A short-term solar irradiance prediction method and device

Abstract

The present invention discloses a short-term solar irradiance prediction method and device. The prediction method includes: data acquisition; using the ICEEMDAN algorithm to decompose the original irradiance sequence into multi-scale modal components, combining the multi-scale modal components to construct a multi-dimensional radiation feature sequence capable of reflecting the change characteristics of irradiance; based on the multi-dimensional radiation feature sequence and the multi-dimensional meteorological feature sequence, constructing a two-dimensional radiation feature matrix and a two-dimensional meteorological feature matrix according to time steps; introducing a residual attention mechanism to reconstruct the two-dimensional meteorological feature matrix to obtain a new meteorological feature matrix; respectively extracting the time series features of the two-dimensional radiation feature matrix and the new meteorological feature matrix and fusing them; using the fused time series features as the input of a multi-layer perceptron to predict the short-term solar irradiance. The present invention can improve the prediction accuracy of short-term solar irradiance.

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic power generation, and relates to a method and device for short-term solar irradiance prediction. Background Technique

[0002] The output power of a photovoltaic power generation system has volatility and intermittency, which is not conducive to the safe and stable operation of the power system during photovoltaic grid connection. Solar irradiance is the main factor affecting the power generation of photovoltaic power, and accurate prediction of solar irradiance helps to accurately predict the power generation of photovoltaic power.

[0003] In recent years, artificial intelligence methods mainly based on machine learning methods and deep learning methods have been widely used in the field of solar radiation prediction. Machine learning methods such as artificial neural networks and random forests, and deep learning methods such as convolutional neural networks and long short-term memory networks.

[0004] However, the existing short-term solar irradiance prediction methods are difficult to capture the volatility and mutability of the irradiance sequence, resulting in low prediction accuracy. Research shows that there is a close connection between solar irradiance and meteorological characteristics. Reasonable utilization of meteorological characteristics can improve the prediction accuracy. Most of the existing prediction methods manually select meteorological characteristics according to the correlation coefficient or experience, which is rather cumbersome and ignores the different degrees of influence of different meteorological characteristics on the prediction task, lacking rationality, and thus affecting the prediction accuracy. Summary of the Invention

[0005] Object of the Invention: The object of the present invention is to propose a method and device for short-term solar irradiance prediction, which can improve the prediction accuracy of short-term solar irradiance.

[0006] Technical Solution: In the first aspect of the present invention, a method for short-term solar irradiance prediction is provided, including:

[0007] Data acquisition, including radiation data and meteorological data. The radiation data is the total horizontal irradiance, including an irradiance sequence, and the meteorological data includes a multi-dimensional meteorological feature sequence;

[0008] Using the ICEEMDAN algorithm, decomposing the original irradiance sequence into multi-scale modal components, combining the multi-scale modal components, and constructing a multi-dimensional radiation feature sequence that can reflect the change characteristics of irradiance;

[0009] Based on the multi-dimensional radiation feature sequence and the multi-dimensional meteorological feature sequence, constructing a two-dimensional radiation feature matrix and a two-dimensional meteorological feature matrix according to time steps;

[0010] Introducing a residual attention mechanism to reconstruct the two-dimensional meteorological feature matrix to obtain a new meteorological feature matrix;

[0011] Respectively extracting the time series features of the two-dimensional radiation feature matrix and the new meteorological feature matrix, and fusing them;

[0012] The fused time series features are used as the input of the multi-layer perceptron to predict the short-term solar irradiance.

[0013] Furthermore, the collected meteorological data includes solar zenith angle, temperature, cloud type, dew point temperature, wind direction, wind speed, relative humidity, and precipitable water.

[0014] Furthermore, the ICEEMDAN algorithm is used to decompose the original irradiance sequence, including:

[0015] Define the original irradiance sequence as s;

[0016] Based on the sequence s, construct a new sequence:

[0017] s i = s + α0E1(w i )

[0018] where s i is the new sequence constructed after adding i groups of white noise, w i is the i groups of white noise added to the sequence s, and E k (·) represents the k-th order modal component generated by the empirical mode decomposition algorithm;

[0019] Calculate the first group of residuals R1:

[0020] R1 = <M(s i )>

[0021] where <·> represents taking the average of the whole; M(·) is the local mean of the sequence generated by the empirical mode decomposition algorithm;

[0022] Calculate the first modal component IMF1:

[0023] IMF1 = s - R1

[0024] Based on the obtained first modal component IMF1, continue to add white noise, and use local mean decomposition to calculate the second group of residuals R2 and the second modal component IMF2:

[0025] R2 = <M(R1 + α1E2(w i ))>

[0026] IMF2 = R1 - R2

[0027] And so on, the k-th group of residuals R k and the k-th modal component IMF k are:

[0028] R k = <M(R k-1+α k-1 E k (w i ))>

[0029] IMF k =R k-1 -R k

[0030] Repeat the above calculation process of the residual and modal components until the calculation ends to obtain all the modal components and the final residual;

[0031] α in the above formula k is expressed as:

[0032]

[0033] where ε0 is the reciprocal of the signal-to-noise ratio between the Gaussian white noise sequence with a mean of 0 added for the first time and the original irradiance sequence to be analyzed; std represents the standard deviation;

[0034] Merge the modal components and residuals of different modes obtained by decomposition to obtain a multi-dimensional radiation feature sequence that can reflect the change characteristics of irradiance.

[0035] Furthermore, use the empirical mode decomposition algorithm to decompose the original irradiance sequence, including: continuously (1) finding the mean of the upper and lower envelope lines of the sequence; (2) subtracting the mean envelope line from the original sequence; (3) iterating repeatedly until the obtained sequence satisfies the two constraint conditions of the intrinsic mode function; at this time, an IMF component is obtained, and the local mean refers to the part obtained by subtracting this IMF from the original sequence; based on the obtained i groups of local means, take their overall average to obtain the first group of residuals R1.

[0036] Furthermore, use the residual attention mechanism to reconstruct the two-dimensional meteorological feature matrix, including:

[0037] Obtain the attention weight matrix based on the two-dimensional meteorological feature matrix:

[0038] A = σ(W2(δ(W1X + b1)) + b2)

[0039] where A represents the attention weight matrix obtained based on the two-dimensional meteorological feature; σ is the sigmoid activation function; W1, W2 represent the updatable weight matrices; δ is the ReLU activation function; is the two-dimensional meteorological feature matrix, T is the number of time steps, F is the number of input features, representing the measured values of F meteorological features in the past T time steps; b1, b2 are the biases corresponding to the updatable weight matrices;

[0040] Add the attention weight to the two-dimensional meteorological feature matrix:

[0041] Xatt = A⊙X

[0042] where X att is the meteorological feature matrix after introducing the attention weight; ⊙ represents the Hadamard product;

[0043] Introduce residual connection:

[0044] X' = X + X att

[0045] X' is the new meteorological feature matrix reconstructed by the residual attention mechanism.

[0046] Furthermore, use the stacked long short-term memory network to extract the temporal features of the two-dimensional radiation feature matrix and the temporal features of the new meteorological feature matrix respectively.

[0047] Furthermore, use the concatenate operation to fuse the temporal features of the two-dimensional radiation feature matrix and the temporal features of the new meteorological feature matrix.

[0048] The second aspect of the present invention provides a short-term solar irradiance prediction device, including:

[0049] A data acquisition module for acquiring radiation data and meteorological data, where the radiation data is the total horizontal irradiance, including an irradiance sequence, and the meteorological data includes a multi-dimensional meteorological feature sequence;

[0050] A data processing module for decomposing the original irradiance sequence into multi-scale modal components through the ICEEMDAN algorithm, merging the multi-scale modal components, and constructing a multi-dimensional radiation feature sequence that can reflect the change characteristics of irradiance; based on the multi-dimensional radiation feature sequence and the multi-dimensional meteorological feature sequence, constructing a two-dimensional radiation feature matrix and a two-dimensional meteorological feature matrix according to time steps; and using the residual attention mechanism to reconstruct the two-dimensional meteorological feature matrix to obtain a new meteorological feature matrix;

[0051] A temporal feature extraction module for respectively extracting the temporal features of the two-dimensional radiation feature matrix and the new meteorological feature matrix and fusing them;

[0052] A multi-layer perceptron for predicting the short-term solar irradiance with the fused temporal features as the input.

[0053] The third aspect of the present invention provides a short-term solar irradiance prediction device, including:

[0054] One or more processors; and one or more memories;

[0055] Wherein one or more programs are stored in the one or more memories, and when the one or more programs are executed by the one or more processors, the prediction method of the first aspect is implemented.

[0056] In the fourth aspect of the present invention, a computer storage medium is provided. Computer instructions are stored in the computer storage medium. When the computer instructions are executed, the prediction method of the first aspect is implemented.

[0057] In the fifth aspect of the present invention, a computer program product is provided. When the computer program product runs on a computer, the computer is enabled to execute the prediction method of the first aspect.

[0058] Beneficial effects: Compared with the prior art, the present invention has the following remarkable advantages:

[0059] (1) By using the ICEEMDAN algorithm to decompose the original irradiance sequence, more multi-scale modal components characterizing the change characteristics of irradiance can be obtained, reducing the influence of the volatility and mutability of the original irradiance sequence on solar radiation prediction, and improving the accuracy and reliability of the prediction results.

[0060] (2) By introducing a residual attention mechanism to reconstruct the two-dimensional meteorological feature matrix, the importance of different meteorological features in prediction can be fully considered, and at the same time, the loss of original meteorological feature information can be avoided, improving the accuracy of solar radiation prediction, and having high robustness and feasibility. Description of the Drawings

[0061] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required to be used in the embodiments of the present invention. Obviously, the following described drawings are only the embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0062] Figure 1 is the flow block diagram of the prediction method of the present invention;

[0063] Figure 2 is the structural block diagram of the prediction device of the present invention;

[0064] Figure 3 is the decomposition result diagram of the original irradiance sequence;

[0065] Figure 4 is the schematic diagram of implementing residual attention on the original meteorological features;

[0066] Figure 5 is the comparison diagram of RMSE and MAE error indexes of different models;

[0067] Figure 6 is the schematic diagram of model curve fitting under different degrees of irradiance fluctuation. Detailed Embodiments

[0068] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are not all of the embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative efforts belong to the protection scope of the present invention.

[0069] As Figure 1 shown is a flow block diagram of a short-term solar irradiance prediction method provided by an embodiment of the present application. The prediction method specifically includes the following steps:

[0070] (1) Data collection, including radiation data and meteorological data;

[0071] The radiation data is an irradiance sequence including the global horizontal irradiance (GHI), and the meteorological data is a multi-dimensional meteorological feature sequence including the solar zenith angle, temperature, cloud type, dew point temperature, wind direction, wind speed, relative humidity, and precipitable water.

[0072] (2) Using the ICEEMDAN (Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise) algorithm, decompose the original irradiance sequence into multi-scale modal components, merge the multi-scale modal components, and construct a multi-dimensional radiation feature sequence that can reflect the change characteristics of the irradiance;

[0073] This step (2) specifically includes:

[0074] Define the original irradiance sequence as s;

[0075] Based on the sequence s, construct a new sequence:

[0076] s i = s + α0E1(w i )

[0077] where s i is the new sequence constructed after adding i groups of white noise, w i is the i groups of white noise added to the sequence s, and E k (·) represents the k-th order modal component decomposed by the empirical mode decomposition algorithm;

[0078] Calculate the first group of residuals R1:

[0079] R1 = <M(s i )>

[0080] Among them, <·> represents taking the average of the whole; M(·) is the local mean of the sequence generated by the empirical mode decomposition algorithm. Specifically, the original irradiance sequence is decomposed using the empirical mode decomposition algorithm: continuously (1) finding the mean of the upper and lower envelope lines of the sequence; (2) subtracting the mean envelope line from the original sequence; (3) iterating repeatedly until the obtained sequence satisfies the two constraint conditions of the intrinsic mode function; at this time, an IMF component is obtained, and the local mean refers to the part obtained by subtracting this IMF from the original sequence; based on the obtained i groups of local means, taking the average of the whole, the first group of residuals R1 is obtained.

[0081] The first mode component IMF1 is calculated as:

[0082] IMF1 = s - R1

[0083] On the basis of the obtained first mode component IMF1, white noise is continuously added, and the second group of residuals R2 and the second mode component IMF2 are calculated using local mean decomposition:

[0084] R2 = <M(R1 + α1E2(w i ))>

[0085] IMF2 = R1 - R2

[0086] And so on, the k-th group of residuals R k and the k-th mode component IMF k are:

[0087] R k = <M(R k-1 + α k-1 E k (w i ))>

[0088] IMF k = R k-1 - R k

[0089] Repeat the above calculation process of residuals and mode components until the calculation is completed to obtain all mode components and the final residuals;

[0090] α k in the above formula is expressed as:

[0091]

[0092] Among them, ε0 is the reciprocal of the signal-to-noise ratio between the first added Gaussian white noise sequence with a mean of 0 and the original irradiance sequence to be analyzed; std represents the standard deviation;

[0093] Merge the mode components and residuals of different modes obtained by decomposition to obtain a multi-dimensional radiation feature sequence that can reflect the change characteristics of irradiance.

[0094] (3) Based on the multi-dimensional radiation feature sequence and the multi-dimensional meteorological feature sequence, construct a two-dimensional radiation feature matrix and a two-dimensional meteorological feature matrix according to time steps;

[0095] (4) Use the residual attention (RA) mechanism to reconstruct the two-dimensional meteorological feature matrix to obtain a new meteorological feature matrix;

[0096] This step (4) specifically includes:

[0097] Obtain the attention weight matrix based on the two-dimensional meteorological feature matrix:

[0098] A = σ(W2(δ(W1X + b1)) + b2)

[0099] Among them, A represents the attention weight matrix obtained based on the two-dimensional meteorological feature; σ is the sigmoid activation function; W1, W2 represent updatable weight matrices; δ is the ReLU activation function; is the two-dimensional meteorological feature matrix, T is the number of time steps, F is the number of input features, representing the measured values of F meteorological features in the past T time steps; b1, b2 are the biases corresponding to the updatable weight matrices;

[0100] Such as Figure 3 shown, add the attention weight to the two-dimensional meteorological feature matrix:

[0101] X att = A⊙X

[0102] Among them, X att is the meteorological feature matrix after introducing the attention weight; ⊙ represents the Hadamard product;

[0103] Introduce the residual connection:

[0104] X' = X + X att

[0105] X' is the new meteorological feature matrix reconstructed by the residual attention.

[0106] (4) Use the stacked long short-term memory network (LSTM) to extract the temporal features of the two-dimensional radiation feature matrix and the temporal features of the new meteorological feature matrix respectively;

[0107] Use the concatenate operation to fuse the temporal features of the multi-two-dimensional radiation feature matrix and the temporal features of the new meteorological feature matrix.

[0108] (5) Use the fused temporal features as the input of the multi-layer perceptron (MLP) to predict the short-term total horizontal irradiance.

[0109] The embodiment of the present application also provides a short-term solar irradiance prediction device, including:

[0110] A data acquisition module, configured to acquire radiation data and meteorological data, where the radiation data is the total horizontal irradiance, including an irradiance sequence, and the meteorological data includes a multi-dimensional meteorological feature sequence;

[0111] A data processing module, configured to decompose the original irradiance sequence into multi-scale modal components through the ICEEMDAN algorithm, merge the multi-scale modal components, and construct a multi-dimensional radiation feature sequence capable of reflecting the irradiance change characteristics; based on the multi-dimensional radiation feature sequence and the multi-dimensional meteorological feature sequence, construct a two-dimensional radiation feature matrix and a two-dimensional meteorological feature matrix according to time steps; and reconstruct the two-dimensional meteorological feature matrix by using a residual attention mechanism to obtain a new meteorological feature matrix;

[0112] A time series feature extraction module, configured to extract the time series features of the two-dimensional radiation feature matrix and the new meteorological feature matrix respectively, and fuse them;

[0113] A multi-layer perceptron, configured to predict the short-term solar irradiance by using the fused time series features as input.

[0114] The embodiment of the present application also provides another short-term solar irradiance prediction device, including:

[0115] One or more processors; and one or more memories;

[0116] Wherein one or more programs are stored in one or more memories, and when the one or more programs are executed by the one or more processors, the prediction method in the above embodiment is implemented.

[0117] The embodiment of the present application also provides a computer storage medium, in which computer instructions are stored, and when the computer instructions are executed, the prediction method in the above embodiment is implemented.

[0118] The computer storage medium may be, for example, various media such as a USB flash drive, a mobile hard disk, a ROM memory, a RAM memory, a magnetic disk, or an optical disc that can be used to store computer instructions.

[0119] The embodiment of the present application also provides a computer program product, when the computer program product runs on a computer, it enables the computer to execute the above related steps and implement the prediction method in the above embodiment.

[0120] The prediction device, computer storage medium, and computer program product provided by the embodiment of the present application are all used to execute the prediction method provided above, so they have the same beneficial effects as the prediction method.

[0121] The implementation process of the prediction method of the present invention for short-term solar radiation prediction will be introduced in detail below in combination with specific examples.

[0122] Radiation data and meteorological data of Nanjing City, Jiangsu Province are selected for testing. Specifically, the data from January 1, 2016 to December 31, 2020 are used, and the time interval for data collection is 1 hour, with a total of 43,800 samples. Among them, the data from January 1, 2016 to December 31, 2019 are used as training samples to train the model, and the data from January 1, 2020 to December 31, 2020 are used as test samples to evaluate the performance of the model.

[0123] The collected data includes radiation data and meteorological data. The radiation data is the total horizontal irradiance, and the meteorological data includes solar zenith angle, temperature, cloud type, dew point temperature, wind direction, wind speed, relative humidity, and precipitable water.

[0124] Using the ICEEMDAN algorithm, the original irradiance sequence is decomposed into multi-scale modal components.

[0125] As Figure 2 shown, the decomposition results include 15 intrinsic mode functions (IMFs) and a residual (Res). Among them, the high-frequency components correspond to the components with large volatility and mutability in the original irradiance sequence, and the low-frequency components correspond to the components with strong regularity in the original irradiance sequence. Combining the 15 intrinsic mode functions and the residual can obtain a 16-dimensional feature sequence that can reflect the change characteristics of irradiance.

[0126] Based on the multi-dimensional radiation feature sequence and the multi-dimensional meteorological feature sequence, two-dimensional radiation feature matrices and two-dimensional meteorological feature matrices are constructed according to the time steps. During actual testing, the number of time steps is set to 48, that is, the multi-dimensional radiation features and meteorological features in the past 48 hours.

[0127] The residual attention mechanism is used to reconstruct the two-dimensional meteorological feature matrix to obtain a new meteorological feature matrix.

[0128] Using the stacked long short-term memory network (LSTM), the temporal features of the two-dimensional radiation feature matrix and the temporal features of the new meteorological feature matrix are extracted respectively; using the concatenate operation, the temporal features of the two-dimensional radiation feature matrix and the temporal features of the new meteorological feature matrix are fused.

[0129] The fused temporal features are used as the input of the multi-layer perceptron (MLP) to predict the short-term total horizontal irradiance, and the final prediction result is output.

[0130] To verify the performance of the prediction method proposed in the embodiments of the present application, the prediction effect of the model established according to the prediction method is evaluated based on test samples. The selected model evaluation indicators are root mean square error (RMSE), mean absolute error (MAE), and correlation coefficient (R). The calculation formulas are as follows:

[0131]

[0132]

[0133]

[0134] where n represents the total number of test samples, and y i represent the predicted value and the actual value of the i-th sample respectively, and y a represent the predicted mean and the actual mean respectively.

[0135] To further evaluate the prediction performance of the prediction model, a total of six comparison models are set, namely MLP, LSTM, LSTM-ANN, Bi-LSTM, CNN-Bi-LSTM, and ICEEMDAN-LSTM. Among them, LSTM-ANN adds an ANN part on the basis of LSTM, aiming to increase the network depth and thus improve the non-linear fitting ability of the model; the proposed model ICEEMDAN-LSTM does not implement residual attention on the original meteorological features, but directly uses them as the input of the model.

[0136] Table 1 shows the prediction errors of each model when predicting the total horizontal irradiance 1 hour in advance, Figure 5 The figure shows the comparison of the RMSE and MAE errors between the model constructed according to the present invention and other models.

[0137] Table 1 Comparison of prediction results of different models

[0138]

[0139] As can be seen from Table 1, regardless of which evaluation index is used, the prediction method of the present invention has the highest prediction accuracy. Moreover, compared with the method that does not decompose the original irradiance sequence using ICEEMDAN, the method that decomposes the original irradiance sequence using ICEEMDAN has a significantly improved prediction effect, which can be clearly seen in Figure 5 . The prediction method of the present invention implements residual attention on the original meteorological features on the basis of decomposing the original irradiance sequence, and obtains the smallest prediction error.

[0140] To further evaluate the prediction performance of the prediction model, Figure 6It shows the fitting of the predicted curves of each model and the actual irradiance curves under different weather conditions. It can be seen that for the case of small fluctuations in irradiance, the predicted value curves of most models can fit well with the actual value curves. In particular, the predicted value curve based on ICEEMDAN-RA-LSTM of the present invention has the smallest deviation from the actual value curve in the rising and falling stages of the irradiance value, and the curve fitting effect is the best. For the case of large fluctuations in irradiance, there are large deviations between the predicted value curves of most models and the actual value curves. However, for the predicted value curve corresponding to the present invention, its overall trend is very close to the change trend of the actual irradiance value. This is because the decomposition result based on ICEEMDAN can obtain the irradiance components with significant volatility, making the model applicable to different weather conditions and predictions under different degrees of irradiance value fluctuations.

[0141] In summary, the present invention uses ICEEMDAN to construct a multi-dimensional feature sequence that can reflect the change characteristics of the original irradiance sequence, thereby capturing the volatility and mutation of the original irradiance sequence, and is applicable to short-term solar irradiance prediction under different weather conditions and different degrees of irradiance fluctuations, and obtains good prediction performance. The present invention can fully consider the importance of different meteorological characteristics during prediction, and at the same time avoid the loss of original meteorological feature information, improve the accuracy of solar radiation prediction, and has high robustness and feasibility. The prediction results of the present invention can be used for photovoltaic power prediction, and further ensure the safe and stable operation of the power system during large-scale photovoltaic grid connection.

[0142] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any change or replacement scheme that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A short-term solar irradiance prediction method, characterized in that, Including: Data collection, including radiation data and meteorological data. The radiation data is the total horizontal irradiance, including an irradiance sequence, and the meteorological data includes a multi-dimensional meteorological feature sequence; Using the ICEEMDAN algorithm, decomposing the original irradiance sequence into multi-scale modal components, combining the multi-scale modal components, and constructing a multi-dimensional radiation feature sequence that can reflect the change characteristics of irradiance; Based on the multi-dimensional radiation feature sequence and the multi-dimensional meteorological feature sequence, constructing a two-dimensional radiation feature matrix and a two-dimensional meteorological feature matrix according to time steps; Introducing a residual attention mechanism to reconstruct the two-dimensional meteorological feature matrix to obtain a new meteorological feature matrix; Respectively extracting the time series features of the two-dimensional radiation feature matrix and the new meteorological feature matrix and fusing them; Using the fused time series features as the input of a multi-layer perceptron to predict the short-term solar irradiance; Using the ICEEMDAN algorithm to decompose the original irradiance sequence, including: Defining the original irradiance sequence as s; Constructing a new sequence based on sequence s: s i = s + α0E1(w i ) Among them, s i is the new sequence constructed after adding i groups of white noise, and w i is the i groups of white noise added to the sequence s. E k (·) represents the k-th order modal component generated by decomposition using the empirical mode decomposition algorithm; Calculating the first group of residuals R1: R1 = <M(s i )> Where <·> represents taking the average of the whole; M(·) is the local mean of the sequence generated by the empirical mode decomposition algorithm; Calculating the first modal component IMF1: IMF1 = s - R1 Based on the obtained first modal component IMF1, continue to add white noise, and use local mean decomposition to calculate the second group of residuals R2 and the second modal component IMF2: R2 = <M(R1+α1E2(w i ))> IMF2 = R1 - R2 And so on, the k-th group of residuals R k and the k-th mode component IMF k are as follows: R k = <M(R k-1 + α k-1 E k (w i ))> IMF k = R k-1 - R k Repeating the above calculation process of residuals and modal components until the calculation ends to obtain all modal components and the final residuals; α in the above formula k is expressed as: Where ε0 is the reciprocal of the signal-to-noise ratio between the first added Gaussian white noise sequence with a mean of 0 and the original irradiance sequence to be analyzed; std represents the standard deviation; Combining the modal components and residuals of different modes obtained by decomposition to obtain a multi-dimensional radiation feature sequence that can reflect the change characteristics of irradiance; Using the empirical mode decomposition algorithm to decompose the original irradiance sequence, including: continuously (1) finding the mean of the upper and lower envelope lines of the sequence; (2) subtracting the mean envelope line from the original sequence; (3) repeatedly iterating until the obtained sequence meets the two constraint conditions of the intrinsic mode function; at this time, an IMF component is obtained, and the local mean refers to the part obtained by subtracting this IMF from the original sequence; based on the obtained i groups of local means, taking the average of the whole to obtain the first group of residuals R1; Using the residual attention mechanism to reconstruct the two-dimensional meteorological feature matrix, including: Obtaining an attention weight matrix based on the two-dimensional meteorological feature matrix: A = σ(W2(δ(W1X + b1)) + b2) Among them, A represents the attention weight matrix obtained based on two-dimensional meteorological features; σ is the sigmoid activation function; W1 and W2 represent updatable weight matrices; δ is the ReLU activation function; is the two-dimensional meteorological feature matrix, T is the number of time steps, F is the number of input features, representing the measured values of F meteorological features in the past T time steps; b1 and b2 are the biases corresponding to the updatable weight matrices; Adding attention weights to the two-dimensional meteorological feature matrix: X att = A ⊙ X Among them, X att is the meteorological feature matrix after introducing the attention weight; ⊙ represents the Hadamard product; Introducing a residual connection: X' = X + X att X' is the new meteorological feature matrix reconstructed by the residual attention.

2. The short-term solar irradiance prediction method according to claim 1, characterized in that The collected meteorological data includes solar zenith angle, temperature, cloud type, dew point temperature, wind direction, wind speed, relative humidity, and precipitable water.

3. The short-term solar irradiance prediction method according to claim 1, characterized in that Using a stacked long short-term memory network to respectively extract the time series features of the two-dimensional radiation feature matrix and the new meteorological feature matrix.

4. The short-term solar irradiance prediction method according to claim 1, wherein Using the concatenate operation to fuse the time series features of the two-dimensional radiation feature matrix and the new meteorological feature matrix.

5. An apparatus for implementing the short-term solar irradiance prediction method according to any one of claims 1 to 4, characterized in that, Including: A data acquisition module for acquiring radiation data and meteorological data. The radiation data is the total horizontal irradiance, including an irradiance sequence, and the meteorological data includes a multi-dimensional meteorological feature sequence; A data processing module for decomposing the original irradiance sequence into multi-scale modal components by the ICEEMDAN algorithm, merging the multi-scale modal components, and constructing a multi-dimensional radiation feature sequence capable of reflecting the change characteristics of irradiance; Based on the multi-dimensional radiation feature sequence and the multi-dimensional meteorological feature sequence, constructing a two-dimensional radiation feature matrix and a two-dimensional meteorological feature matrix according to time steps; And reconstructing the two-dimensional meteorological feature matrix by using a residual attention mechanism to obtain a new meteorological feature matrix; A time series feature extraction module for respectively extracting the time series features of the two-dimensional radiation feature matrix and the new meteorological feature matrix and fusing them; A multi-layer perceptron for predicting short-term solar irradiance with the fused time series features as the input.

6. A short-term solar irradiance prediction device, characterized in that, Including: One or more processors; And one or more memories; Wherein one or more programs are stored in the one or more memories, and when the one or more programs are executed by the one or more processors, the prediction method described in any one of claims 1 to 4 is implemented.

7. A computer storage medium storing computer instructions, characterized in that, When the computer instructions are executed, the prediction method described in any one of claims 1 to 4 is implemented.

Images