Short-term photovoltaic power generation prediction method and system based on multi-period clustering and signal reconstruction

Abstract

The invention provides a short-term photovoltaic power generation prediction method and system based on multi-period clustering and signal reconstruction, and the method comprises the steps: obtaining key historical meteorological data, dividing the key historical meteorological data according to the geographic position and time, carrying out the dimension reduction and clustering of the divided meteorological features, and obtaining the preprocessed meteorological features; similar distance screening is carried out on the preprocessed meteorological features and the meteorological features of the day to be measured, the meteorological features with the similar distances smaller than a set value are selected and sorted, and similar days are obtained; photovoltaic output data corresponding to meteorological characteristics of similar days are obtained, secondary decomposition and reconstruction are carried out on the photovoltaic output data to obtain subsequences with different characteristics, and the photovoltaic output data subsequences are combined with the corresponding meteorological characteristics to obtain a plurality of subdata sets; and transmitting the sub-data set to the multi-modal prediction model to obtain a final predicted value of the photovoltaic output of the to-be-tested sunlight. According to the invention, the accuracy and reliability of photovoltaic output prediction are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of renewable energy prediction, and in particular relates to a short-term photovoltaic power generation prediction method and system based on multi-period clustering and signal reconstruction. Background Art

[0002] In addressing the dual challenges of global climate change and energy security, the development and utilization of renewable energy has become an inevitable choice. Solar energy, with its inexhaustible nature, can effectively reduce dependence on traditional energy sources and plays a vital role in the transformation of the global energy structure. However, photovoltaic power generation is characterized by volatility and intermittency, with its output affected by a variety of factors, including weather conditions, geographic location, and seasonal variations. Therefore, accurate photovoltaic power generation forecasting is crucial for maintaining stable grid operation and reducing backup power costs.

[0003] The output characteristics of photovoltaic power generation systems are often affected by meteorological conditions, exhibiting irregular dynamic fluctuations. This makes model predictions sensitive to the choice of dataset. Therefore, when processing data, researchers often consider methods such as screening for similar days, clustering datasets, or decomposing the input data signal to improve the quality of the dataset. Currently, the conventional method for selecting similar days is based on weather conditions. By directly comparing meteorological factors, historical days with weather conditions similar to those on the predicted day are selected as similar days. In addition, there are methods based on statistical feature matching to select similar days. However, the selection of typical similar days usually focuses only on time period characteristics or the similarity of weather characteristics, ignoring the inherent laws and connections between different time periods and weather.

[0004] The methods for photovoltaic power generation prediction mainly include the use of physical models, statistical models and artificial intelligence models for prediction. The physical model mainly makes predictions by solving relevant physical equations and combining meteorological data and photovoltaic system parameters. Although this model has the potential for high accuracy in theory, in practical applications it is often difficult to achieve the expected results due to its complex calculations and the requirement for high precision of input parameters. Statistical models usually establish prediction models by analyzing the characteristics of time series. The calculations are relatively simple, but they are easily affected by outliers and noise. Therefore, error accumulation problems are prone to occur when dealing with complex photovoltaic power generation systems. In contrast, artificial intelligence models have been more widely used in the field of prediction in recent years. Because of their powerful nonlinear modeling capabilities, they often perform better than traditional linear models when dealing with highly complex environmental data.

[0005] Common single models include XGBoost, SVM, GRU, and LSTM, each with its own strengths in different prediction tasks. However, single models often perform poorly when processing nonlinear data, making prediction accuracy difficult to guarantee. In contrast, hybrid prediction models can outperform baseline single prediction models under various conditions, but this relies heavily on understanding data characteristics and selecting the right model.

[0006] There are currently multiple data decomposition methods, including EMD and EEMD. However, EMD is prone to problems such as modal stacking. While EEMD addresses the inherent issues of EMD by adding white noise, it still carries the risk of residual noise. VMD, on the other hand, adaptively decomposes the signal into multiple modal functions, effectively avoiding the modal aliasing and endpoint effects of EMD. Summary of the Invention

[0007] To address the problems existing in the prior art, the present invention provides a short-term photovoltaic power generation forecasting method and system using multi-period clustering and signal reconstruction. This method addresses the limitations of similar day selection by segmenting key historical meteorological data by geographic location and time. Dimensionality reduction, clustering, and similarity distance screening enable more accurate selection of similar days. Furthermore, to address the shortcomings of signal decomposition, a secondary decomposition and reconstruction of photovoltaic output data is performed to better account for complex dynamic changes in the signal. Furthermore, to address the drawbacks of single-model prediction, a multimodal prediction model is employed to improve overall prediction accuracy. This invention effectively addresses the shortcomings of traditional methods in similar day selection, signal processing, and prediction accuracy, thereby improving the accuracy and reliability of photovoltaic output forecasts.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a short-term photovoltaic power generation prediction method based on multi-period clustering and signal reconstruction, the specific steps of which are as follows:

[0009] Obtain key historical meteorological data, divide the key historical meteorological data according to geographical location and time period, perform dimensionality reduction and clustering on the divided meteorological features, and obtain the pre-processed weather categories for each time period;

[0010] After preprocessing, under the same weather category in each period, the divided meteorological characteristics are screened with the meteorological characteristics of the test day for similarity distance, and the meteorological characteristics with similarity distance less than the set value are selected and sorted to obtain similar days;

[0011] Obtain PV output data corresponding to meteorological characteristics on similar days, perform secondary decomposition and reconstruction on the PV output data to obtain subsequences with different characteristics, and merge the PV output data subsequences with the corresponding meteorological characteristics to obtain multiple sub-datasets;

[0012] The sub-datasets are transferred to the hybrid prediction model to obtain the final predicted value of photovoltaic output on the test day.

[0013] Furthermore, in the step of obtaining key historical meteorological data, dividing the key historical meteorological data according to geographical location and time, and performing dimensionality reduction and clustering on the divided meteorological features to obtain pre-processed meteorological features, specifically:

[0014] According to the sunrise and sunset times in different geographical locations, the key historical meteorological data are divided into three time periods: rising period, peak period, and falling period. The PCA-NRBO-K-means++ algorithm is used to reduce the dimension of the meteorological characteristics of the above three time periods and cluster them to obtain the weather category distribution in different time periods.

[0015] Furthermore, in the same weather category of each period after preprocessing, the divided meteorological characteristics are screened with the meteorological characteristics of the day to be measured by similarity distance, and meteorological characteristics with similarity distances less than a set value are selected and sorted to obtain the similar day step:

[0016] The DTW algorithm is used to obtain the similarity distance between the meteorological characteristic data in different time periods and the meteorological characteristics of the measured time. In the time periods where the similarity distance is less than 5, the meteorological characteristic data are sorted and reorganized to obtain similar days. The meteorological characteristics of similar days are then merged with the corresponding photovoltaic output data to form a similar day data set of multiple time periods.

[0017] Furthermore, in the step of obtaining the photovoltaic output data corresponding to the meteorological characteristics of similar days, performing secondary decomposition and reconstruction on the photovoltaic output data and merging it with the corresponding meteorological characteristics to obtain a sub-dataset:

[0018] The photovoltaic output data is decomposed and reconstructed by using the VMD method, the photovoltaic output data is decomposed into intrinsic mode function components, and the components with entropy values greater than 0.5 are reconstructed to obtain the first high-frequency term;

[0019] The CEEMDAN method is used to perform secondary decomposition and reconstruction on the first high-frequency term, and the secondary decomposition components and the components that are not reconstructed after the primary decomposition are obtained to obtain the secondary decomposition components. Among the secondary decomposition components, the components with entropy values greater than 0.5 are reconstructed as the second high-frequency term; the components with entropy values between 0.2 and 0.5 are reconstructed as low-frequency terms; and the components with entropy values less than 0.2 are reconstructed as trend terms.

[0020] The reconstructed photovoltaic output data is obtained by adding the second high-frequency term, the low-frequency term, and the trend term. The difference between the undecomposed and reconstructed photovoltaic output data and the reconstructed photovoltaic output data is the residual term.

[0021] The photovoltaic output data of the second high-frequency item, low-frequency item, trend item and residual item are combined with the corresponding meteorological characteristics to obtain a high-frequency item dataset, a low-frequency item dataset, a trend item dataset and a residual item dataset.

[0022] Furthermore, the sub-dataset is transferred to the hybrid prediction model to obtain the final predicted value of the photovoltaic output on the test day:

[0023] The hybrid prediction model includes the CNN-GRU-Attention model and the XGBoost model;

[0024] The second high-frequency item dataset uses the CNN-GRU-Attention model for short-term photovoltaic power generation forecasting;

[0025] The low-frequency item dataset, trend item dataset, and residual item dataset are used to perform short-term photovoltaic power generation forecasting using the XGBoost model;

[0026] The predicted values of the second high-frequency term, low-frequency term, trend term, and residual term are combined to obtain the final predicted value of the photovoltaic output on the day to be measured.

[0027] Furthermore, unfiltered historical meteorological data and photovoltaic output data were obtained as historical initial data, the historical initial data were cleaned, and the Pearson and Spearman correlation analysis methods were used to screen the key meteorological factors affecting photovoltaic output;

[0028] The data corresponding to meteorological factors with absolute values of correlation coefficients greater than 0.4 are selected as key historical meteorological data.

[0029] The present invention also provides a short-term photovoltaic power generation prediction system, which runs the above-mentioned short-term photovoltaic power generation prediction method of multi-period clustering and signal reconstruction, including:

[0030] The meteorological feature preprocessing module is used to obtain key historical meteorological data, divide the key historical meteorological data according to geographical location and time period, reduce the dimension and cluster the divided meteorological features, and obtain the weather category for each time period after preprocessing;

[0031] The similar day screening module is used to screen the meteorological characteristics after classification and the meteorological characteristics of the test day under the weather category of each period after preprocessing, select the meteorological characteristics with a similarity distance less than the set value and sort them to obtain similar days;

[0032] The photovoltaic output data reconstruction module is used to obtain photovoltaic output data corresponding to the meteorological characteristics of similar days, perform secondary decomposition and reconstruction on the photovoltaic output data to obtain subsequences with different characteristics, and merge the photovoltaic output data subsequences with the corresponding meteorological characteristics to obtain multiple sub-datasets;

[0033] The photovoltaic output prediction module is used to transfer the sub-datasets to the hybrid prediction model to obtain the final predicted value of the photovoltaic output on the test day.

[0034] The present invention also provides a terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the above-mentioned short-term photovoltaic power generation prediction method based on multi-period clustering and signal reconstruction are implemented.

[0035] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-mentioned short-term photovoltaic power generation prediction method based on multi-period clustering and signal reconstruction.

[0036] The present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements the steps of the above-mentioned short-term photovoltaic power generation prediction method based on multi-period clustering and signal reconstruction.

[0037] Compared with the prior art, the present invention has at least the following beneficial effects:

[0038] This paper provides a short-term photovoltaic power generation forecasting method based on multi-period clustering and signal reconstruction. This method significantly improves the selection of similar days and proposes a similar day selection method based on multi-period clustering. Traditional similar day selection methods often fail to effectively integrate the intrinsic relationship between time periods and weather types, resulting in low data selection quality and thus affecting forecast accuracy. This paper selects similar time periods from the same clustering results of different time periods for reconstruction. This not only takes into account the differences between time periods but also takes into account the uniformity of the same clustering results, thereby more comprehensively capturing the inherent patterns in the data.

[0039] Specifically, this method first divides historical meteorological data into three time periods: rising, peak, and falling, based on geographic location and sunrise and sunset times. It then uses the PCA-NRBO-K-means++ algorithm to reduce the dimensionality and cluster the meteorological characteristics of each time period. PCA is used for dimensionality reduction and denoising, effectively reducing data redundancy and noise and improving clustering accuracy. NRBO is used to optimize the initial centroid selection of K-means++, finding the optimal solution through a global search. This avoids the traditional K-means++ sensitivity to initial point selection and improves clustering effectiveness.

[0040] During the similar day screening phase, the present invention uses the DTW algorithm to calculate the similarity distance between the time period under test and the historical time periods in each clustering result. Time periods with similarity distances less than a set threshold are sorted and reorganized to obtain similar days. This method considers not only the numerical similarity of meteorological characteristics but also the similarity of their temporal series morphology, thereby improving the accuracy and reliability of similar day selection.

[0041] The present invention provides a short-term photovoltaic power generation prediction method based on multi-period clustering and signal reconstruction. This method addresses the problems of spectral aliasing and modal confusion in signal decomposition, and proposes a multi-level photovoltaic output data processing method. The photovoltaic output data is secondary decomposed and reconstructed into subsequences with different characteristics, thereby improving the accuracy and stability of the photovoltaic output data decomposition and being able to better cope with complex dynamic changes in the signal.

[0042] Specifically, the present invention reconstructs the photovoltaic output signal into high-frequency terms, low-frequency terms, trend terms, and residual terms by combining the VMD-FE and CEEMDAN-FE models. The VMD method decomposes the signal into a series of intrinsic mode function components through variational mode decomposition, effectively avoiding the problem of spectral aliasing. The CEEMDAN method performs a secondary decomposition of the high-frequency terms through adaptive noise complete set empirical mode decomposition, further improving the accuracy and stability of the decomposition. By combining these two methods, the present invention can better cope with complex dynamic changes in the signal and provide a more detailed means for signal feature extraction and analysis.

[0043] During the signal reconstruction phase, the present invention combines and reconstructs the decomposed subsequences to obtain reconstructed photovoltaic output data. Simultaneously, the difference between the undecomposed and reconstructed photovoltaic output data is calculated as a residual term for subsequent error compensation. This multi-level photovoltaic output data processing method not only improves the accuracy of signal decomposition but also enhances the model's robustness to noise and irregularities.

[0044] This paper provides a short-term photovoltaic power generation forecasting method based on multi-period clustering and signal reconstruction. This method addresses the overfitting and underfitting issues that can arise when using a single model to predict different subsequences after decomposition. Instead, it proposes a hybrid forecasting model. This model uses different forecasting models based on the different characteristics of the subsequences, thereby improving the accuracy and reliability of the overall forecast.

[0045] Specifically, the present invention uses the CNN-GRU-Attention model to predict high-frequency items. The CNN-GRU-Attention model combines the advantages of convolutional neural networks (CNNs), gated recurrent units (GRUs), and attention mechanisms (Attention). It can effectively extract local features and long-range dependencies of time series data and focus on key information, thereby improving the accuracy and stability of predictions.

[0046] For low-frequency terms, trend terms, and residual terms, the present invention uses the XGBoost model for prediction. The XGBoost model is an ensemble learning algorithm based on gradient boosted decision trees (GBDT), with powerful feature extraction capabilities and high prediction accuracy. By using the XGBoost model to predict low-frequency terms, trend terms, and residual terms, we can further capture the inherent laws and trends in the data and improve the accuracy and reliability of the prediction.

[0047] Finally, the present invention combines the predicted values of the high-frequency term, low-frequency term, trend term, and residual term to obtain the final predicted value of the photovoltaic output on the day to be measured. This multimodal prediction model not only improves the accuracy and reliability of the prediction, but also enhances the model's adaptability to data of varying characteristics and complexity.

[0048] The short-term photovoltaic power generation forecasting system of the present invention includes a meteorological feature preprocessing module, which is responsible for acquiring key historical meteorological data, dividing it according to geographic location and time period, and performing dimensionality reduction and clustering on the divided meteorological features to obtain preprocessed weather categories for each time period. By acquiring key historical meteorological data and screening key meteorological factors that affect photovoltaic output, the present invention can remove redundant and irrelevant data, improving data quality and effectiveness. Data corresponding to meteorological factors with absolute values of correlation coefficients greater than 0.4 are selected as key historical meteorological data, ensuring data accuracy and relevance. Utilizing the PCA-NRBO-K-means++ algorithm to perform dimensionality reduction and clustering on meteorological features, the present invention can further refine useful information from the data, reduce the dimensionality and complexity of the data, and improve the accuracy and efficiency of clustering. PCA is used for dimensionality reduction and denoising, effectively reducing data redundancy and noise; NRBO is used to optimize the initial centroid selection of K-means++, improving the clustering's ability to search for the global optimal solution. Through processing by the meteorological feature preprocessing module, the present invention can obtain more accurate and useful meteorological feature data, providing a solid foundation for subsequent similar day screening and photovoltaic output forecasting.

[0049] The similar day screening module of the present invention is responsible for screening the meteorological characteristics after preprocessing and the meteorological characteristics of the day to be measured by similarity distance, selecting meteorological characteristics with a similarity distance less than a set value and sorting them to obtain similar days. By using the DTW algorithm to calculate the similarity distance, the present invention can accurately measure the similarity between meteorological feature time series, even if they are stretched and offset. This makes it possible to more comprehensively consider the morphology and characteristics of the time series when looking for similar time periods, thereby improving the accuracy and reliability of similar day selection. By setting a similarity distance threshold and sorting and reorganizing similar time periods, the present invention can ensure that the obtained similar days have a high similarity with the day to be measured, thereby improving the accuracy of subsequent photovoltaic output forecasts. The design of the similar day screening module also takes into account the differences in time periods and the unity of the same clustering results. By selecting similar time periods from the same clustering results of different time periods for reconstruction, the inherent patterns in the data are more comprehensively captured, further improving the accuracy and reliability of similar day selection.

[0050] The photovoltaic output prediction module of the present invention is responsible for transferring sub-datasets into a hybrid prediction model to obtain the final predicted photovoltaic output value for the day to be measured. Using the hybrid prediction model for prediction, the present invention fully leverages the advantages of different models to improve prediction accuracy and reliability. The CNN-GRU-Attention model is used to predict high-frequency items, effectively extracting local features and long-range dependencies in time series data. The XGBoost model is used to predict low-frequency items, trend items, and residual items, offering powerful feature extraction capabilities and high prediction accuracy. Through the design of an adaptive hybrid prediction model, the present invention can use different prediction models based on the different characteristics of the sub-sequences, avoiding the overfitting or underfitting issues that may arise with a single model. This adaptive prediction strategy enables the model to perform better when dealing with sub-sequences with different characteristics, further improving overall prediction accuracy. The photovoltaic output prediction module also considers the merging of prediction results and the processing of residual items. The predicted values for the high-frequency items, low-frequency items, trend items, and residual items are merged to obtain the final predicted photovoltaic output value for the day to be measured. Furthermore, the residual item is used for error compensation, improving prediction accuracy and stability. This design that comprehensively considers prediction results and error compensation enables the photovoltaic output prediction module of the present invention to have higher prediction accuracy and reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0052] Figure 1A flowchart of the method of the present invention;

[0053] Figure 2 This is the technical roadmap for feature vector dimensionality reduction and clustering of the present invention;

[0054] Figure 3 Flowchart for improving the selection of initial centroids for clustering using the NRBO algorithm of the present invention;

[0055] Figure 4 The number distribution of various weather types in different time periods in the embodiment;

[0056] Figure 5 is the fuzzy entropy value of each component after multi-signal processing in the embodiment;

[0057] Figure 6 This is the structural diagram of GRU;

[0058] Figures 7 to 9 1 is a comparison chart of the predicted values and true values of several control group models in the examples. DETAILED DESCRIPTION

[0059] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0060] This paper proposes a short-term photovoltaic power generation forecasting method based on multi-period clustering and signal reconstruction. Through systematic data processing and analysis, combined with deep learning techniques, it predicts photovoltaic output in different weather conditions. This method integrates correlation analysis, similar day selection, signal decomposition and reconstruction, and hybrid deep learning model prediction. This significantly improves the accuracy and robustness of photovoltaic power generation forecasts, providing strong support for the efficient utilization of renewable energy and optimized grid scheduling.

[0061] See also Figure 1 The present invention provides a short-term photovoltaic power generation prediction method based on multi-period clustering and signal reconstruction, comprising the following steps:

[0062] S1. Cleaning and Correlation Analysis. The historical initial data was cleaned, and the Pearson and Spearman correlation analysis methods were used to screen out key meteorological factors affecting PV output (those with absolute correlation coefficients greater than 0.4 using both analysis methods). The filtered key historical meteorological data was combined with the PV output data to define them as new historical data, which was then uniformly normalized.

[0063] Among them, the historical initial data are unfiltered historical meteorological data and photovoltaic output data.

[0064] The existing data set is preprocessed, including three measures: missing data processing, abnormal data processing and data normalization.

[0065] Missing data handling: When there are missing values in the data set, the missing value is filled using the average of the valid surrounding values of the current value. The span of the neighboring points represents the range of the interval extending upward and downward from the position of the missing value. The valid values within this interval will be used to calculate the average value. This can avoid over-reliance on the value of a single neighboring point and also avoid excessive consideration of irrelevant data. The formula is as follows:

[0066]

[0067] where x i,t is the load value at hour i.

[0068] Abnormal data processing: Use the Three Sigma principle to identify outliers in the data set. Since historical data is continuous over time, it is not possible to simply remove abnormal data, mark them as missing values, and then use the median or mean instead.

[0069] Data normalization: Standardize the pre-clustered data and normalize the pre-predicted data.

[0070]

[0071] Where x1 and x2 are the normalized and standardized data values respectively; x is the original data; x min is the minimum value of the data; x max is the maximum value; μ and σ are the mean and standard deviation of the data respectively.

[0072] Correlation analysis: The correlation coefficient formula of double correlation analysis can be expressed as:

[0073]

[0074] Where r p is the Pearson correlation coefficient; x i with y i Respectively represent the observed values of meteorological factor variable X and photovoltaic output Y; μ x With μ y is the average of these two variables. s is the Spearman correlation coefficient; R(x i ) and R(y i ) represent the rank of the i-th observation value of the meteorological factor variable X and the photovoltaic output Y respectively; n represents the number of observations.

[0075] S2. Dataset Segmentation and Clustering. Based on the sunrise and sunset times at different geographic locations, the historical data was divided into three time periods: rising, peak, and falling. Meteorological feature vectors were extracted for each time period and clustered using PCA (Principal Component Analysis)-NRBO (Newton-Raphson Optimization)-K-means++ after dimensionality reduction to distinguish different weather categories within each time period. The specific meanings of the symbols in the meteorological feature vectors are shown in Table 1.

[0076] Table 1 Specific meanings of symbols in meteorological characteristic vectors

[0077]

[0078] The technical roadmap for feature vector dimensionality reduction and clustering in each period is as follows: Figure 2 shown.

[0079] PCA-NRBO-K-means++: In photovoltaic power generation forecasting, partitioning meteorological feature vectors into multiple highly similar populations can identify potential correlations in the data. After reducing the feature vector dimensionality, the principal components of the data are extracted. The NRBO algorithm is then used to optimize the initial centroid selection of the K-means++ algorithm and perform clustering.

[0080] The meteorological eigenvectors are organized into the following matrix:

[0081]

[0082] Among them, X ij Represents the elements in the i-th row and j-th column of the matrix. The X matrix contains n-dimensional eigenvectors, and each eigenvector has m historical samples. Standardize the X matrix to get X * :

[0083]

[0084] Where, X ij * is the standardized data of row i and column j; X j * is the standard deviation of the jth column; is the mean of the jth column.

[0085] After normalization, the matrix X * Perform eigenvalue decomposition on the correlation coefficient matrix and rotate the data to project it onto the principal component axis:

[0086] R=Q∧Q T

[0087] Z=Q T X *

[0088] Where R is the correlation coefficient matrix; ∧ is a diagonal matrix containing eigenvalues λ1,λ2,…,λ n ; Q is the eigenvector matrix; Z is the data matrix projected onto the principal component axis. Calculate the cumulative contribution rate of the principal component. The cumulative contribution rate of the principal component is:

[0089]

[0090] Where, v h is the contribution rate of the hth principal component; h is the hth eigenvalue; v sk is the cumulative contribution rate of the first k principal components. The principal components with a cumulative explained variance exceeding 90% are retained to obtain the meteorological characteristic dataset X′ after dimensionality reduction.

[0091] The flowchart of NRBO algorithm to improve the selection of initial centroid of clustering is as follows Figure 3 As shown in Figure 2, the intra-cluster squared error (SSE) is used as the fitness function of clustering. The intra-cluster squared error calculation formula is:

[0092]

[0093] Where G s is the sum of squared errors within the cluster. The smaller the SSE value, the closer the data points within the cluster are. k is the total number of clusters, which is set to 4 here. C i is the i-th cluster; / / x-μ i / / 2 From the data point x to the centroid μ of its cluster i The square of the Euclidean distance.

[0094] The algorithm flow is implemented as follows:

[0095] Input the randomly selected centroid C1 from K-means++ into the NRBO algorithm as the initial centroid. Initialize the NRBO algorithm parameters and population. Parameter settings are: population size pop = 30; upper and lower bounds lb = min(data), ub = max(data); Max-IT = 50, where Max-IT is the maximum number of iterations.

[0096]

[0097] Where, is the position of the jth dimension of the nth population; N p is the number of individuals in the population; lb and ub are the upper and lower limits of the parameters to be optimized respectively; rand is a random number between (0,1).

[0098] Calculate fitness. According to the set fitness function, calculate the fitness value of each centroid individual, and find the best fitness value and the worst fitness value and their corresponding position X b and X w .

[0099] Use the NRBO rule to search for a new solution, and the new position is:

[0100]

[0101] Where, IT is the current iteration number; n is the individual number; is the solution at the next iteration; r1 and r2 are both random numbers between (0,1); The specific expression is:

[0102]

[0103] Where, are the three positions generated by updating the current position to enhance the ability of local search and global search; Rh is the step size factor used to guide the population in the right direction; δ is the adaptive coefficient. The corresponding formula is as follows:

[0104]

[0105] Where Y w With Y b are positions used to further enhance the search performance; Δx is the search range; r3 and r4 are both random numbers between (0,1); a1 and a2 are different unequal integers randomly selected from the population.

[0106] The purpose of introducing TAO is to improve the effectiveness of NRBO in dealing with real-world problems. It is based on the improved and enhanced calculation formula proposed by Ahmadianfar et al. When the value of rand is less than DF, the solution is generated.

[0107]

[0108] Where α and β are consistent random numbers between (-1, 1) and (-0.5, 0.5), respectively; μ1 and μ2 are random numbers that can diversify the population. A random number rand between (0, 1) is compared with 0.5. If rand ≥ 0.5, then μ1 = μ2 = 1, otherwise: μ1 = 3μ2 = 3rand;

[0109] When IT = Max-IT, output k best positions X bAs the final centroid position, we calculate the distance D(X) between each sample in the dataset and the centroid, and divide it into the cluster with the shortest distance corresponding to the centroid. Finally, we get the data corresponding to different weather categories in each time period.

[0110]

[0111] After clustering each time period, the number distribution of various weather types in different time periods in the embodiment is as follows: Figure 4 As shown in the figure, the distribution of the number of each weather type in each time period is inconsistent. Therefore, the multi-period clustering method of selecting similar days can better capture the impact of weather changes on photovoltaic power generation and improve the accuracy of forecasts.

[0112] S3. Selection and reconstruction of similar time periods. Use the Dynamic Time Warping (DTW) algorithm to calculate the similarity distance between the time period to be predicted and historical data from the same time period and belonging to the same clustering result. For time periods with a similarity distance less than 5, sort and reorganize the data. The data with the smallest similarity distance to the predicted day is designated as Similarity Day 1. Similarity Days 2 through 20 are generated sequentially to form a multi-period similarity day dataset.

[0113] Among them, the multi-period combined similar day data set is the meteorological characteristics of 20 similar days selected according to similarity and their corresponding photovoltaic output data.

[0114] S4. Primary decomposition and reconstruction of PV output. Using the variational mode decomposition (VMD) method, the PV output sequence A in the similar daily dataset is decomposed into intrinsic mode function (IMF) components. Based on the magnitude and similarity of the fuzzy entropy values of each component, components with entropy values greater than 0.5 are reconstructed to obtain the first high-frequency term, IMF_A.

[0115] The fuzzy entropy value is used to judge the decomposed subsequences in the reconstruction, and it is ensured to be more suitable for classifying and recombining the decomposed components by enhancing the robustness to noise and irregularity.

[0116] S5. Secondary decomposition and reconstruction of photovoltaic output. The CEEMDAN (complete ensemble empirical mode decomposition with adaptive noise) method is used to perform secondary decomposition on the first high-frequency term IMF_A. By comparing the fuzzy entropy values of the new components after the secondary decomposition and the entropy values of the components that were not reconstructed in the primary decomposition, the components with entropy values greater than 0.5 are reconstructed into the second high-frequency term IMF_A', the components with entropy values between 0.2 and 0.5 are reconstructed into the low-frequency term IMF_B, and the components with entropy values lower than 0.2 are reconstructed into the trend term IMF_C. The values of IMF_A', IMF_B, and IMF_C are added together to obtain the reconstructed photovoltaic output sequence A'. The difference between A and A' is calculated as the residual term IMF_E, which is used for error compensation to better characterize the prediction error of the signal and its dynamic characteristics.

[0117] In the embodiment, the fuzzy entropy value of each component after multi-signal processing is as follows: Figure 5 shown.

[0118] Table 2 shows the reconstruction results of each component obtained by using the above multi-level processing method for photovoltaic output.

[0119] Table 2 Reconstruction results of each component

[0120]

[0121] High-frequency components often contain rapidly changing features or noise, while low-frequency and trend components reflect trends or cyclical behavior. By using models tailored to their respective frequency characteristics, we can more accurately capture information about the different characteristics of the series.

[0122] S6. Deep learning prediction based on a hybrid model. The PV output sequences under each modal classification are combined with their corresponding meteorological features to construct datasets for high-frequency, low-frequency, trend, and residual terms. The high-frequency dataset is predicted using the CNN-GRU-Attention model (a convolutional neural network with an attention mechanism and a gated recurrent unit model), while the low-frequency, trend, and residual datasets are predicted using the XGBoost model (extreme gradient boosting tree model).

[0123] In the prediction of high-frequency data sets of photovoltaic power generation, CNN is mainly used to automatically extract important features from the data to capture complex nonlinear relationships. At the same time, its multi-layer structure and parameters enhance the generalization ability of the model, enabling it to adapt to diverse changes in environmental conditions. At the same time, the gated recurrent unit (GRU) is used in the prediction to solve the gradient vanishing problem faced by traditional RNN when processing long sequence data. Finally, the attention mechanism is used to enhance the model's selective attention to input information, improve the ability to capture complex patterns and prediction performance, and enable the model to focus on the most relevant information when making predictions. The structural diagram of GRU is shown below. Figure 6 As shown. The network architecture inside GRU is relatively simple, and its main formula is as follows:

[0124] r t =σ(W r h t-1 +U r X t +b r )

[0125] Z t =σ(W Z h t-1 +U Z X t +b Z )

[0126]

[0127] Where Z t To update the gate, decide which information needs to be transmitted between the previous moment and the current moment; r t is the reset gate, which determines how to combine the current input information with the hidden state at the previous moment; h is the current system state; σ is the Sigmoid function; tanh is the hyperbolic tangent function; is the dot product of the vectors.

[0128] The XGBoost model is used to predict low-frequency items, trend items, and residual item data sets, which has excellent generalization ability, fast running speed, and high prediction accuracy.

[0129] S7. Superposition of the predicted values of each subsequence. Add up all the predicted values to obtain the final predicted value of photovoltaic output for each test day, and evaluate the prediction accuracy. The following three indicators are selected to evaluate the prediction performance of the model:

[0130]

[0131] Where R 2 is the coefficient of determination, which is used to evaluate the degree of fit of the regression model. The closer it is to 1, the better the model fits the data; MAE is the mean absolute error; RMSE is the root mean square error; m is the number of prediction samples; y i and are the actual value and predicted value of point i respectively.

[0132] Example 1:

[0133] Test results

[0134] The data used in this example is measured data from a 30MW photovoltaic power station in Northwest China collected by an on-site renewable energy station. The power generation and meteorological data from January 1, 2019 to July 31, 2020 were selected for analysis, with July 2020 being the month to be measured. Since photovoltaic power generation is only generated during periods of sunlight, data was collected between 5:00 and 19:45 daily, with a sampling interval of 15 minutes, and a total of 60 sampling points per day. The data set includes photovoltaic power generation, total horizontal irradiance, temperature, air pressure, relative humidity, and wind direction. After processing, a total of 34,440 data items were obtained.

[0135] Bivariate correlation analysis was performed to obtain a comprehensive understanding of the data relationships. Finally, horizontal total irradiance, temperature, and relative humidity were selected as input meteorological data, and the rest were considered redundant meteorological features that needed to be deleted.

[0136] To effectively address the impact of weather changes on photovoltaic power generation, a day is divided into multiple time periods to capture the changing characteristics of photovoltaic power generation. If the time intervals are set too short, the eigenvectors in the cluster analysis may not be clear enough, thus affecting the effectiveness of the clustering results. Therefore, after comprehensively considering the volatility of power generation and its relationship with weather changes, the day is divided into three time periods: the rising period (5:00-10:00), the peak period (10:00-15:00), and the falling period (15:00-20:00).

[0137] To validate the effectiveness of this invention, 10 comparison models were constructed. The input data for each model used the same training and prediction settings, with days 1 to 18 of similarity as the training set and the last two days as the test set. The models discussed maintained consistency in some parameter settings. The abbreviations of the different comparison models are shown in Table 2. Two ablation experiments were conducted, analyzing the impact of different similarity selection methods, different decomposition and reconstruction methods, and the presence or absence of error compensation mechanisms on the models to evaluate the model's performance under different conditions.

[0138] Table 2 Abbreviations of different models

[0139]

[0140]

[0141] The VMD-FE model and the VMD-FE-CEEMDAN-FE model implement single-shot decomposition and multi-level signal processing based on similar days selected using the proposed multi-period clustering method, then perform prediction using a hybrid model and perform residual correction. The VMD-FE-CEEMDAN-FE model represents the integrated model proposed in this paper.

[0142] The comparison of the predicted values and the actual values of the control group models in different weather conditions in the embodiment is shown in the figure below. Figures 7-9 shown.

[0143] M1, M2, M3 and M4 were selected as the multivariate comparison benchmark models and compared with the M11 proposed in this invention. The photovoltaic power generation under different weather categories was predicted and the final prediction curve was obtained as follows: Figure 7 As shown in Table 3, the prediction results of M11 are closer to the actual photovoltaic power generation situation. The other models sometimes only capture the trend of photovoltaic output in the prediction and the predicted power fluctuates more than the actual value. The prediction accuracy evaluation is shown in Table 3. Compared with the other models, the average R 2 The results were improved by 0.0723, 0.0475, 0.0366, and 0.0311, respectively. The average MAE decreased by 0.7681 MW, 0.5630 MW, 0.4720 MW, and 0.3309 MW, respectively. The average RMSE decreased by 0.9353 MW, 0.6991 MW, 0.5604 MW, and 0.4654 MW, respectively. The results show that M11 has high forecast accuracy in different weather categories and can better capture the dynamic changes of photovoltaic power.

[0144] Table 3 Prediction accuracy evaluation indicators of different models

[0145]

[0146]

[0147] By gradually removing or replacing different modules of the model, it is easier to analyze the specific contribution of each component to the overall performance, thereby revealing the internal mechanism of the model. In this paper, two ablation experiments were conducted to analyze the impact of different similar day selection methods, different decomposition and reconstruction methods, and the presence or absence of error compensation mechanisms on the model to evaluate the model's performance under different conditions. The multi-period similar day method only considers the similarity of time periods.

[0148] Under different weather conditions, the prediction results are obtained by adopting multiple similar day selection strategies. Figure 8 As shown in the figure. On clear days, because PV output is relatively stable and close to a normal distribution, selecting similar days is relatively simple, resulting in a variety of similar day selection methods achieving high forecast accuracy. However, when faced with fluctuating meteorological conditions, the multi-period similar day method and the conventional similar day selection method perform poorly. Conventional similar day selection methods, when dealing with fluctuating trends, are unable to accurately capture the changing characteristics of meteorological fluctuations due to the large distance between the selected similar days and the test day, resulting in a tendency for forecast deviations.

[0149] The prediction accuracy evaluation is shown in Table 4. The results show that the prediction error is minimized by using the method of multi-period clustering to select similar days. Specifically, this method makes the average R 2 The accuracy improved by 0.0201 and 0.0087, respectively, while the average MAE decreased by 0.2665MW and 0.1231MW, and the average RMSE decreased by 0.3380MW and 0.1952MW, respectively. This improvement in accuracy highlights the importance of similar day selection methods. Using multi-period clustering to select similar days can simultaneously account for differences in time periods and the uniformity of clustering results. This allows for the selection of similar days that better align with the meteorological trends of the day being measured, thereby achieving higher forecast accuracy.

[0150] Table 4. Prediction accuracy evaluation of the first part of the ablation experiment

[0151]

[0152]

[0153] Figure 9 The prediction curves obtained by ablation experiments on the second part of the proposed model using different models are shown. Table 5 shows the prediction accuracy evaluation indicators:

[0154] Table 5. Prediction accuracy evaluation of the second part of the ablation experiment

[0155]

[0156] The results show that a single decomposition and reconstruction of PV output can extract the main trend components and periodic fluctuations, but it may still fail to fully capture subtle changes when dealing with complex climate changes, resulting in a decrease in prediction accuracy. In contrast, multi-level processing of PV output can more effectively remove noise from the data and better adapt to the complexity of PV output, significantly improving the prediction performance, specifically: average R 2 It is improved by 0.0174, and the average MAE and RMSE are reduced by 0.2448MW and 0.2954MW respectively.

[0157] Based on the secondary decomposition and reconstruction, a hybrid model was used to compare the prediction performance with a single model. The results showed that the hybrid model has a greater advantage in prediction accuracy, with an average R 2The average MAE and RMSE increased by 0.0200 and 0.0110, respectively, and the average MAE and RMSE decreased by 0.2623 MW, 0.1542 MW, and 0.3310 MW and 0.1948 MW, respectively. Compared with a single model, the hybrid model better captures the multi-scale characteristics of PV output, thereby improving forecast accuracy. When directly using the CNN-GRU-Attention model for forecasting, it may overfit low-frequency random fluctuations, resulting in a decrease in overall forecast performance.

[0158] After error compensation, the prediction accuracy of the model is further improved, and the average R 2 The average MAE and RMSE increased by 0.0062, and the average MAE and RMSE decreased by 0.1077 MW and 0.1429 MW, respectively. This improvement is attributed to the synergistic effect of the hybrid model and the error compensation mechanism after secondary decomposition and reconstruction, which helps to further reduce the prediction error and enhance the model's adaptability to changes in PV output, thereby improving overall prediction performance.

[0159] The following are device embodiments of the present invention, which can be used to perform the method embodiments of the present invention. For details not disclosed in the device embodiments, please refer to the method embodiments of the present invention.

[0160] The present invention also provides a short-term photovoltaic power generation prediction system, which runs the above-mentioned short-term photovoltaic power generation prediction method of multi-period clustering and signal reconstruction, including:

[0161] The meteorological feature preprocessing module is used to obtain key historical meteorological data, divide the key historical meteorological data according to geographical location and time, and perform dimensionality reduction and clustering on the divided meteorological features to obtain preprocessed meteorological features;

[0162] The similar day screening module is used to screen the similarity distance between the pre-processed meteorological characteristics and the meteorological characteristics of the test day, select meteorological characteristics with a similarity distance less than a set value and sort them to obtain similar days;

[0163] The photovoltaic output data reconstruction module is used to obtain photovoltaic output data corresponding to the meteorological characteristics of similar days, perform secondary decomposition and reconstruction on the photovoltaic output data to obtain subsequences with different characteristics, and merge the photovoltaic output data subsequences with the corresponding meteorological characteristics to obtain multiple sub-datasets;

[0164] The photovoltaic output prediction module is used to transfer the sub-datasets to the multimodal prediction model to obtain the final predicted value of the photovoltaic output on the test day.

[0165] In another embodiment of the present invention, a terminal device is provided, comprising a processor and a memory, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is used to execute the program instructions stored in the computer storage medium. The processor may be a central processing unit (CPU), or may be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to implement the corresponding method flow or corresponding function; the processor described in the embodiment of the present invention can implement the operation of a short-term photovoltaic power generation prediction method of multi-period clustering and signal reconstruction.

[0166] In another embodiment of the present invention, the present invention further provides a storage medium, specifically a computer-readable storage medium (Memory), which is a memory device in a terminal device for storing programs and data. It is understandable that the computer-readable storage medium here can include both built-in storage media in the terminal device and, of course, extended storage media supported by the terminal device. The computer-readable storage medium provides a storage space that stores the operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space. These instructions can be one or more computer programs (including program codes). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the short-term photovoltaic power generation prediction method of multi-period clustering and signal reconstruction in the above embodiment.

[0167] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0168] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0169] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0170] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0171] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A short-term photovoltaic power generation prediction method based on multi-period clustering and signal reconstruction, characterized in that: The specific steps are as follows: Obtain key historical meteorological data, divide the key historical meteorological data into time periods according to geographical location, perform dimensionality reduction and clustering on the divided meteorological features, and obtain the weather categories for each time period after preprocessing; After preprocessing, under the same weather category in each period, the divided meteorological characteristics are screened with the meteorological characteristics of the test day for similarity distance, and the meteorological characteristics with similarity distance less than the set value are selected and sorted to obtain similar days; Obtain PV output data corresponding to meteorological characteristics on similar days, perform secondary decomposition and reconstruction on the PV output data to obtain subsequences with different characteristics, and merge the PV output data subsequences with the corresponding meteorological characteristics to obtain multiple sub-datasets; The sub-datasets are transferred to the hybrid prediction model to obtain the final predicted value of photovoltaic output on the test day.

2. The short-term photovoltaic power generation prediction method based on multi-period clustering and signal reconstruction according to claim 1 is characterized in that: Obtain key historical meteorological data, divide the key historical meteorological data into time periods according to geographical location, perform dimensionality reduction and clustering on the divided meteorological features, and obtain the weather categories for each time period after preprocessing. Specifically: According to the sunrise and sunset times in different geographical locations, the key historical meteorological data are divided into three time periods: rising period, peak period, and falling period. The PCA-NRBO-K-means++ algorithm is used to reduce the dimension of the meteorological characteristics of the above three time periods and cluster them to obtain the distribution of weather types in different time periods.

3. The short-term photovoltaic power generation prediction method based on multi-period clustering and signal reconstruction according to claim 1 is characterized in that After preprocessing, under the same weather category in each period, the divided meteorological characteristics are screened with the meteorological characteristics of the test day by similarity distance, and meteorological characteristics with similarity distance less than the set value are selected and sorted to obtain the steps of similar days: The DTW algorithm is used to obtain the similarity distance between the meteorological characteristic data in different time periods and the meteorological characteristics of the measured time. In the time periods where the similarity distance is less than 5, the meteorological characteristic data are sorted and reorganized to obtain similar days. The meteorological characteristics of similar days are then merged with the corresponding photovoltaic output data to form a similar day data set of multiple time periods.

4. The short-term photovoltaic power generation prediction method based on multi-period clustering and signal reconstruction according to claim 1 is characterized in that: The steps of obtaining the photovoltaic output data corresponding to the meteorological characteristics of similar days, performing secondary decomposition and reconstruction on the photovoltaic output data, and merging it with the corresponding meteorological characteristics to obtain the sub-dataset are as follows: The photovoltaic output data is decomposed and reconstructed by using the VMD method, the photovoltaic output data is decomposed into intrinsic mode function components, and the components with entropy values greater than 0.5 are reconstructed to obtain the first high-frequency term; The CEEMDAN method is used to perform secondary decomposition and reconstruction on the first high-frequency term, and the secondary decomposition components and the components that are not reconstructed after the primary decomposition are obtained to obtain secondary decomposition components. Among the secondary decomposition components, the components with entropy values greater than 0.5 are reconstructed into second high-frequency terms; Components with entropy values between 0.2 and 0.5 are reconstructed as low-frequency terms; components with entropy values less than 0.2 are reconstructed as trend terms; The reconstructed photovoltaic output data is obtained by adding the second high-frequency term, the low-frequency term, and the trend term. The difference between the undecomposed and reconstructed photovoltaic output data and the reconstructed photovoltaic output data is the residual term. The photovoltaic output data of the second high-frequency item, low-frequency item, trend item and residual item are combined with the corresponding meteorological characteristics to obtain a high-frequency item dataset, a low-frequency item dataset, a trend item dataset and a residual item dataset.

5. The short-term photovoltaic power generation prediction method based on multi-period clustering and signal reconstruction according to claim 4 is characterized in that: The sub-dataset is transferred to the hybrid prediction model to obtain the final predicted value of the photovoltaic output on the test day: The hybrid prediction model includes the CNN-GRU-Attention model and the XGBoost model; The second high-frequency item dataset uses the CNN-GRU-Attention model for short-term photovoltaic power generation forecasting; The low-frequency item dataset, trend item dataset, and residual item dataset are used to perform short-term photovoltaic power generation forecasting using the XGBoost model; The predicted values of the second high-frequency term, low-frequency term, trend term, and residual term are combined to obtain the final predicted value of the photovoltaic output on the day to be measured.

6. The short-term photovoltaic power generation prediction method based on multi-period clustering and signal reconstruction according to claim 1 is characterized in that: Unfiltered historical meteorological data and photovoltaic output data were obtained as historical initial data. The historical initial data were cleaned and the key meteorological factors affecting photovoltaic output were screened using the Pearson and Spearman correlation analysis methods. The data corresponding to meteorological factors with absolute values of correlation coefficients greater than 0.4 are selected as key historical meteorological data.

7. A short-term photovoltaic power generation prediction system, characterized in that: Running a short-term photovoltaic power generation prediction method based on multi-period clustering and signal reconstruction according to any one of claims 1 to 6, comprising: The meteorological feature preprocessing module is used to obtain key historical meteorological data, divide the key historical meteorological data according to geographical location and time period, reduce the dimension and cluster the divided meteorological features, and obtain the weather category of each time period after preprocessing; The similar day screening module is used to screen the meteorological characteristics after classification and the meteorological characteristics of the test day under the same weather category in each period after preprocessing, select meteorological characteristics with similarity distance less than the set value and sort them to obtain similar days; The photovoltaic output data reconstruction module is used to obtain photovoltaic output data corresponding to the meteorological characteristics of similar days, perform secondary decomposition and reconstruction on the photovoltaic output data to obtain subsequences with different characteristics, and merge the photovoltaic output data subsequences with the corresponding meteorological characteristics to obtain multiple sub-datasets; The photovoltaic output prediction module is used to transfer the sub-datasets to the hybrid prediction model to obtain the final predicted value of the photovoltaic output on the test day.

8. A terminal device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the short-term photovoltaic power generation prediction method based on multi-period clustering and signal reconstruction are implemented as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the short-term photovoltaic power generation prediction method based on multi-period clustering and signal reconstruction are implemented as claimed in any one of claims 1 to 6.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the short-term photovoltaic power generation prediction method based on multi-period clustering and signal reconstruction are implemented as claimed in any one of claims 1 to 6.

Images