Photovoltaic power generation power prediction method

Abstract

The invention relates to a photovoltaic power generation power prediction method, which comprises the steps of collecting and processing real-time multi-source meteorological data to obtain a meteorological feature set; according to historical photovoltaic power generation data, time sequence information is obtained, dynamic time features are extracted, and a time dimension feature set is obtained; splicing and fusing the meteorological feature set and the time dimension feature set to obtain a fused feature set; the fused feature set is input into a power prediction model, a photovoltaic power generation power prediction value is obtained, the power prediction model is constructed through a double attention mechanism and an LSTM network and is obtained through training of a training set, and the training set comprises historical multi-source meteorological data, photovoltaic power generation data and corresponding photovoltaic power generation power values. Through the multi-source data fusion and deep learning method, the accuracy and stability of photovoltaic power generation power prediction are effectively improved, and important support is provided for efficient operation of a photovoltaic power station and power grid dispatching.

Description

Technical Field

[0001] The present invention relates to the field of electric power technology, and in particular to a photovoltaic power generation power prediction method. Background Art

[0002] As a core technology in the renewable energy sector, photovoltaic power generation is crucial for achieving energy transition. The accuracy of its power forecast directly impacts grid dispatch, energy management, and power plant operational efficiency, and is a crucial step in promoting the large-scale application of clean energy. However, existing photovoltaic power forecasting methods have significant limitations when dealing with complex meteorological conditions and dynamic temporal changes. Traditional models often rely on single features or simple statistical methods, making it difficult to fully capture the nonlinear relationship between meteorological factors and time series. This results in insufficient forecast accuracy, particularly under conditions of severe weather fluctuations or data anomalies.

[0003] The current core challenge lies in effectively integrating multidimensional meteorological information with the dynamic characteristics of time series. Specifically, meteorological factors such as irradiance, temperature, and humidity affect photovoltaic power generation to varying degrees, exhibiting complex patterns over time. Furthermore, outliers and missing values ​​in historical data further complicate modeling. These factors make it difficult for existing methods to accurately extract key features and effectively integrate them, thus impacting the reliability of forecast results. Therefore, the key issue in photovoltaic power generation forecasting is how to construct a forecasting model that comprehensively considers the impact of different meteorological factors at different times, fully exploits the deep connections between time and meteorological dimensions, and effectively handles data anomalies and missing values. Summary of the Invention

[0004] The purpose of this invention is to provide a photovoltaic power generation power prediction method, which effectively improves the accuracy and stability of photovoltaic power generation power prediction through multi-source data fusion and deep learning methods, and provides important support for the efficient operation of photovoltaic power stations and grid scheduling.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A photovoltaic power generation power prediction method, comprising:

[0007] Collect and process real-time multi-source meteorological data to obtain meteorological feature sets;

[0008] Based on historical photovoltaic power generation data, time series information is obtained and dynamic time features are extracted to obtain a time dimension feature set;

[0009] Splicing and fusing the meteorological feature set and the time dimension feature set to obtain a fused feature set;

[0010] The fused feature set is input into a power prediction model to obtain a photovoltaic power generation power prediction value, wherein the power prediction model is constructed through a dual attention mechanism and an LSTM network and is obtained through training with a training set, wherein the training set includes historical multi-source meteorological data, photovoltaic power generation data and corresponding photovoltaic power generation power values.

[0011] Optionally, real-time multi-source meteorological data is collected and processed to obtain a meteorological feature set including:

[0012] Collecting real-time multi-source meteorological data including irradiance, temperature, and humidity information, parsing the real-time multi-source meteorological data in a standardized format to obtain an initial meteorological data set;

[0013] Completing and smoothing the initial meteorological data set to obtain a complete meteorological data set;

[0014] Perform dimensionality reduction processing on the complete meteorological data set to obtain the meteorological feature set.

[0015] Optionally, completing and smoothing the initial meteorological dataset to obtain a complete meteorological dataset includes:

[0016] Determine whether the initial meteorological data set has missing values. If so, complete the irradiance, temperature, and humidity information using a linear interpolation method to obtain a completed meteorological data set.

[0017] Smoothing the completed meteorological data set using a mean filtering method to obtain a smoothed meteorological data set;

[0018] It is determined whether the data at each time point in the smoothed meteorological dataset is complete. If incomplete, the smoothed meteorological dataset is re-completed using a nearest neighbor interpolation method to obtain the complete meteorological dataset.

[0019] Optionally, performing dimensionality reduction processing on the complete meteorological dataset to obtain the meteorological feature set includes:

[0020] preprocessing the complete meteorological data set by a standardization method to eliminate dimensional differences and obtain a standardized meteorological data set;

[0021] The covariance matrix calculation method is used to calculate the covariance matrix between the variables of the standardized meteorological data set to obtain the covariance matrix

[0022] Decomposing the covariance matrix by an eigenvalue decomposition method, obtaining eigenvalues ​​and eigenvectors, and determining principal components;

[0023] Sorting the eigenvalues, selecting the principal components whose cumulative contribution rate reaches a preset threshold, and obtaining a principal component set;

[0024] Using a principal component analysis method, the standardized meteorological data set is projected onto the principal component set to obtain a dimension-reduced feature set;

[0025] If the dimension of the reduced dimension feature set is lower than the preset threshold, the principal component selection criteria are readjusted and the projection is repeated to obtain the optimized reduced dimension feature set;

[0026] By optimizing the dimension reduction feature set, key meteorological features are extracted to generate the meteorological feature set.

[0027] Optionally, obtain time series information and extract dynamic time features. Obtaining the time dimension feature set includes:

[0028] Performing outlier processing and smoothing on the time series information to obtain a smoothed time series;

[0029] Using a long short-term memory network to extract dynamic features of the smoothed time series to obtain initial time feature values;

[0030] Perform dimensionality reduction processing on the initial time feature value to obtain the time dimension feature set.

[0031] Optionally, concatenating and fusing the meteorological feature set and the time dimension feature set to obtain a fused feature set includes:

[0032] Using a matrix splicing operation, the meteorological feature set and the time dimension feature set are spliced ​​to obtain a spliced ​​feature set;

[0033] The meteorological features and time features in the spliced ​​feature set are weightedly calculated through an attention mechanism model to obtain the fused feature set.

[0034] Optionally, inputting the fused feature set into a power prediction model to obtain a photovoltaic power prediction value includes:

[0035] Encoding the fused feature set through an LSTM encoder to generate a hidden layer vector;

[0036] Calculating attention weights for the meteorological factors and time factors of the fused feature set through the dual attention mechanism, and combining them with the hidden layer vector to obtain a new vector;

[0037] The new vector is decoded by the LSTM network to obtain a photovoltaic power generation power prediction value.

[0038] Optionally, calculating the attention weights of the meteorological factors and the time factors of the fused feature set through the dual attention mechanism and combining them with the hidden layer vector to obtain a new vector includes:

[0039] The attention weight of the time factor is calculated through the self-attention mechanism module to generate the context vector of the time dimension;

[0040] The attention mechanism module calculates the attention weight of meteorological factors and generates a context vector of the meteorological dimension;

[0041] The context vectors of the time dimension and the weather dimension are fused with the hidden layer vector to obtain a new vector.

[0042] The beneficial effects of the present invention are as follows: the method of the present invention first collects real-time multi-source meteorological data and pre-processes it, and extracts key meteorological features through principal component analysis; at the same time, it obtains time series information from historical photovoltaic power generation data, and uses a long short-term memory network to extract dynamic time features. The meteorological features and time features are then spliced ​​together, and weighted fused through an attention mechanism to obtain a fused feature set. Finally, the fused feature set is input into a power prediction model constructed based on a dual attention mechanism and an LSTM network. The model is trained with a training set containing historical multi-source meteorological data, photovoltaic power generation data, and corresponding power values, thereby achieving accurate prediction of photovoltaic power generation power. The present invention effectively improves the accuracy and stability of photovoltaic power generation power prediction through multi-source data fusion and deep learning methods, providing important support for the efficient operation of photovoltaic power stations and grid scheduling. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0044] Figure 1 This is a flow chart of a photovoltaic power generation power prediction method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0046] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0047] like Figure 1As shown, this embodiment provides a photovoltaic power generation power prediction method, including:

[0048] Collect and process real-time multi-source meteorological data to obtain meteorological feature sets;

[0049] Based on historical photovoltaic power generation data, time series information is obtained and dynamic time features are extracted to obtain a time dimension feature set;

[0050] The meteorological feature set and the time dimension feature set are spliced ​​and fused to obtain a fused feature set;

[0051] The fused feature set is input into the power prediction model to obtain the photovoltaic power generation power prediction value. The power prediction model is constructed through a dual attention mechanism and an LSTM network, and is obtained through training with a training set. The training set includes historical multi-source meteorological data, photovoltaic power generation data, and the corresponding photovoltaic power generation power values.

[0052] Furthermore, real-time multi-source meteorological data is collected and processed to obtain meteorological feature sets including:

[0053] Collect real-time multi-source meteorological data including irradiance, temperature, and humidity information, parse the real-time multi-source meteorological data in a standardized format, and obtain the initial meteorological data set;

[0054] Completing and smoothing the initial meteorological dataset to obtain a complete meteorological dataset;

[0055] Perform dimensionality reduction on the complete meteorological dataset to obtain the meteorological feature set.

[0056] Specifically, data can be extracted from weather stations, satellite data, and numerical forecast models, and parsed into a unified CSV format, including timestamps, irradiance (W / m 2 ), temperature (℃), humidity (%).

[0057] Furthermore, the initial meteorological dataset is supplemented and smoothed to obtain a complete meteorological dataset including:

[0058] Determine whether the initial meteorological dataset has missing values. If so, use linear interpolation to complete the irradiance, temperature, and humidity information to obtain the completed meteorological dataset.

[0059] The completed meteorological data set is smoothed using the mean filtering method to obtain a smoothed meteorological data set;

[0060] Determine whether the data at each time point in the smoothed meteorological dataset is complete. If incomplete, use the nearest neighbor interpolation method to recompile the smoothed meteorological dataset to obtain a complete meteorological dataset.

[0061] In one possible implementation, suppose a weather station records hourly data. For a particular day, the data is as follows: 08:00, 600, 25, 70; 09:00, missing; and 10:00, 650, 26, 68. The initial dataset needs to be imputed because the data at 09:00 is missing. Linear interpolation is a simple and effective method for addressing missing values. Specifically, the irradiance at 09:00 can be calculated using the values ​​of 600 at 08:00 and 650 at 10:00. Assuming linear temporal variation, the imputed value is 625. Similarly, for temperature and humidity, the imputed dataset is: 08:00, 600, 25, 70; 09:00, 625, 25.5, 69; and 10:00, 650, 26, 68. This imputed dataset improves its integrity, ensuring continuity in subsequent analysis. Mean filtering is performed on the imputed dataset to smooth the data and reduce noise. This method is simple and efficient, suitable for short-term absences, and ensures the reliability of the complete meteorological dataset.

[0062] Furthermore, the complete meteorological dataset is subjected to dimensionality reduction to obtain the meteorological feature set including:

[0063] The complete meteorological dataset is preprocessed by standardization method to eliminate dimensional differences and obtain a standardized meteorological dataset;

[0064] Using the covariance matrix calculation method, the covariance matrix between variables is calculated for the standardized meteorological data set to obtain the covariance matrix

[0065] By using the eigenvalue decomposition method, the covariance matrix is ​​decomposed to obtain the eigenvalues ​​and eigenvectors and determine the principal components;

[0066] Sort by eigenvalues, select the principal components whose cumulative contribution rate reaches the preset threshold, and obtain the principal component set;

[0067] The principal component analysis method is used to project the standardized meteorological data set into the principal component set to obtain the dimension-reduced feature set;

[0068] If the dimension of the reduced dimension feature set is lower than the preset threshold, the principal component selection criteria are readjusted and the projection is repeated to obtain the optimized reduced dimension feature set;

[0069] By optimizing the dimensionality reduction feature set, key meteorological features are extracted and a meteorological feature set is generated.

[0070] In one embodiment, assuming the analysis shows a high correlation between irradiance and temperature, the three-dimensional data can be reduced to two dimensions while retaining 90% of the variance.

[0071] Furthermore, time series information is obtained and dynamic time features are extracted. The time dimension feature set obtained includes:

[0072] Perform outlier processing and smoothing on the time series information to obtain a smoothed time series;

[0073] Long short-term memory network is used to extract dynamic features of smooth time series and obtain initial time feature values;

[0074] Perform dimensionality reduction on the initial time feature values ​​to obtain the time dimension feature set.

[0075] Specifically, the initial time series is extracted by obtaining continuous power generation data from historical photovoltaic power generation data, typically in hours or minutes. For example, at a photovoltaic power plant, technicians extracted hourly power generation data for the past year from a database, generating a series of 8,760 data points. This series reflects the temporal changes in power generation and provides a basis for subsequent analysis. It is important to note that data extraction requires timestamp alignment to prevent missing values ​​from impacting analysis. Identifying and handling outliers is critical to ensuring data quality. Outliers may be sudden changes in power generation caused by equipment failure, sensor errors, or extreme weather. Specifically, assuming a preset threshold of ±3 standard deviations of the normal power generation mean, any hourly power generation outside this range is marked as an outlier. In one embodiment, a power plant discovered abnormally high power generation in the early morning of a certain day, which was later identified as a sensor failure. A median filtering algorithm was applied to the outlier set, replacing the outlier with the median of the adjacent data points. For example, for an outlier, the median of the five data points before and after the outlier was selected to generate a smoothed time series. This method effectively eliminates sudden changes while preserving the overall trend of the series. Smoothed time series are used to extract periodic features that reflect regular changes in photovoltaic power generation, such as daily cycles or seasonal cycles.

[0076] In one possible implementation, a long-short-term memory network (LSTM) is used to train smoothed time series to extract dynamic temporal features. Photovoltaic power generation is affected by weather and seasons, and the data exhibits nonlinear variations. The network captures the temporal dependence of power generation through memory units. For example, if the training data contains power generation for 30 consecutive days, the network learns the pattern of decreased power generation on rainy days and generates an initial temporal feature set containing features such as the daily rate of change of power generation and peak time. This dynamic feature extraction helps capture complex temporal patterns. If the initial temporal feature set is too high in dimensionality, such as 100 features, principal component analysis (PCA) is used to reduce the dimensionality. After analysis, features that explain 80% of the variance are retained, reducing the number to 10 key features, such as average daily power generation and fluctuation amplitude.

[0077] Furthermore, the meteorological feature set and the time dimension feature set are spliced ​​and fused to obtain a fused feature set including:

[0078] Matrix splicing operation is used to splice the meteorological feature set and the time dimension feature set to obtain a spliced ​​feature set;

[0079] The meteorological features and time features in the spliced ​​feature set are weightedly calculated through the attention mechanism model to obtain the fused feature set.

[0080] Specifically, a matrix concatenation operation is used to combine the dimensionality reduction feature set with the time dimension feature set to obtain a concatenated feature set. If the dimension of the concatenated feature set exceeds a preset threshold, the concatenated feature set is screened using a feature selection algorithm to obtain an optimized concatenated feature set. If the dimension does not exceed the preset threshold, the concatenated feature set is directly determined to be the optimized concatenated feature set. The meteorological and temporal features in the optimized fusion feature set are weighted using an attention mechanism model to obtain a weighted fusion feature set. Based on the weighted fusion feature set, the feature values ​​are normalized using standardization to obtain the final fusion feature set. The quality of the final fusion feature set is assessed using a cross-validation algorithm to determine its stability.

[0081] Feature selection algorithms such as recursive feature elimination can evaluate the importance of features. For example, in the six-dimensional feature set of a power station, analysis found that a certain time code contributed little to the prediction, so after elimination, the five-dimensional optimized fusion feature set was retained. If the dimension is lower than the threshold, such as six dimensions, the spliced ​​feature set is used directly. This screening ensures feature efficiency. In one embodiment, the attention mechanism model performs weighted calculations on the optimized spliced ​​feature set. The attention mechanism assigns weights to each feature to highlight important features. For example, in the five-dimensional feature set of a power station, the irradiance principal component is more critical in the prediction, and the attention model assigns it a higher weight to generate a weighted fusion feature set. This weighting can focus on core information.

[0082] Furthermore, the fused feature set is input into the power prediction model to obtain the photovoltaic power prediction value including:

[0083] Encode the fused feature set through the LSTM encoder to generate the hidden layer vector;

[0084] The attention weights of the meteorological factors and time factors of the fusion feature set are calculated through the dual attention mechanism, and the new vector is obtained by combining the hidden layer vector.

[0085] The new vector is decoded through the LSTM network to obtain the predicted value of photovoltaic power generation.

[0086] Furthermore, the dual attention mechanism is used to calculate the attention weights of the meteorological factors and time factors of the fusion feature set, and combined with the hidden layer vector, a new vector is obtained, including:

[0087] The attention weight of the time factor is calculated through the self-attention mechanism module to generate the context vector of the time dimension;

[0088] The attention mechanism module calculates the attention weight of meteorological factors and generates a context vector of the meteorological dimension;

[0089] The context vectors of the time dimension and the weather dimension are fused with the hidden layer vector to obtain a new vector.

[0090] Specifically, in one possible implementation, the input vector contains hourly photovoltaic power generation and meteorological data, such as temperature and irradiance. Assuming that the meteorological data for a particular day includes a temperature of 25 degrees Celsius and an irradiance of 600 watts per square meter at 8:00 AM, the LSTM encoder analyzes this data and generates a hidden layer vector that captures the underlying pattern of power variation. This encoding approach ensures that subsequent modules can process based on rich temporal information. Specifically, the self-attention mechanism module calculates attention weights for different moments and generates a context vector in the time dimension. The self-attention mechanism compares data at each moment in the time series to determine which moments are more important for prediction. For example, in a day's power generation data, 12:00 PM may be assigned a higher attention weight due to the highest irradiance. Assuming the power is 800 kilowatts at 10:00 AM and 1200 kilowatts at 12:00 AM, the self-attention mechanism calculates a weight of 0.7 for 12:00 AM and 0.3 for 10:00 AM. The resulting context vector highlights the high power characteristics of noon. This approach ensures that the model focuses on key time points. In one embodiment, the attention mechanism module calculates attention weights for different meteorological factors and generates a context vector in the meteorological dimension. Different meteorological factors have varying degrees of impact on photovoltaic power generation. The attention mechanism can highlight key factors. For example, irradiance and temperature have a greater impact on power, while humidity has a smaller impact. For example, assuming the irradiance at a certain moment is 700 watts / square meter, the temperature is 28 degrees Celsius, and the humidity is 60%, the attention mechanism might assign a weight of 0.5 to irradiance, 0.4 to temperature, and 0.1 to humidity. The resulting meteorological context vector emphasizes the impact of irradiance and temperature. This approach allows the model to focus more on key meteorological factors. The context vectors for both the time and meteorological dimensions are then fused with the hidden layer vector to generate a new vector. For example, the time context vector highlights the power characteristics at 12:00, the meteorological context vector emphasizes irradiance, and the hidden layer vector contains the overall temporal pattern. The fusion module concatenates these three into a single high-dimensional vector. This fusion approach ensures the synergy of multi-source information, providing comprehensive features for subsequent decoding. Preferably, the decoder decodes the new vector to generate a preliminary prediction of photovoltaic power generation. The decoder typically uses a fully connected layer or LSTM network to map the fused vector to the predicted value. For example, the fused vector contains a high irradiance feature at 12 o'clock, and the decoder may predict that the power at that moment is 1250 kilowatts.

[0091] The encoder provides the temporal basis, the self-attention mechanism highlights key moments, the attention mechanism focuses on key meteorological factors, and then integrates multi-dimensional information. The decoder generates the final prediction. The logical progression of each component ensures the rigor and efficiency of the prediction process.

[0092] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A photovoltaic power generation power prediction method, characterized in that: include: Collect and process real-time multi-source meteorological data to obtain meteorological feature sets; Based on historical photovoltaic power generation data, time series information is obtained and dynamic time features are extracted to obtain a time dimension feature set; Splicing and fusing the meteorological feature set and the time dimension feature set to obtain a fused feature set; The fused feature set is input into a power prediction model to obtain a photovoltaic power generation power prediction value, wherein the power prediction model is constructed through a dual attention mechanism and an LSTM network and is obtained through training with a training set, wherein the training set includes historical multi-source meteorological data, photovoltaic power generation data and corresponding photovoltaic power generation power values.

2. The photovoltaic power generation prediction method according to claim 1, characterized in that: Collect and process real-time multi-source meteorological data to obtain meteorological feature sets including: Collecting real-time multi-source meteorological data including irradiance, temperature, and humidity information, parsing the real-time multi-source meteorological data in a standardized format to obtain an initial meteorological data set; Completing and smoothing the initial meteorological data set to obtain a complete meteorological data set; Perform dimensionality reduction processing on the complete meteorological data set to obtain the meteorological feature set.

3. The photovoltaic power generation prediction method according to claim 2, characterized in that: Completing and smoothing the initial meteorological dataset to obtain a complete meteorological dataset includes: Determine whether the initial meteorological data set has missing values. If so, complete the irradiance, temperature, and humidity information using a linear interpolation method to obtain a completed meteorological data set. Smoothing the completed meteorological data set using a mean filtering method to obtain a smoothed meteorological data set; It is determined whether the data at each time point in the smoothed meteorological dataset is complete. If incomplete, the smoothed meteorological dataset is re-completed using a nearest neighbor interpolation method to obtain the complete meteorological dataset.

4. The photovoltaic power generation prediction method according to claim 2, characterized in that: Performing dimensionality reduction processing on the complete meteorological data set to obtain the meteorological feature set includes: preprocessing the complete meteorological data set by a standardization method to eliminate dimensional differences and obtain a standardized meteorological data set; The covariance matrix calculation method is used to calculate the covariance matrix between the variables of the standardized meteorological data set to obtain the covariance matrix Decomposing the covariance matrix by an eigenvalue decomposition method, obtaining eigenvalues ​​and eigenvectors, and determining principal components; Sorting by the size of the eigenvalues, selecting the principal components whose cumulative contribution rate reaches a preset threshold, and obtaining a principal component set; Using a principal component analysis method, the standardized meteorological data set is projected onto the principal component set to obtain a dimension-reduced feature set; If the dimension of the reduced dimension feature set is lower than the preset threshold, the principal component selection criteria are readjusted and the projection is repeated to obtain the optimized reduced dimension feature set; By optimizing the dimension reduction feature set, key meteorological features are extracted to generate the meteorological feature set.

5. The photovoltaic power generation prediction method according to claim 1, characterized in that: Obtain time series information and extract dynamic time features. Obtaining the time dimension feature set includes: Performing outlier processing and smoothing on the time series information to obtain a smoothed time series; Using a long short-term memory network to extract dynamic features of the smoothed time series to obtain initial time feature values; Perform dimensionality reduction processing on the initial time feature value to obtain the time dimension feature set.

6. The photovoltaic power generation prediction method according to claim 1, characterized in that: The meteorological feature set and the time dimension feature set are spliced ​​and fused to obtain a fused feature set, including: Using a matrix splicing operation, the meteorological feature set and the time dimension feature set are spliced ​​to obtain a spliced ​​feature set; The meteorological features and time features in the spliced ​​feature set are weightedly calculated through an attention mechanism model to obtain the fused feature set.

7. The photovoltaic power generation prediction method according to claim 1, characterized in that: Inputting the fused feature set into the power prediction model to obtain the photovoltaic power prediction value includes: Encoding the fused feature set through an LSTM encoder to generate a hidden layer vector; Calculating attention weights for the meteorological factors and time factors of the fused feature set through the dual attention mechanism, and combining them with the hidden layer vector to obtain a new vector; The new vector is decoded by the LSTM network to obtain a photovoltaic power generation power prediction value.

8. The photovoltaic power generation prediction method according to claim 7, characterized in that: The attention weights of the meteorological factors and time factors of the fused feature set are calculated by the dual attention mechanism, and combined with the hidden layer vector to obtain a new vector including: The attention weight of the time factor is calculated through the self-attention mechanism module to generate the context vector of the time dimension; The attention mechanism module calculates the attention weight of meteorological factors and generates a context vector of the meteorological dimension; The context vectors of the time dimension and the weather dimension are fused with the hidden layer vector to obtain a new vector.

Images