediction module, and a fusion module; using the sunlight intensity data set and the power sequence data set as a training set to train the neural network A network model, wherein the sunshine intensity data set is input into the illumination prediction module, the power sequence data set is input into the power prediction module, and the output results of the illumination prediction module and the power prediction module are input into the fusion module together deal with; S3. After the data is processed by the fusion module, the result is output as the predicted photovoltaic power generation power of the photovoltaic power plant; Wherein, the structure of the illumination prediction module and the power prediction module in step S2 is the same, both of which are two-layer encoder and one-layer decoder. Sequence processing is output to the fusion module as the final result. 2. The photovoltaic power prediction method based on frequency domain decomposition as claimed in claim 1, wherein the encoder comprises: a frequency learning module, a cycle-trend decomposition module and a forward propagation module, and residual Differential connection, after the data is input into the encoder, it is processed by the above modules in turn and then output. 3. the photovoltaic power generation power prediction method based on frequency domain decomposition as claimed in claim 2, is characterized in that, described coder process comprises: The data first enters the frequency learning module and is processed to obtain time-series data. The time-series data is residually connected with unprocessed data and sent to the cycle-trend decomposition module. The cycle-trend decomposition module decomposes the input data to obtain The periodic component and trend component in the time series are discarded, and the periodic component is output after discarding the trend component. It is also connected with the signal input to this module by residual link and sent to the forward propagation module; the data output by the forward propagation module is fused with the input data After that, it enters a period-trend decomposition module, discards the trend component, and finally outputs the period component as the result of the encoder. 4. The method for predicting photovoltaic power generation based on frequency-domain decomposition as claimed in claim 1, wherein the two-layer encoder is divided into a first-layer encoder and a second-layer encoder, both of which are identical in structure, The result processed by the two-layer encoder is decomposed into the value and key input to the decoder. 5. The method for predicting photovoltaic power generation based on frequency domain decomposition as claimed in claim 1, wherein the decoder comprises: frequency learning module, cycle-trend decomposition module, frequency domain attention module and forward propagation module , the residual connection between each module, the data is input to the decoder and then processed by the above modules in sequence and then output. 6. the photovoltaic power generation power prediction method based on frequency domain decomposition as claimed in claim 5, is characterized in that, described decoder process comprises: First initialize a periodic component and a trend component, input the initialized periodic component into the frequency learning module, and make a residual link between the results before and after the transformation, and then send the connection result to the cycle-trend decomposition module, and the output of the cycle-trend decomposition module The trend component is connected to the residual of the initialization trend, and the result is denoted as A; the periodic component output by the cycle-trend decomposition module is sent to the frequency domain attention module together with the sequence output by the encoder; The frequency-domain attention module performs residual linking of the processed result with the output of the cycle-trend decomposition module, and sends it to the next cycle-trend decomposition module; the cycle-trend decomposition module decomposes it into period quantity and trend Quantity, periodic quantity enters the forward propagation module, the trend quantity is connected with the result A by residual error, and the result is recorded as B; The results of the input and output of the forward propagation module are connected by residuals and then sent to the next cycle-trend decomposition module. The cycle-trend decomposition module decomposes it into periodic quantities and trend quantities, and connects the trend quantities with B as residuals. The result is recorded as C; Finally, the periodic quantity decomposed by the cycle-trend decomposition module is fused with the trend quantity C, and the result is output to the fusion module. 7. The photovoltaic power prediction method based on frequency domain decomposition as claimed in claim 1, wherein the results obtained by the illumination prediction module and the power prediction module are input into the fusion module, and the illumination prediction data and the power prediction data are respectively The weighted average is processed by the Gaussian window, and the results are input to the forward propagation module located in the fusion module, and the final output result data is processed by the fully connected layer. 8. The method for predicting photovoltaic power generation based on frequency domain decomposition as claimed in claim 7, wherein the Gaussian window processing is specifically:

Among them, n is the serial number of discrete sequence data points, and M is the width of the window. 9. the photovoltaic power generation power prediction method based on frequency domain decomposition as claimed in claim 7, is characterized in that, described weighted average formula is:

Among them, L(n) is the output value, that is, the value put into the fully connected layer, and l(n) is the original sequence data point.

Images