Photovoltaic output prediction method based on multivariate feature time-frequency synchronous analysis and attention mechanism

Abstract

The invention discloses a photovoltaic output prediction method based on multivariate feature time-frequency synchronous analysis and an attention mechanism, and the method comprises the following steps: obtaining the output data and meteorological data of a target photovoltaic station, and extracting the most relevant weather features of the output through employing a time-phased MIC correlation analysis method; mVMD multivariate variational mode decomposition is adopted to carry out time-frequency synchronous analysis on the output time sequence in combination with related meteorological characteristics, and a frequency-aligned multivariate intrinsic mode is obtained; an Informer network model based on a Huber loss function and an improved attention mechanism is constructed; according to the corresponding output IMF modal components and the related meteorological modal components, training and predicting the Informer network model to obtain prediction results of different output IMF modal components; and superposing different output IMF mode component prediction results to obtain a photovoltaic output prediction result. According to the method, the MVMD is combined with the Informer network model, so that the accuracy of photovoltaic output prediction can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic power generation, relates to photovoltaic output prediction, and specifically relates to a photovoltaic output prediction method based on multivariate feature time-frequency synchronization analysis and attention mechanism. Background Art

[0002] With the rapid development of the world economy, the demand for energy consumption is getting higher and higher. The shortage of traditional resources and the need for environmental protection have led to a sharp increase in the demand for renewable energy. Among all renewable energy sources, solar energy, as the most typical one, has attracted widespread attention due to its abundance, almost ubiquitous availability, and clean and renewable characteristics, and has gradually developed from an independent power generation mode to a grid-connected power generation mode. However, photovoltaic power generation is easily affected directly or indirectly by a variety of meteorological factors in the region, such as solar irradiance, temperature, and relative humidity, resulting in photovoltaic output having certain characteristics such as randomness, intermittency, and uncertainty. These uncertainties will reduce real-time control performance, reduce system economy, and endanger the stable operation of the power system, thus posing major challenges to the management and operation of power and energy systems. In order to overcome these shortcomings, accurate photovoltaic power prediction is required.

[0003] In terms of prediction methods, photovoltaic power prediction methods can be divided into three categories: 1) Methods based on statistics and probability, such as autoregressive models (AR) and probabilistic methods. Statistical methods are more suitable for ultra-short-term and short-term predictions. 2) Methods based on machine learning, such as k-nearest neighbor (kNN), support vector machine (SVM), gradient boosting decision tree, etc. These methods are also suitable for short-term predictions. 3) Deep learning methods, which can be divided into three methods: convolutional neural network (CNN) method, recurrent neural network (RNN) method, and a hybrid model of the two, where RNN usually includes long short-term memory network (LSTM) and gated recurrent unit (GRU), which can automatically learn from input to.

[0004] However, in the photovoltaic output forecast, traditional deep learning models have difficulty in effectively distinguishing and focusing on meteorological features that play a key role in photovoltaic power changes because they do not introduce an attention mechanism. As a result, the model's performance is relatively limited when dealing with multi-variable and multi-scale features, and it is unable to fully capture complex time dependencies and key features, which in turn affects the accuracy and robustness of the forecast. Summary of the invention

[0005] Purpose of the invention: In order to overcome the deficiencies in the prior art, a photovoltaic output prediction method based on multivariate feature time-frequency synchronization analysis and attention mechanism is provided to improve the accuracy of photovoltaic output prediction.

[0006] Technical solution: To achieve the above purpose, the present invention provides a photovoltaic output prediction method based on multivariate feature time-frequency synchronization analysis and attention mechanism, comprising the following steps:

[0007] S1: Obtain output data and meteorological data of the target photovoltaic station, with a data sampling interval of 15 minutes, and pre-process the collected data;

[0008] The most relevant weather features were extracted using the time-division MIC correlation analysis method;

[0009] S2: MVMD multivariate mode decomposition is used to perform time-frequency synchronization analysis on the output time series combined with relevant meteorological characteristics to obtain frequency-aligned multivariate eigenmodes;

[0010] S3: Construct an Informer network model based on Huber loss function and improved attention mechanism;

[0011] S4: According to the corresponding output IMF modal components and their related meteorological modal components, multiple informer network models are trained and predicted to obtain the prediction results of different output IMF modal components;

[0012] S5: Superimpose the prediction results of IMF modal components with different outputs to obtain the photovoltaic output prediction result.

[0013] Furthermore, the time-divided MIC correlation analysis method in step S1 includes full-time correlation analysis and output period correlation analysis. Full-time correlation analysis covers data for 24 hours a day, which helps to reveal some common and overall weather influencing factors. Given that photovoltaic output is mainly concentrated during the day, especially during the period with the strongest solar irradiance, output period correlation analysis focuses on these key periods and can more directly identify the weather factors that are most critical to photovoltaic output.

[0014] Furthermore, the MIC in step S1 represents the maximum information coefficient, which is used to measure the strength of the correlation between two variables. The formula for its correlation analysis involves the calculation of entropy, and the formula for the time-division MIC correlation analysis method is as follows:

[0015]

[0016] In the formula, X and Y are two variables, L x and L y Represents the number of bins of X and Y, I(XL ,Y M ) represents the mutual information between the Lth sample of X and the Mth sample of Y.

[0017] Furthermore, the core idea of ​​the MVMD multivariate variational mode decomposition in step S2 is to decompose the input multivariate time series data into its inherent multivariate modulation oscillation, and its model is:

[0018]

[0019] In the formula, k is the number of separation modes set in the algorithm; c is the number of original signals input to the algorithm; w k is the center frequency of each mode after MVMD decomposition; x c (t) is the set of input multivariate time series characteristic signals; u k,c (t) is the set of decomposed time series feature signals.

[0020] Furthermore, in order to simplify the calculation process in step S2, the Lagrange multiplier method is used to construct and transform the constrained variation problem of the MVMD model into an unconstrained variation problem. The form of the Lagrange function is as follows:

[0021]

[0022] In the formula, ε is the penalty factor, whose value is a constant; λ(t) is the time-dependent Lagrange multiplier; in the process of transformation, the introduction of penalty factors and Lagrange multipliers will affect the optimization process of the objective function, which may lead to deviations in the final variational mode decomposition results; therefore, the penalty factor or Lagrange multiplier must be properly adjusted in real time during the decomposition operation;

[0023] Solve the above equations to obtain the solutions of each variable:

[0024]

[0025] Where ∧ is Fourier transform; τ is the update coefficient; the required decomposition mode is obtained through continuous iterative update of the algorithm.

[0026] Furthermore, the existing attention mechanism calculates the scaled dot product after inputting (Query, Key), and its self-attention "sparseness" score presents a long-tail distribution, that is, only a small part of the query matrix Q contributes more to the attention. In the Informer model of the present invention, the attention mechanism is improved. The attention calculation method of the improved attention mechanism in step S3 is:

[0027]

[0028] In the formula, is a sparse matrix of the same size as Q that has been filtered by attention and contains only the query matrices that contribute the most to attention.

[0029] Furthermore, in order to improve the accuracy of model training without sacrificing training speed during training, the improved Informer network model of the Huber loss function is used to train the photovoltaic output prediction model. The Huber loss function has hyperparameters, which can be adapted to different types of output prediction model training by setting appropriate hyperparameters δ. When the prediction deviation is less than δ, a loss calculation method similar to MSE is used to improve the model convergence speed when the gradient of the loss is small; when the prediction deviation is greater than δ, a linear error similar to MAE is used to reduce the impact of large loss values ​​caused by outliers on model training, thereby improving the robustness of the model to outliers. Therefore, the calculation formula of the Huber loss function in step S3 is:

[0030]

[0031] In the formula, m is the number of training samples, y i is the actual output value, is the predicted value of the model output, and δ is the hyperparameter of the Huber loss function.

[0032] Beneficial effects: Compared with the prior art, the present invention realizes the time-frequency synchronization analysis of photovoltaic output by combining multivariate variational mode decomposition and the Informer model based on the attention mechanism, and significantly improves the prediction accuracy. The multivariate time series synchronization analysis of MVMD ensures the consistency of the prediction results of each mode, so that the features at each time scale can be effectively captured. In addition, the improved attention mechanism can effectively screen key features and enhance the model's ability to capture long-term dependencies. The use of the Huber loss function improves the robustness of the model to outliers, reduces the impact of errors, and thus improves the overall prediction effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is a flow chart of the prediction method of the present invention;

[0034] Figure 2 It is a framework diagram of the prediction method of the present invention;

[0035] Figure 3 It is the IMF diagram after MVMD decomposition;

[0036] Figure 4 It is the IMF frequency distribution diagram after MVMD decomposition;

[0037] Figure 5 It is a comparison chart of the actual and predicted values ​​of photovoltaic output on the day to be predicted;

[0038] Figure 6 It is the box plot of the photovoltaic output error on the day to be predicted. DETAILED DESCRIPTION

[0039] 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.

[0040] like Figure 1 and Figure 2 As shown, the present invention provides a photovoltaic output prediction method based on multivariate feature time-frequency synchronization analysis and attention mechanism, comprising the following steps:

[0041] S1: Obtain the output data and meteorological data of the target photovoltaic station, with a data sampling interval of 15 minutes, and preprocess the collected data. The focus of preprocessing includes the compensation 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 the prediction model;

[0042] The time-division MIC correlation analysis method is used to extract the most relevant weather features for output: The time-division MIC correlation analysis method includes full-time correlation analysis and output time correlation analysis. Full-time correlation analysis covers data for 24 hours a day, which helps to reveal some common and overall weather influencing factors. Given that photovoltaic output is mainly concentrated during the day, especially during the period with the strongest solar irradiance, the output time correlation analysis focuses on these key periods and can more directly identify the weather factors that are most critical to photovoltaic output;

[0043] MIC stands for Maximum Information Coefficient, which is used to measure the strength of the correlation between two variables. The formula for its correlation analysis involves the calculation of entropy. The formula for the time-division MIC correlation analysis method is as follows:

[0044]

[0045] In the formula, X and Y are two variables, L x and L y Represents the number of bins of X and Y, I(X L ,Y M ) represents the mutual information between the Lth sample of X and the Mth sample of Y.

[0046] S2: MVMD multivariate mode decomposition is used to perform time-frequency synchronization analysis on the output time series combined with relevant meteorological characteristics to obtain frequency-aligned multivariate eigenmodes;

[0047] The core idea of ​​MVMD multivariate variational mode decomposition is to decompose the input multivariate time series data into its inherent multivariate modulation oscillation. Its model is:

[0048]

[0049] In the formula, k is the number of separation modes set in the algorithm; c is the number of original signals input to the algorithm; w k is the center frequency of each mode after MVMD decomposition; x c (t) is the set of input multivariate time series characteristic signals; u k,c (t) is the set of decomposed time series feature signals.

[0050] In order to simplify the calculation process, the Lagrange multiplier method is used to construct the constrained variation problem of the MVMD model into an unconstrained variation problem. The form of the Lagrange function is as follows:

[0051]

[0052] In the formula, ε is the penalty factor, whose value is a constant; λ(t) is the time-dependent Lagrange multiplier; in the process of transformation, the introduction of penalty factors and Lagrange multipliers will affect the optimization process of the objective function, which may lead to deviations in the final variational mode decomposition results; therefore, the penalty factor or Lagrange multiplier must be properly adjusted in real time during the decomposition operation. The adjustment process is as follows: First, the penalty factor ε is adjusted to control the difference between the decomposed modes. If the decomposition result deviates too much from the constraint condition, ε is appropriately increased to strengthen the constraint; if the decomposition result gradually meets the constraint condition, ε can be reduced to increase the flexibility of decomposition. Secondly, the update of the Lagrange multiplier λ(t) is carried out through a recursive relationship, and it is dynamically adjusted in each iteration according to the difference between the current decomposition result and the original signal. The update formula is:

[0053]

[0054] Where τ is the update step size. This formula shows that the Lagrange multiplier will be adjusted in each iteration according to the difference between the current decomposed signal and the original signal, thereby continuously optimizing the decomposition result during the iteration process to ensure the accuracy and convergence of the modal decomposition;

[0055] Solve the above equations to obtain the solutions of each variable:

[0056]

[0057] Where ∧ is Fourier transform; τ is the update coefficient; through the continuous iterative update of the algorithm, the required decomposition mode is obtained, that is, the corresponding output IMF modal component and its related meteorological modal component.

[0058] S3: Construct an Informer network model based on Huber loss function and improved attention mechanism;

[0059] The existing attention mechanism calculates the scaled dot product after inputting (Query, Key), and its self-attention "sparseness" score presents a long-tail distribution, that is, only a small part of the query matrix Q contributes more to the attention. In the Informer model of the present invention, the attention mechanism is improved, and the attention calculation method of the improved attention mechanism is:

[0060]

[0061] In the formula, is a sparse matrix of the same size as Q that has been filtered by attention and contains only the query matrices that contribute the most to attention.

[0062] The Informer network model in this invention continues the encoder and decoder structure in Transformer:

[0063] The encoder receives long sequence input and obtains feature representation through the multi-head probabilistic sparse self-attention module and the self-attention distillation module. The multi-head probabilistic sparse self-attention mechanism uses a sparse matrix to replace the original attention matrix, which greatly reduces the computing power requirements while maintaining good performance.

[0064] The decoder receives long sequence input, interacts with the encoded features through multi-head attention, and finally directly predicts the output target part. The one-shot generative prediction method used by Informer in the present invention is faster than the step-by-step reasoning method and has comparable effect.

[0065] In order to improve the accuracy of model training without sacrificing training speed during training, the improved Informer network model of the Huber loss function is used to train the photovoltaic output prediction model. The Huber loss function has hyperparameters, which can be adapted to different types of output prediction model training by setting appropriate hyperparameters δ. When the prediction deviation is less than δ, a loss calculation method similar to MSE is used to improve the model convergence speed when the gradient of the loss is small; when the prediction deviation is greater than δ, a linear error similar to MAE is used to reduce the impact of large loss values ​​caused by outliers on model training and improve the robustness of the model to outliers. Therefore, the calculation formula of the Huber loss function is:

[0066]

[0067] In the formula, m is the number of training samples, y i is the actual output value, is the predicted value of the model output, δ is the hyperparameter of the Huber loss function, and in this embodiment, δ is set to 0.5.

[0068] S4: Based on the corresponding k output IMF modal components and their related meteorological modal components obtained by MVMD, k Informer network models are constructed for training and prediction to obtain the prediction results of different output IMF modal components;

[0069] S5: Superimpose the prediction results of IMF modal components with different outputs to obtain the photovoltaic output prediction result.

[0070] In order to verify the effectiveness and practical effect of the method of the present invention, this embodiment is verified by experiments, specifically:

[0071] The experimental part selected the photovoltaic output and weather historical data of a certain area in Zhejiang, China as experimental data. Among them, the historical photovoltaic output and meteorological data are the historical data of the whole year from 0:00 on January 1, 2023 to 23:55 on December 31, 2023, with a sampling frequency of 15min / point, and the output data unit is MW.

[0072] First, the MIC correlation analysis is carried out in different time periods, namely the full-time correlation analysis and the output period correlation analysis. The MIC correlation coefficients are shown in Table 1.

[0073] Table 1 Photovoltaic output MIC correlation coefficient table

[0074]

[0075]

[0076] As shown in Table 1, cloud cover and photovoltaic output show moderate correlation in both the full time period and the output period. In contrast, the correlation in the output period is slightly higher, indicating that cloud cover has a more significant impact during photovoltaic power generation. Solar irradiance is highly correlated with photovoltaic output in both the full time period and the output period, especially in the full time period, the correlation coefficient is more than 0.9, which reflects that solar irradiance is the decisive factor for photovoltaic output. At the same time, the correlation between meteorological variables and photovoltaic output in different seasons is significantly different. For example, the correlation between relative humidity and photovoltaic output is significantly enhanced in summer (July-September), while it is relatively weak in winter. This may be related to the fact that the high humidity environment in summer will reduce the power generation efficiency of photovoltaic panels. After comprehensive consideration and analysis, total cloud cover, irradiance, relative humidity and surface temperature are selected as relevant weather characteristics of photovoltaic output, so as to carry out the next step of MVMD decomposition and prediction process.

[0077] Next, MVMD decomposition is performed in combination with relevant weather characteristics and output time series to obtain frequency-aligned multivariate intrinsic modes. Considering that when the number of decompositions exceeds 7, the central frequencies of IMFs will be too similar, and too many decompositions will increase the computational burden of the model, the number of MVMD decompositions K is 7. Its multivariate modal diagram and FFT frequency diagram are shown as follows: Figure 3 , Figure 4 shown.

[0078] according to Figure 3 , photovoltaic output, total cloud cover, irradiance, relative humidity and surface temperature are input into MVMD as five channels for decomposition, and each channel consists of 7 IMF features. Unlike VMD, the IMF of each channel is a characteristic decomposition diagram of multidimensional data in time and space dimensions. IMF1 is the main part of the time series decomposition of each channel, which retains the main characteristic information. IMF2 to IMF7 of each channel are different frequency components, and their frequencies decrease successively, representing lower frequency components or long-term changes on different time scales. At the same time, Figure 3 , Figure 4 It can be seen that the same IMF number of different characteristic channels has the same frequency distribution, which shows that since MVMD can perform multivariate decomposition, the center frequencies of IMF1~IMF7 obtained by channel decomposition of photovoltaic power and its meteorological factors are the same, thus realizing time-frequency synchronization analysis and ensuring that the prediction model is established at the same center frequency, that is, the same IMF number of each different channel represents a different time scale, which provides rich information for the model at each time scale and helps to improve the prediction accuracy of IMFs with different outputs.

[0079] In the prediction stage, an experimental data set is constructed for each output IMF combined with its weather multivariate intrinsic mode of the same frequency, and it is input into the Informer prediction model based on the Huber loss function and attention mechanism for training and prediction to obtain the prediction results of each output mode, and the total prediction result of photovoltaic output is obtained by superposition. In this experiment, 80% of the experimental data set is used as the training set, 10% as the validation set, and 10% as the prediction set.

[0080] In the comparative experiment, the undecomposed Informer prediction model, CNN-LSTM prediction model, WaveNet prediction model, and the MVMD-CNN-LSTM prediction model, MVMD-WaveNet prediction model and the proposed MVMD-Informer prediction model after MVMD decomposition were introduced for comparative experiments, and then the differences between the proposed method and other methods were studied. Figure 5 The error box plot is shown in Figure 6 shown.

[0081] according to Figure 5 , Figure 6 It can be seen that the prediction curve of the proposed MVMD-Informer prediction model is closest to the true value and has the best prediction effect. For the prediction effect of different neural networks, the prediction effect after adding MVMD is better than that without adding it. This shows that MVMD can realize the time-frequency synchronization analysis of photovoltaic power and meteorological series and complete multivariate decomposition, so it is more in line with the actual operation of distributed photovoltaic stations and improves the prediction accuracy. At the same time, Informer is more advantageous than other models in terms of prediction accuracy and model size. The evaluation indicators of the prediction results are shown in Table 2.

[0082] Table 2 Evaluation indicators of prediction results

[0083]

[0084] According to the evaluation index data table 2, the MVMD-Informer model proposed in the present invention performs best in predicting photovoltaic output, with the lowest RMSE (0.939MW), MAE (0.758MW) and MAPE (4.31%). In particular, compared with the Informer model, MVMD-Informer reduced RMSE by 62.48%, MAE by 52.18%, and MAPE by 20.92%. Compared with the CNN-LSTM model, the performance improvement is more significant, with RMSE reduced by 81.67%, MAE reduced by 79.92%, and MAPE reduced by 43.40%. This means that the MVMD-Informer model combines the time series decomposition capability of MVMD and the self-attention mechanism of Informer, thereby more effectively capturing the dynamic characteristics and long-term dependencies of photovoltaic output.

Claims

1. A photovoltaic output prediction method based on multivariate feature time-frequency synchronization analysis and attention mechanism, characterized in that: The steps include: S1: Obtain the output data and meteorological data of the target PV station, and use the time-division MIC correlation analysis method to extract the weather characteristics most relevant to the output; S2: MVMD multivariate mode decomposition is used to perform time-frequency synchronization analysis on the output time series combined with relevant meteorological characteristics to obtain frequency-aligned multivariate eigenmodes; S3: Construct an Informer network model based on Huber loss function and improved attention mechanism; S4: According to the corresponding output IMF modal components and their related meteorological modal components, the Informer network model is trained and predicted to obtain the prediction results of different output IMF modal components; S5: Superimpose the prediction results of IMF modal components with different outputs to obtain the photovoltaic output prediction result.

2. A photovoltaic output prediction method based on multivariate feature time-frequency synchronization analysis and attention mechanism according to claim 1, characterized in that: The time-division MIC correlation analysis method in step S1 includes full-time correlation analysis and output time period correlation analysis.

3. The photovoltaic output prediction method based on multivariate feature time-frequency synchronization analysis and attention mechanism according to claim 1 is characterized in that: The formula of the time-division MIC correlation analysis method in step S1 is as follows: In the formula, X and Y are two variables, L x and L y Represents the number of bins of X and Y, I(X L ,Y M ) represents the mutual information between the Lth sample of X and the Mth sample of Y.

4. The photovoltaic output prediction method based on multivariate feature time-frequency synchronization analysis and attention mechanism according to claim 1 is characterized in that: In step S2, the MVMD multivariate variational modal decomposition decomposes the input multivariate time series data into its inherent multivariate modulation oscillation, and its model is: In the formula, k is the number of separation modes set in the algorithm; c is the number of original signals input to the algorithm; w k is the center frequency of each mode after MVMD decomposition; x c (t) is the set of input multivariate time series characteristic signals; u k,c (t) is the set of time series characteristic signals after decomposition.

5. The photovoltaic output prediction method based on multivariate feature time-frequency synchronization analysis and attention mechanism according to claim 4 is characterized in that: In step S2, the Lagrange multiplier method is used to construct and transform the constrained variation problem of the MVMD model into an unconstrained variation problem. The form of the Lagrange function is as follows: In the formula, ε is the penalty factor, and its value is a constant; λ(t) is the time-dependent Lagrange multiplier. In the process of transformation, the introduction of penalty factors and Lagrange multipliers will affect the optimization process of the objective function, which may lead to deviations in the final variational mode decomposition results. Therefore, the penalty factors or Lagrange multipliers must be properly adjusted in real time during the decomposition operation. Solve the above equations to obtain the solutions of each variable: Where ∧ is Fourier transform; τ is the update coefficient; the required decomposition mode is obtained through continuous iterative update of the algorithm.

6. A photovoltaic output prediction method based on multivariate feature time-frequency synchronization analysis and attention mechanism according to claim 5, characterized in that: The update of the Lagrange multiplier λ(t) in step S2 is performed through a recursive relationship, and is dynamically adjusted in each iteration according to the difference between the current decomposition result and the original signal. The update formula is: where τ is the update step size.

7. The photovoltaic output prediction method based on multivariate feature time-frequency synchronization analysis and attention mechanism according to claim 1 is characterized in that: The attention calculation method of the improved attention mechanism in step S3 is: In the formula, is a sparse matrix of the same size as Q that has been filtered by attention and contains only the query matrices that contribute the most to attention.

8. The photovoltaic output prediction method based on multivariate feature time-frequency synchronization analysis and attention mechanism according to claim 1 is characterized in that: The calculation formula of the Huber loss function in step S3 is: In the formula, m is the number of training samples, y i is the actual output value, is the predicted value of the model output, and δ is the hyperparameter of the Huber loss function.

9. The photovoltaic output prediction method based on multivariate feature time-frequency synchronization analysis and attention mechanism according to claim 1 is characterized in that: The Informer network model in step S3 includes an encoder and a decoder structure: The encoder receives long sequence input and obtains feature representation through a multi-head probabilistic sparse self-attention module and a self-attention distillation module; The multi-head probabilistic sparse self-attention mechanism uses a sparse matrix to replace the original attention matrix, which greatly reduces the computing power requirements while maintaining good performance. The decoder receives long sequence input, interacts with the encoded features through multi-head attention, and finally directly predicts the output target part.

Images