Solar radiation prediction method and system fusing time domain-space-frequency domain characteristics

Abstract

The invention discloses a solar radiation prediction method and system fusing time domain-space-frequency domain characteristics, and relates to the technical field of energy. Comprising the following steps: acquiring multi-site solar radiation historical data, and decomposing the multi-site solar radiation historical data into multiple groups of intrinsic mode function data; analyzing characteristics of the intrinsic mode function data of the target site, and classifying the intrinsic mode function data into high-frequency component data and low-frequency component data; constructing a high-frequency prediction model based on the high-frequency component data, constructing a low-frequency prediction model based on the low-frequency component data, and performing targeted training on the constructed high-frequency prediction model and low-frequency prediction model to obtain a prediction result of each intrinsic mode function data; and performing linear superposition on the prediction result of each eigenmode function data, and filling a negative value with zero to obtain a final prediction result. According to the method, the time domain-space-frequency domain data features can be fully captured, and comprehensive utilization and accurate and reliable prediction of various information contained in the solar radiation data are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of energy, and in particular, to a solar radiation prediction method and system integrating time-domain - space - frequency domain features. Background Art

[0002] Converting solar energy into electrical energy can effectively reduce the dependence on traditional fossil fuels and plays a crucial role in alleviating the energy crisis and promoting the transformation of the global energy structure. However, due to the dynamic characteristics of solar radiation, such as significant instability and strong volatility, solar energy is usually regarded as an intermittent energy source. In this case, integrating this energy into the traditional power grid usually poses great challenges to the stability and security of the operation of the energy system. One of the effective measures to solve this problem is to implement accurate solar radiation prediction. At the same time, this processing method can promote the optimal design of solar energy systems. Therefore, accurate and reliable solar radiation prediction is of great significance.

[0003] Currently, solar radiation prediction methods are mainly divided into three categories, namely, methods based on statistical models, methods based on machine learning, and methods based on deep learning. Among them, statistical models usually belong to linear models, but solar radiation data usually exhibits strong non-linear characteristics. In this case, their prediction accuracy will inevitably be greatly reduced. Machine learning models can be divided into shallow learning models and deep learning models according to the complexity of the structure. In contrast, the latter can learn deeper features in solar radiation data through high-dimensional representation, so it has better generalization ability and applicability. In deep learning models, they can be further divided into single models and hybrid models. Among them, single models are usually limited to capturing one type of non-linear feature, so they cannot effectively process data with multiple non-linear features. Although traditional hybrid models can capture complex patterns in time series, they often ignore model diversity, and due to the redundant representation of the same features by model components, it is easy to generate the problem of multicollinearity. In addition, these methods do not fully utilize frequency domain information and are also difficult to explain the non-steady characteristics in solar radiation data, thus limiting their prediction accuracy. To solve the non-steady problem, various signal preprocessing techniques have been developed currently, including univariate and multivariate decomposition techniques. Among them, univariate decomposition techniques cannot consider the coupling relationship between multiple variables and often have problems such as endpoint effects and mode mixing. Although multivariate signal decomposition techniques have achieved the extension from univariate data to multivariate data, they still face the problems existing in univariate decomposition techniques, and their ability to preprocess multivariate data is limited.

[0004] Therefore, it is an urgent problem for those skilled in the art to propose a solar radiation prediction method and system integrating time-domain - space - frequency domain features to solve the difficulties existing in the prior art. Summary of the Invention

[0005] In view of this, the present invention provides a solar radiation prediction method and system that integrates time-domain, space-domain, and frequency-domain features, which is conducive to fully capturing the data features of the time-domain, space-domain, and frequency-domain, and realizes the comprehensive utilization and accurate and reliable prediction of various information contained in solar radiation data.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A solar radiation prediction method that integrates time-domain, space-domain, and frequency-domain features, comprising the following steps:

[0008] S1. Obtain data: Obtain historical solar radiation data of multiple stations, and decompose the historical solar radiation data of multiple stations into multiple groups of intrinsic mode function data;

[0009] S2. Data classification: Analyze the characteristics of the intrinsic mode function data of the target station, and classify the intrinsic mode function data into high-frequency component data and low-frequency component data;

[0010] S3. Model prediction: Based on the high-frequency component data, construct a high-frequency prediction model, and based on the low-frequency component data, construct a low-frequency prediction model. Targeted training is performed on the constructed high-frequency prediction model and low-frequency prediction model to obtain the prediction results of each intrinsic mode function data;

[0011] S4. Result acquisition: Linearly superimpose the prediction results of each intrinsic mode function data and fill in negative values with zeros to obtain the final prediction result.

[0012] Optionally, the specific content of obtaining the historical solar radiation data of multiple stations and decomposing the historical solar radiation data of multiple stations into multiple groups of intrinsic mode function data in S1 is:

[0013] Synchronously decompose the historical solar radiation data of the target station and the corresponding time points of adjacent stations into multiple groups of intrinsic mode function data through multivariate fast iterative filtering.

[0014] Optionally, the specific content of analyzing the characteristics of the intrinsic mode function data of the target station and classifying the intrinsic mode function data into high-frequency component data and low-frequency component data in S2 is:

[0015] Use permutation entropy to analyze the characteristics of the intrinsic mode function data of the target station, and classify the intrinsic mode function data with an entropy value greater than the entropy value of the original data as high-frequency component data, and classify the intrinsic mode function data with an entropy value less than or equal to the entropy value of the original data as low-frequency component data.

[0016] Optionally, the specific content of constructing a high-frequency prediction model based on the high-frequency component data, constructing a low-frequency prediction model based on the low-frequency component data, and performing targeted training on the constructed high-frequency prediction model and low-frequency prediction model to obtain the prediction results of each intrinsic mode function data in S3 is:

[0017] Based on the high-frequency component data, a high-frequency prediction model that combines cross-attention and conditional parametric convolution is constructed, and based on the low-frequency component data, a low-frequency prediction model that combines cross-attention and long short-term memory network is constructed;

[0018] Targeted training is performed on the constructed high-frequency prediction model and low-frequency prediction model to obtain the trained high-frequency prediction model and low-frequency prediction model;

[0019] Use the trained high-frequency prediction model and low-frequency prediction model to perform predictions and output the corresponding prediction results.

[0020] Optionally, the specific content of linearly superimposing the prediction results of each intrinsic mode function data and padding negative values with zeros in S4 to obtain the final prediction result is as follows:

[0021] Linearly superimpose the prediction results of each intrinsic mode function data to obtain an ensemble sequence As follows:

[0022]

[0023] In the formula, u S,j represents the predicted value of the jth intrinsic mode function of the target site S, j is a positive integer, k represents the number of modes. Subsequently, by padding the meaningless negative solar radiation data in the ensemble sequence with zeros, the final prediction result of the target site is obtained.

[0024] A solar radiation prediction system that fuses time-domain - space-domain - frequency-domain features, applying a solar radiation prediction method that fuses time-domain - space-domain - frequency-domain features according to any one of the above, includes: a data acquisition module, a data classification module, a model prediction module, and a result acquisition module;

[0025] The data acquisition module, connected to the input end of the data classification module, is used to acquire the historical solar radiation data of multiple sites and decompose the historical solar radiation data of multiple sites into multiple groups of intrinsic mode function data;

[0026] The data classification module, connected to the input end of the model prediction module, is used to analyze the characteristics of the intrinsic mode function data of the target site and classify the intrinsic mode function data into high-frequency component data and low-frequency component data;

[0027] The model prediction module, connected to the input end of the result acquisition module, is used to construct a high-frequency prediction model based on the high-frequency component data, a low-frequency prediction model based on the low-frequency component data, perform targeted training on the constructed high-frequency prediction model and low-frequency prediction model, and obtain the prediction results of each intrinsic mode function data;

[0028] A result acquisition module, which is used to linearly superimpose the prediction results of each intrinsic mode function data and fill the negative values with zeros to obtain the final prediction result.

[0029] As can be seen from the above technical solutions, compared with the prior art, the present invention provides a solar radiation prediction method and system that fuses time-domain, space-domain, and frequency-domain features, and has the following beneficial effects:

[0030] (1) The method of the present invention has stronger generalization ability and robustness compared with other methods. Compared with traditional signal preprocessing techniques, MvFIF has stronger decomposition ability and higher modal alignment, significantly alleviates the modal aliasing problem, and effectively captures the spatio-temporal and time-frequency correlations between multivariate data;

[0031] (2) The present invention uses EFA to divide the subsequence corresponding to the target station into two components with different frequency distributions, which is beneficial to selecting a suitable model for classification prediction. Then, combined with model diversity, two cross-attention assisted algorithms are designed as predictors. In two hybrid model architectures, two algorithms with different modeling principles (i.e., CPC and LSTM) are used as basic predictors for classification prediction, and CA is added to each of these two basic predictors to consider the time-frequency important information in the components, further improving their prediction ability;

[0032] (3) Through the signal decomposition of MvFIF, the subsequence classification of EFA, and the CA-CPC model and CA-LSTM model of the present invention, the data characteristics of the time-domain, space-domain, and frequency-domain can be fully captured, realizing the comprehensive utilization of various information contained in solar radiation data and accurate and reliable prediction. Description of the Drawings

[0033] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.

[0034] Figure 1 It is a flowchart of a solar radiation prediction method that fuses time-domain, space-domain, and frequency-domain features provided by the present invention;

[0035] Figure 2 It is a schematic diagram of the measured solar radiation data of each meteorological station provided by the embodiment of the present invention;

[0036] Figure 3Schematic diagram of the intrinsic mode functions (IMFs) of each site obtained by MvFIF decomposition provided by the embodiments of the present invention. Among them, 3a is the schematic diagram of the intrinsic mode function of Site C1, 3b is the schematic diagram of the intrinsic mode function of Site C2, 3c is the schematic diagram of the intrinsic mode function of Site C3, and 3d is the schematic diagram of the intrinsic mode function of Site C4;

[0037] Figure 4 Schematic diagram of the CA-CPC model structure of the present invention;

[0038] Figure 5 Schematic diagram of the CA-LSTM model structure of the present invention;

[0039] Figure 6 Loss curve diagram of the training process of the models corresponding to each IMF provided by the embodiments of the present invention;

[0040] Figure 7 Schematic diagram of the final predicted values of each IMF provided by the embodiments of the present invention. Among them, 7a is the schematic diagram of the final predicted value of IMF1, 7b is the schematic diagram of the final predicted value of IMF2, 7c is the schematic diagram of the final predicted value of IMF3, and 7d is the schematic diagram of the final predicted value of IMF4;

[0041] Figure 8 Schematic diagram of the final prediction result of the target site provided by the embodiments of the present invention;

[0042] Figure 9 Schematic diagram of the predicted values of all the participating comparison methods provided by the embodiments of the present invention. Among them, 9a is the schematic diagram of the predicted values of Method 2, Method 3, Method 4, Method 5 and the method of the present invention, and 9b is the schematic diagram of the predicted values of Method 6, Method 7, Method 8, Method 9 and the method of the present invention;

[0043] Figure 10 Comparison chart of the evaluation indicators of all the participating comparison methods provided by the embodiments of the present invention. Among them, 10a is the comparison chart of the RSME evaluation indicators of all the participating comparison methods, 10b is the comparison chart of the MAE evaluation indicators of all the participating comparison methods, and 10c is the comparison chart of the RAE evaluation indicators of all the participating comparison methods. Detailed implementation manners

[0044] 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 only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0045] Refer to Figure 1As shown, the present invention discloses a solar radiation prediction method that fuses time-domain, space-domain, and frequency-domain features, including the following steps:

[0046] S1. Data acquisition: Obtain historical solar radiation data from multiple stations, and decompose the historical solar radiation data from multiple stations into multiple groups of intrinsic mode function data;

[0047] S2. Data classification: Analyze the characteristics of the intrinsic mode function data of the target station, and classify the intrinsic mode function data into high-frequency component data and low-frequency component data;

[0048] S3. Model prediction: Based on the high-frequency component data, construct a high-frequency prediction model, and based on the low-frequency component data, construct a low-frequency prediction model. Perform targeted training on the constructed high-frequency prediction model and low-frequency prediction model to obtain the prediction results of each group of intrinsic mode function data;

[0049] S4. Result acquisition: Linearly superimpose the prediction results of each group of intrinsic mode function data and fill in the negative values with zeros to obtain the final prediction result.

[0050] Further, the specific content of obtaining the historical solar radiation data from multiple stations and decomposing the historical solar radiation data from multiple stations into multiple groups of intrinsic mode function data in S1 is as follows:

[0051] Synchronously decompose the historical solar radiation data at the corresponding time points of the target station and adjacent stations into multiple groups of intrinsic mode function data through multivariate fast iterative filtering.

[0052] Further, the specific content of analyzing the characteristics of the intrinsic mode function data of the target station and classifying the intrinsic mode function data into high-frequency component data and low-frequency component data in S2 is as follows:

[0053] Use permutation entropy to analyze the characteristics of the intrinsic mode function data of the target station, and classify the intrinsic mode function data with an entropy value greater than the entropy value of the original data as high-frequency component data, and classify the intrinsic mode function data with an entropy value less than or equal to the entropy value of the original data as low-frequency component data.

[0054] Further, the specific content of constructing a high-frequency prediction model based on the high-frequency component data, constructing a low-frequency prediction model based on the low-frequency component data, and performing targeted training on the constructed high-frequency prediction model and low-frequency prediction model to obtain the prediction results of each group of intrinsic mode function data in S3 is as follows:

[0055] Based on the high-frequency component data, construct a high-frequency prediction model that mixes cross-attention and conditional parametric convolution. Based on the low-frequency component data, construct a low-frequency prediction model that mixes cross-attention and long short-term memory network;

[0056] Targeted training is performed on the constructed high-frequency prediction model and low-frequency prediction model to obtain the trained high-frequency prediction model and low-frequency prediction model;

[0057] Use the trained high-frequency prediction model and low-frequency prediction model to perform predictions and output the corresponding prediction results.

[0058] Furthermore, in S4, the specific content of linearly superimposing the prediction results of each intrinsic mode function data and filling negative values with zeros to obtain the final prediction result is as follows:

[0059] Linearly superimpose the prediction results of each intrinsic mode function data to obtain an ensemble sequence As follows:

[0060]

[0061] In the formula, u S,j represents the predicted value of the j-th intrinsic mode function of the target site S, j is a positive integer, k represents the number of modes. Subsequently, by filling the meaningless negative solar radiation data in the ensemble sequence with zeros, the final prediction result of the target site is obtained.

[0062] In a specific embodiment, taking the solar radiation data measured at four sites (named C1, C2, C3, and C4 respectively) in a certain area as an example, specific prediction operations and effect verification of the present invention are carried out. The present invention uses the solar radiation measurement data with a sampling frequency of 1 / 3600 Hz as experimental data. The data collection time period for each site is from 0:00 on March 1, 2023 to 23:00 on May 1, 2023. Each site contains 1488 data points, that is Among them, represents the data of site C i at time point C. The measured data of each site is as Figure 2 shown. Assuming that site C1 is the target site, the historical data is used as the training set, and the future data is used as the test set.

[0063] Through multivariate fast iterative filtering (MvFIF), the historical solar radiation data of the target site and adjacent sites at corresponding time points are synchronously decomposed into multiple groups of intrinsic mode functions (IMFs) with the same decomposition layer number and almost the same frequency domain distribution. It should be noted that only the corresponding decomposition results of the target site (i.e., the site to be predicted) are retained for subsequent analysis. The parameter settings of MvFIF are: the stopping parameter is taken as 0.001; the maximum number of iterations is taken as 200. The data of each site is decomposed into 4 IMFs through decomposition, and the decomposition results are as Figure 3As shown, where 3a is the schematic diagram of the intrinsic mode function of Site C1, 3b is the schematic diagram of the intrinsic mode function of Site C2, 3c is the schematic diagram of the intrinsic mode function of Site C3, and 3d is the schematic diagram of the intrinsic mode function of Site C4. From Figure 3 It can be seen that the IMFs with the same decomposition level at different sites have a high degree of consistency, that is, they have similar change trends. To prove the superiority of MvFIF, time-frequency decomposition comparisons were made with noise-assisted multivariate empirical mode decomposition (NA-MEMD) and empirical mode decomposition (EMD), and the parameter settings of these decomposition techniques are listed in Table 1. Among them, the bandwidth of each IMF is defined as the mean ± standard deviation of the corresponding instantaneous frequency; the bandwidth overlap rate is defined as the ratio of the overlapping bandwidth of two adjacent IMFs to the total bandwidth of the two. The lower the bandwidth overlap rate, the less severe the mode mixing. Different from the above frequency-domain analysis, the time-domain analysis measures the mode consistency between the IMF of the target site and the corresponding IMF of the adjacent site through the Spearman correlation coefficient. The value range of the correlation coefficient is [-1, 1], and the larger the absolute value, the stronger the mode consistency. Tables 2 and 3 list the bandwidth overlap rate and mode consistency analysis of different decomposition techniques respectively. It can be seen from Table 2 that the degree of mode mixing of MvFIF is the lowest, with a maximum overlap rate of only 3.16%, while the mode mixing of NA-MEMD and EMD is more severe, with maximum overlap rates reaching 53.50% and 47.43% respectively. Table 3 shows that MvFIF also has the highest level of mode consistency in the time domain. Specifically, the correlation coefficients of the MvFIF decomposition results are between 0.9484 and 0.9996, and the average value of the correlation coefficients varies between 0.9754 and 0.9992, significantly higher than the other two decomposition methods. Therefore, MvFIF shows excellent performance in both the time domain and the frequency domain for preprocessing multivariate signals.

[0064] Table 1 Parameter settings of different decomposition techniques

[0065]

[0066] Table 2 Bandwidth overlap rate between adjacent IMFs

[0067]

[0068] Table 3 Time-domain correlation analysis of IMFs between the target site and other sites

[0069]

[0070] Calculate the permutation entropy of the original data of the target site and the corresponding IMFs, and classify the IMFs with entropy values greater than that of the original data as high-frequency components, and the IMFs with entropy values less than or equal to that of the original data as low-frequency components.

[0071] Through permutation entropy calculation, the entropy value of the original data is 0.5780, while the entropy values of IMF1 to IMF4 are 0.7751, 0.5789, 0.5305, and 0.4565 respectively. Therefore, IMF1 and IMF2 can be regarded as high-frequency components, and a high-frequency prediction model (CA-CPC model) that combines cross-attention and conditional parametric convolution is selected as the predictor for IMF1 and IMF2. For IMF3 and IMF4, their entropy values are lower than that of the original data, so they are considered low-frequency components, and a low-frequency prediction model (CA-LSTM model) that combines cross-attention and long short-term memory network is constructed accordingly.

[0072] For IMF1 and IMF2, the corresponding CA-CPC models are constructed, while for IMF3 and IMF4, the corresponding CA-LSTM models are constructed, where the CA-CPC structure is as Figure 4 shown, and the CA-LSTM structure is as Figure 5 shown. Figure 4 and Figure 5 In it, FFT is the fast Fourier transform; Q, K, and V are the query vector, key vector, and value vector respectively; W Q , W K , W V are the weight matrices corresponding to Q, K, and V respectively; A is the attention weight matrix; Softmax(·), Relu(·), and Sigmoid(·) are three different types of activation functions. Figure 6 shows the training process of the models corresponding to each IMF. It can be seen that all the loss function curves converge rapidly. When the number of training epochs reaches 100, the mean square error (MSE) values of the corresponding models are 0.1324, 0.0360, 0.0252, and 0.0349 respectively, which indicates that the training process has high stability. Then, taking single-step prediction as an example, each trained model is used to perform the corresponding prediction, and the prediction results of each IMF are as Figure 7 shown, where 7a is the schematic diagram of the final predicted value of IMF1, 7b is the schematic diagram of the final predicted value of IMF2, 7c is the schematic diagram of the final predicted value of IMF3, and 7d is the schematic diagram of the final predicted value of IMF4.

[0073] The prediction results of each intrinsic mode function data are linearly superimposed to obtain the integrated sequence as follows:

[0074]

[0075] In the formula, u S,jIt represents the predicted value of the j-th eigenmode function of the target site S, where j is a positive integer and k represents the number of modes. Subsequently, by zero-padding the meaningless negative solar radiation data in the integrated sequence, the final prediction result of the target site is obtained.

[0076] The final prediction result is obtained by linearly superimposing the predicted values of each IMF and zero-padding the negative values, as Figure 8 shown. To comprehensively evaluate the prediction effect, three evaluation indicators are used to measure the error between the actual data and the predicted data, namely the root mean square error (RMSE), the mean absolute error (MAE), and the relative absolute error (RAE). Among them, the lower the RMSE, MAE, and RAE, the higher the prediction accuracy. The calculation formulas are as follows:

[0077]

[0078] In the formula, n represents the number of predicted data points; represents the predicted data value at the i-th time point; x i represents the actual data value at the i-th time point; represents the average value of the actual data.

[0079] To prove the advantages of the present invention, eight different prediction methods are used for comparison. Table 4 specifically illustrates the methods involved, including five decomposition-based methods and four non-decomposition-based methods. Table 5 lists the evaluation indicators of all relevant methods, and Table 6 illustrates the improvement of the present invention in terms of prediction accuracy. Figure 9 shows the predicted values of all the participating comparison methods. Among them, 9a is a schematic diagram of the predicted values of Method 2, Method 3, Method 4, Method 5, and the method of the present invention; 9b is a schematic diagram of the predicted values of Method 6, Method 7, Method 8, Method 9, and the method of the present invention; Figure 10 A visual comparison of the relevant evaluation indicators is carried out. Among them, 10a is a comparison chart of the RSME evaluation indicators of all the participating comparison methods, 10b is a comparison chart of the MAE evaluation indicators of all the participating comparison methods, and 10c is a comparison chart of the RAE evaluation indicators of all the participating comparison methods. Based on this, the main observation results are as follows:

[0080] (1) Analyses based on decomposition methods show that the method based on MvFIF has higher prediction accuracy. Compared with the method based on NA-MEMD (i.e., Method 4), the RMSE, MAE, and RAE are reduced by 34.50%, 36.27%, and 36.27% respectively. The improvement of the method based on univariate decomposition (such as Method 5) may be more significant, and the corresponding error metrics are reduced by 43.19%, 42.78%, and 42.79% respectively. This can be attributed to the fact that MvFIF can effectively extract the spatio-temporal and time-frequency correlations between multiple stations and significantly suppress mode mixing. These advantages can ensure that the decomposed IMFs are more realistic, thus facilitating subsequent model prediction.

[0081] (2) The comparison between the methods with and without the introduction of EFA shows that this scheme is an effective attempt to improve prediction accuracy. For example, compared with the methods without the introduction of EFA (such as Method 2 and Method 3), the improvements of the method proposed in the present invention in terms of RMSE are 10.84% and 20.99% respectively. This EFA-based scheme can divide the decomposed subsequences corresponding to the target station into two components with different frequency distributions, thereby selecting appropriate predictors. This selection mechanism can enable each predictor involved in the prediction to give full play to its respective advantages, thus obtaining higher prediction accuracy.

[0082] (3) Analyses of these methods without the assistance of decomposition techniques show that the models of the present invention (i.e., the CA-CPC model and the CA-LSTM model) are superior to single models (i.e., LSTM and GRU). For example, the RMSE of LSTM is 104.2943, while the RMSE of the CA-CPC model and the CA-LSTM model are 91.5675 and 93.8112 respectively, and their values are reduced by 12.20% and 10.05% respectively. The reason for the reduction is that CA can incorporate frequency domain information into the identification of important features, thereby improving the feature extraction ability.

[0083] (4) The present invention is a combination of MvFIF-based signal decomposition, EFA-based subsequence classification, and CA-CPC and CA-LSTM-based classification prediction. By leveraging the advantages of each module, it has significant advantages in terms of prediction accuracy. For example, compared with Method 9, the RMSE, MSE, and RAE of the present invention are reduced by 58.83%, 62.17%, and 62.18% respectively. The present invention can make full use of data features from the time-space-frequency perspective, which helps to comprehensively utilize various information in solar radiation data, thereby greatly improving prediction accuracy.

[0084] Table 4 Descriptions of Different Methods

[0085] Method Decomposition technique Predictor The present invention MvFIF CA-CPC; CA-LSTM Method 2 MvFIF CA-CPC Method 3 MvFIF CA-LSTM Method 4 NA-MEMD CA-CPC; CA-LSTM Method 5 EMD CA-CPC; CA-LSTM Method 6 / CA-CPC Method 7 / CA-LSTM Method 8 / LSTM Method 9 / GRU

[0086] Table 5 All evaluation indicators related to the method

[0087] Method RMSE MAE RAE The present invention 44.0282 28.3752 0.1158 Method 2 49.3832 31.0206 0.1266 Method 3 55.7257 35.1082 0.1433 Method 4 67.2210 44.5263 0.1817 Method 5 77.5049 49.5924 0.2024 Method 6 91.5675 66.8929 0.2730 Method 7 93.8112 65.0833 0.2657 Method 8 104.2943 72.5352 0.2961 Method 9 106.9494 75.0159 0.3062

[0088] Table 6 Comparison of evaluation indicators between the present invention and other methods

[0089]

[0090]

[0091] and Figure 1 Corresponding to the method described above, an embodiment of the present invention further provides a solar radiation prediction system that fuses time-domain, space-domain, and frequency-domain features, which is used to Figure 1 The specific implementation of the method in includes: a data acquisition module, a data classification module, a model prediction module, and a result acquisition module;

[0092] The data acquisition module is connected to the input end of the data classification module, and is used to acquire historical solar radiation data of multiple stations, and decompose the historical solar radiation data of multiple stations into multiple groups of intrinsic mode function data;

[0093] The data classification module is connected to the input end of the model prediction module, and is used to analyze the characteristics of the intrinsic mode function data of the target station, and classify the intrinsic mode function data into high-frequency component data and low-frequency component data;

[0094] The model prediction module is connected to the input end of the result acquisition module, and is used to construct a high-frequency prediction model based on the high-frequency component data and a low-frequency prediction model based on the low-frequency component data, and perform targeted training on the constructed high-frequency prediction model and low-frequency prediction model to obtain the prediction results of each intrinsic mode function data;

[0095] The result acquisition module is used to linearly superimpose the prediction results of each intrinsic mode function data and fill the negative values with zeros to obtain the final prediction result.

[0096] In this specification, each embodiment is described in a progressive manner. The key point of each embodiment is the difference from other embodiments. The same or similar parts between each embodiment can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0097] The foregoing description of the disclosed embodiments enables those skilled in the art to practice or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A solar radiation prediction method integrating time-domain, space-frequency domain features, characterized in that: The following steps are involved: S1. Acquiring data: acquiring historical solar radiation data of multiple sites, and decomposing the historical solar radiation data of multiple sites into multiple groups of intrinsic mode function data; S2. Data classification: Analyze the characteristics of the intrinsic mode function data of the target site and classify the intrinsic mode function data into high-frequency component data and low-frequency component data; S3. Model prediction: construct a high-frequency prediction model based on the high-frequency component data and a low-frequency prediction model based on the low-frequency component data, perform targeted training on the constructed high-frequency prediction model and low-frequency prediction model, and obtain the prediction results of each eigenmode function data; S4. Result acquisition: Linearly superimpose the prediction results of each eigenmode function data and fill negative values ​​with zero to obtain the final prediction result.

2. The solar radiation prediction method integrating time-domain, space-domain and frequency-domain features according to claim 1, characterized in that: In S1, the historical solar radiation data of multiple sites are obtained, and the specific contents of decomposing the historical solar radiation data of multiple sites into multiple groups of intrinsic mode function data are as follows: The historical solar radiation data of the target site and adjacent sites at corresponding time points are synchronously decomposed into multiple groups of intrinsic mode function data through multivariate rapid iterative filtering.

3. The solar radiation prediction method integrating time-domain, space-domain and frequency-domain features according to claim 1, characterized in that: In S2, the characteristics of the intrinsic mode function data of the target site are analyzed, and the specific contents of classifying the intrinsic mode function data into high-frequency component data and low-frequency component data are as follows: The permutation entropy is used to analyze the eigenmode function data characteristics of the target site, and the eigenmode function data with a value greater than the original data entropy is classified as high-frequency component data, and the eigenmode function data with a value less than or equal to the original data entropy is classified as low-frequency component data.

4. The solar radiation prediction method integrating time-domain, space-frequency domain features according to claim 1, characterized in that: In S3, a high-frequency prediction model is constructed based on the high-frequency component data, and a low-frequency prediction model is constructed based on the low-frequency component data. The constructed high-frequency prediction model and low-frequency prediction model are trained in a targeted manner, and the specific content of the prediction result of each intrinsic mode function data is obtained as follows: Based on the high-frequency component data, a high-frequency prediction model that is a mixture of cross-attention and conditional parameterized convolution is constructed. Based on the low-frequency component data, a low-frequency prediction model that is a mixture of cross-attention and long short-term memory network is constructed. Performing targeted training on the constructed high-frequency prediction model and low-frequency prediction model to obtain trained high-frequency prediction model and low-frequency prediction model; Use the trained high-frequency prediction model and low-frequency prediction model to perform predictions and output the corresponding prediction results.

5. The solar radiation prediction method integrating time-domain, space-frequency domain features according to claim 1, characterized in that: In S4, the prediction results of each eigenmode function data are linearly superimposed and negative values ​​are filled with zeros, and the specific content of the final prediction result is: The prediction results of each eigenmode function data are linearly superimposed to obtain an integrated sequence As follows: In the formula, u S,j It represents the predicted value of the jth eigenmode function of the target site S, where j is a positive integer and k represents the number of modes. Subsequently, the final prediction result of the target site is obtained by zero-filling the meaningless negative solar radiation data in the integrated sequence.

6. A solar radiation prediction system integrating time-space-frequency characteristics, characterized in that: A solar radiation prediction method integrating time-domain, space-domain and frequency-domain features according to any one of claims 1 to 5 is applied, comprising: a data acquisition module, a data classification module, a model prediction module and a result acquisition module; A data acquisition module is connected to the input end of the data classification module, and is used to acquire the historical solar radiation data of multiple sites, and decompose the historical solar radiation data of multiple sites into multiple groups of intrinsic mode function data; A data classification module is connected to the input end of the model prediction module, and is used to analyze the characteristics of the intrinsic mode function data of the target site and classify the intrinsic mode function data into high-frequency component data and low-frequency component data; The model prediction module is connected to the input end of the result acquisition module, and is used to construct a high-frequency prediction model based on the high-frequency component data and a low-frequency prediction model based on the low-frequency component data, and to perform targeted training on the constructed high-frequency prediction model and low-frequency prediction model to obtain the prediction result of each intrinsic mode function data; The result acquisition module is used to linearly superimpose the prediction results of each eigenmode function data and fill negative values ​​with zero to obtain the final prediction result.

Images