Ultra-short-term photovoltaic power prediction method based on multivariable phase space reconstruction and SENet-BiLSTM

Abstract

The invention relates to an ultra-short-term photovoltaic power prediction method based on multivariable phase space reconstruction and SENet-BiLSTM, and the method comprises the steps: analyzing the correlation strength of a meteorological factor and photovoltaic power generation power through a Pearson correlation coefficient, and carrying out the selection to obtain an input feature; performing multivariable phase space reconstruction for the input features to obtain a corresponding multivariable phase space reconstruction matrix; performing training test on the constructed SENet-BiLSTM model by using the multivariable phase space reconstruction matrix to obtain a photovoltaic power prediction model; and inputting the current meteorological factor data into the photovoltaic power prediction model, and outputting to obtain a photovoltaic power prediction result. Compared with the prior art, the accuracy of photovoltaic power prediction can be improved, the stability of prediction performance and prediction precision are improved, and the influence of weather fluctuation on a prediction result is effectively reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power generation prediction, and in particular to an ultra-short-term photovoltaic power prediction method based on multivariable phase space reconstruction and SENet-BiLSTM. Background Art

[0002] With the increasing global attention to sustainable development and clean energy, photovoltaic power generation, as a green energy source, has been widely used and promoted. However, the output power of photovoltaic power generation is affected by many factors, including weather conditions, seasonal changes, geographical location, etc. The changes in these factors make the power prediction of photovoltaic power generation complex and challenging. In photovoltaic power generation systems, accurate power prediction can not only improve the stability and reliability of the power grid, but also effectively optimize energy management and scheduling.

[0003] In recent years, with the rise of artificial intelligence technology, hybrid prediction methods based on artificial intelligence methods such as neural networks and machine learning have become an effective means to improve the accuracy of photovoltaic prediction. However, the main reason affecting the accuracy of photovoltaic power prediction is the randomness and volatility of photovoltaic power generation. Solar radiation fluctuates due to meteorological conditions, which hinders the further improvement of photovoltaic power prediction accuracy. Relevant researchers preprocess data through single variable phase space reconstruction and use long-term and short-term neural networks for prediction. Phase space reconstruction can better mine the chaotic characteristics of photovoltaic data, and long-term and short-term neural networks also have certain advantages in processing long sequence data. However, these methods do not consider the impact of meteorological factors on the prediction results, and LSTM (long short-term memory) only considers the impact of past information. It is difficult to consider the strength of the impact of different features on the prediction results, resulting in low photovoltaic power prediction accuracy in practical applications. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide an ultra-short-term photovoltaic power prediction method based on multivariable phase space reconstruction and SENet (squeeze-and-excitation networks)-BiLSTM (bidirectional long short-term memory), which can effectively improve the accuracy of photovoltaic power prediction.

[0005] The purpose of the present invention can be achieved by the following technical solution: An ultra-short-term photovoltaic power prediction method based on multivariable phase space reconstruction and SENet-BiLSTM, comprising the following steps:

[0006] S1. Analyze the correlation between meteorological factors and photovoltaic power generation through Pearson correlation coefficient, and select input features;

[0007] S2, reconstruct the multivariate phase space according to the input features to obtain the corresponding multivariate phase space reconstruction matrix;

[0008] S3. Use the multivariable phase space reconstruction matrix to train and test the constructed SENet-BiLSTM model to obtain a photovoltaic power prediction model;

[0009] S4. Input the current meteorological factor data into the photovoltaic power prediction model, and output the photovoltaic power prediction result.

[0010] Furthermore, the input features in step S1 include meteorological factors and corresponding photovoltaic power generation powers whose correlation is greater than or equal to a preset threshold.

[0011] Furthermore, the calculation formula of the Pearson correlation coefficient in step S1 is:

[0012]

[0013] in, They are respectively the photovoltaic power time series X i , meteorological factor time series Y i The average value of , N is the number of time series data.

[0014] Furthermore, the step S2 specifically utilizes the improved CC method to perform multivariate phase space reconstruction on each input feature, and solves the phase space reconstruction parameters m and τ of each feature.

[0015] Furthermore, step S2 includes the following process:

[0016] S21, the photovoltaic power generation time series and n meteorological factors form a D-dimensional multivariate time series {p i ,i=1,2,...,D}, where p i =[p i (1),p i (2),...,p i (N)] T , N is the number of time series data, and the delay time τ of each variable time series is calculated separately i and embedding dimension m i , we get the phase space of the D-dimensional multivariate time series:

[0017] P i (t) = [p i (t),p i (t-τ1),...,p i(t-(m i -1)τ i )]

[0018] t=(m i -1)τ i +1,(m i -1)τ i +2,...,N

[0019] S22. After calculating m and τ corresponding to the power and each meteorological factor time series, determine the reconstructed phase space of the corresponding D-dimensional multivariate time series.

[0020] Furthermore, the reconstructed phase space of the D-dimensional multivariate time series in step S22 is specifically:

[0021] P r (t)=[p1(t),p1(t-τ1),…,p1(t-(m1-1)τ1),

[0022] p2(t),p2(t-τ2),…,p2(t-(m2-1)τ2),

[0023] …

[0024] p D (t),p D (t-τ D ),…,p D (t-(m D -1)τ D )]

[0025]

[0026] Where d is the dimension of the reconstructed phase space of the D-dimensional feature time series.

[0027] Furthermore, step S3 includes the following process:

[0028] S31, dividing the multivariate phase space reconstruction matrix into a training set and a test set;

[0029] S32. Use the training set to train the constructed SENet-BiLSTM model, and use the test set to test the constructed SENet-BiLSTM model to determine the optimal parameters of the model to obtain a photovoltaic power prediction model.

[0030] Furthermore, the SENet-BiLSTM model includes a SENet network and a BiLSTM network, wherein a multivariate phase space matrix is ​​input into the SENet network for multi-channel attention weight allocation, and a feature matrix is ​​output;

[0031] The feature matrix is ​​input into the BiLSTM network, and the photovoltaic power prediction result is output by combining the current moment with the past and future bidirectional historical information.

[0032] Furthermore, the working process of the SENet network includes:

[0033] After the multivariate phase space matrix X is convolved to obtain the feature matrix U, U is squeezed, that is, the feature matrix of each channel is globally averaged pooled, and each two-dimensional channel is converted into a value with a global receptive field. Its output dimension is equal to the number of channels of the original feature matrix. The formula for the squeezing operation is:

[0034]

[0035] Among them, Z c is the extrusion result of the cth channel; H and W are the height and width of the feature matrix;

[0036] Each channel is adaptively weighted according to its importance, so that it can selectively emphasize important feature information and suppress less important feature information. The excitation operation is used to limit the model complexity and enhance the generalization ability of the model through two fully connected layers and two activation functions. The expression of the excitation operation is:

[0037] s=F ex (z,W)=σ(W2δ(W1z))

[0038] Where s is the output after the excitation operation, σ(·) and δ(·) are activation functions, W1(·) and W2(·) are two fully connected layer operations respectively;

[0039] Use the Scale operation to weight the attention weight obtained above to the features of each channel. Multiply each channel by the weight coefficient to weight the attention weight of the channel into the feature matrix of each channel. The weighting formula of the attention weight is:

[0040]

[0041] Among them, s c is the weight of the cth channel; u c is the feature map of the cth channel; is the weighted output.

[0042] Furthermore, the working process of the BiLSTM network includes:

[0043] The feature matrix output by SENet Input into BiLSTM, which is composed of multiple forward and reverse LSTMs. The expressions of variables inside LSTM are:

[0044]

[0045] Among them, x t is the input at the current moment; W is the input weight matrix of each gate; b is the bias; C t is the state of the memory storage unit; subscripts c, f, i, o represent the cell state, forget gate, input gate, and output gate; σ(·) is the activation function;

[0046] The final photovoltaic power prediction value is calculated according to the following formula:

[0047]

[0048] in, is the forward LSTM hidden layer state; is the backward LSTM hidden layer state; W t 、V t are the respective output weights; y t is the output result; b y is the offset.

[0049] Compared with the prior art, the present invention has the following advantages:

[0050] The present invention uses the Pearson correlation coefficient to analyze the correlation between meteorological factors and photovoltaic power generation, and takes the meteorological factor data and power with higher correlation as input features; by reconstructing the phase space of each input feature, the reconstructed multidimensional phase space matrix is ​​input into the SENet and BiLSTM combined model based on the channel attention mechanism for training and testing to obtain a photovoltaic power prediction model. On the one hand, the data preprocessing method of multivariate phase space reconstruction can effectively mine the chaotic characteristics of photovoltaic data, reduce the impact of the volatility problem of the original photovoltaic power generation on the prediction accuracy, and thus improve the subsequent photovoltaic power prediction accuracy; on the other hand, the SENet-BiLSTM combined model based on the channel attention mechanism is used to assign corresponding weights according to the importance of the input photovoltaic power features, which helps to improve the ability of the LSTM neural network to learn the importance of photovoltaic power data features, thereby achieving more accurate photovoltaic power prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0052] Figure 2 It is a schematic diagram of the application process of the embodiment;

[0053] Figure 3 It is the Pearson correlation coefficient analysis diagram in the embodiment;

[0054] Figure 4a ,4b 4c and 4d are graphs showing the photovoltaic output power prediction results under different weather conditions in the embodiment. DETAILED DESCRIPTION

[0055] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0056] Example

[0057] like Figure 1 As shown, an ultra-short-term photovoltaic power prediction method based on multivariable phase space reconstruction and SENet-BiLSTM includes the following steps:

[0058] S1. Analyze the correlation between meteorological factors and photovoltaic power generation through Pearson correlation coefficient, and select input features;

[0059] S2, reconstruct the multivariate phase space according to the input features to obtain the corresponding multivariate phase space reconstruction matrix;

[0060] S3. Use the multivariable phase space reconstruction matrix to train and test the constructed SENet-BiLSTM model to obtain a photovoltaic power prediction model;

[0061] S4. Input the current meteorological factor data into the photovoltaic power prediction model, and output the photovoltaic power prediction result.

[0062] This embodiment applies the above solution, wherein the process of establishing a photovoltaic power prediction model is as follows: Figure 2 As shown, including:

[0063] 1) Use the Pearson correlation coefficient to analyze the correlation between meteorological factors and actual photovoltaic power generation, and select factors with stronger correlation as input variables. The Pearson correlation coefficient formula is:

[0064]

[0065] Where: They are respectively the photovoltaic power time series X i , meteorological factor time series Y i The average value of; N is the number of time series data.

[0066] 2) According to the photovoltaic power generation and meteorological factor data in step 1), the multivariate phase space of each input feature is reconstructed using the improved CC method to solve the phase space reconstruction parameters m and τ of each feature, specifically:

[0067] 2.1) The photovoltaic power generation time series and n meteorological factors are combined into a D-dimensional multivariate time series {p i ,i=1,2,...,D}, where pi =[p i (1),p i (2),...,p i (N)] T , N is the number of time series data. Calculate the delay time τ of each univariate time series respectively i and embedding dimension m i , we get the phase space of the D-dimensional multivariate time series:

[0068] P i (t) = [p i (t),p i (t-τ1),...,p i (t-(m i -1)τ i )] (2)

[0069] Where: t=(m i -1)τ i +1,(m i -1)τ i +2,...,N.

[0070] 2.2) After calculating the m and τ corresponding to the power and meteorological factor time series, the corresponding D-dimensional multivariate time series reconstructed phase space is obtained:

[0071]

[0072] Where:

[0073] The dimension of the reconstructed phase space of this D-dimensional characteristic time series is d:

[0074]

[0075] 3) According to the multivariate phase space reconstruction matrix obtained in step 1) and step 2), input it into SENet to implement multi-channel attention weight allocation. The specific steps are:

[0076] 3.1) The collected data are divided into training set and test set, and the multivariate phase space matrix is ​​used as the model input matrix.

[0077] 3.2) After the multivariate phase space matrix X is convolved to obtain the feature matrix U, U is squeezed, that is, the feature matrix of each channel is globally averaged and pooled, and each two-dimensional channel is converted into a value with a global receptive field, and its output dimension is equal to the number of channels of the original feature matrix. This operation can shield the spatial distribution information and better utilize the correlation between channels. The squeezing operation formula is:

[0078]

[0079] Where: Z c is the squeezing result of the cth channel; H and W are the rows and columns of the feature matrix.

[0080] 3.3) Each channel is adaptively weighted according to its importance, so that it can selectively emphasize important feature information and suppress less important feature information. The excitation operation is used to limit the model complexity and enhance the generalization ability of the model through two fully connected layers and two activation functions. The formula of the excitation operation is:

[0081] s=F ex (z,W)=σ(W2δ(W1z)) (6)

[0082] Where: s is the output after the excitation operation; σ(·) and δ(·) are activation functions; W1(·) and W2(·) are two fully connected layer operations respectively.

[0083] 3.4) Use the Scale operation to weight the attention weights obtained above to the features of each channel. Multiply each channel by the weight coefficient to weight the channel attention weights to the feature matrix of each channel. The weighted calculation formula is:

[0084]

[0085] Among them, s c is the weight of the cth channel; u c is the feature map of the cth channel; is the weighted output.

[0086] 4) According to the features extracted in step 3), the features are input into BiLSTM to implement nonlinear fitting, and a SENet-BiLSTM combined model is established to achieve ultra-short-term prediction of photovoltaic power. The specific steps are as follows:

[0087] 4.1) SENet output matrix Entering BiLSTM, BiLSTM is composed of multiple LSTMs in the forward and reverse directions, which fully considers the bidirectional historical information of the current moment and the past and future. The expressions of the variables inside LSTM are:

[0088]

[0089] Where: x t is the current input; W is the input weight matrix of each gate; b is the bias; C t is the state of the memory storage unit; subscripts c, f, i, o represent the cell state, forget gate, input gate and output gate; σ(·) is the activation function.

[0090] 4.2) The final predicted value is calculated according to the following formula:

[0091]

[0092] Where: is the forward LSTM hidden layer state; is the backward LSTM hidden layer state; W t 、V t are the respective output weights; y t is the output result; b y is the offset.

[0093] To verify the effectiveness of this solution, this embodiment compares the predicted value output by the ultra-short-term photovoltaic power prediction model based on multivariate phase space reconstruction and SENet-BiLSTM with the actual value, and selects photovoltaic power data and meteorological factor data from January 1, 2020 to March 31, 2020 provided by a photovoltaic power station for simulation experiments, and the sampling interval of the data set is 5 minutes. Since the output of the photovoltaic system is zero at night, this embodiment only uses the data from 7:00-19:00 every day, that is, 145 sampling points every day for simulation experiments, and divides the collected data into training set and test set according to 8:2.

[0094] First, the Pearson correlation coefficient of step 1) is used to analyze the correlation between meteorological factors and the actual photovoltaic power generation, and meteorological factors with stronger correlation are selected as input variables. The results of the correlation analysis are shown in Figure 3 As shown. The photovoltaic power in this area has a strong positive correlation with global horizontal radiation, global oblique radiation, diffuse horizontal radiation and diffuse oblique radiation, and a relatively weak correlation with temperature, relative humidity and wind direction. Therefore, the four meteorological factors of global horizontal radiation, global oblique radiation, diffuse horizontal radiation and diffuse oblique radiation are selected as the input variables of the model together with power.

[0095] According to step 3) and step 4), an ultra-short-term photovoltaic power prediction model based on multivariable phase space reconstruction and SENet-BiLSTM is constructed. The key parameters of the example are set as follows:

[0096] The number of hidden nodes in the two pre-convolutional layers of SENet is set to 32 and 64 respectively, the number of input layer nodes is the number of phase space matrix columns, the number of hidden layer nodes of BiLSTM is 6, and the number of output neurons is 1. In order to objectively reflect the prediction effect of the model, the mean absolute error (MAE), root mean square error (RMSE) and determination coefficient R are used. 2 As evaluation indicators, the numerical units of MAE and RMSE are MW, R 2No units. The results are shown in Table 1.

[0097] Table 1

[0098]

[0099] pass Figure 4a , 4b , 4c, it can be seen that the prediction curve of the proposed prediction method does not show obvious fluctuation error when the power fluctuates, and has certain stability and adaptability. It can be seen from Table 1. Compared with the BiLSTM prediction method, the proposed prediction method has higher prediction accuracy after data preprocessing. The prediction results are better in weather with smaller fluctuations. In weather with drastic power fluctuations, the prediction accuracy is significantly improved compared with the general prediction model. It can be seen that through the above example analysis, the ultra-short-term photovoltaic power prediction model based on multivariate phase space reconstruction and SENet-BiLSTM proposed in the present invention has achieved good prediction results with high accuracy and stability.

[0100] In summary, this scheme uses the Pearson correlation coefficient to analyze the correlation between meteorological factors and actual photovoltaic power generation, and takes meteorological factors and power with higher correlation as input variables of the model at the same time; uses the improved CC method to reconstruct the multivariate phase space of each input feature, solves the phase space reconstruction parameters m and τ of each feature, and inputs the obtained multivariate phase space reconstruction matrix into SENet and BiLSTM successively, establishes the SENet-BiLSTM combined prediction model, and realizes the ultra-short-term prediction of photovoltaic power. The prediction performance of this scheme is more stable and more accurate than the traditional prediction method, which effectively reduces the prediction volatility of photovoltaic power generation and improves the prediction performance.

Claims

1. An ultra-short-term photovoltaic power prediction method based on multivariable phase space reconstruction and SENet-BiLSTM, characterized in that: The following steps are involved: S1. Analyze the correlation between meteorological factors and photovoltaic power generation through Pearson correlation coefficient, and select input features; S2, reconstruct the multivariate phase space according to the input features to obtain the corresponding multivariate phase space reconstruction matrix; S3. Use the multivariable phase space reconstruction matrix to train and test the constructed SENet-BiLSTM model to obtain a photovoltaic power prediction model; S4. Input the current meteorological factor data into the photovoltaic power prediction model, and output the photovoltaic power prediction result.

2. The ultra-short-term photovoltaic power prediction method based on multivariable phase space reconstruction and SENet-BiLSTM according to claim 1 is characterized in that: The input features in step S1 include meteorological factors whose correlation is greater than or equal to a preset threshold and the corresponding photovoltaic power generation power.

3. The ultra-short-term photovoltaic power prediction method based on multivariable phase space reconstruction and SENet-BiLSTM according to claim 1 is characterized in that: The calculation formula of the Pearson correlation coefficient in step S1 is: in, They are respectively the photovoltaic power time series X i , meteorological factor time series Y i The average value of , N is the number of time series data.

4. The ultra-short-term photovoltaic power prediction method based on multivariable phase space reconstruction and SENet-BiLSTM according to claim 1 is characterized in that: The step S2 specifically uses the improved CC method to reconstruct the multivariate phase space of each input feature, and solves the phase space reconstruction parameters m and τ of each feature.

5. The ultra-short-term photovoltaic power prediction method based on multivariable phase space reconstruction and SENet-BiLSTM according to claim 4 is characterized in that: The step S2 includes the following process: S21, the photovoltaic power generation time series and n meteorological factors form a D-dimensional multivariate time series {p i ,i=1,2,...,D},where p i =[p i (1), p i (2), ..., p i (N)] T , N is the number of time series data, and the delay time τ of each variable time series is calculated separately i and embedding dimension m i , we get the phase space of the D-dimensional multivariate time series: P i (t)=[p i (t),p i (t-τ1),...,p i (t-(m i -1)τ i )] t=(m i -1)t i +1,(m i -1)t i +2,...,N S22. After calculating m and τ corresponding to the power and each meteorological factor time series, determine the reconstructed phase space of the corresponding D-dimensional multivariate time series.

6. The ultra-short-term photovoltaic power prediction method based on multivariable phase space reconstruction and SENet-BiLSTM according to claim 5 is characterized in that: The reconstructed phase space of the D-dimensional multivariate time series in step S22 is specifically: P r (t)=[p1(t),p1(t-τ1),…,p1(t-(m1-1)τ1), p2(t), p2(t-τ2),…, p2(t-(m2-1)τ2), … p D (t),p D (t-τ D ),…,p D (t-(m D -1)τ D )] Where d is the dimension of the reconstructed phase space of the D-dimensional feature time series.

7. The ultra-short-term photovoltaic power prediction method based on multivariable phase space reconstruction and SENet-BiLSTM according to claim 1 is characterized in that: Step S3 The process includes: S31, dividing the multivariate phase space reconstruction matrix into a training set and a test set; S32. Use the training set to train the constructed SENet-BiLSTM model, and use the test set to test the constructed SENet-BiLSTM model to determine the optimal parameters of the model to obtain a photovoltaic power prediction model.

8. The ultra-short-term photovoltaic power prediction method based on multivariable phase space reconstruction and SENet-BiLSTM according to claim 7 is characterized in that: The SENet-BiLSTM model includes a SENet network and a BiLSTM network, wherein a multivariate phase space matrix is ​​input into the SENet network to perform multi-channel attention weight allocation, and a feature matrix is ​​output; The feature matrix is ​​input into the BiLSTM network, and the photovoltaic power prediction result is output by combining the current moment with the past and future bidirectional historical information.

9. The ultra-short-term photovoltaic power prediction method based on multivariable phase space reconstruction and SENet-BiLSTM according to claim 8, characterized in that: The working process of the SENet network includes: After the multivariate phase space matrix X is convolved to obtain the feature matrix U, U is squeezed, that is, the feature matrix of each channel is globally averaged pooled, and each two-dimensional channel is converted into a value with a global receptive field. Its output dimension is equal to the number of channels of the original feature matrix. The formula for the squeezing operation is: Among them, Z c is the extrusion result of the cth channel; H and W are the height and width of the feature matrix; Each channel is adaptively weighted according to its importance, so that it can selectively emphasize important feature information and suppress less important feature information. The excitation operation is used to limit the model complexity and enhance the generalization ability of the model through two fully connected layers and two activation functions. The expression of the excitation operation is: s=F ex (z,W)=σ(W2δ(W1z)) Where s is the output after the excitation operation, σ(·) and δ(·) are activation functions, W1(·) and W2(·) are two fully connected layer operations respectively; Use the Scale operation to weight the attention weight obtained above to the features of each channel. Multiply each channel by the weight coefficient to weight the attention weight of the channel into the feature matrix of each channel. The weighting formula of the attention weight is: Among them, s c is the weight of the cth channel; u c is the feature map of the cth channel; is the weighted output.

10. The ultra-short-term photovoltaic power prediction method based on multivariable phase space reconstruction and SENet-BiLSTM according to claim 9, characterized in that: The working process of the BiLSTM network includes: The feature matrix output by SENet Input into BiLSTM, which is composed of multiple forward and reverse LSTMs. The expressions of variables inside LSTM are: Among them, x t is the current input; W is the input weight matrix of each gate; b is the bias; C t is the state of the memory storage unit; subscripts c, f, i, o represent the cell state, forget gate, input gate, and output gate; σ(·) is the activation function; The final photovoltaic power prediction value is calculated according to the following formula: in, is the forward LSTM hidden layer state; is the backward LSTM hidden layer state; W t 、V t are the respective output weights; y t is the output result; b y is the offset.

Images