A multi-view clustering ultra-short-term photovoltaic irradiance prediction method and system based on adaptive weighted missing completion

Abstract

The application discloses a kind of multi-view clustering ultra-short-term photovoltaic irradiance prediction methods and systems of self-adaptive weighting missing completion, comprising: the multi-modal historical data set of target photovoltaic power station is preprocessed, and data set is formed;Construct the irradiance prediction framework of missing data multi-view clustering, including pre-training module, clustering module and regression prediction module;Pre-training module is used to encode, decode and missing value completion to input data, and complete data is output to regression prediction module, and the low-level feature after completion is output to clustering module;Clustering module obtains clustering result based on low-level feature;Regression prediction module trains corresponding regression prediction submodel respectively to each kind of completion data, and outputs irradiance prediction result;The prediction framework is jointly trained;Call regression prediction submodel, and input sample is predicted.The application significantly improves the precision and robustness of ultra-short-term photovoltaic power generation irradiance prediction under the condition of incomplete data.

Description

Technical Field

[0001] This invention belongs to the field of multimodal data prediction technology, and relates to an adaptive weighted missing completion multi-view clustering ultra-short-term photovoltaic irradiance prediction method and system. Background Technology

[0002] With the continuous expansion of clean energy applications, the proportion of photovoltaic (PV) power generation in the power system is constantly increasing. The output power of PV power plants is closely related to solar irradiance, which is affected by a variety of factors such as cloud cover, humidity, wind speed, wind direction, solar altitude angle, and cloud movement, exhibiting significant randomness, volatility, and non-stationarity. Therefore, accurate ultra-short-term PV irradiance forecasting is of great significance for grid dispatching, PV power plant operation control, and the safe and stable operation of the power system.

[0003] Existing methods for predicting photovoltaic irradiance mainly include physical modeling methods, traditional machine learning methods, and deep learning methods. Physical modeling methods typically rely on numerical weather prediction, radiative transfer mechanisms, and related prior knowledge, offering a degree of interpretability, but are heavily dependent on specialized knowledge and external meteorological data. Traditional machine learning methods can establish a mapping relationship between multivariate time-series data and irradiance, but still have limitations in modeling complex nonlinear relationships. While deep learning methods have advantages in time-series feature extraction and nonlinear relationship modeling, under complex weather conditions, a single data source or a single prediction model is still insufficient to fully characterize the irradiance variation patterns.

[0004] In actual photovoltaic power plant operation, multimodal data such as multivariate time series data and ground-based cloud image data are often missing due to factors such as sensor failure, communication anomalies, image acquisition equipment obstruction, or data storage errors. Existing methods typically first use mean imputation, linear interpolation, or reconstruction models to complete the missing data before inputting the completed data into the prediction model. Such methods usually separate the missing data completion, feature learning, and prediction modeling processes, making it difficult to dynamically adjust according to the effectiveness and reliability of different data perspectives in different samples. This easily introduces completion bias, which in turn affects the accuracy and stability of subsequent prediction results.

[0005] Furthermore, the characteristics of photovoltaic irradiance variation differ significantly under different weather patterns. Irradiance changes are relatively stable under clear skies, while fluctuations are more pronounced under cloudy, alternating cloudy / sunny, or rapidly moving cloud conditions. Using a single prediction model to uniformly model all samples is insufficient to adequately adapt to the irradiance variation characteristics under different weather patterns. Multi-view clustering methods can utilize complementary information between multivariate time-series data and ground-based cloud image data to segment weather patterns; however, existing methods typically lack effective adaptive completion mechanisms and clustering consistency constraints under data gaps, leading to insufficient reliability of clustering and prediction results. Summary of the Invention

[0006] Purpose of the invention: The purpose of this invention is to provide an adaptive weighted missing data completion multi-view clustering ultra-short-term photovoltaic irradiance prediction method that can solve the problems of low accuracy and insufficient model adaptability caused by incomplete multivariate time series data due to missing multivariate time series data and ground cloud image data.

[0007] Technical solution: The present invention provides an adaptive weighted missing-filling multi-view clustering method for predicting ultra-short-term photovoltaic irradiance, comprising the following steps:

[0008] Collect multimodal historical datasets of the target photovoltaic power station, including multivariate time series data and corresponding ground cloud image data; preprocess the data to construct a dataset; and divide the dataset into training set, validation set and test set according to a predetermined ratio.

[0009] An irradiance prediction framework based on multi-view clustering of missing data is constructed. The framework includes a pre-training module, a clustering module, and a regression prediction module. The pre-training module comprises a CNN module, a BiLSTM encoder, a bidirectional prediction mechanism, a ScoreNet MLP, and a BiLSTM decoder. Multivariate time-series data without missing values ​​from the training set are input into the corresponding BiLSTM encoder. Simultaneously, ground-based cloud image data without missing values ​​are first processed by the CNN module to extract spatial features before being input into the BiLSTM encoder. The BiLSTM encoder outputs low-level features, which are then sequentially processed by the bidirectional prediction mechanism and ScoreNet. The MLP performs adaptive weighted imputation on missing data to obtain imputed low-level features. These imputed low-level features are then input into a BiLSTM decoder to obtain imputed data. The decoded imputed data is compared with the corresponding original data in the training set to calculate reconstruction loss and prediction loss. The clustering module performs multi-view clustering learning based on the imputed low-level features to obtain clustering results, and divides the imputed low-level features into K classes according to the clustering results. The regression prediction module receives the imputed data and its corresponding clustering results, and trains regression prediction sub-models 1 to K for classes 1 to K respectively. During prediction, the soft-assignment probabilities of the input sample in each view are averaged, and the class with the highest average probability is determined as the class to which the input sample belongs. Then, the corresponding regression prediction sub-model is called to output the irradiance prediction result.

[0010] Set training parameters, perform joint training on the prediction framework, evaluate the performance of the training results, and save the optimal prediction framework parameters.

[0011] The historical input data of the target photovoltaic power station is fed into the trained prediction framework. Based on the cluster category to which the input sample belongs, the corresponding regression prediction sub-model is called to output the prediction result of future photovoltaic power generation irradiance.

[0012] Furthermore, the multivariate time series data includes nine variables: azimuth, cloud cover, horizontal diffuse irradiance, direct normal irradiance, global horizontal irradiance, relative humidity, wind direction, wind speed, and zenith angle.

[0013] Furthermore, the data is preprocessed to construct a dataset, specifically as follows:

[0014] Multivariate time series data and their corresponding ground cloud image data are converted into formats and finally integrated and stored into a single HDF5 file format;

[0015] For the multivariate time series data, a time series curve is plotted, and outlier judgment thresholds are set according to the physical value ranges of each variable, such as azimuth, wind direction, zenith angle, cloud cover, and relative humidity. The value ranges of wind speed, horizontal diffuse irradiance, direct normal irradiance, and global horizontal irradiance are all set to be no less than 0 and less than the upper limit threshold. The upper limit threshold is set according to the historical distribution of the corresponding variable in the training set.

[0016] Values ​​exceeding the corresponding threshold range among the nine variables are identified as outliers and removed. Missing positions resulting from the removal of outliers are filled using linear interpolation.

[0017] Furthermore, the low-level features are sequentially augmented with a bidirectional prediction mechanism and ScoreNet MLP to adaptively weight and complete the missing data, resulting in the augmented low-level features; specifically including:

[0018] Record the current perspective that needs to be completed as the target perspective. Other perspectives used to generate cross-predictions for the target perspective are denoted as source perspectives. ,in ;

[0019] First, from the perspective of each source Generate cross-prediction:

[0020] ;

[0021] in, Indicates the perspective from the source Generate target perspective through cross-prediction The forecast indicates that, From the perspective of source To the target perspective The predictor MLP structure, For the first Each training set sample from the source perspective Low-level features;

[0022] Meanwhile, target perspective By leveraging its own low-level features within complete historical time periods, an LSTM network is used to generate autoregressive predictions for missing time periods.

[0023] Subsequently, from the perspective of the source and target perspective Based on its own historical data, each target perspective is dynamically assigned a sample-level confidence score using its own ScoreNet MLP:

[0024] ;

[0025] in, For the first From the perspective of the source of each training set sample and target perspective From its own historical data to the target perspective Credibility score, For the first A ScoreNet MLP is implemented for each target perspective; masked-softmax is used to normalize all confidence scores to obtain sample-level weights for cross-prediction and autoregressive prediction. :

[0026] Finally, the target perspective is obtained by weighted summation. Final predicted value:

[0027] ;

[0028] in, To obtain the target perspective through weighted summation The final predicted value, For the first In each training set sample, the target perspective Autoregressive candidate predictions generated based on their own historical visible time period The corresponding sample-level weights; the final completed low-level features Defined as:

[0029] ;

[0030] in, For the first The first target perspective A view-level missing mask for each training set sample, where 1 represents a complete view and 0 represents a missing view. For the first Each training set sample from the source perspective Low-level features.

[0031] Furthermore, the reconstruction loss and prediction loss are calculated, specifically including:

[0032] Completed low-level features Input the corresponding BiLSTM decoder to obtain the reconstructed sequence ; Reconstruction losses Defined as:

[0033] ;

[0034] in, Indicates the first An independent perspective Spatial features obtained from multivariate time series data corresponding to each sample, or from ground-based cloud image data extracted by CNN. The total number of viewpoints. The total number of samples in the training set. For the first A decoder with an independent perspective. For decoder parameters, For the first The first target perspective Viewpoint-level missing mask for each training set sample; cross-prediction loss is calculated using the following formula:

[0035] ;

[0036] in, For cross-prediction loss, This indicates element-wise multiplication. This is a time-point missing mask, where 1 indicates the missing number. The training sample of the th training sample A single, independent point in time, visible only at the point in time. The above calculation error.

[0037] Furthermore, the decoded and completed data is compared with the corresponding original data in the training set, specifically including:

[0038] Low-level features after completion The high-level features are obtained by stacking the feature MLPs of the learnable weight matrix network used to transform low-level features into instance-level high-level representations, and the clustering probabilities are obtained through the label MLP; cosine similarity and InfoNCE form are used to implement feature contrast loss in the high-level feature space.

[0039] Total Feature Contrast Loss for:

[0040] ;

[0041] in, The total number of viewpoints. From the perspective Perspective Feature contrast loss between;

[0042] Implement label contrastive loss and regularization in the semantic label space:

[0043] ;

[0044] in, For the first The first perspective and the first Label contrast loss between different perspectives For the first All samples in the first perspective belong to the first... A vector composed of the soft clustering probabilities of each cluster. The view index variable takes the value of or , For the number of clusters, For another sample cluster index, For the current cluster index, From the perspective The Middle The average soft assignment probability vector of each cluster From the perspective The Middle The average soft assignment probability vector of each cluster For temperature parameters;

[0045] The total label contrast loss is expressed as: ;

[0046] in, For total label contrast loss, , For the first Perspective 1 Cluster average probability.

[0047] Furthermore, the clustering module performs multi-view clustering learning based on the completed low-level features to obtain clustering results; specifically including:

[0048] A fine-tuning strategy is implemented to run K-means on the high-level features of each viewpoint to obtain pseudo-labels; the label MLP outputs semantic labels, and the pseudo-labels are aligned with the semantic labels predicted by the network using the Hungarian algorithm before fine-tuning.

[0049] Calculate the cross-entropy fine-tuning loss :

[0050] ;

[0051] in, The total number of viewpoints. The total number of samples in the training set. For the first The first perspective The training set sample belongs to the th training set sample. The soft assignment probability of each cluster, This represents the pseudo-label indicator variable after matching; if the first... From the perspective of the first The matched pseudo-label of the sample belongs to the first... Each cluster, then ;otherwise, , To avoid numerical instability; the final cluster label is the average hard decision of the soft assignment probabilities of all views:

[0052] ;

[0053] in, For the first The final hard clustering label for each sample, For the first The first perspective The sample belongs to the first The soft assignment probability of each cluster.

[0054] Based on the same inventive concept, the present invention provides an adaptive weighted missing-complete multi-view clustering ultra-short-term photovoltaic irradiance prediction system, comprising:

[0055] The data processing unit is used to collect multimodal historical datasets of the target photovoltaic power station, including multivariate time series data and corresponding ground cloud image data; preprocess the data to construct a dataset; and divide the dataset into training set, validation set and test set according to a predetermined ratio.

[0056] A prediction framework building unit is used to construct an irradiance prediction framework based on multi-view clustering of missing data. The prediction framework includes a pre-training module, a clustering module, and a regression prediction module. The pre-training module includes a CNN module, a BiLSTM encoder, a bidirectional prediction mechanism, a ScoreNet MLP, and a BiLSTM decoder. Multivariate time-series data without missing values ​​from the training set are input into the corresponding BiLSTM encoder. Simultaneously, ground-based cloud image data without missing values ​​are first processed by the CNN module to extract spatial features before being input into the BiLSTM encoder. The BiLSTM encoder outputs low-level features, which are then sequentially passed through the bidirectional prediction mechanism and ScoreNet. The MLP performs adaptive weighted imputation on missing data to obtain imputed low-level features. These imputed low-level features are then input into a BiLSTM decoder to obtain imputed data. The decoded imputed data is compared with the corresponding original data in the training set to calculate reconstruction loss and prediction loss. The clustering module performs multi-view clustering learning based on the imputed low-level features to obtain clustering results, and divides the imputed low-level features into K classes according to the clustering results. The regression prediction module receives the imputed data and its corresponding clustering results, and trains regression prediction sub-models 1 to K for classes 1 to K respectively. During prediction, the soft-assignment probabilities of the input sample in each view are averaged, and the class with the highest average probability is determined as the class to which the input sample belongs. Then, the corresponding regression prediction sub-model is called to output the irradiance prediction result.

[0057] A joint training unit is used to set training parameters, jointly train the prediction framework, evaluate the performance of the training results, and save the optimal prediction framework parameters.

[0058] The prediction unit is used to feed the historical input data of the target photovoltaic power station into the trained prediction framework, call the corresponding regression prediction sub-model according to the cluster category of the input sample, and output the prediction result of future photovoltaic power generation irradiance.

[0059] Based on the same inventive concept, the present invention also provides a non-volatile storage medium for storing a computer program, wherein the computer program implements the method described when executed by a processor.

[0060] Based on the same inventive concept, the present invention also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the method described.

[0061] Beneficial effects: Compared with the prior art, the significant technical effects of the present invention are as follows:

[0062] (1) This invention integrates missing data recovery and multi-view clustering tasks into a unified end-to-end framework. By jointly optimizing the missing value recovery objective, the representation learning objective, the clustering consistency objective, and the clustering fine-tuning objective, it achieves mutual promotion between representation learning and clustering performance, effectively solving the problem of the disconnect between imputation and clustering in existing methods. Among them, the missing value recovery objective achieves ScoreNet adaptive weighted imputation through reconstruction loss and cross-prediction loss. The representation learning objective uses instance-level contrastive learning in the high-level feature space to mine common semantics across perspectives. The clustering consistency objective implements label contrastive loss and regularization term to constrain the probability distribution of clustering from different perspectives in the semantic label space. The clustering fine-tuning objective feeds back high-level clustering information to semantic labels through K-means pseudo-label Hungarian algorithm matching and cross-entropy fine-tuning loss. It achieves mutual promotion between representation learning and clustering performance, effectively solving the problem of the disconnect between imputation and clustering in existing methods.

[0063] (2) The present invention designs a multi-layer feature learning framework, which corresponds to the multi-layer feature learning part in step S4 of the specific implementation. Missing data recovery and structure preservation are completed in the low-level feature space, common semantic mining is completed in the high-level feature space, and clustering consistency constraints are completed in the semantic feature space. Each objective is optimized independently in different feature spaces, avoiding objective conflicts, while preserving the cross-dimensional correlation of multivariate time series to the greatest extent.

[0064] (3) The ScoreNet MLP and masked-softmax weighted fusion mechanism proposed in this invention can dynamically evaluate the credibility according to the actual situation of the sample, realize the adaptive completion of missing values, significantly reduce the uncertainty introduced by traditional data augmentation or simple imputation, and improve the quality and stability of representation learning.

[0065] (4) The overall method of the present invention does not rely on data augmentation technology. It directly uses the complementary information of the original multivariate time series and the ground cloud map to complete the comparative learning, which reduces the computational complexity and avoids the significant subjectivity and uncertainty that data augmentation technology may produce, thus providing more reliable technical support for the grid dispatch and safe operation of photovoltaic power plants.

[0066] (5) After clustering, the present invention assigns a dedicated regression prediction sub-model to each category according to the meteorological model category to which the sample belongs, so that the prediction framework can perform refined modeling for the non-stationary characteristics of different weather patterns, thereby significantly improving the accuracy and robustness of photovoltaic irradiance prediction under incomplete data conditions. Attached Figure Description

[0067] Figure 1 This is a flowchart of the method described in this invention;

[0068] Figure 2 This is a structural diagram of the irradiance prediction framework based on multi-view clustering of missing data in an embodiment of the present invention;

[0069] Figure 3 This is a flowchart of the irradiance prediction framework training method based on multi-view clustering of missing data in an embodiment of the present invention. Detailed Implementation

[0070] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. However, the described embodiments are only some embodiments of the present invention, and the protection scope of the present invention is not limited to the described embodiments.

[0071] This invention addresses the problem of incomplete multimodal data caused by missing multivariate time series data and ground-based cloud maps, leading to low accuracy in photovoltaic (PV) irradiance prediction and insufficient model adaptability to complex meteorological models. It proposes an end-to-end joint optimization method that combines adaptive weighted imputation of missing values ​​with multi-view clustering. This method constructs an irradiance prediction framework based on multi-view clustering of missing data, achieving end-to-end joint optimization of adaptive weighted imputation of missing values ​​and multi-view clustering. Based on the clustering results, an independent regression prediction sub-model is trained for each sample class. Finally, historical input data from the target PV power plant is fed into the trained prediction framework, and the corresponding regression prediction sub-model is invoked according to the cluster category of the sample, outputting the PV irradiance prediction results for the next 0-4 hours with a time resolution of 10 minutes. The proposed multi-view clustering prediction framework treats each variable column and the ground-based cloud map as independent views. It employs an encoder combining a Bidirectional Long Short-Term Memory (BiLSTM) network and a Convolutional Neural Network (CNN) to extract low-level features. A ScoreNet Multi-Layer Perceptron (MLP) tailored to each target viewpoint utilizes historical information from both the source and target views to achieve sample-level adaptive weighted fusion and completion of the target viewpoint. Multi-level information comparison learning and cluster consistency optimization are then performed on the completed low-level features, and fine-tuning is achieved using K-means clustering combined with the Hungarian algorithm. Finally, dedicated regression prediction sub-models are trained for different categories of samples. Compared with existing technologies, this invention integrates missing data completion and multi-perspective clustering tasks into a unified end-to-end framework. Through the ScoreNet adaptive weighted fusion mechanism and multi-level feature learning, it avoids the problems of missing data completion and clustering being disconnected and multi-objective conflicts. Under the condition of incomplete data, it significantly improves the accuracy and robustness of ultra-short-term photovoltaic power generation irradiance prediction, and provides reliable technical support for the grid dispatch and safe operation of photovoltaic power plants.

[0072] like Figure 1 As shown, the method of the present invention includes the following steps:

[0073] S1: Collect multimodal historical data of the target photovoltaic power station, including multivariate time series data and its corresponding ground-based cloud image data at specific time points. The multivariate time series data includes nine variables: azimuth, cloud cover, horizontal diffuse irradiance, direct normal irradiance, global horizontal irradiance, relative humidity, wind direction, wind speed, and zenith angle. In this embodiment, the time span for the multimodal historical data is three years. In this embodiment, the original acquisition time interval for the multivariate time series data and its corresponding ground-based cloud image data is 10 minutes. This time interval is only used to illustrate the actual acquisition situation of the dataset in this embodiment and does not constitute a limitation on the applicable time resolution of the method of this invention.

[0074] S2: Preprocess the multimodal historical data collected in S1 to create a dataset. Specifically, this includes:

[0075] First, the multivariate time series data and their corresponding ground cloud image data from the collected multimodal historical data are converted into a single HDF5 file format and then integrated and stored. Specifically, the ground cloud image data corresponding to the multivariate time series data is assumed to have no outliers and is not subjected to outlier detection, removal, or filling. For the multivariate time series data, a time series curve was plotted, and outlier thresholds were set based on the physical value ranges of variables such as azimuth, wind direction, zenith angle, cloud cover, and relative humidity: the physical value ranges for azimuth and wind direction were set to 0° to 360°, for zenith angle to 0° to 180°, and for cloud cover and relative humidity to 0° to 100%. The value ranges for wind speed, horizontal diffuse irradiance, direct normal irradiance, and global horizontal irradiance were all set to be no less than 0 and less than the upper limit threshold. The upper limit threshold was set as follows: for cases where the values ​​of wind speed, horizontal diffuse irradiance, direct normal irradiance, and global horizontal irradiance were abnormally large, the upper limit threshold was set based on the historical distribution of the corresponding variable in the training set, preferably using the 99th percentile. Values ​​exceeding the corresponding threshold ranges for the nine variables were identified as outliers and removed. Missing positions resulting from outlier removal were filled using linear interpolation. Setting the threshold too wide can lead to the retention of anomalous noise, affecting the stability and accuracy of the prediction framework training; setting the threshold too narrow may inadvertently remove real meteorological fluctuations, weakening the non-stationary characteristics of the time series. Therefore, this embodiment identifies outliers by using the aforementioned physical value range threshold and historical distribution upper limit threshold to achieve a balance between outlier removal and the preservation of real meteorological characteristics.

[0076] S3: Split the dataset. Specifically, this includes:

[0077] The dataset created in step S2 is divided into training set, validation set and test set according to a predetermined ratio. In this example, it is divided into training set, validation set and test set in a ratio of 6:2:2.

[0078] S4: Construct an irradiance prediction framework based on multi-view clustering of missing data; such as Figure 2 As shown, the prediction framework includes a pre-training module, a clustering module, and a regression prediction module, which are sequentially connected and work together to complete missing data completion, clustering, and irradiance prediction. The pre-training module uses multivariate time series data from the training set. The training set uses ground-based cloud image data as input. The multivariate time-series data in the training set contains missing time-series data, and the ground-based cloud image data contains missing image data. The multivariate time-series data without missing values ​​is input into the corresponding BiLSTM encoder, while the ground-based cloud image data without missing values ​​is first processed by a CNN to extract spatial features. Then the input is fed into a BiLSTM encoder, which outputs low-level features. and The low-level features are then adaptively weighted and imputed using a bidirectional prediction mechanism and ScoreNet MLP to complete the missing data, resulting in the imputed low-level features. and The padded low-level features are input into the BiLSTM decoder to obtain the padded data. and (i.e., the completed multivariate time series data and ground-based cloud image data); the decoded data are compared with the corresponding original data in the training set to calculate reconstruction loss and prediction loss; the reconstruction loss is calculated only for samples without missing data, and the prediction loss is calculated only for visible time points. The completed data is output to the regression prediction module, and the completed low-level features are output to the clustering module. The clustering module, based on the completed low-level features, performs multi-view clustering learning using feature MLP, label MLP, feature contrast loss, label contrast loss, and fine-tuning loss to obtain clustering results, and accordingly divides the completed low-level features into K classes. The regression prediction module receives the completed data and its corresponding clustering results, and trains regression prediction sub-models 1 to K for each class. During prediction, the soft-assignment probabilities of the input sample across each view are averaged, and the class with the highest average probability is determined as the class of the input sample. Then, the corresponding regression prediction sub-model is called to output the irradiance prediction result.

[0079] Specifically, it includes the following steps:

[0080] S41. In the irradiance prediction task, the multivariate time series data is divided into 9 perspectives, and the ground-based cloud image data is divided into 1 perspective, for a total of 10 perspectives. Each time series variable and the ground-based cloud image data are considered as independent perspectives. For each training set sample after splitting the dataset, only the multivariate time series data without missing values ​​and its corresponding ground-based cloud image data in the training set sample are input into the corresponding BiLSTM encoder and CNN for feature extraction. For the perspective of multivariate time series data, the BiLSTM encoder is used; for the ground-based cloud image perspective, spatial features are first extracted from each image using CNN, and then input into the BiLSTM encoder to obtain temporal representation, thus obtaining the low-level features of each independent perspective. For samples containing missing values, the prediction loss is calculated only for the non-missing parts. Specifically, the prediction loss for the multivariate time series perspective is calculated based on the non-missing time series data, and the prediction loss for the ground-based cloud image perspective is calculated based on the spatial features extracted by CNN from the non-missing images. The low-level feature generation formula is:

[0081] ;

[0082] in, Indicates the first An independent perspective Spatial features obtained from multivariate time series data corresponding to each training set sample, or from ground-based cloud image data extracted by CNN. For the first A BiLSTM encoder with an independent viewpoint. For BiLSTM encoder parameters, These are low-level features. This selective mechanism ensures that only reliable information is used, avoiding the pollution of the encoding process by missing data.

[0083] S42. In this step, the current viewpoint to be completed is recorded as the target viewpoint. Other perspectives used to generate cross-predictions for the target perspective are denoted as source perspectives. ,in For missing perspectives, this prediction framework considers two sources of prediction simultaneously: first, the cross-prediction of source perspectives within the same missing time period; and second, the autoregressive prediction of the target perspective itself over the missing time period within a complete historical time period. First, each source perspective... Generate cross-prediction:

[0084] ;

[0085] in, From the perspective of source To the target perspective The predictor MLP structure, For the first Each training set sample from the source perspective Low-level features, Indicates the perspective from the source Generate target perspective through cross-prediction The prediction indicates that...

[0086] Meanwhile, target perspective Leveraging its low-level features within complete historical timeframes, an LSTM network is used to generate autoregressive predictions for missing timeframes, denoted as... This autoregressive prediction takes low-level features from historical time periods as input and outputs predicted values ​​for missing time periods. Subsequently, from the source perspective... and target perspective Its own historical data is dynamically assigned sample-level confidence scores through the ScoreNet MLP, which is designed specifically for each target perspective:

[0087] ;

[0088] in, For the first From the perspective of the source of each training set sample and target perspective From its own historical data to the target perspective Credibility score, For the first A ScoreNet MLP is specifically designed for each target perspective. Masked-softmax is used to normalize all confidence scores, yielding sample-level weights for cross-prediction and autoregressive prediction. :

[0089] ;

[0090] in, For the first From the perspective of the source of each training set sample To the target perspective The credibility score; For the first From the perspective of the source of each training set sample To the target perspective The credibility score; To exclude the target perspective From a perspective other than that, it is used for normalizing the sum of the denominators; As an indicator function, when the source perspective is on the training set samples The value is 1 if there are no missing views, and 0 otherwise, used to mask missing views; For the first Source perspective of each training set sample The missing indicator variable, For the first Source perspective of each training set sample The missing indicator variable, is the numerical stability constant.

[0091] This sample-level weight By exponentially scaling the ScoreNet confidence score and then normalizing it using softmax, dynamic weighting is achieved for two types of prediction results: first, the cross-prediction of the source perspective within the same missing time period; and second, the autoregressive prediction of the target perspective itself within the historical complete time period for the missing time period. Finally, a weighted summation is used to obtain the prediction for the target perspective. Final predicted value:

[0092] ;

[0093] in, To obtain the target perspective through weighted summation The final predicted value, For the first In each training set sample, the target perspective Autoregressive candidate predictions generated based on their own historical visible time period The corresponding sample-level weights. The final completed low-level features. Defined as:

[0094] ;

[0095] in, For the first The first target perspective A viewpoint-level missing mask for each training set sample, where 1 represents a complete viewpoint and 0 represents a missing viewpoint.

[0096] Calculate the reconstruction loss and cross-prediction loss, and then apply the completed low-level features. Input the corresponding BiLSTM decoder to obtain the reconstructed sequence Reconstruction losses Defined as:

[0097] ;

[0098] in, Indicates the first An independent perspective Spatial features obtained from multivariate time series data corresponding to each sample, or from ground-based cloud image data extracted by CNN. The total number of viewpoints. The total number of samples in the training set. For the first A decoder with an independent perspective. These are the decoder parameters. The reconstruction loss is only applied to complete viewpoint samples. Calculations are performed to match the characteristics of incomplete data. Simultaneously, cross-prediction loss is calculated to enhance cross-perspective consistency.

[0099] ;

[0100] in, For cross-prediction loss, This indicates element-wise multiplication. This is a time-point missing mask, where 1 indicates the missing number. The training sample of the th training sample A single, independent point in time, visible only at the point in time. The above calculation error.

[0101] To ensure that the completed data serves both the clustering task and provides reliable input for subsequent downstream regression predictions that train dedicated regression prediction sub-models for each weather category based on the clustering results and generate the final photovoltaic power generation prediction, this invention first conducts a separate pre-training phase. In this phase, only the reconstruction loss and cross-prediction loss are optimized, and adaptive completion of missing values ​​is learned end-to-end, resulting in high-quality low-level features and reconstructed sequences with completeness from all perspectives after adaptive weighted fusion completion. This pre-training phase, which optimizes the missing value completion module separately before overall end-to-end joint training, avoids the distortion problem caused by the completed values ​​excessively serving clustering during joint training. Simultaneously, the completed low-level features obtained through pre-training provide a good initial state for the subsequent clustering phase, effectively improving the stability of the clustering phase. The completed dataset is complete multivariate time series data obtained through ScoreNet adaptive weighted fusion. After entering the clustering phase, the encoder and decoder of the reconstruction module and the imputation-related parameters of the prediction module can continue to participate in joint optimization, thus ensuring that the entire framework remains a strictly end-to-end learning process.

[0102] S43. After missing data completion, all perspectives participate in comparative learning, and the distinction between the target perspective and the source perspective is no longer made; in this step, and This represents two different perspectives participating in comparative learning, among which In the completed low-level features The high-level features are obtained by stacking feature MLPs, which are dedicated to transforming low-level features into instance-level high-level representations, on top of each other. Clustering probabilities are obtained through label MLP. , This is the normalization function. and These are the learnable weight matrices for the feature MLP and the label MLP, respectively. The elements in the matrices represent the connection weights between the completed low-level features and high-level features, and the clustering probabilities. and At the start of model training, non-zero random values ​​are used for initialization, with the corresponding bias term initialized to 0. These bias terms are then iteratively updated using backpropagation and the Adam optimization algorithm based on the loss function. In the high-level feature space, cosine similarity and InfoNCE-based feature contrast loss are implemented. Represents the dot product of vectors. Representing the Euclidean norm:

[0103] ;

[0104] in, Cosine similarity is used to measure the similarity between high-level feature vectors from different perspectives of the same sample. This is a high-level instance-level representation obtained by further mapping from the completed low-level features using feature MLP;

[0105] ;

[0106] in, From the perspective Perspective Feature contrast loss between For another sample index, For perspective indexing, The cosine similarity function is used. For training set samples From the perspective High-level feature vectors, For training set samples From the perspective High-level feature vectors, This refers to the temperature parameter.

[0107] This feature contrastive loss aims to maximize the high-level feature similarity of the same sample from different perspectives, while minimizing the similarity between different samples. Total Feature Contrast Loss for:

[0108] ;

[0109] Implement label contrastive loss and regularization in the semantic label space:

[0110] ;

[0111] in, For the first The first perspective and the first Label contrast loss between different perspectives For the first All samples in the first perspective belong to the first... A vector composed of the soft clustering probabilities of each cluster. The view index variable takes the value of or , For the number of clusters, For another sample cluster index, For the current cluster index, From the perspective The Middle The average soft assignment probability vector of each cluster From the perspective The Middle The average soft assignment probability vector of each cluster This refers to the temperature parameter.

[0112] Label contrast loss ensures that the clustering probability distribution of the same sample remains consistent across different perspectives.

[0113] ;

[0114] in, For total label contrast loss, , For the first Perspective 1 Cluster average probability. This multi-layer feature comparison module performs instance comparison learning in the high-level feature space and label comparison learning in the semantic label space, optimizing independently in different feature spaces to avoid target conflicts, while simultaneously mining cross-perspective common semantics and cluster consistency.

[0115] S44. In this step... This indicates any participating perspective in the clustering, without the meaning of a target perspective or source perspective. A fine-tuning strategy is implemented, running K-means on the high-level features of each perspective to obtain pseudo-labels. These pseudo-labels are then aligned with the network's predicted labels using the Hungarian algorithm and fine-tuned. The reconstructed sequence output from the decoder is used as the final imputation data. Based on the clustering results, a dedicated regression prediction sub-model is trained for each category to achieve ultra-short-term photovoltaic irradiance prediction. The final clustering label is obtained by averaging the soft-assignment probabilities of all perspectives using a hard decision. Specifically, this includes:

[0116] ;

[0117] in, For the first The first perspective The center of each cluster, For the first The first perspective Pseudo-labels for each sample. Semantic labels are output by the Label MLP:

[0118] ;

[0119] in, For the first The first perspective Semantic labels of training set samples, For the first The first perspective The training set sample belongs to the th training set sample. The probability of each cluster. The Hungarian algorithm is used to solve for the optimal matching matrix. To achieve alignment between K-means pseudo-labels and label MLP predicted labels:

[0120] ;

[0121] ;

[0122] The optimization problem aims to find the optimal matching matrix. This is achieved by minimizing the matching cost between the K-means pseudo-labels and the predicted cluster labels, thus realizing the optimal alignment between them. The first number predicted by the labeled MLP Cluster and the first The matching cost between K-means pseudo-label clusters. When If the value is true, it indicates that the two clusters match; otherwise, it indicates that they do not match. This constraint ensures a one-to-one correspondence between the labeled MLP predicted clusters and the K-means pseudo-label clusters. The cross-entropy fine-tuning loss is then calculated. :

[0123] ;

[0124] in, For the first The first perspective The training set sample belongs to the th training set sample. The soft assignment probability of each cluster, This represents the pseudo-label indicator variable after matching. If the... From the perspective of the first The matched pseudo-label of the sample belongs to the first... Each cluster, then ;otherwise, , To avoid numerical instability, the final cluster labels are the average hard decision of the soft assignment probabilities across all views.

[0125] ;

[0126] in, For the first The final hard clustering label for each sample, For the first The first perspective The sample belongs to the first The soft assignment probability of each cluster.

[0127] The hard decision-making process obtains the final cluster label by averaging the probabilities of soft assignments across all perspectives. This fine-tuning mechanism further enhances the accuracy of semantic features by utilizing high-level clustering information.

[0128] Perform downstream filling and irradiance prediction, and reconstruct the sequence output by the decoder. As the final data imputation: for missing locations, the reconstructed values ​​are used for replacement; for locations without missing data, the original multivariate time series data and their corresponding ground-based cloud image data are retained. This yields a complete multivariate time series.

[0129] Based on the sample categories obtained from clustering results, a dedicated regression prediction sub-model is assigned to each category in the regression prediction stage of the prediction framework. This dedicated regression prediction sub-model, as a downstream regression prediction module, together with the clustering stage, constitutes a two-stage refined prediction strategy. The clustering framework performs adaptive completion of missing data and multi-view clustering to classify samples according to weather patterns, while the dedicated regression prediction sub-model is trained independently for each weather pattern category, thereby achieving refined modeling of irradiance variation patterns under different weather patterns. Each regression prediction sub-model is trained by minimizing the MAE loss function.

[0130] ;

[0131] in, Mean absolute error (MAE) is used to characterize the average absolute error between the predicted and actual values. The smaller the value, the higher the prediction accuracy of the regression prediction sub-model. Indicates the first The actual irradiance value at the predicted time. Indicates time The corresponding category regression prediction sub-model predicted value, This represents the total number of samples in the training set. During the validation and testing phases, for each sample, the corresponding regression prediction sub-model is invoked based on its cluster category to output the predicted photovoltaic irradiance for the next 0 to 4 hours.

[0132] S5: Set training parameters, perform joint training on the prediction framework, evaluate the performance of the training results, and save the optimal prediction framework parameters.

[0133] likeFigure 3 As shown, the training process of the prediction framework in this embodiment of the invention includes: first, training the pre-training module to enable it to learn the adaptive completion capability of missing data; then, inputting the completed low-level features output by the pre-training module into the clustering module for comparative learning clustering to learn common semantic features between different perspectives and form a preliminary clustering structure; then, fine-tuning the clustering results to further improve the stability of the clustering results; after fine-tuning, outputting the final clustering results and dividing the sample categories according to the clustering results; then, training the corresponding regression prediction sub-models for different categories; finally, saving the trained prediction framework, and the training process ends.

[0134] Specifically, it includes the following steps:

[0135] Before training, the general parameters, contrastive learning parameters, and clustering parameters of the prediction framework are set. Specifically, in this embodiment, the Adam optimization algorithm is used to update the prediction framework parameters. The initial learning rate is set to 3×10^-4, the batch size is set to 256, and the weight decay coefficient is set to 0. The low-level feature dimension is set to 64, the high-level feature dimension is set to 128, the number of clusters is set to 3, and a fixed random seed is used to ensure the reproducibility of training results. The temperature coefficient of the feature contrastive loss is set to 0.5, and the temperature coefficient of the label contrastive loss is set to 1, where the temperature coefficient is used to adjust the smoothness of the sample similarity distribution in contrastive learning. The weights of the feature contrastive loss and the label contrastive loss are both set to 0.5, the weights of the reconstruction loss and the cross-prediction loss are both set to 0.1, and the weight of the semantic label space regularization term is set to 0.5. The number of repeated initializations for K-means clustering is set to 100 to reduce the impact of random selection of initial cluster centers on the clustering results.

[0136] The prediction framework adopts a phased training strategy, which includes pre-training module training, clustering module training, and regression prediction sub-model training in sequence; among them, clustering module training includes two sequentially connected steps: contrastive learning training and iterative fine-tuning.

[0137] First, the pre-training module is trained for a maximum of 200 epochs, optimizing only the reconstruction loss and cross-prediction loss. In each epoch, the validation loss is monitored on the validation set: if the learning rate is not lower than the historical minimum for 10 consecutive epochs, the learning rate is multiplied by 0.5; if there is no improvement for 20 consecutive epochs, the learning rate is stopped early and the historical best weights are loaded to obtain high-quality missing completion initialization.

[0138] Subsequently, the contrastive learning training in the clustering module is carried out for a fixed 100 rounds, with the total feature contrastive loss and the total label contrastive loss as the main components, while retaining the reconstruction and cross-prediction losses with smaller weights, to achieve the learning of cross-perspective common semantics and clustering consistency.

[0139] Then, iterative fine-tuning is performed in the clustering module, running for a fixed 100 rounds. Every 10 rounds, K-means is run on high-level features to generate pseudo-labels. The K-means pseudo-labels are then globally aligned with the predicted cluster labels output by the label MLP using the Hungarian algorithm. Fine-tuning is then performed primarily using cross-entropy, supplemented by a small amount of reconstruction, prediction, and contrastive losses, ultimately improving clustering accuracy. The above training process uses the same missing mask mechanism to ensure that the completion process of the pre-training module, the contrastive learning process of the clustering module, and the iterative fine-tuning process are logically consistent. This allows the pre-training module to first acquire reliable completion capabilities, and then the clustering module learns high-level representations and optimizes the clustering results, thus significantly improving training stability. This phased, progressively stacked loss design avoids conflicts between different objectives, ensuring that the completion module remains stably optimized throughout the training process. It also avoids early noise interference with high-level features and prevents oscillating convergence caused by poor pseudo-label quality.

[0140] Subsequently, in the regression prediction phase, MAE is used as the loss function for backpropagation and gradient descent optimization. Based on this, an adaptive learning rate scheduling and early stopping mechanism based on validation loss is introduced. Specifically, the historical minimum validation loss is used as a benchmark: if the validation loss does not fall below the historical minimum for 10 consecutive epochs, the current learning rate is multiplied by 0.5; if the validation loss still does not fall below the historical minimum for 20 consecutive epochs, the early stopping mechanism is triggered, immediately stopping training and saving the prediction framework parameters at the point of minimum validation loss. This mechanism effectively prevents the prediction framework from overfitting due to excessive training epochs, while ensuring that the prediction framework converges to the optimal generalization state.

[0141] Input the multimodal historical data of the target power plant and begin iterative training.

[0142] Visualize the training set loss, validation set loss, and test set loss after each training session. Print out the loss value of the regression prediction sub-model after each training session to facilitate the assessment of its convergence and aid in parameter tuning.

[0143] The regression prediction sub-model is evaluated using error metrics such as root mean square error (RMSE), mean absolute error (MAE), normalized root mean square error (nRMSE), and normalized mean absolute error (nMAE).

[0144] The formulas for calculating each error index are as follows:

[0145] ;

[0146] ;

[0147] ;

[0148] ;

[0149] ;

[0150] in, Indicates the first The actual irradiance value corresponding to each predicted time. Indicates the first The predicted value of the categorical regression prediction sub-model at each prediction time. This represents the total number of predicted and actual value pairs involved in error calculation during the evaluation phase. This represents the mean of all true irradiance values ​​used in the evaluation. RMSE measures the overall squared deviation of the prediction error and is more sensitive to larger prediction errors. MAE measures the mean absolute deviation between the predicted and observed values, reflecting the average prediction error level of the regression prediction sub-model. nRMSE and nMAE are the normalized forms of RMSE and MAE, respectively, used to eliminate the influence of different data scales on the evaluation results and facilitate horizontal comparisons between different models. The smaller the values ​​of the above error indicators, the closer the prediction results of the regression prediction sub-model are to the actual observed values, and the better the prediction performance.

[0151] After training is complete, the parameters of the prediction framework and multiple error metrics are saved, and the prediction framework with the lowest overall error metrics is selected for prediction.

[0152] Save the trained prediction framework in .path file format for easy subsequent prediction.

[0153] S6: Input the historical input data of the target photovoltaic power station into the trained prediction framework, call the corresponding regression prediction sub-model according to the cluster category of the input sample, and output the photovoltaic power generation irradiance prediction results for the next 0 to 4 hours with a time resolution of 10 minutes.

[0154] Specifically, it includes the following steps:

[0155] Use torch library functions to load the .path file, and then use the prediction framework after loading to make predictions.

[0156] The prediction requires inputting historical time series data for a day corresponding to the frequency. The prediction framework will eventually output the prediction results of the target power plant's photovoltaic power generation irradiance for the next 0 to 4 hours with a time resolution of 10 minutes.

[0157] The performance of the prediction results was evaluated using a test set. The prediction range was from 0 to 4 hours in the future, with a total of 24 prediction steps. The evaluation metrics included mean absolute error (MAE), root mean square error (RMSE), normalized mean absolute error (nMAE), and normalized root mean square error (nRMSE). The evaluation results are shown in Table 1.

[0158] Table 1 Performance Indicators of Prediction Results

[0159]

[0160] As shown in Table 1, in the 24-step irradiance prediction of the next 0 to 4 hours, the MAE, RMSE, nMAE and nRMSE of the prediction framework are 89.0098, 150.4408, 0.2331 and 0.3940, respectively, indicating that the prediction framework can achieve ultra-short-term photovoltaic irradiance prediction.

[0161] The method described in this invention addresses the problems of insufficient photovoltaic irradiance prediction accuracy and weak model adaptability to complex meteorological models caused by incomplete multimodal data, particularly when multivariate time series data and ground-based cloud image data are missing. This invention utilizes a multi-view clustering method based on contrastive learning to accurately classify sample meteorological models even with missing data, and accordingly configures a dedicated regression prediction sub-model for each sample class, thereby significantly improving prediction accuracy and robustness.

[0162] Based on the same inventive concept, the present invention provides an adaptive weighted missing-complete multi-view clustering ultra-short-term photovoltaic irradiance prediction system, comprising:

[0163] The data processing unit is used to collect multimodal historical datasets of the target photovoltaic power station, including multivariate time series data and corresponding ground cloud image data; preprocess the data to construct a dataset; and divide the dataset into training set, validation set and test set according to a predetermined ratio.

[0164] A prediction framework building unit is used to construct an irradiance prediction framework based on multi-view clustering of missing data. The prediction framework includes a pre-training module, a clustering module, and a regression prediction module. The pre-training module includes a CNN module, a BiLSTM encoder, a bidirectional prediction mechanism, a ScoreNet MLP, and a BiLSTM decoder. Multivariate time-series data without missing values ​​from the training set are input into the corresponding BiLSTM encoder. Simultaneously, ground-based cloud image data without missing values ​​are first processed by the CNN module to extract spatial features before being input into the BiLSTM encoder. The BiLSTM encoder outputs low-level features, which are then sequentially passed through the bidirectional prediction mechanism and ScoreNet. The MLP performs adaptive weighted imputation on missing data to obtain imputed low-level features. These imputed low-level features are then input into a BiLSTM decoder to obtain imputed data. The decoded imputed data is compared with the corresponding original data in the training set to calculate reconstruction loss and prediction loss. The clustering module performs multi-view clustering learning based on the imputed low-level features to obtain clustering results, and divides the imputed low-level features into K classes according to the clustering results. The regression prediction module receives the imputed data and its corresponding clustering results, and trains regression prediction sub-models 1 to K for classes 1 to K respectively. During prediction, the soft-assignment probabilities of the input sample in each view are averaged, and the class with the highest average probability is determined as the class to which the input sample belongs. Then, the corresponding regression prediction sub-model is called to output the irradiance prediction result.

[0165] A joint training unit is used to set training parameters, jointly train the prediction framework, evaluate the performance of the training results, and save the optimal prediction framework parameters.

[0166] The prediction unit is used to feed the historical input data of the target photovoltaic power station into the trained prediction framework, call the corresponding regression prediction sub-model according to the cluster category of the input sample, and output the prediction result of future photovoltaic power generation irradiance.

[0167] Based on the same inventive concept, the present invention also provides a non-volatile storage medium for storing a computer program, wherein the computer program implements the method described when executed by a processor.

[0168] Based on the same inventive concept, the present invention also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the method described.

[0169] In summary, the method described in this invention addresses the problems of incomplete multimodal data due to missing multivariate time series data and ground-based cloud image data, as well as insufficient accuracy and robustness in ultra-short-term photovoltaic irradiance prediction under incomplete multimodal data conditions. It constructs a prediction method for multi-view information fusion of multivariate time series data and ground-based cloud image data, treating them as multiple independent perspectives for unified modeling. Through selective coding and ScoreNet adaptive weighted fusion, it achieves end-to-end joint optimization of sample-level adaptive weighted completion and clustering under missing data conditions. Furthermore, it combines multi-view clustering to identify different meteorological patterns and establishes dedicated regression prediction sub-models for different categories of samples based on the clustering results, thereby significantly improving the accuracy, robustness, and stability of photovoltaic irradiance prediction under complex weather conditions.

Claims

1. A multi-view clustering method for predicting ultra-short-term photovoltaic irradiance using adaptive weighted missing completion, characterized in that, Includes the following steps: Collect multimodal historical datasets of the target photovoltaic power station, including multivariate time series data and corresponding ground cloud image data; preprocess the data to construct a dataset; and divide the dataset into training set, validation set and test set according to a predetermined ratio. An irradiance prediction framework based on multi-view clustering of missing data is constructed. The framework includes a pre-training module, a clustering module, and a regression prediction module. The pre-training module comprises a CNN module, a BiLSTM encoder, a bidirectional prediction mechanism, a ScoreNet MLP, and a BiLSTM decoder. Multivariate time-series data without missing values ​​from the training set are input into the corresponding BiLSTM encoder. Simultaneously, ground-based cloud image data without missing values ​​are first processed by the CNN module to extract spatial features before being input into the BiLSTM encoder. The BiLSTM encoder outputs low-level features, which are then sequentially processed by the bidirectional prediction mechanism and ScoreNet. The MLP performs adaptive weighted imputation on missing data to obtain imputed low-level features. These imputed low-level features are then input into a BiLSTM decoder to obtain imputed data. The decoded imputed data is compared with the corresponding original data in the training set to calculate reconstruction loss and prediction loss. The clustering module performs multi-view clustering learning based on the imputed low-level features to obtain clustering results, and divides the imputed low-level features into K classes according to the clustering results. The regression prediction module receives the imputed data and its corresponding clustering results, and trains regression prediction sub-models 1 to K for classes 1 to K respectively. During prediction, the soft-assignment probabilities of the input sample in each view are averaged, and the class with the highest average probability is determined as the class to which the input sample belongs. Then, the corresponding regression prediction sub-model is called to output the irradiance prediction result. Set training parameters, perform joint training on the prediction framework, evaluate the performance of the training results, and save the optimal prediction framework parameters. The historical input data of the target photovoltaic power station is fed into the trained prediction framework. Based on the cluster category to which the input sample belongs, the corresponding regression prediction sub-model is called to output the prediction result of future photovoltaic power generation irradiance.

2. The method according to claim 1, characterized in that, The multivariate time series data includes nine variables: azimuth, cloud cover, horizontal diffuse irradiance, direct normal irradiance, global horizontal irradiance, relative humidity, wind direction, wind speed, and zenith angle.

3. The method according to claim 2, characterized in that, The data is preprocessed to form a dataset, specifically as follows: Multivariate time series data and their corresponding ground cloud image data are converted into formats and finally integrated and stored into a single HDF5 file format; For the multivariate time series data, a time series curve is plotted, and outlier judgment thresholds are set according to the physical value ranges of each variable, such as azimuth, wind direction, zenith angle, cloud cover, and relative humidity. The value ranges of wind speed, horizontal diffuse irradiance, direct normal irradiance, and global horizontal irradiance are all set to be no less than 0 and less than the upper limit threshold. The upper limit threshold is set according to the historical distribution of the corresponding variable in the training set. Values ​​exceeding the corresponding threshold range among the nine variables are identified as outliers and removed. Missing positions resulting from the removal of outliers are filled using linear interpolation.

4. The method according to claim 1, characterized in that, The low-level features are sequentially augmented using a bidirectional prediction mechanism and ScoreNet MLP to adaptively weight and complete the missing data, resulting in the augmented low-level features; specifically including: Record the current perspective that needs to be completed as the target perspective. Other perspectives used to generate cross-predictions for the target perspective are denoted as source perspectives. ,in ; First, from each source perspective Generate cross-prediction: ； in, Indicates the perspective from the source Target perspective is generated through cross-prediction. The forecast indicates that, From the perspective of source To the target perspective The predictor MLP structure, For the first Each training set sample from the source perspective Low-level features; Meanwhile, target perspective By leveraging its own low-level features within complete historical time periods, an LSTM network is used to generate autoregressive predictions for missing time periods. Subsequently, from the perspective of the source and target perspective Based on its own historical data, each target perspective is dynamically assigned a sample-level confidence score using its own ScoreNet MLP: ； in, For the first From the perspective of the source of each training set sample and target perspective From its own historical data to the target perspective Credibility score, For the first A ScoreNet MLP is implemented for each target perspective; masked-softmax is used to normalize all confidence scores to obtain sample-level weights for cross-prediction and autoregressive prediction. : Finally, the target perspective is obtained by weighted summation. Final predicted value: ； in, To obtain the target perspective through weighted summation The final predicted value, For the first In each training set sample, the target perspective Autoregressive candidate predictions generated based on their own historical visible time period The corresponding sample-level weights; the final completed low-level features Defined as: ； in, For the first The first target perspective A view-level missing mask for each training set sample, where 1 represents a complete view and 0 represents a missing view. For the first Each training set sample from the source perspective Low-level features.

5. The method according to claim 1, characterized in that, The calculation of reconstruction loss and prediction loss includes: Completed low-level features Input the corresponding BiLSTM decoder to obtain the reconstructed sequence ; Reconstruction losses Defined as: ； in, Indicates the first An independent perspective Spatial features obtained from multivariate time series data corresponding to each sample, or from ground-based cloud image data extracted by CNN. The total number of viewpoints. The total number of samples in the training set. For the first A decoder with an independent perspective. For decoder parameters, For the first The first target perspective Viewpoint-level missing mask for each training set sample; cross-prediction loss is calculated using the following formula: ； in, For cross-prediction loss, This indicates element-wise multiplication. This is a time-point missing mask, where 1 indicates the missing number. The training sample of the th training sample A single, independent point in time, visible only at the point in time. The above calculation error.

6. The method according to claim 1, characterized in that, The decoded and completed data is compared with the corresponding original data in the training set, specifically including: Low-level features after completion The high-level features are obtained by stacking the feature MLPs of the learnable weight matrix network used to transform low-level features into instance-level high-level representations, and the clustering probabilities are obtained through the label MLP; cosine similarity and InfoNCE form are used to implement feature contrast loss in the high-level feature space. Total Feature Contrast Loss for: ； in, The total number of viewpoints. From the perspective Perspective Feature contrast loss between; Implement label contrastive loss and regularization in the semantic label space: ； in, For the first The first perspective and the first Label contrast loss between different perspectives For the first All samples in the first perspective belong to the first... A vector composed of the soft clustering probabilities of each cluster. The view index variable takes the value of or , For the number of clusters, For another sample cluster index, For the current cluster index, From the perspective The Middle The average soft assignment probability vector of each cluster From the perspective The Middle The average soft assignment probability vector of each cluster For temperature parameters; The total label contrast loss is expressed as: ; in, For total label contrast loss, , For the first Perspective 1 Cluster average probability.

7. The method according to claim 1, characterized in that, The clustering module performs multi-view clustering learning based on the completed low-level features to obtain clustering results; specifically, it includes: A fine-tuning strategy is implemented to run K-means on the high-level features of each viewpoint to obtain pseudo-labels; the label MLP outputs semantic labels, and the pseudo-labels are aligned with the semantic labels predicted by the network using the Hungarian algorithm before fine-tuning. Calculate the cross-entropy fine-tuning loss : ； in, The total number of viewpoints. The total number of samples in the training set. For the first The first perspective The training set sample belongs to the th training set sample. The soft assignment probability of each cluster, This represents the pseudo-label indicator variable after matching; if the first... From the perspective of the first The matched pseudo-label of the sample belongs to the first... Each cluster, then ;otherwise, , To avoid numerical instability; the final cluster label is the average hard decision of the soft assignment probabilities of all views: ； in, For the first The final hard clustering label for each sample, For the first The first perspective The sample belongs to the first The soft assignment probability of each cluster.

8. An adaptive weighted missing-complete multi-view clustering ultra-short-term photovoltaic irradiance prediction system, characterized in that, include: The data processing unit is used to collect multimodal historical datasets of the target photovoltaic power station, including multivariate time series data and corresponding ground cloud image data; The data is preprocessed to form a dataset; and the dataset is divided into a training set, a validation set, and a test set according to a predetermined ratio. A prediction framework building unit is used to construct an irradiance prediction framework based on multi-view clustering of missing data. The prediction framework includes a pre-training module, a clustering module, and a regression prediction module. The pre-training module includes a CNN module, a BiLSTM encoder, a bidirectional prediction mechanism, a ScoreNet MLP, and a BiLSTM decoder. Multivariate time-series data without missing values ​​from the training set are input into the corresponding BiLSTM encoder. Simultaneously, ground-based cloud image data without missing values ​​are first processed by the CNN module to extract spatial features before being input into the BiLSTM encoder. The BiLSTM encoder outputs low-level features, which are then sequentially passed through the bidirectional prediction mechanism and ScoreNet. The MLP performs adaptive weighted imputation on missing data to obtain imputed low-level features. These imputed low-level features are then input into a BiLSTM decoder to obtain imputed data. The decoded imputed data is compared with the corresponding original data in the training set to calculate reconstruction loss and prediction loss. The clustering module performs multi-view clustering learning based on the imputed low-level features to obtain clustering results, and divides the imputed low-level features into K classes according to the clustering results. The regression prediction module receives the imputed data and its corresponding clustering results, and trains regression prediction sub-models 1 to K for classes 1 to K respectively. During prediction, the soft-assignment probabilities of the input sample in each view are averaged, and the class with the highest average probability is determined as the class to which the input sample belongs. Then, the corresponding regression prediction sub-model is called to output the irradiance prediction result. A joint training unit is used to set training parameters, jointly train the prediction framework, evaluate the performance of the training results, and save the optimal prediction framework parameters. The prediction unit is used to feed the historical input data of the target photovoltaic power station into the trained prediction framework, call the corresponding regression prediction sub-model according to the cluster category of the input sample, and output the prediction result of future photovoltaic power generation irradiance.

9. A non-volatile storage medium, characterized in that, Used to store a computer program, wherein the computer program, when executed by a processor, implements the method as described in any one of claims 1-7.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the method described in any one of claims 1-7.

Images