Photovoltaic output prediction method based on multivariate disturbance feature extraction and auxiliary task learning

Abstract

The invention discloses a photovoltaic output prediction method based on multivariate disturbance feature extraction and auxiliary task learning. The method comprises the following steps: obtaining position information, panel data, output data and meteorological data of a target photovoltaic station; vMD variational mode decomposition is adopted to decompose historical photovoltaic output, and an output disturbance IMFs mode and a clear sky-like IMF mode are obtained; k-means clustering is carried out on the IMFs modals representing the output disturbance, and a power change feature cluster is constructed; spearman correlation analysis is carried out on different power change feature clusters in combination with weather factors; and combining different power change feature clusters and related meteorology thereof as auxiliary task training data, and inputting the auxiliary task training data into a CNN-GRU hybrid prediction model which introduces shared features and a cross-task attention mechanism for training to obtain a final photovoltaic output prediction result. According to the method, the overall precision of photovoltaic output prediction can be improved, the robustness is high, and data support can be provided for operation management of a photovoltaic station.

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic power generation, relates to photovoltaic output prediction technology, and specifically relates to a photovoltaic output prediction method based on multivariate disturbance feature extraction and auxiliary task learning. Background Art

[0002] As a clean and renewable energy source, photovoltaic power generation plays an increasingly important role in the global energy transformation and renewable energy development process. As the scale of photovoltaic power generation continues to expand, accurate output prediction is particularly important. Traditional photovoltaic output prediction methods are mainly based on deep learning, machine learning and numerical weather forecasting to predict future photovoltaic output. However, changes in lighting conditions often have a significant impact on the output timing, especially when the short-term sunlight changes caused by weather changes, the photovoltaic power will fluctuate greatly.

[0003] In the past few years, researchers have actively explored various methods to predict photovoltaic output in order to accurately predict power fluctuations. With the development of artificial intelligence technology, photovoltaic output prediction methods based on data-driven deep learning and machine learning have been widely used. However, existing photovoltaic output prediction methods mainly rely on deep learning, machine learning, and numerical weather forecasting, but have significant deficiencies in dealing with short-term light fluctuations caused by drastic weather changes. These methods are usually difficult to accurately capture the characteristics of fluctuations and have poor performance in predicting sudden disturbances. In addition, due to the interference caused by strong volatility, they find it difficult to effectively deal with drastic changes in weather factors, resulting in limited accuracy and stability in output prediction. Summary of the invention

[0004] Purpose of the invention: In order to overcome the deficiencies in the prior art, a photovoltaic output prediction method based on multivariate disturbance feature extraction and auxiliary task learning is provided, which can effectively improve the accuracy of photovoltaic output prediction.

[0005] Technical solution: To achieve the above purpose, the present invention provides a photovoltaic output prediction method based on multivariate disturbance feature extraction and auxiliary task learning, comprising the following steps:

[0006] S1: Obtain the target PV station location information, panel data, output data and meteorological data;

[0007] The VMD variational modal decomposition is used to decompose the historical PV output to obtain the output disturbance IMFs mode and the clear sky-like IMF mode.

[0008] S2: Perform K-means clustering on the IMFs modes representing the output disturbance to construct the power change feature cluster;

[0009] S3: Spearman correlation analysis of different power change characteristic clusters combined with weather factors;

[0010] S4: Different power change feature clusters and their related meteorological features are combined as auxiliary task training data and input into the CNN-GRU hybrid prediction model that introduces shared features and cross-task attention mechanism for training. The shared features and cross-task attention mechanism are used to reduce the impact of strong meteorological volatility on output forecasting, and the final photovoltaic output forecast result is obtained.

[0011] Furthermore, in step S1, VMD variational modal decomposition is used to decompose the historical photovoltaic output. When performing VMD decomposition, the sum of the expected bandwidths of each power IMF mode is required to be minimum, and the constraint condition is that the sum of all decomposed modes is the same as the original output characteristic signal sequence. The specific formula of the constrained variational problem is:

[0012]

[0013] Where f is the original time series, δ(t) is the Dirac distribution function, {u k}:={u1,…,u K},{w k}:={w1,…,w K} are the abbreviations of all modes and their center frequencies; e -jωkt for u k The corresponding center frequency w k The index term of ;

[0014] By introducing the Lagrange operator, the inequality constraint is transformed into an equality constraint and solving the above equation, the mode u is obtained. k The solution formula is:

[0015]

[0016] Where α is a quadratic penalty factor, which is used to balance the trade-off between the objective function and the degree of constraint violation. By penalizing the constraints, the algorithm is forced to converge to a solution that satisfies the constraints. λ is the Lagrange multiplication operator.

[0017] Get the center frequency w k The formula is:

[0018]

[0019] Furthermore, the method for determining the clear-sky-like IMF in step S1 is:

[0020] After obtaining the IMF, by setting the thresholds of the standard deviation and the rate of change, if the standard deviation and the rate of change of the IMF are both lower than the set thresholds, it is determined to be a clear sky IMF. The determination formula is:

[0021] Std(IMF)<Threshold std and ROC(IMF)<Threshold roc

[0022] Where Std(IMF) is the standard deviation of IMF, which is used to measure the dispersion of IMF data; Threshold std is the standard deviation threshold. When the standard deviation of IMF is less than this threshold, it is considered that the fluctuation of IMF is small. ROC(IMF) is the rate of change of IMF, which is used to measure the speed of change of IMF data. roc is the change rate threshold. When the change rate of IMF is less than this threshold, it is considered that the change of IMF is relatively slow.

[0023] Furthermore, in step S2, K-means clustering is performed on the IMFs modes characterizing the output disturbance, and the silhouette coefficient and the Davis-Boulding index are used as evaluation indicators. After the disturbance IMFs are clustered by K-means, the cluster center line is drawn, and the power change characteristic indicators of each output disturbance cluster center are calculated to construct an output change characteristic cluster.

[0024] Furthermore, the K-means clustering method in step S2 is:

[0025] When clustering, initialize the cluster center, determine the number of clusters K, assign samples D to calculate the Euclidean distance between the remaining data points and the cluster center Ci, find the shortest distance and assign all samples to the cluster corresponding to the cluster center Ci. The Euclidean distance calculation is as follows:

[0026]

[0027] Where: Ci is the i-th cluster center, m is the dimension of the data object, x j , C ij is the j-th attribute value of x and Ci;

[0028] Update the cluster center and calculate the average value and square error of all points in each cluster. Update the cluster center and repeat the above steps. The square error calculation is as follows:

[0029]

[0030] The silhouette coefficient and Davis-Boulding index are used as evaluation indicators. When the cluster center no longer changes or the maximum number of iterations is reached, the loop stops, the clustering results are updated, the evaluation indicators of each cluster number K are recorded, the optimal value of the evaluation indicator is selected as the cluster number, and its cluster center line is drawn.

[0031] Furthermore, in step S2, after performing K-means mutation clustering, cluster center lines of different time lengths for each cluster are drawn, and the mutation duration, slope and average value of the cluster center are calculated as characteristic indicators to construct a mutation feature sequence.

[0032] Furthermore, the calculation formula of the Spearman correlation coefficient in the Spearman correlation analysis of step S3 is as follows:

[0033]

[0034] In the formula, ρ represents the Spearman correlation coefficient, d i represents the difference between each pair of corresponding ranks (i.e., the difference between the rankings corresponding to the IMF variable and the weather characteristic variable), and n represents the number of data pairs.

[0035] Furthermore, if the meteorological conditions in step S4 are highly volatile, directly modeling and predicting the original output will make it difficult for the model to capture the impact of relevant meteorological characteristic disturbances on the output during the training process. Therefore, it is considered to establish prediction tasks according to different power change feature clusters and their related meteorological conditions, and use them as auxiliary tasks to enhance the target task's learning ability for features. Specifically, it includes:

[0036] First, the target task is set to predict photovoltaic output, and the auxiliary task is to predict the components of different power change feature clusters and their related meteorological combinations. The two use the same input features and the same structure of CNN feature extraction network layer;

[0037] Next, the representations of the target task and auxiliary tasks pass through the shared fully connected layer in GRU to achieve sharing of learned feature representations among multiple tasks;

[0038] Finally, by adding an attention module between the feature representation of the target task and the feature representation of the auxiliary task, the feature weight of the target photovoltaic output prediction task can be dynamically adjusted with the help of the auxiliary task. The attention mechanism calculation formula of the attention module is:

[0039] y=(σ(W+b))☉z where x is the feature vector output by each task, W is the weight matrix, b is the bias vector, σ represents the Sigmoid activation function, z is the input feature of each task, and ⊙ represents element-by-element multiplication.

[0040] Furthermore, the dynamic adjustment of feature weights in step S4 is achieved through a cross-task attention mechanism; the attention module calculates weights based on the feature outputs of the target task and the auxiliary task so as to adjust the features during the training process. The expression of this process is as follows:

[0041]

[0042] Where: αi is the weight of the i-th feature, Wa is the weight matrix, and They represent the feature representations of the target task and the auxiliary task respectively, Concat represents the connection of the feature representations of the target task and the auxiliary task, b is the bias vector, and the Softmax function is used to normalize the result into a weighted probability.

[0043] Beneficial effects: Compared with the prior art, the present invention significantly improves the accuracy of photovoltaic output prediction by combining multivariate disturbance feature extraction and auxiliary task learning. Different types of output disturbance features are identified and grouped using variational mode decomposition (VMD) and K-means clustering, enabling the model to more accurately capture photovoltaic fluctuations caused by weather changes. Through shared features and cross-task attention mechanisms, multi-task learning further improves the robustness and adaptability of the prediction model, effectively improving the accuracy of photovoltaic output prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 is a flow chart of the prediction method of the present invention;

[0045] Figure 2 It is the CNN-GRU network structure diagram;

[0046] Figure 3 It is the number of perturbation IMF clusters at different k;

[0047] Figure 4 It is the VMD breakdown diagram of PV output;

[0048] Figure 5 It is the K-means clustering evaluation index graph;

[0049] Figure 6 It is the centerline diagram of the perturbation mode K-means clustering;

[0050] Figure 7 It is a comparison chart between the output prediction value and the actual value. DETAILED DESCRIPTION

[0051] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0052] like Figure 1 As shown, the present invention provides a photovoltaic output prediction method based on multivariate disturbance feature extraction and auxiliary task learning, comprising the following steps:

[0053] S1: Obtain the target PV station location information, panel data, output data and meteorological data, with a data sampling interval of 5 minutes;

[0054] Preprocess the collected data. The focus of preprocessing includes the filling of missing values ​​and the replacement of outliers. For missing values, interpolation is used to complete them to ensure the continuity and integrity of the data. For outliers, abnormal data is identified through box plots and replaced according to historical means to reduce the interference of noise on subsequent analysis and model training, and improve data quality and the accuracy of prediction models;

[0055] The VMD variational modal decomposition is used to decompose the historical PV output to obtain the output disturbance IMFs mode and the clear sky-like IMF mode.

[0056] The VMD variational modal decomposition is used to decompose the historical photovoltaic output. When performing VMD decomposition, the sum of the expected bandwidth of each power IMF mode is required to be minimum, and the constraint condition is that the sum of all decomposed modes is the same as the original output characteristic signal sequence. The specific formula of the constrained variation problem is:

[0057]

[0058] Where f is the original time series, δ(t) is the Dirac distribution function, {u k}:={u1,…,u K},{w k}:={w1,…,w K} are the abbreviations of all modes and their center frequencies; e -jωkt for u k The corresponding center frequency w k The index term of ;

[0059] By introducing the Lagrange operator, the inequality constraint is transformed into an equality constraint and solving the above equation, the mode u is obtained. k The solution formula is:

[0060]

[0061] Where α is a quadratic penalty factor, which is used to balance the trade-off between the objective function and the degree of constraint violation. By penalizing the constraints, the algorithm is forced to converge to a solution that satisfies the constraints. λ is the Lagrange multiplication operator.

[0062] Get the center frequency w k The formula is:

[0063]

[0064] The determination method of quasi-clear sky IMF is:

[0065] After obtaining the IMF, by setting the thresholds of the standard deviation and the rate of change, if the standard deviation and the rate of change of the IMF are both lower than the set thresholds, it is determined to be a clear sky IMF. The determination formula is:

[0066] Std(IMF)<Threshold std and ROC(IMF)<Threshold roc

[0067] Where Std(IMF) is the standard deviation of IMF, which is used to measure the dispersion of IMF data; Threshold std is the standard deviation threshold. When the standard deviation of IMF is less than this threshold, it is considered that the fluctuation of IMF is small. ROC(IMF) is the rate of change of IMF, which is used to measure the speed of change of IMF data. roc is the change rate threshold. When the change rate of IMF is less than this threshold, it is considered that the change of IMF is relatively slow.

[0068] S2: Perform K-means clustering on the IMFs modes representing the output disturbance to construct the power change feature cluster;

[0069] K-means clustering is performed on the IMFs modes that characterize the output disturbance, and the silhouette coefficient and Davis-Boulding index are used as evaluation indicators. After K-means clustering of the disturbance IMFs, the cluster center line is drawn, and the power change characteristic indicators of each output disturbance cluster center are calculated to construct the output change characteristic cluster.

[0070] The K-means clustering method is:

[0071] When clustering, initialize the cluster center, determine the number of clusters K, assign samples D to calculate the Euclidean distance between the remaining data points and the cluster center Ci, find the shortest distance and assign all samples to the cluster corresponding to the cluster center Ci. The Euclidean distance calculation is as follows:

[0072]

[0073] Where: Ci is the i-th cluster center, m is the dimension of the data object, x j , C ij is the j-th attribute value of x and Ci;

[0074] Update the cluster center and calculate the average value and square error of all points in each cluster. Update the cluster center and repeat the above steps. The square error calculation is as follows:

[0075]

[0076] The silhouette coefficient and Davis-Boulding index are used as evaluation indicators. When the cluster center no longer changes or the maximum number of iterations is reached, the loop stops, the clustering results are updated, the evaluation indicators of each cluster number K are recorded, the optimal value of the evaluation indicator is selected as the cluster number, and its cluster center line is drawn.

[0077] After performing K-means mutation clustering, the cluster center lines of different time lengths for each cluster were drawn, and the mutation duration, slope, and average value of the cluster center were calculated as characteristic indicators to construct a mutation feature sequence.

[0078] S3: Spearman correlation analysis is performed on different power change characteristic clusters combined with weather factors. The Spearman correlation coefficient calculation formula is as follows:

[0079]

[0080] In the formula, ρ represents the Spearman correlation coefficient, d i represents the difference between each pair of corresponding ranks (i.e., the difference between the rankings corresponding to the IMF variable and the weather characteristic variable), and n represents the number of data pairs.

[0081] S4: Different power change feature clusters and their related meteorological features are combined as auxiliary task training data and input into the CNN-GRU hybrid prediction model that introduces shared features and cross-task attention mechanism for training. The shared features and cross-task attention mechanism are used to reduce the impact of strong meteorological volatility on output forecasting, and the final photovoltaic output forecast result is obtained;

[0082] If the meteorological conditions are highly volatile, directly modeling and predicting the original output will make it difficult for the model to capture the impact of relevant meteorological characteristic disturbances on the output during the training process. Therefore, it is considered to establish prediction tasks based on different power change feature clusters and their related meteorological conditions, and use them as auxiliary tasks to enhance the target task's ability to learn features. The network structure diagram of CNN-GRU is shown in the figure below: Figure 2 As shown, specifically including:

[0083] First, the target task is set to predict photovoltaic output, and the auxiliary task is to predict the components of different power change feature clusters and their related meteorological combinations. The two use the same input features and the same structure of CNN feature extraction network layer;

[0084] Next, the representations of the target task and auxiliary tasks pass through the shared fully connected layer in GRU to achieve sharing of learned feature representations among multiple tasks;

[0085] Finally, by adding an attention module between the feature representation of the target task and the feature representation of the auxiliary task, the feature weight of the target photovoltaic output prediction task can be dynamically adjusted with the help of the auxiliary task. The attention mechanism calculation formula of the traditional attention module is:

[0086] y=(σ(Wx+b))☉z where x is the feature vector output by each task, W is the weight matrix, b is the bias vector, σ represents the Sigmoid activation function, z is the input feature of each task, and ⊙ represents element-by-element multiplication.

[0087] In the present invention, dynamic adjustment of feature weights is achieved through a cross-task attention mechanism. The attention module calculates weights based on the feature outputs of the target task and the auxiliary task so as to adjust the features during the training process. The expression of this process is as follows:

[0088]

[0089] Where: αi is the weight of the i-th feature, Wa is the weight matrix, and Represent the feature representation of the target task and auxiliary task respectively, Concat represents the connection of the feature representation of the target task and auxiliary task, b is the bias vector, and the Softmax function is used to normalize the result into a weighted probability. By introducing weight calculation between the feature representation of the target task and the auxiliary task, it is possible to dynamically adjust the contribution ratio of features between different tasks, thereby enhancing the adaptability of the target task to the disturbance characteristics and improving the accuracy of photovoltaic output prediction.

[0090] In order to verify the effectiveness and effect of the method of the present invention, this embodiment is verified by experiments, as follows:

[0091] The experimental part selected the output and weather historical data of photovoltaic stations in a certain area of ​​Zhejiang, China as experimental data.

[0092] First, VMD decomposition is performed on the historical output time series to extract and analyze different modes of output changes. In VMD, the parameter alpha controls the balance between the smoothness of decomposition and data fitting. In order to better fit the data, the fitting coefficient alpha is set to 10. When performing VMD operations, the experiment first sets the mode number k to 3, 5, 7, 9, 11, etc., and brings the perturbation IMF obtained with different k into K-means clustering for clustering operations. The number of perturbation IMF clusters with different k is as follows: Figure 3 shown.

[0093] Depend on Figure 3 It can be seen that as the number of VMD decompositions k increases, the number of subsequent clusters tends to be stable starting from k=7. It can be seen that the increase in the number of decompositions has no change or improvement on the clustering effect. Therefore, this experiment sets the modal number k to 7. The VMD decomposition diagram of photovoltaic output is shown in Figure 4 shown.

[0094] according to Figure 4 It can be seen that VMD decomposes the output into 7 characteristic IMFs. Among them, IMF1 is the quasi-clear sky curve of the day, which represents the output curve of the sunny day that is not affected by the weather. IMF4 and IMF5 represent the high-amplitude, low-frequency disturbances in the output caused by changes in weather clouds, while IMF2 and IMF6 represent the high-amplitude, medium-frequency disturbances in the output caused by changes in weather clouds, and IMF3 and IMF7 represent the low-amplitude, high-frequency disturbances in the output caused by changes in weather clouds and contained in the photovoltaic system itself. Residue represents the noise part of the photovoltaic system.

[0095] Next, the K-means clustering operation is performed on the perturbation IMF modes other than the clear-sky IMF. The SC index and DBI index of the cluster are as follows: Figure 5 As shown, the cluster center line is Figure 6 shown.

[0096] according to Figure 5 As shown in the figure, when K = 3, the DBI index and SC index are optimal, and the clustering effect is the best. That is, the disturbance IMF is mainly divided into three categories, and the prediction model is constructed and trained with the same type of IMF time series data in subsequent predictions. Figure 6 The clustering result is the same as the conclusion of VMD decomposition. By identifying different disturbance feature classes, we can explore the changing rules and periodic impact of PV output disturbances, and further analyze the weather change pattern in combination with historical meteorological data to improve the prediction accuracy of PV output.

[0097] After obtaining the K-means clustering results, the clear-sky IMF curve and the cluster center lines of disturbance clusters 1, 2, and 3 were subjected to Spearman correlation analysis with the weather historical data. The correlation analysis results are shown in Table 1.

[0098] Table 1 Spearman correlation analysis results

[0099]

[0100] From the data in Table 1, it can be analyzed that the Spearman correlation between the clear-sky-like IMF and the irradiance is very high, indicating that there is a significant correlation between the clear-sky-like IMF and the irradiance; while the Spearman correlation between the perturbation IMF cluster 1, perturbation IMF cluster 2, and perturbation IMF cluster 3 and each meteorological parameter is relatively low, indicating that their correlation with the meteorological parameters is weak. Specifically, the Spearman correlation coefficient of perturbation IMF cluster 2 is low in all meteorological parameters, and the connection with each meteorological parameter is the weakest. In contrast, the Spearman correlation between perturbation IMF cluster 1 and irradiance and surface temperature is high, and the correlation between perturbation IMF cluster 3 and humidity is also relatively high, that is, the meteorological factors have the greatest impact on perturbation IMF cluster 3. On this basis, the optimal meteorological factors are selected for each power cluster to construct the experimental data set.

[0101] In the prediction stage, different power change feature clusters and their related meteorological combinations are used as auxiliary task training data and input into the CNN-GRU hybrid prediction model, where the target task is set as the photovoltaic output prediction task, and the auxiliary task is the component prediction task of different power change feature clusters and their related meteorological combinations.

[0102] When conducting the experiment, the above content was organized into an experimental data set. The last three days of data in the experimental data set were used as test data, and the remaining data were used as training data. In the comparative experiment, the unclustered VMD-CNN-LSTM prediction model, VMD-CNN-GRU prediction model, and the CNN-GRU prediction model that was directly predicted without decomposition were introduced to conduct comparative tests with the proposed prediction model, and then the differences between the proposed method and other methods were studied. The actual and predicted values ​​of the output are shown in Figure 2. Figure 7 shown.

[0103] according to Figure 7 The prediction curve of the prediction model proposed in the present invention is closest to the true value and has the best prediction effect. This shows that the model of the present invention has higher precision and accuracy in predicting photovoltaic output, and has a better fitting effect with the true value. Its evaluation indicators are shown in Table 2.

[0104] Table 2 Evaluation indicators of prediction results

[0105]

[0106] According to the evaluation index data table provided, the proposed model performs best in predicting photovoltaic output, with the smallest RMSE (27.01W), MAE (10.90W) and MAPE (6.94%), showing the highest prediction accuracy and precision. The VMD-CNN-GRU model is slightly better than the VMD-CNN-LSTM and CNN-GRU models in MAE and MAPE, but still lags behind the proposed model. Comprehensive analysis shows that the proposed model performs best in predicting photovoltaic output, with high precision and accuracy.

[0107] In summary, the photovoltaic output prediction method based on multivariate disturbance feature extraction and auxiliary task learning proposed in the present invention can improve the overall accuracy of photovoltaic output prediction, has high robustness, and can provide data support for photovoltaic station operation and management.

Claims

1. A photovoltaic output prediction method based on multivariate disturbance feature extraction and auxiliary task learning, characterized in that: The steps include: S1: Obtain the location information, panel data, output data and meteorological data of the target PV station, use VMD variational mode decomposition to decompose the historical PV output, and obtain the output disturbance IMFs mode and quasi-clear sky IMF mode; S2: Perform K-means clustering on the IMFs modes representing the output disturbance to construct the power change feature cluster; S3: Spearman correlation analysis of different power change characteristic clusters combined with weather factors; S4: Different power change feature clusters and their related meteorological features are combined as auxiliary task training data and input into the CNN-GRU hybrid prediction model that introduces shared features and cross-task attention mechanism for training to obtain the final photovoltaic output prediction result.

2. A photovoltaic output prediction method based on multivariate disturbance feature extraction and auxiliary task learning according to claim 1, characterized in that: In step S1, VMD variational modal decomposition is used to decompose the historical photovoltaic output. When performing VMD decomposition, the sum of the expected bandwidth of each power IMF mode is required to be minimum, and the constraint condition is that the sum of all decomposed modes is the same as the original output characteristic signal sequence. The specific formula of the constrained variational problem is: Where f is the original time series, δ(t) is the Dirac distribution function, {u k }:={u1,…,u K },{w k }:={w1,…,w K } are the abbreviations of all modes and their center frequencies; e -jωkt for u k The corresponding center frequency w k The index term of ; By introducing the Lagrange operator, the inequality constraint is transformed into an equality constraint and solving the above equation, the mode u is obtained. k The solution formula is: Where α is a quadratic penalty factor, which is used to balance the trade-off between the objective function and the degree of constraint violation. By penalizing the constraints, the algorithm is forced to converge to a solution that satisfies the constraints. λ is the Lagrange multiplication operator. Get the center frequency w k The formula is:

3. The photovoltaic output prediction method based on multivariate disturbance feature extraction and auxiliary task learning according to claim 1 is characterized in that: The method for determining the clear-sky-like IMF in step S1 is: After obtaining the IMF, by setting the thresholds of the standard deviation and the rate of change, if the standard deviation and the rate of change of the IMF are both lower than the set thresholds, it is determined to be a clear sky IMF. The determination formula is: Std(IMF)<Threshold std and ROC(IMF)<Threshold roc Where Std(IMF) is the standard deviation of IMF, which is used to measure the dispersion of IMF data; Threshold std is the standard deviation threshold. When the standard deviation of IMF is less than this threshold, it is considered that the fluctuation of IMF is small. ROC(IMF) is the rate of change of IMF, which is used to measure the speed of change of IMF data. roc is the change rate threshold. When the change rate of IMF is less than this threshold, it is considered that the change of IMF is relatively slow.

4. The photovoltaic output prediction method based on multivariate disturbance feature extraction and auxiliary task learning according to claim 1 is characterized in that: In the step S2, K-means clustering is performed on the IMFs modes representing the output disturbance, and the silhouette coefficient and the Davis-Boulding index are used as evaluation indicators. After the disturbance IMFs are clustered by K-means, the cluster center line is drawn, and the power change characteristic indicators of each output disturbance cluster center are calculated to construct an output change characteristic cluster.

5. A photovoltaic output prediction method based on multivariate disturbance feature extraction and auxiliary task learning according to claim 4, characterized in that: The K-means clustering method in step S2 is: When clustering, initialize the cluster center, determine the number of clusters K, assign samples D to calculate the Euclidean distance between the remaining data points and the cluster center Ci, find the shortest distance and assign all samples to the cluster corresponding to the cluster center Ci. The Euclidean distance calculation is as follows: Where: Ci is the i-th cluster center, m is the dimension of the data object, x j , C ij is the j-th attribute value of x and Ci; Update the cluster center and calculate the average value and square error of all points in each cluster. Update the cluster center and repeat the above steps. The square error calculation is as follows: The silhouette coefficient and Davis-Boulding index are used as evaluation indicators. When the cluster center no longer changes or the maximum number of iterations is reached, the loop stops, the clustering results are updated, the evaluation indicators of each cluster number K are recorded, the optimal value of the evaluation indicator is selected as the cluster number, and its cluster center line is drawn.

6. The photovoltaic output prediction method based on multivariate disturbance feature extraction and auxiliary task learning according to claim 1 is characterized in that: In step S2, after performing K-means mutation clustering, cluster center lines of different time lengths are drawn for each cluster, and the mutation duration, slope and average value of the cluster center are calculated as characteristic indicators to construct a mutation feature sequence.

7. The photovoltaic output prediction method based on multivariate disturbance feature extraction and auxiliary task learning according to claim 1 is characterized in that: The calculation formula of the Spearman correlation coefficient in the Spearman correlation analysis of step S3 is as follows: In the formula, ρ represents the Spearman correlation coefficient, d i represents the difference between each pair of corresponding levels, and n represents the number of data pairs.

8. The photovoltaic output prediction method based on multivariate disturbance feature extraction and auxiliary task learning according to claim 1 is characterized in that: The step S4 specifically includes: First, the target task is set to predict photovoltaic output, and the auxiliary task is to predict the components of different power change feature clusters and their related meteorological combinations. The two use the same input features and the same structure of CNN feature extraction network layer; Next, the representations of the target task and auxiliary tasks pass through the shared fully connected layer in GRU to achieve sharing of learned feature representations among multiple tasks; Finally, by adding an attention module between the feature representation of the target task and the feature representation of the auxiliary task, the feature weight of the target PV output prediction task can be dynamically adjusted with the help of the auxiliary task.

9. The photovoltaic output prediction method based on multivariate disturbance feature extraction and auxiliary task learning according to claim 8 is characterized in that: The attention mechanism calculation formula of the attention module in step S4 is: y=(σ(Wx+b))⊙z Where: x is the feature vector output by each task, W is the weight matrix, b is the bias vector, σ represents the Sigmoid activation function, z is the input feature of each task, and ⊙ represents element-by-element multiplication.

10. The photovoltaic output prediction method based on multivariate disturbance feature extraction and auxiliary task learning according to claim 8, characterized in that: The dynamic adjustment of feature weights in step S4 is achieved through a cross-task attention mechanism; the attention module calculates weights based on the feature outputs of the target task and the auxiliary task so as to adjust the features during the training process. The expression of this process is as follows: Where: α i is the weight of the i-th feature, W a is the weight matrix, and They represent the feature representations of the target task and the auxiliary task respectively, Concat represents the connection of the feature representations of the target task and the auxiliary task, b is the bias vector, and the Softmax function is used to normalize the result into a weighted probability.

Images