Photovoltaic power generation power clustering method fusing time-frequency fusion features

Abstract

The invention relates to a photovoltaic power generation power clustering method fusing time-frequency fusion features, and the method comprises the steps: carrying out the preprocessing of a synchronously collected original photovoltaic power sequence and meteorological data, and obtaining a preprocessed photovoltaic power sequence set and preprocessed meteorological data; performing time domain feature extraction and frequency domain feature extraction on each single-day photovoltaic power sequence to obtain a time domain feature vector and a frequency domain feature vector; performing feature fusion and principal component analysis dimensionality reduction on the time domain feature vectors and the frequency domain feature vectors corresponding to all the single-day photovoltaic power sequences, and outputting a dimensionality-reduced time-frequency fusion feature matrix; and inputting the clustering result into the improved variational Bayesian Gaussian mixture model, and outputting the clustering result. The clustering result and the preprocessed meteorological data can be jointly used as input of a subsequent photovoltaic power prediction model. According to the method, more reliable classification input can be provided for the photovoltaic power prediction model, and improvement of prediction precision is effectively supported.

Description

Technical Field

[0001] This invention belongs to the field of intelligent meteorological data analysis technology, specifically involving a photovoltaic power generation clustering method that integrates time-frequency fusion features. Background Technology

[0002] In photovoltaic (PV) power generation systems, high-precision power forecasting is crucial for stable grid dispatch, optimized electricity market transactions, and the operation and maintenance management of the power plant itself. PV output is directly affected by various meteorological factors such as solar irradiance, temperature, and cloud cover, and its volatility and intermittency pose significant challenges to forecasting. To improve the accuracy of forecasting models, an effective approach is to first divide the complex and variable historical meteorological data into several representative "meteorological models," and then construct or adjust forecasting models for each model. Therefore, accurate clustering of meteorological data is a key preliminary step in optimizing the input of PV power forecasting models and improving their predictive performance.

[0003] In existing technologies, some studies attempt to improve performance by optimizing the structure of the clustering model itself, but often fail to fully consider the essential characteristics of the dynamic evolution of meteorological data over time. These methods typically rely only on single-dimensional statistical features (such as mean and variance) like irradiance and temperature to construct clustering indices, making it difficult to comprehensively and precisely characterize the trend evolution of meteorological conditions over time and the periodic fluctuations in frequency. This one-sidedness in feature representation leads to weak discriminative and generalization abilities in the clustering results, failing to provide reliable input with high discriminative power for subsequent power prediction models. Furthermore, traditional methods are susceptible to noise interference and overfitting during training, and their clustering stability and accuracy still have considerable room for improvement under complex real-world meteorological conditions. Summary of the Invention

[0004] To address the aforementioned problems in the existing technology, this invention provides a photovoltaic power generation clustering method that integrates time-frequency fusion features. The technical problem to be solved by this invention is achieved through the following technical solution: This invention provides a photovoltaic power generation clustering method that integrates time-frequency fusion features, comprising: The collected raw photovoltaic power sequence is preprocessed to obtain a preprocessed photovoltaic power sequence set. The preprocessing includes: time alignment, outlier correction and missing value supplementation. For each daily photovoltaic power sequence in the preprocessed photovoltaic power sequence set, time domain feature extraction and frequency domain feature extraction are performed respectively to obtain time domain feature vector and frequency domain feature vector; For all daily photovoltaic power sequences, feature fusion and principal component analysis are performed on the time-domain and frequency-domain feature vectors to reduce the dimensionality, and the reduced time-frequency fusion feature matrix is ​​output. The dimensionality-reduced time-frequency fusion feature matrix is ​​input into the improved variational Bayesian-Gaussian mixture model. The Dirichlet prior concentration parameter of the model is adaptively adjusted according to the joint entropy of the dimensionality-reduced time-frequency fusion feature matrix. Model inference is performed by maximizing the lower bound of evidence, and clustering results are output. The clustering results include the meteorological model category to which each daily photovoltaic power sequence belongs. The clustering results are used to distinguish different meteorological models and optimize the input feature structure of the photovoltaic power generation prediction model to improve prediction accuracy.

[0005] Compared with the prior art, the beneficial effects of the present invention are as follows: To address the shortcomings of existing meteorological clustering methods, which fail to fully consider the dynamic time-frequency characteristics of meteorological data, resulting in low clustering accuracy and poor generalization ability, this invention provides a photovoltaic power generation clustering method that integrates time-frequency fusion features. By constructing time-frequency fusion features and combining them with adaptive variational Bayesian clustering, the accuracy and practicality of meteorological model classification are significantly improved. At the feature level, the method integrates morphological statistical features in the time domain and fluctuating energy distribution features in the frequency domain to comprehensively depict the dynamic changes in power under the influence of meteorology, overcoming the shortcomings of traditional single features and insufficient expressive power. At the model level, an adaptive mechanism for prior parameters based on time-frequency joint entropy is introduced, eliminating the need for manual pre-setting of the number of categories during the clustering process and automatically suppressing noise and overfitting based on data complexity, thereby improving the robustness and generalization ability of the model. The final output of high-discrimination meteorological model labels provides a more reliable classification input for photovoltaic power prediction models, effectively supporting the improvement of prediction accuracy. Attached Figure Description

[0006] Figure 1 This is a flowchart illustrating the photovoltaic power generation clustering method that integrates time-frequency fusion features provided in an embodiment of the present invention. Figure 2 This is a comparison chart of clustering effects using only GMM provided in an embodiment of the present invention; Figure 3 This is a comparison chart of clustering effects using only VBGMM provided in this embodiment of the invention; Figure 4 This is a comparison chart of the clustering effects of the method proposed in this invention, provided in an embodiment of this invention; Figure 5 This describes the calculation of various evaluation indicators under different clustering methods provided in the embodiments of the present invention. Detailed Implementation

[0007] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0008] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0009] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0010] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, disclosure, and appended claims in carrying out the claimed invention. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0011] The photovoltaic power generation clustering method that integrates time-frequency fusion features, proposed in this invention, will now be described in detail with reference to the accompanying drawings.

[0012] Figure 1 This is a flowchart illustrating the photovoltaic power generation clustering method that integrates time-frequency fusion features provided in an embodiment of the present invention. Figure 1 As shown, the method includes steps S110-140; specifically: S110: Preprocess the acquired raw photovoltaic power sequence to obtain a preprocessed photovoltaic power sequence set. The preprocessing includes time alignment, outlier correction, and missing value supplementation.

[0013] Here, the method further includes: simultaneously collecting meteorological data while collecting the original photovoltaic power sequence; performing the preprocessing on the meteorological data to obtain preprocessed meteorological data; performing feature filtering on the preprocessed meteorological data to obtain a filtered subset of meteorological features; the filtered subset of meteorological features is used together with the clustering result as input features of the photovoltaic power generation prediction model.

[0014] Specifically, raw photovoltaic power sequences and meteorological data are typically obtained through Supervisory Control and Data Acquisition (SCADA) systems deployed in photovoltaic power plants or dedicated weather stations. For example, the raw photovoltaic power sequence is the real-time active power output recorded by the power plant's inverters or meters at 5-minute or 15-minute intervals; variables such as global horizontal irradiance (GHI), ambient temperature, relative humidity, wind speed, and wind direction are synchronously recorded by meteorological sensors within the station. The raw photovoltaic power sequence and meteorological data are linked through timestamps.

[0015] After acquiring the data, due to potential slight clock deviations or incomplete overlap in sampling times between different monitoring devices (electricity meters, irradiance meters, temperature and humidity sensors, etc.), the power data and all meteorological variable data need to be resampled or interpolated to a uniform, regular timestamp sequence to eliminate analytical errors caused by time misalignment. Subsequently, robust algorithms such as Huber regression are used for outlier correction. This method identifies anomalies based on physical and statistical laws. For example, photovoltaic power should be zero or within a reasonable range at night; negative values ​​or positive values ​​far exceeding the rated capacity are considered anomalies. For meteorological data, judgments are based on physical conditions or reasonable ranges. For identified outliers (whether power or meteorological variables), robust algorithms such as Huber regression are uniformly used for correction. This method uses squared loss for small fluctuations conforming to the model (considered normal) to maintain data accuracy; and uses linear loss for large outliers (considered anomalies) to suppress their impact, thus achieving smooth and reasonable correction and avoiding information loss caused by simple removal. Finally, missing values ​​are added to the outlier-corrected data. Specifically, for short-term missing data (such as a single point or several consecutive minutes), linear interpolation is used to fill in the gaps using valid data points before and after the missing data; for long-term missing data (such as several consecutive hours), machine learning-based prediction methods are used to fill in the gaps by utilizing the physical correlation between variables. The final result is a preprocessed photovoltaic power sequence set and preprocessed meteorological data.

[0016] S120: Perform time-domain feature extraction and frequency-domain feature extraction on each daily photovoltaic power sequence in the preprocessed photovoltaic power sequence set to obtain time-domain feature vectors and frequency-domain feature vectors.

[0017] It should be noted that meteorological models (such as sunny, cloudy, and rainy) are defined and changed based on the "day". The power sequence within a day carries the complete weather impact characteristics of that day.

[0018] Specifically, S120 includes: in the time domain, calculating the number of extreme points, trend slope, coefficient of variation, skewness, and kurtosis of each daily photovoltaic power sequence, and combining them into a time domain feature vector; in the frequency domain, performing fast Fourier transform, frequency band division, and normalized energy ratio sequentially on each daily photovoltaic power sequence to obtain a frequency domain feature vector.

[0019] In the time domain, the number of extreme points, trend slope, coefficient of variation, skewness, and kurtosis of each daily photovoltaic power sequence are calculated and combined into a time domain feature vector. This includes: calculating the first-order difference sequence of each daily photovoltaic power sequence and counting the number of sign changes of adjacent difference values ​​in the first-order difference sequence as the number of extreme points; performing linear regression fitting on the daily photovoltaic power sequence and the corresponding time index, and using the regression coefficient obtained as the trend slope; and combining the number of extreme points, trend slope, coefficient of variation, skewness, and kurtosis to form the time domain feature vector.

[0020] For example, the number of extreme points can be calculated based on the following formula: ; in, , For indicator functions, The length of the power sequence. For the first Power value at any time For the first Power value at any given time.

[0021] The trend slope can be calculated based on the following formula: ; ; in, The slope The average of the time indices. The mean of the power series. for.

[0022] Here, in the frequency domain, a Fast Fourier Transform is sequentially performed on each daily photovoltaic power sequence, followed by frequency band division and normalized energy ratio, to obtain a frequency domain feature vector. This includes: based on the physical characteristics of photovoltaic power fluctuations, dividing the spectrum obtained from the transformation of each daily photovoltaic power sequence into a first frequency band, a second frequency band, and a third frequency band according to frequency from low to high; calculating the frequency band energy within each frequency band to obtain the first frequency band energy, the second frequency band energy, and the third frequency band energy; dividing the first frequency band energy, the second frequency band energy, and the third frequency band energy by the total energy to obtain the first normalized energy ratio, the second normalized energy ratio, and the third normalized energy ratio; the total energy is the sum of the first frequency band energy, the second frequency band energy, and the third frequency band energy; and combining the first normalized energy ratio, the second normalized energy ratio, and the third normalized energy ratio to form the frequency domain feature vector.

[0023] The first frequency band is a frequency band with a frequency lower than f1, corresponding to a fluctuation component with a period longer than T1, used to reflect the trend component in the daily photovoltaic power sequence determined by the stable change of solar irradiance under clear sky conditions; the second frequency band is a frequency band with a frequency between f1 and f2, corresponding to a fluctuation component with a period between T1 and T2, used to reflect the quasi-periodic fluctuation component in the daily photovoltaic power sequence caused by cloud movement and intermittent shading; the third frequency band is a frequency band with a frequency higher than f2, corresponding to a fluctuation component with a period shorter than T2, used to reflect the random disturbance component in the daily photovoltaic power sequence caused by intermittent precipitation or measurement noise; where f1<f2,T1> T2>0.

[0024] It should be noted that after the fast Fourier transform (FFT) of each daily photovoltaic (PV) power sequence, its spectrum includes all frequency components from 0 to the Nyquist frequency. Subsequently, the entire frequency range is artificially divided into three distinct frequency bands: the first band primarily captures the overall trend throughout the day caused by changes in the sun's position, as well as the slow changes in large-scale weather systems, appearing as a smooth single-peak curve under clear skies. The second band mainly reflects regular power fluctuations caused by cloud movement and intermittent shading, appearing as multiple dips and recoveries on the power curve. The third band is mainly related to measurement noise, short-term strong disturbances (such as the initial drop of rain), or rapidly changing local meteorological conditions, appearing as high-frequency spikes on the power curve. Ultimately, these three normalized energy proportions together constitute a three-dimensional frequency domain feature vector. This vector summarizes the dynamic behavior pattern of the daily PV power sequence in the frequency dimension in a compact and physically clear manner: the proportion of high and low frequencies represents a stable trend, the proportion of high and medium frequencies represents periodic shading, and the proportion of high and high frequencies represents disordered disturbances.

[0025] For example, for meteorological power sequences Perform Fast Fourier Transform The spectrum is obtained, and the frequency domain is divided into three segments—low frequency, mid frequency, and high frequency—based on the frequency bandwidth. The frequency band energy is then calculated for each segment. Reconstruct the normalized energy ratio As a frequency domain feature. This refers to the frequency index, which is the number of each frequency point in the spectrum obtained after the Fast Fourier Transform, corresponding to different frequency components. Refers to meteorological power series The length (number of samples) is the total number of original data points participating in the Fast Fourier Transform. It is the first frequency band energy. It is the second frequency band energy. It is third frequency band energy.

[0026] S130: Perform feature fusion and principal component analysis to reduce the dimensionality of the time-domain and frequency-domain feature vectors corresponding to all daily photovoltaic power sequences, and output the dimensionality-reduced time-frequency fusion feature matrix.

[0027] Specifically, S130 includes: concatenating the time-domain feature vectors and frequency-domain feature vectors corresponding to all daily photovoltaic power sequences in the feature dimension to obtain the original fused feature matrix; standardizing the original fused feature matrix to obtain the standard feature matrix; calculating the covariance matrix of the standard feature matrix; performing eigenvalue decomposition and principal component selection on the covariance matrix of the standard feature matrix to construct the projection matrix; and reducing the dimensionality of the standard feature matrix based on the projection matrix to generate the dimensionality-reduced time-frequency fused feature matrix, wherein each row in the dimensionality-reduced time-frequency fused feature matrix represents a dimensionality-reduced time-frequency fused feature vector corresponding to a daily photovoltaic power sequence.

[0028] For example, there are 3 samples, each with a time-domain feature vector of 5 and a frequency-domain feature vector of 3. The 5-dimensional time feature and the 3-dimensional frequency feature of each sample are concatenated along the feature dimension to obtain an 8-dimensional original fused feature vector. These three 8-dimensional vectors are stacked to obtain a 3×8-dimensional original fused feature matrix. The mean and standard deviation of each column (each feature) of the 3×8-dimensional original fused feature matrix are calculated, and Z-score standardization is performed to obtain the standard feature matrix. Eliminate the influence of dimensions. Calculate the standard characteristic matrix. covariance matrix Solve for the eigenvalues ​​and eigenvectors of the covariance matrix, and select the first... The eigenvector matrix is ​​formed by the eigenvectors corresponding to the largest eigenvalues. Through formula Obtain the dimensionality-reduced fusion feature matrix .

[0029] S140: Input the dimensionality-reduced time-frequency fusion feature matrix into the improved variational Bayesian-Gaussian mixture model, and adaptively adjust the Dirichlet prior concentration parameter of the model according to the joint entropy of the dimensionality-reduced time-frequency fusion feature matrix. Perform model inference by maximizing the lower bound of evidence, and output clustering results. The clustering results include the meteorological model category to which each daily photovoltaic power sequence belongs. The clustering results are used to distinguish different meteorological models and optimize the input feature structure of the photovoltaic power generation prediction model to improve prediction accuracy.

[0030] Specifically, S140 includes: calculating the time-frequency joint entropy based on the distribution of each feature component in the time-frequency fusion feature matrix; mapping the time-frequency joint entropy to the concentration parameter of the Dirichlet process prior through a preset mapping function; initializing the model's Dirichlet prior using the concentration parameter, and performing variational Bayes iterative inference with the goal of maximizing the lower bound of evidence, successively updating the posterior distribution approximation of the model parameters; terminating the optimization process and obtaining the inference result when the iterative inference meets the preset convergence condition; wherein, the preset convergence condition is: the absolute value of the difference between the lower bound values ​​of evidence calculated in two consecutive iterations is less than a preset threshold, or the number of iterations reaches the preset maximum number of iterations; determining the cluster category with the maximum posterior probability corresponding to each daily photovoltaic power sequence in the inference result as the clustering result of that sequence.

[0031] For example, the following steps are performed within the improved variational Bayesian-Gaussian mixture model: (1) Initialize VBGMM parameters: data From Dirichlet's a priori ,in For concentration parameters, mean Follows a normal distribution Inverse covariance matrix Follows the Weishat distribution ; (2) Adaptive adjustment of prior concentration parameters driven by time-frequency joint entropy: To enhance the model's adaptability to different meteorological complexities, a prior adjustment mechanism driven by time-frequency entropy is introduced to adjust the fused feature vector for each day. Calculate the joint time-frequency entropy ;in The probability components are normalized time-frequency features, and the Dirichlet prior concentration is dynamically adjusted based on the joint entropy. , This refers to the initial baseline value of the prior concentration. This refers to the adjustment coefficient. The larger the entropy, the greater the change in entropy. The more significant the adjustment, the more pronounced the effect; when weather fluctuations are complex (high entropy), the more effective the adjustment. This preserves more of the mixed components; when the sequence is stable (low entropy), Reduce and suppress redundant components to avoid over-clustering; in, It is the joint entropy of time-frequency features, characterizing the time-frequency complexity of meteorological sequences. The higher the entropy value, the more complex the meteorological fluctuations, and the model automatically increases the prior strength to improve separability; the lower the entropy value, the lower the prior strength to avoid over-clustering. (3) Variational Bayesian Inference and ELBO Optimization: With the objective of maximizing the lower bound of evidence (ELBO), the posterior distributions of the mixing coefficients, mean, and covariance are iteratively optimized, as shown in the formula: .in Lower bound of evidence; Latent variables Z and model parameters The approximate posterior distribution; It is a joint probability distribution; Indicates based on distribution Expectations; (4) After iterative convergence, the adaptive clustering results are output, and the meteorological data are divided into three categories: sunny (high proportion of low frequency energy and stable trend), cloudy (prominent medium frequency energy), and rainy (high proportion of high frequency energy and many extreme points).

[0032] To address the problem that existing meteorological clustering methods fail to fully consider the dynamic time-frequency characteristics of meteorological data, resulting in low clustering accuracy and poor generalization ability, this invention provides a photovoltaic power generation clustering method that integrates time-frequency fusion features. By innovatively constructing a time-frequency fusion feature system and driving an improved adaptive variational Bayesian-Gaussian mixture model for clustering, significant technical effects are achieved in the meteorological model classification stage of photovoltaic power generation prediction. Specifically, at the feature extraction level, the method simultaneously extracts complementary dynamic information from the time and frequency domains: time-domain statistical features (such as extreme points and trend slopes) accurately characterize the morphological jumps and overall trends of the power sequence; while by dividing frequency bands through fast Fourier transform and calculating the normalized energy ratio, the fluctuation patterns dominated by different meteorological processes (such as stable trends, cloud cover, and rapid disturbances) are quantitatively analyzed from the frequency dimension. This multi-dimensional fusion feature expression, from... This method fundamentally overcomes the limitations of traditional methods that rely on single-dimensional statistics, achieving a comprehensive and detailed characterization of complex meteorological models. At the model level, by calculating the joint entropy of time-frequency fusion features, the method dynamically adjusts the Dirichlet prior concentration parameter of the variational Bayesian-Gaussian mixture model, enabling the model to adaptively determine the number of clusters based on the complexity of the data itself. This avoids the subjectivity of manual pre-setting and enhances the model's robustness to noise and its ability to fit complex data distributions. Subsequently, features optimized by principal component analysis are input into this adaptive model for inference, resulting in clustering results with clear boundaries and high class discrimination. Overall, this method generates highly discriminative meteorological model labels through the close integration of "comprehensive feature representation" and "adaptive intelligent clustering," providing downstream photovoltaic power prediction models with structurally optimized and highly reliable input features, thus providing a solid guarantee for fundamentally improving the accuracy of power prediction models.

[0033] To verify the effectiveness of this invention, simulation software is used for verification.

[0034] Figure 2This is a comparison chart of the clustering effects of the proposed method and existing methods provided in this embodiment of the invention. The photovoltaic power generation data used here comes from the Alice Springs photovoltaic power station in Australia, covering the entire year of 2016. This dataset contains various meteorological and environmental variables required for photovoltaic power prediction. Specifically, it includes the following features: photovoltaic output power (Power), wind speed (WS), temperature (Temp), relative humidity (Humidity), global horizontal radiation (GHR), diffuse horizontal radiation (DHR), wind direction (WD), rainfall (Rain), actual ground temperature (RGT), and actual downward radiation temperature (RDT). To eliminate redundant information and identify the most informative variables, the Pearson correlation coefficient method and the Recursive Feature Elimination Algorithm and Cross-Validation (RFECV) algorithm are combined, and Light GBM is used as the base learner to achieve adaptive feature subset selection. The final optimal feature subset is determined to be GHR, RGT, temperature, humidity, DHR, and RDT. Since the photovoltaic output is always zero at night, the analysis only considers the data from 07:00 to 19:00. The dataset has a time resolution of 5 minutes. To improve data quality, Huber regression is used for outlier correction, and a tree-based model is combined with linear interpolation to fill in missing values.

[0035] Here, the SC, CHI, DBI, Dunn, and Gap indices are used to evaluate the clustering results to ensure that the clustering quality meets the input requirements for photovoltaic power generation prediction. The calculation method is as follows: 1)SC: ,in For the sample The average distance to all other samples in the same cluster, i.e. ; This refers to the cluster where the sample belongs, i.e., the current cluster. For the sample The average distance to all samples within the nearest heterogeneous cluster, i.e. ), To be compatible with the current cluster For adjacent heterogeneous clusters, the closer they are to 1, the better the clustering effect. 2) CHI: ,in global mean , Let be the intra-cluster mean of the k-th cluster. K This represents the number of clusters. n The total number of samples; a larger CHI indicates more significant inter-cluster differences and smaller intra-cluster differences, resulting in better clustering performance. 3) DBI: , For the number of clusters, DBI refers to the inter-class / intra-class evaluation index between cluster k and cluster l; the smaller the DBI, the better the balance between intra-cluster compactness and inter-cluster separation, and the better the clustering effect. 4) The expression for calculating the Dunn coefficient is: ; For the first There are m clusters, where the numerator is the minimum sample distance between any two different clusters; the denominator is the largest cluster diameter (the maximum distance between any two samples within a cluster); the larger the Dunn, the higher the separation between clusters, the better the compactness within clusters, and the better the clustering effect. 5) Gap coefficient: , The expectation operation involves averaging the statistic within the parentheses over multiple random data experiments. This refers to the sum of squared deviations within a cluster of random data. The sum of squared intra-cluster deviations of the actual data. Gap ( K The larger the value, and the more it satisfies ( When the standard deviation of the reference error is used, the corresponding The larger the Gap value, the more significant the clustering structure of the actual data is compared to that of the random data, indicating a better clustering effect.

[0036] like Figure 2-4 As shown, existing methods include Gaussian Mixture Model (GMM) and Variational Bayesian Gaussian Mixture Model (VBGMM). Based on... Figure 2-4 It can be concluded that the points of different colors in the figure represent samples from different clusters. If points of the same color are densely clustered and points of different colors are widely distributed, it indicates a good clustering effect. When only GMM or VBGMM is used, the data clusters show obvious overlap, such as the blurred boundaries between the yellow and green clusters and the loose distribution of samples within the clusters. However, after introducing the time-series and frequency-domain fusion features of PCA, the separation of each cluster is significantly improved, the compactness of samples within the same cluster is significantly enhanced, and the clustering of the blue cluster and the boundary distinction between the yellow and green clusters are optimized. This shows that multi-dimensional feature processing can effectively enhance the separability of data. Combined with the VBGMM algorithm, more accurate cluster partitioning can be achieved, and the clustering effect is significantly improved. Figure 5 This describes the calculation of various evaluation indicators under different clustering methods provided in the embodiments of the present invention. For example... Figure 5As shown in the figure, this chart compares the performance of six clustering methods using five clustering evaluation metrics: CHI, SC, Gap, Dunn, and DBI. Higher values ​​for CHI, SC, Gap, and Dunn indicate better clustering results, while a lower DBI value indicates better performance. The results show that the VBGMM-TS-FFT method, which incorporates time-frequency feature fusion, performs best across all metrics. Specifically, CHI reaches 231.41, indicating the most significant inter-cluster differences; SC is 0.39, indicating a more reasonable intra-cluster and inter-cluster structure; Gap and Dunn are both at their maximum values; and DBI is at its minimum value. In contrast, traditional methods such as VBGMM, FCM, and KMeans++ show significantly weaker performance across all metrics, demonstrating that time-frequency feature fusion effectively improves the overall performance of clustering methods.

[0037] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A photovoltaic power generation power clustering method fusing time-frequency fusion features, characterized in that, The method comprises: The method comprises: The method comprises: The method comprises: The method comprises:

2. The photovoltaic power generation power clustering method of claim 1, wherein, The method comprises: The method comprises: The method comprises: 3.The photovoltaic power generation power clustering method of fusing time-frequency fusion features according to claim 2, characterized in that, The method comprises: The method comprises: The method comprises: The method comprises: 4.The photovoltaic power generation power clustering method of fusing time-frequency fusion features according to claim 2, characterized in that, The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: The method comprises: divide the first frequency band energy, the second frequency band energy and the third frequency band energy by a total energy to obtain a first normalized energy ratio, a second normalized energy ratio and a third normalized energy ratio; the total energy is a sum of the first frequency band energy, the second frequency band energy and the third frequency band energy; combine the first normalized energy ratio, the second normalized energy ratio and the third normalized energy ratio to form the frequency domain feature vector.

5. The photovoltaic power clustering method of claim 4, wherein, the first frequency band is a frequency band with a frequency lower than f1, corresponding to a fluctuation component with a period longer than T1, and used to reflect a trend component in the single-day photovoltaic power sequence determined by the smooth change of solar irradiance under clear sky conditions; the second frequency band is a frequency band with a frequency between f1 and f2, corresponding to a fluctuation component with a period between T1 and T2, and used to reflect a quasi-periodic fluctuation component in the single-day photovoltaic power sequence caused by cloud movement and intermittent shading; the third frequency band is a frequency band with a frequency higher than f2, corresponding to a fluctuation component with a period shorter than T2, and used to reflect a random disturbance component in the single-day photovoltaic power sequence caused by array precipitation or measurement noise; wherein f1 < f2, T1 > T2 > 0. 6.The photovoltaic power generation power clustering method of fusing time-frequency fusion features according to claim 1, characterized in that, perform feature fusion and principal component analysis dimension reduction on the time domain feature vectors and the frequency domain feature vectors corresponding to all single-day photovoltaic power sequences, and output a reduced time-frequency fusion feature matrix, including: concatenate the time domain feature vectors and the frequency domain feature vectors corresponding to all single-day photovoltaic power sequences in the feature dimension to obtain an original fusion feature matrix; perform standardization processing on the original fusion feature matrix to obtain a standard feature matrix; calculate the covariance matrix of the standard feature matrix; perform eigenvalue decomposition and principal component number selection on the covariance matrix of the standard feature matrix to construct a projection matrix; perform dimension reduction on the standard feature matrix based on the projection matrix to generate the reduced time-frequency fusion feature matrix, wherein each row in the reduced time-frequency fusion feature matrix represents a reduced time-frequency fusion feature vector corresponding to a single-day photovoltaic power sequence. 7.The photovoltaic power generation power clustering method of fusing time-frequency fusion features according to claim 1, characterized in that, input the reduced time-frequency fusion feature matrix into an improved variational Bayesian Gaussian mixture model, and adaptively adjust the Dirichlet prior concentration parameter of the model according to the joint entropy of the reduced time-frequency fusion feature matrix, perform model inference by maximizing the lower bound of evidence, and output the clustering result, including: calculate the time-frequency joint entropy based on the distribution of each feature component in the time-frequency fusion feature matrix; map the time-frequency joint entropy to the concentration parameter of the Dirichlet process prior through a preset mapping function; initialize the Dirichlet prior of the model using the concentration parameter, and perform variational Bayesian iterative inference to update the posterior distribution approximation of the model parameters step by step with the goal of maximizing the lower bound of evidence. When the iterative inference satisfies a preset convergence condition, the optimization process is terminated, and an inference result is obtained; the preset convergence condition is that an absolute value of a difference between lower bound values of evidence obtained by two consecutive iterations is less than a preset threshold, or a number of iterations reaches a preset maximum number of iterations; In the inference result, a cluster category corresponding to each single-day photovoltaic power sequence and having a maximum posterior probability is determined as a clustering result of the sequence. 8.The photovoltaic power generation power clustering method of fusing time-frequency fusion features according to claim 1, characterized in that, The method further includes: When the original photovoltaic power sequence is collected, meteorological data is synchronously collected; The preprocessing is performed on the meteorological data, and preprocessed meteorological data is obtained; Feature screening is performed on the preprocessed meteorological data, and a screened meteorological feature subset is obtained; the screened meteorological feature subset is used as an input feature of the photovoltaic power generation prediction model together with the clustering result.

Images