Photovoltaic power generation power prediction method and device, electronic equipment and storage medium

Abstract

The application relates to the technical field of photovoltaic power generation, in particular to a photovoltaic power generation power prediction method and device, electronic equipment and a storage medium. Multi-source heterogeneous data such as historical power generation power, NWP, satellite remote sensing and cloud height corresponding to a prediction time point are acquired and preprocessed, the preprocessed data is input into four deep learning sub-models, namely a historical power analysis model, a numerical prediction analysis model, a cloud area analysis model and a cloud height analysis model, corresponding future power prediction sequences are respectively output, the sequences are fused, and a photovoltaic power generation power prediction result is obtained. Through comprehensive multi-source data and multi-model analysis fusion, the application can more accurately capture local cloud cluster changes, mine deep features of multi-modal data and consider the influence of physical processes, thereby effectively improving the accuracy and reliability of photovoltaic power generation power prediction.

Description

Technical Field

[0001] This application relates to the field of photovoltaic power generation technology, and in particular to a photovoltaic power generation prediction method, device, electronic device and storage medium. Background Technology

[0002] In the fields of energy forecasting, especially photovoltaic power forecasting, accurate forecasting plays a crucial role in ensuring the stable operation and efficient allocation of energy systems. Traditional photovoltaic power forecasting methods rely on numerical weather prediction, which suffers from low spatiotemporal resolution and imprecise spatiotemporal perception, making it difficult to accurately capture the rapid movement of local cloud clusters that trigger sudden power changes. Furthermore, current deep learning models often only perform simple data fusion on a basic level, lacking in-depth feature extraction. Energy forecasting involves diverse data types, such as meteorological data, satellite cloud imagery, and historical power data, all of which contain rich information. However, existing deep learning models mostly simply stitch these data together without fully exploring the deep correlations and features between different data points, failing to achieve effective feature extraction and thus limiting the improvement of forecast accuracy. Meanwhile, the purely data-driven "black box" model fails to effectively construct the physical process of "cloud cover - reduced irradiance - reduced power output", resulting in poor interpretability; single temporal convolutional networks (TCN) or long short-term memory (LSTM) models are difficult to simultaneously capture the long-term temporal dependence of power, spatial characteristics of cloud images, and the future impact of meteorological conditions.

[0003] In addition, although there have been predictive studies based on satellite cloud images, most of them simply input a single channel cloud image into a convolutional neural network (CNN) and combine it with temporal processing, without fully exploring temporal and physical information (such as cloud height). Furthermore, there are no reports on a scheme that designs deep learning sub-models for the four dimensions of power time series, future weather, two-dimensional cloud area, and three-dimensional cloud height and effectively integrates them. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention provides a photovoltaic power generation prediction method, device, electronic equipment, and storage medium. By leveraging deep feature extraction and fusion, it achieves collaborative modeling of historical patterns, future weather, cloud planar distribution, and vertical structure, thereby improving prediction accuracy and interpretability.

[0005] A first aspect of this application provides a method for predicting photovoltaic power generation, the method comprising: Acquire multi-source heterogeneous data corresponding to the predicted time point and preprocess the multi-source heterogeneous data; the multi-source heterogeneous data includes historical power generation data, NWP data, satellite remote sensing data and cloud height data; Preprocessed multi-source heterogeneous data is input into deep learning sub-models, and each sub-model outputs a future power prediction sequence. The deep learning sub-models include a historical power analysis model, a numerical weather prediction analysis model, a cloud region analysis model, and a cloud height analysis model. The historical power analysis model is used to predict the future power prediction sequence corresponding to the historical power generation data. The numerical weather prediction analysis model is used to predict the future power prediction sequence corresponding to the NWP data. The cloud region analysis model is used to predict the future power prediction sequence corresponding to the fusion of historical power generation data and satellite remote sensing data. The cloud height analysis model is used to predict the future power prediction sequence corresponding to the fusion of satellite remote sensing data and cloud height data. The future power prediction sequence is fused to obtain the photovoltaic power generation prediction result.

[0006] In an optional implementation, fusing the future power prediction sequence to obtain the photovoltaic power prediction result includes: adaptively fusing the future power prediction sequences output by the historical power analysis model, the numerical forecast analysis model, the cloud region analysis model, and the cloud height analysis model through a Fusion fusion layer; the Fusion fusion layer adopts a parameter matrix of shape [4, 16], where the number 4 represents the number of sub-models in the deep learning sub-model, and the number 16 represents the dimension of the time series prediction value; the Fusion fusion layer automatically adjusts the weight allocation after training and achieves adaptive accuracy contribution allocation based on data-driven methods to fuse the future power prediction sequence output by the deep learning sub-model.

[0007] In an optional implementation, the method further includes: Freeze the Fusion fusion layer, using parameters all of 1 and unlearnable. Instead of the original fusion layer, only the historical power analysis model, the numerical forecast analysis model, the cloud region analysis model, and the cloud height analysis model are trained; The historical power analysis model, the numerical forecast analysis model, the cloud region analysis model, and the cloud height analysis model are frozen and replaced with a trainable Fusion fusion layer. The weight allocation parameters of the Fusion fusion layer are then trained.

[0008] In one optional implementation, the historical power analysis model is constructed based on the iTransformer framework, the numerical forecast analysis model is constructed based on the FA-MLP structure, the cloud region analysis model is constructed based on U-Net, TCN structure and cross-attention, and the cloud height analysis model is constructed based on Prior_Attention and TCN structure.

[0009] In an optional implementation, the method further includes: Within the grid area, the maximum value in the preprocessed tbb_13 channel data is selected as the ground temperature reference; Calculate the difference between the value of each pixel within the grid region and the ground temperature reference. ; Based on the decreasing factor, using the formula Calculate pixel coordinates The cloud height values ​​corresponding to the grid points; where Represents pixel coordinates The cloud height value corresponding to the grid point. It is a decreasing factor.

[0010] In an optional implementation, the preprocessing of the multi-source heterogeneous data includes: The instantaneous power generation value collected at time t is used as the historical power generation data, and the historical power generation data is normalized using the following formula: ; in, This represents the normalized historical power generation data. This represents the instantaneous power generation value collected at time t. This represents the inverter's on-time capacity at time t; The satellite remote sensing data is normalized according to the channel type using the following formula: ; in, This represents normalized satellite remote sensing data. Represents the pixel coordinates of the c channel. grayscale value, This indicates the upper limit of the effective grayscale data in channel c. This indicates the lower limit of the valid grayscale data in channel c; The NWP data is normalized using the following formula: ; in, This represents the normalized NWP data. This indicates the corresponding weather model when generating NWP data. The values ​​of meteorological elements, This represents the reasonable maximum value of the corresponding meteorological element in the NWP data, and this represents the reasonable minimum value of the corresponding meteorological element in the NWP data.

[0011] In an optional implementation, the preprocessing of the multi-source heterogeneous data further includes: The normalized historical power generation data is refined from a time interval of 1 step / 15 minutes to 1 step / 5 minutes through interpolation. When the time interval of the satellite remote sensing data is determined to be 15 minutes, the normalized satellite remote sensing data is refined from a time interval of 1 step / 15 minutes to 1 step / 5 minutes through interpolation. The normalized NWP data is refined from a time interval of 1 step / 1 hour to 1 step / 5 minutes through interpolation.

[0012] A second aspect of this application provides a photovoltaic power generation prediction device, the device comprising: The data processing module is used to acquire multi-source heterogeneous data corresponding to the prediction time point and preprocess the multi-source heterogeneous data; the multi-source heterogeneous data includes historical power generation data, NWP data, satellite remote sensing data and cloud height data; A separate prediction module is used to input preprocessed multi-source heterogeneous data into deep learning sub-models and output future power prediction sequences for each sub-model. The deep learning sub-models include a historical power analysis model, a numerical weather prediction analysis model, a cloud region analysis model, and a cloud height analysis model. The historical power analysis model is used to predict the future power prediction sequence corresponding to the historical power generation data. The numerical weather prediction analysis model is used to predict the future power prediction sequence corresponding to the NWP data. The cloud region analysis model is used to predict the future power prediction sequence corresponding to the fusion of historical power generation data and satellite remote sensing data. The cloud height analysis model is used to predict the future power prediction sequence corresponding to the fusion of satellite remote sensing data and cloud height data. The fusion prediction module is used to fuse the future power prediction sequence to obtain the photovoltaic power generation prediction result.

[0013] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the photovoltaic power generation prediction method.

[0014] A fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the photovoltaic power generation prediction method described above.

[0015] In summary, the photovoltaic power generation prediction method, apparatus, electronic equipment, and storage medium provided in this application have at least one of the following beneficial effects: 1. Acquire multi-source heterogeneous data corresponding to the prediction time point, including satellite remote sensing data and cloud height data. By combining these data with other data (historical power generation data, NWP data) and inputting them into different deep learning sub-models for analysis, it is possible to perceive spatiotemporal changes more precisely. Compared with traditional methods that rely on low-resolution numerical weather forecasting, it can more accurately capture the phenomenon of rapid movement of local cloud clusters that cause power mutations.

[0016] 2. Four deep learning sub-models were designed: a historical power analysis model, a numerical weather prediction (NWP) analysis model, a cloud region analysis model, and a cloud height analysis model. Each sub-model processes different types of data or data fusions: the historical power analysis model specifically processes historical power generation data to uncover its inherent patterns and predict future power; the numerical weather prediction (NWP) analysis model focuses on NWP data, extracting power-related features for prediction; the cloud region analysis model takes historical power generation data and satellite remote sensing data as input, and uses its internal mechanisms to uncover the deep correlations and features between these two types of data, achieving effective collaborative feature extraction rather than simple concatenation; the cloud height analysis model processes the fused data of satellite remote sensing data and cloud height data, similarly delving into the feature relationships between them. By designing sub-models for different data combinations, the deep correlations and features between different data can be fully explored, achieving effective collaborative feature extraction and thus improving prediction accuracy.

[0017] 3. Satellite remote sensing data and cloud height data were introduced, and cloud region analysis models and cloud height analysis models were designed. Satellite remote sensing data can reflect the distribution and shading of clouds, while cloud height data is related to the impact of clouds on irradiance. By processing and analyzing the relevant data through these two models, the physical process of "cloud shading - reduced irradiance - reduced power output" can be simulated and reflected to a certain extent. Compared with purely data-driven "black box" models, this improves the interpretability of the models.

[0018] 4. Multiple deep learning sub-models are employed to process different types of data or for data fusion. Each sub-model can be optimized for specific tasks and data. For example, the historical power analysis model can better capture the long-term temporal dependence of historical power generation data; the cloud region analysis model can extract the spatial features of cloud images using satellite remote sensing data; and the numerical weather prediction analysis model can process NWP data to consider the future impact of meteorological conditions. Finally, the future power prediction sequences output by each sub-model are fused, combining the advantages of different models to simultaneously capture the long-term temporal dependence of power, the spatial features of cloud images, and the future impact of meteorological conditions.

[0019] 5. In addition to acquiring satellite remote sensing data (involving two-dimensional cloud area information), cloud height data (involving three-dimensional cloud height information) was also acquired, combined with historical power generation data (power time series information) and NWP data (future weather information). Four deep learning sub-models were designed, each targeting different dimensions of data or data fusion. Finally, the future power prediction sequences output by each sub-model were fused, achieving effective processing and fusion of four dimensions: power time series, future weather, two-dimensional cloud area, and three-dimensional cloud height. Attached Figure Description

[0020] Figure 1 This is a schematic flowchart illustrating a photovoltaic power generation prediction method according to an embodiment of this application; Figure 2 This is satellite remote sensing data of a channel with a center wavelength of 1.61 micrometers at a time step, as shown in an embodiment of this application. Figure 3 This is a schematic diagram of the architecture of a cloud region analysis model shown in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an attention module in a cloud region analysis model, as shown in an embodiment of this application. Figure 5 This is a feature map of the hidden layer 3-head cross-attention scores of a cloud region analysis model after training, as shown in an embodiment of this application. Figure 6 This is a schematic diagram of the prior-attention module in a cloud height analysis model, as shown in an embodiment of this application. Figure 7 This is a flowchart illustrating the step-by-step training logic of various sub-models and the Fusion fusion layer, as shown in an embodiment of this application. Figure 8 This is a functional block diagram of a photovoltaic power generation prediction model shown in an embodiment of this application; Figure 9 This is a schematic diagram of the structure of an electronic device shown in an embodiment of this application. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] The following will clearly and completely describe the concept, specific structure, and technical effects of the present invention in conjunction with embodiments and accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the scope of protection of the present invention. Furthermore, all connections / linkages involved in the patent do not simply refer to direct contact between components, but rather to the ability to form a better connection structure by adding or reducing connecting accessories according to specific implementation conditions. The various technical features in this invention can be combined interactively without contradicting each other.

[0023] Reference Figure 1 The diagram shown is a flowchart illustrating a photovoltaic power generation prediction method according to an embodiment of this application. The photovoltaic power generation prediction method includes the following steps.

[0024] To facilitate understanding of the inventive concept of this application, the photovoltaic power generation prediction method operates in a specific hardware environment. The electronic device is equipped with at least a central processing unit (CPU) and a graphics processing unit (GPU). The CPU is an Intel Xeon E5-2670, possessing powerful multi-threaded computing capabilities, enabling rapid processing of various computational tasks in data preprocessing, such as data cleaning and format conversion, while also meeting the needs of multi-task scheduling during model training. The GPU is an NVIDIA RTX 3090 with 24GB of video memory, providing ample memory space for the training and inference of the deep learning model, accelerating the forward and backward propagation processes, and significantly shortening training time. Furthermore, 32GB of RAM ensures the stability and smoothness of the system when processing large amounts of data, running complex programs, and performing multiple tasks simultaneously, avoiding program crashes or slow operation due to insufficient memory.

[0025] In terms of software environment, Ubuntu 20.04 was chosen as the operating system. Its stability and security provide a solid foundation for the operation of the entire technical solution, effectively managing hardware resources and ensuring the normal operation of each software component. Python 3.8 was used as the programming language, leveraging its rich libraries and concise syntax to facilitate data processing, model development, and result analysis. PyTorch 2.7 was adopted as the deep learning framework, providing robust support for model building, training, and optimization, and facilitating the implementation of various deep learning algorithms and functions, such as automatic differentiation and model parallelism. Furthermore, relevant data processing libraries (such as Pandas and NumPy) were used for data reading, cleaning, and transformation, and data visualization libraries (such as Matplotlib) were used to visualize the prediction results.

[0026] S11, acquire multi-source heterogeneous data corresponding to the predicted time point, and preprocess the multi-source heterogeneous data.

[0027] When a photovoltaic power generation forecast command is received, the corresponding forecast time point is determined, and multi-source heterogeneous data corresponding to the forecast time point is obtained, which may include, but is not limited to: historical power generation data, NWP (numerical weather forecast) data, satellite remote sensing data, and cloud height data.

[0028] Specifically, electronic devices can acquire historical power generation data 96 time steps prior to the prediction time, with 15-minute intervals, starting 5-10 minutes before the prediction time; and acquire NWP data (Natural Power Plant) data for the 24-hour irradiance and temperature predicted by the meteorological model at the prediction time, with 1-hour intervals, obtained through the API interface provided by the open-meteo official website https: / / open-meteo.com / ; and acquire satellite remote sensing data within 5 hours prior to the prediction time, starting 40 minutes before the prediction time, with data type being full-disk hemispherical data captured by satellites using sensors with center wavelengths of 1.61 micrometers and 10.4 micrometers, with 15-minute or 10-minute intervals, referring to... Figure 2 As shown. The satellite remote sensing data includes solar angle data, which includes solar azimuth and solar altitude angles. Using the Python 3 plugin PySolar, the azimuth and altitude angles of the sun relative to the target point can be obtained by inputting the latitude, longitude, and time zone of the target point.

[0029] It should be noted that the NWP data are all derived from the results of official meteorological models. The models use daily 00:00 forecasts of weather conditions for the next 24 hours, therefore the collected NWP data represents future NWP data. "Future" here refers to the timestamps of the data used, which are relative to the time point in the algorithm model's inference in this embodiment.

[0030] Simultaneously, electronic devices can also utilize the tbb_13 channel data of the Himawari L1 Gridded data for inversion calculations to obtain cloud height data. Specifically, this includes the following steps: A) Select the highest value of the preprocessed data within the grid area. As a reference for ground temperature.

[0031] in, This is the maximum value in the preprocessed tbb_13 channel data, expressed in Kelvin. For example, 0 Kelvin represents absolute 0 degrees. The highest value is the maximum value of a single frame of target area data captured by the tbb_13 channel sensor in the historical time frame of Himawari L1 Gridded data.

[0032] B) Calculate the value of each pixel within the grid. The difference .

[0033] C) According to the decreasing factor Cloud height can be calculated using the following formula: ; in, Represents pixel coordinates Cloud height values ​​(in km) corresponding to grid points, decrease factor The value is -5.5 K / km.

[0034] It should be noted that the value of the decreasing factor, -5.5K / km, is an empirical value specified after comparative testing of the embodiments of this application.

[0035] Once the cloud height values ​​are obtained, they are used to generate a mask to segment the cloud planar distribution data into height intervals. For example, low clouds are defined as 0 ≤ [missing value]. <2km; medium cloud range is 2≤ <7km; Gao Yunwei ≥7km. And generate a binary mask, for example, the cloud mask is... .

[0036] In some embodiments, after acquiring historical power generation data, NWP data, and satellite remote sensing data, the electronic device can normalize this data. Specifically, for historical power generation data, the electronic device can normalize it using the following formula: ; in, This represents the normalized historical power generation data (hereinafter referred to as historical normalized power). express t Instantaneous power generation value (kW) collected at all times. express t Inverter operating capacity at any time (kW).

[0037] For satellite remote sensing data, each channel is normalized using the following formula and value range: ; in, This represents normalized satellite remote sensing data. Represents the pixel coordinates of the c channel. grayscale value, This indicates the upper limit of the effective grayscale data in channel c. This represents the effective lower limit value of the grayscale data in channel c. The normalized value range table for remote sensing data is shown in Table 1 below.

[0038] Table 1:

[0039] For solar azimuth data, electronic devices can convert it into a two-dimensional vector for normalization using the following formula: ; ; ; in, This indicates the original solar azimuth angle.

[0040] For solar altitude angle data, electronic devices can linearly normalize it using the following formula: ; in, This represents the original solar altitude angle.

[0041] For NWP data, electronic devices can be normalized using the following formula and value range: ; in, This represents the normalized NWP data (hereinafter referred to as future NWP data). This represents the value of the meteorological element corresponding to the NWP data generated by the large meteorological model. This represents the reasonable maximum value of the corresponding meteorological element in the NWP data. This represents the reasonable minimum value of the corresponding meteorological element in the NWP data. The normalized value range table for NWP data is shown in Table 2 below.

[0042] Table 2:

[0043] In some embodiments, the electronic device also needs to perform time-series alignment on multi-source data. Specifically, this involves refining the time-series granularity using interpolation, including: NWP data: 1 step / hour → interpolation → 1 step / 5min; RSD (Remote Sensing Desktop, a general-purpose remote sensing data processing platform) satellite remote sensing data (15-minute interval): 1 step / 15min → interpolation → 1 step / 5min; where remote sensing data with a 10-minute interval does not require interpolation.

[0044] Historical PV power generation data: 1 step / 15min → interpolation → 1 step / 5min.

[0045] Since the historical power analysis model, numerical weather prediction analysis model, cloud region analysis model, and cloud height analysis model require different data types and time series intervals for their input data, the specific input data types and time series alignment tables are shown in Tables 3 to 5 below. The four request inference combinations can cover any 15-minute time point for the inference request. Specifically, Table 3 shows the data type and time series alignment table for the historical power analysis model, Table 4 shows the data type and time series alignment table for the numerical weather prediction analysis model, and Table 5 shows the data type and time series alignment table for the cloud region and cloud height models.

[0046] Table 3:

[0047] Table 4:

[0048] Table 5:

[0049] It should be noted that solar angle data does not require time-series alignment; it can be calculated incidentally using a Python plugin while processing other time-series data within the program. Cloud height data does not require normalization; time-series alignment is completed during RSD time-series alignment, and since cloud height data is derived from RSD data inversion, the calculated cloud height data naturally includes time-series alignment attributes.

[0050] S12 inputs the preprocessed multi-source heterogeneous data into the deep learning sub-model and outputs the future power prediction sequence of each sub-model.

[0051] The deep learning sub-models include a historical power analysis model, a numerical forecast analysis model, a cloud region analysis model, and a cloud height analysis model. The historical power analysis model is used to predict the future power prediction sequence corresponding to the historical power generation data. The numerical forecast analysis model is used to predict the future power prediction sequence corresponding to the NWP data. The cloud region analysis model is used to predict the future power prediction sequence corresponding to the fusion of the historical power generation data and the satellite remote sensing data. The cloud height analysis model is used to predict the future power prediction sequence corresponding to the fusion of the satellite remote sensing data and the cloud height data.

[0052] The historical normalized power sequence contains detailed power changes over a past period, while the future NWP data covers information on temperature and irradiance across 16 time steps. By inputting the historical normalized power sequence into the historical power analysis model, the output of the model is a 16-dimensional future normalized power prediction sequence, providing accurate predictive references for future power conditions. In this embodiment, the historical power analysis model is constructed using the iTransformer framework (a model designed for multivariate time series forecasting tasks). By leveraging the powerful tool of iTransformer, the model delves into the long-term temporal correlations of power, uncovering hidden patterns through in-depth mining of historical data, thus laying a solid foundation for future power prediction.

[0053] Simultaneously, future NWP data is input into the numerical weather prediction (NWMR) model, enabling the model to output a 16-dimensional future normalized power prediction sequence. In this embodiment, a numerical weather prediction model is constructed using a FA-MLP structure. Functionally, a feature gating mechanism is employed to calculate the contribution ratio (score) of temperature and irradiance to the final prediction data at each time step in the historical period. The input [16, 2] NWP data is then weighted and summed at each historical step using the score, achieving optimal contribution weighting of the two meteorological data at different historical time steps. The weighted result is then converted into a power prediction result by an MLP (Multilayer Perceptron). The output of this module is crucial for improving the final prediction accuracy.

[0054] Simultaneously, normalized 1-channel cloud grid data (i.e., satellite remote sensing data), solar angle grid data (i.e., solar angle data obtained through the PySolar plugin and processed into grid data format internally), and 5 hours of historical power data are input into the cloud region analysis model after being encoded by location and time (PE+TE). This data is then referenced together. Figure 3 In this embodiment, a cloud region analysis model is constructed using U-Net (a CNN architecture for image segmentation) + TCN + cross-attention. The cloud region analysis model first concatenates the encoded cloud image and solar angle data, then performs downsampling (DS) and batch normalization (BN) operations via U-Net to output feature maps of three different resolutions. Each feature map is processed by a parameter-discretionary TCN3D (three-dimensional temporal convolutional network) to extract temporal features, resulting in the output RSD_hidden (also known as RSD_TE_PE, remote sensing data with temporal and location encoding information). Historical power data is processed by TCN1D (one-dimensional temporal convolutional network) to extract temporal features, outputting PV_hidden (also known as PV_TE_PE, power data with temporal and location encoding information). Next, the three RSD_hidden values ​​at different resolutions are used as keys and values, respectively, to perform 1D-to-2D cross-attention calculations with the PV_hidden value used as the query, outputting three ff_out values. The cross-attention calculation process is as follows: Figure 4 As shown, the three ff_out values ​​are concatenated along the channel dimension (CH) (i.e., cat in CH), and then subjected to time series analysis by a multilayer perceptron (MLP) composed of fully connected layers (FC) via residual connections. Finally, a linear regression layer outputs the final power prediction values ​​for the next 16 steps. (Refer to...) Figure 5As shown. Figure 5 The image shows the feature maps of the hidden layer's 3-head cross-attention scores for the trained cloud region analysis model. It uses a 43,264 km² cloud region at the Shenzhen Pinghu station center as an example, processed from 5:15 AM to 10:00 AM on May 8, 2024. The cloud region analysis model accurately perceives the spatiotemporal motion of cloud regions and deepens the model's understanding of power values ​​over historical time steps. This is significant for improving the accuracy of power prediction models based on historical data.

[0055] Meanwhile, the historical CFR (Cloud Cover Rate) gridded data of the planar cloud cover (i.e., the normalized data of remote sensing data with a center wavelength of 1.64 micrometers is defined as CFR data, with a value range of 0~1, representing the degree of cloud coverage in the region) and 5 layers were used. Historical cloud height data and solar angle grid data are input into the cloud height analysis model. The CFR is segmented by height using a mask, outputting CFR distribution data at 5 different heights. Please refer to the above. Figure 6 In this embodiment, a cloud height analysis model is constructed using the Prior_Attention + TCN structure. At any time step t, the prior attention grid-weight (2D grid weight data) is calculated using the projection positions of the intersection points of the sun and the power station line with the five height layers on the pixel grid. Then, LoRA (low-rank adaptation fine-tuning block) is inserted for fine-tuning to obtain Q*K (in the cross-attention calculation). Figure 6 The score shown is then used to perform attention calculations with the corresponding height of the CFR data, which serves as the attention mechanism's value part V. The core calculation logic is as follows: First, generate the prior grid weights and initialize a zero-based matrix project_pad[H,W]. Using the intersection of the line connecting the sun and the station at each altitude level as the center, generate a 3×3 ray projection region and fill it with nine learnable parameters learnable_patch33, each initially set to 1. Initialize the transmittance parameter rate (dimension [5, 9]) with all 1s. Extract the coordinates of the nine parameters for each layer from coord_param_height and fill them according to a specific formula, which is expressed as: ; The padded `project_pad` is mapped to Q*K. Next, cross-attention is calculated, and the score, which incorporates the `grid_LoRA`, is used as the result of the query and key calculation. As the value, a multi-head cross-attention operation with head = 5 is used to simulate the attention operation of 5 height layers. Finally, a feedforward neural network (FF) with layer normalization (LN) outputs the normalized power prediction value for the next 16 steps. By incorporating the three-dimensional structure of clouds and the geometric relationship of sunlight into the cloud height analysis model, the shading loss is quantified in a physical way, thereby predicting power conditions more accurately.

[0056] It should be noted that the prior attention score initially calculated by the cloud height analysis model is called grid-weight, and the grid-weight after incorporating LoRA is called score (i.e., Q*K).

[0057] S13, the future power prediction sequence is fused to obtain the photovoltaic power generation prediction result.

[0058] Furthermore, after the prediction results are output through multiple sub-models, that is, the corresponding future power prediction sequences are output through the historical power analysis model, numerical forecast analysis model, cloud region analysis model and cloud height analysis model respectively, the electronic device can fuse these four future power prediction sequences through a learnable Fusion fusion layer to obtain the final photovoltaic power generation prediction result.

[0059] In this embodiment, the Fusion fusion layer uses a parameter matrix of shape [4, 16], where the number 4 represents the number of sub-models and the number 16 represents the dimension of the time-series predictions. The Fusion fusion layer implements a data-driven, adaptive accuracy contribution allocation mechanism, which ultimately leads to a significant improvement in the overall prediction accuracy of the fused model.

[0060] The Fusion layer is a structure with a defined dimension. After sufficient training, it automatically adjusts weight allocation based on the characteristics of each sub-model's prediction results, thereby effectively fusing the predictions of multiple sub-models and improving the overall accuracy and reliability of the predictions. In some embodiments, the electronic device can employ a segmented freeze training method to train each sub-model and the Fusion layer. This aims to prevent premature overfitting of sub-models during training, which could devastate the training of the Fusion layer, while also ensuring the integrity of each sub-model's semantic function, thus enhancing interpretability. Specifically, in stage one, the fusion layer is first frozen, using a parameter set to 1 and non-learnable. To replace the original fusion layer, training is then performed only on the four sub-models. In Phase Two, the operation changes. The four sub-models are frozen and replaced with a trainable Fusion fusion layer. The subsequent training focuses on the contribution allocation parameters of this fusion layer, and training is performed only on it.

[0061] For details, please refer to the following: Figure 7 During training, the data usage ratio is reasonably allocated within each epoch. For example, 55% of the data is used to train the various sub-models (such as the historical power analysis model, numerical prediction analysis model, cloud region analysis model, and cloud height analysis model) for initial training. During training, data is input into each sub-model, and the parameters within the sub-model are continuously adjusted through the backpropagation algorithm, enabling the sub-models to gradually learn the features and patterns in the data, thereby reducing prediction errors. After the sub-models have completed their initial training, 15% of the data is used to train the Fusion layer, and an early stopping mechanism is implemented during this process. When training the Fusion layer, the outputs of the sub-models on this portion of the data are used as input. By adjusting the parameters of the Fusion layer, it is possible to reasonably allocate the weights of the prediction results of each sub-model, thereby achieving effective fusion. 20% of the data is dedicated to validation operations. Validation data is input into the pre-trained sub-models and the complete model composed of the Fusion layer to obtain prediction results. Then, according to a predefined loss function (such as mean squared error), the difference between the predicted results and the true values ​​is calculated, i.e., the validation loss. The validation loss reflects the model's generalization ability under the current parameters. Early stopping is determined based on the calculated validation loss. If the validation loss stops decreasing, it indicates the model may have converged in the current training phase, and continued training could lead to overfitting. In this case, the early stopping mechanism is triggered, stopping training for the current epoch. If the validation loss continues to decrease, subsequent steps continue. Finally, after meeting the early stopping condition or completing the preset training epochs, the trained model is tested using the last 10% of the data. The test data is input into the model to obtain the final prediction results, and relevant evaluation metrics (such as mean absolute error, root mean square error, etc.) are calculated to comprehensively evaluate the model's performance. Throughout the training process, once the validation loss stops decreasing, the early stopping mechanism is triggered. Simultaneously, the model that has reached its optimal state is saved, i.e., the trained model parameters are saved as a .pth file for later loading and use.

[0062] Through the above optional implementation methods, four dedicated deep learning sub-models capture key information from different dimensions to achieve accurate power prediction under complex cloud conditions; deep intelligent fusion of multi-source heterogeneous data at the feature level is achieved by leveraging mechanisms such as cross-attention and physical prior attention; the cloud height model incorporates physical formulas into the network structure, and the cloud area model outputs correction values, improving decision transparency and interpretability; a segmented freeze training strategy is adopted to effectively address the problems of non-convergence and overfitting in complex model training, ensuring stable training; and high-frequency satellite data can be used to perform excellent short-term prediction of power fluctuations from minutes to hours caused by cloud movement, achieving forward-looking early warning.

[0063] Compared to existing technologies, this application constructs a unified deep learning architecture consisting of iTransformer, FA-MLP, U-Net++, and PA-Attention, enabling end-to-end collaborative modeling and feature mining of power time-series data, future weather forecast data, cloud spatiotemporal imagery data, and cloud height physical parameter data. Based on this architecture, a hybrid deep learning structure of U-Net (spatial feature extraction), TCN (temporal evolution analysis), and cross-attention (modal feature alignment and fusion) is constructed in the cloud region analysis model. This structure is specifically designed for processing multimodal image sequences such as satellite visible light and infrared, effectively capturing the spatiotemporal dynamic evolution of clouds and establishing deep correlations between clouds and heterogeneous modalities such as meteorology and power, further enhancing the model's ability to process and analyze multimodal data. Furthermore, in the cloud height analysis model, the ray tracing physical model is embedded into an attention mechanism as learnable parameters, pioneering a new interpretable approach of "physical prior + data-driven." This cleverly combines the deterministic advantages of the physical model with the ability of deep learning to learn complex patterns from multimodal satellite remote sensing data (such as stereo imaging and infrared brightness temperature), significantly improving the theoretical consistency and practical accuracy of cloud height inversion and providing more reliable technical support for cloud height prediction. In addition, addressing the model heterogeneity problem caused by multimodal satellite remote sensing data, an innovative training strategy supporting the collaboration of complex heterogeneous models—the segmented freezing training strategy—is designed. By selectively unfreezing the parameters of different pillar networks in stages, this ensures that the entire complex system can converge stably and efficiently when fusing multimodal data sources. This effectively avoids training instability caused by differences in data sources and model structures, guaranteeing the smooth progress of model training. This application can be applied to power grid dispatching departments and distributed photovoltaic power station operators, providing them with high-accuracy short-term power prediction at low cost, effectively mitigating the impact of photovoltaic power generation uncertainty on the power grid.

[0064] Reference Figure 8The diagram shown is a functional block diagram of a photovoltaic power generation prediction device according to an embodiment of this application.

[0065] In some embodiments, the photovoltaic power generation prediction device 80 may include multiple functional modules composed of computer program segments. The computer programs for each program segment of the photovoltaic power generation prediction device 80 may be stored in the memory of an electronic device and executed by at least one processor to perform (see details). Figure 1 (Description) The function of photovoltaic power generation prediction. Based on its function, it can be divided into multiple functional modules. These functional modules may include: a data processing module 801, a separate prediction module 802, and a fusion prediction module 803. The module referred to in this application is a series of computer program segments that can be executed by at least one processor and perform a fixed function, stored in memory. In this embodiment, the functions of each module will be detailed in subsequent embodiments.

[0066] The data processing module 801 is used to acquire multi-source heterogeneous data corresponding to the prediction time point and preprocess the multi-source heterogeneous data; the multi-source heterogeneous data includes historical power generation data, NWP data, satellite remote sensing data and cloud height data.

[0067] The separate prediction module 802 is used to input preprocessed multi-source heterogeneous data into a deep learning sub-model and output the future power prediction sequence of each sub-model. The deep learning sub-model includes a historical power analysis model, a numerical forecast analysis model, a cloud region analysis model, and a cloud height analysis model. The historical power analysis model is used to predict the future power prediction sequence corresponding to the historical power generation data. The numerical forecast analysis model is used to predict the future power prediction sequence corresponding to the NWP data. The cloud region analysis model is used to predict the future power prediction sequence corresponding to the fusion of the historical power generation data and the satellite remote sensing data. The cloud height analysis model is used to predict the future power prediction sequence corresponding to the fusion of the satellite remote sensing data and the cloud height data.

[0068] The fusion prediction module 803 is used to fuse the future power prediction sequence to obtain the photovoltaic power generation prediction result.

[0069] It should be understood that the various variations and specific embodiments of the photovoltaic power generation prediction method provided in the above embodiments are also applicable to the photovoltaic power generation prediction device of this embodiment. Through the foregoing detailed description of the photovoltaic power generation prediction method, those skilled in the art can clearly understand the implementation method of the photovoltaic power generation prediction device in this embodiment. For the sake of brevity, it will not be described in detail here.

[0070] See Figure 9The diagram shown is a schematic representation of the structure of an electronic device according to an embodiment of this application. In a preferred embodiment of this application, the electronic device 9 includes a memory 91, at least one processor 92, and at least one communication bus 93.

[0071] Those skilled in the art should understand that Figure 9 The structure of the electronic device shown does not constitute a limitation of the embodiments of this application. It can be a bus structure or a star structure. The electronic device 9 may also include more or fewer other hardware or software than shown, or different component arrangements.

[0072] In some embodiments, the electronic device 9 is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions. Its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), digital processors, and embedded devices. The electronic device 9 may also include user equipment, which includes, but is not limited to, any electronic product capable of human-computer interaction with a user via a keyboard, mouse, remote control, touchpad, or voice control device, such as a personal computer, tablet computer, smartphone, or digital camera.

[0073] In the embodiments provided in this application, it should be understood that the disclosed methods, apparatuses, computer-readable storage media, and electronic devices can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple components or modules may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices, components, or modules may be electrical, mechanical, or other forms.

[0074] The components described as separate parts may or may not be physically separate. The components shown as components may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the components can be selected to achieve the purpose of this embodiment according to actual needs.

[0075] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each component can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0076] If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0077] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0078] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0079] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method for predicting photovoltaic power generation, characterized in that, The method includes: Acquire multi-source heterogeneous data corresponding to the predicted time point and preprocess the multi-source heterogeneous data; the multi-source heterogeneous data includes historical power generation data, NWP data, satellite remote sensing data, and cloud height data; the cloud height data is obtained by inversion calculation using the tbb_13 channel data of Himawari L1 Gridded data, including: selecting the highest value of the preprocessed data within the gridded area. As a ground temperature benchmark, the value of each pixel within the grid is calculated relative to the highest value. The difference According to the decreasing factor, through Calculate cloud height; where, Represents pixel coordinates The cloud height value corresponding to the grid point. It is a decreasing factor; Preprocessed multi-source heterogeneous data is input into deep learning sub-models, and each sub-model outputs a future power prediction sequence. The deep learning sub-models include a historical power analysis model, a numerical weather prediction analysis model, a cloud region analysis model, and a cloud height analysis model. The historical power analysis model is used to predict the future power prediction sequence corresponding to the historical power generation data. The numerical weather prediction analysis model is used to predict the future power prediction sequence corresponding to the NWP data. The cloud region analysis model is used to predict the future power prediction sequence corresponding to the fusion of historical power generation data and satellite remote sensing data. The cloud height analysis model is used to predict the future power prediction sequence corresponding to the fusion of satellite remote sensing data and cloud height data. Specifically, the historical power analysis model is constructed based on the iTransformer framework, the numerical weather prediction analysis model is constructed based on the FA-MLP structure, the cloud region analysis model is constructed based on U-Net, TCN structure, and cross-attention, and the cloud height analysis model is constructed based on Prior_Attention and TCN structure. The photovoltaic power generation prediction result is obtained by fusing the future power prediction sequence through a learnable Fusion fusion layer, including: adaptively fusing the future power prediction sequence output by the historical power analysis model, the numerical forecast analysis model, the cloud region analysis model, and the cloud height analysis model through the Fusion fusion layer; the Fusion fusion layer adopts a parameter matrix of shape [4, 16], where the number 4 represents the number of sub-models in the deep learning sub-model, and the number 16 represents the dimension of the time series prediction value; the Fusion fusion layer automatically adjusts the weight allocation after training and realizes adaptive accuracy contribution allocation based on data-driven approach to fuse the future power prediction sequence output by the deep learning sub-model.

2. The photovoltaic power generation prediction method according to claim 1, characterized in that, The method further includes: Freeze the Fusion fusion layer, using parameters all of 1 and unlearnable. Instead of the original fusion layer, only the historical power analysis model, the numerical forecast analysis model, the cloud region analysis model, and the cloud height analysis model are trained; The historical power analysis model, the numerical forecast analysis model, the cloud region analysis model, and the cloud height analysis model are frozen and replaced with a trainable Fusion fusion layer. The weight allocation parameters of the Fusion fusion layer are then trained.

3. The photovoltaic power generation prediction method according to claim 1, characterized in that, The preprocessing of the multi-source heterogeneous data includes: The instantaneous power generation value collected at time t is used as the historical power generation data, and the historical power generation data is normalized using the following formula: ； in, This represents the normalized historical power generation data. This represents the instantaneous power generation value collected at time t. This represents the inverter's on-time capacity at time t; The satellite remote sensing data is normalized according to the channel type using the following formula: ； in, This represents normalized satellite remote sensing data. Represents the pixel coordinates of the c channel. grayscale value, This indicates the upper limit of the effective grayscale data in channel c. This indicates the lower limit of the valid grayscale data in channel c; The NWP data is normalized using the following formula: ； in, This represents the normalized NWP data. This represents the value of the meteorological element corresponding to the NWP data generated by the large meteorological model. This represents the reasonable maximum value of the corresponding meteorological element in the NWP data. This represents the reasonable minimum value of the corresponding meteorological element in the NWP data.

4. The photovoltaic power generation prediction method according to claim 3, characterized in that, The preprocessing of the multi-source heterogeneous data further includes: The normalized historical power generation data is refined from a time interval of 1 step / 15 minutes to 1 step / 5 minutes through interpolation. When the time interval of the satellite remote sensing data is determined to be 15 minutes, the normalized satellite remote sensing data is refined from a time interval of 1 step / 15 minutes to 1 step / 5 minutes through interpolation. The normalized NWP data is refined from a time interval of 1 step / 1 hour to 1 step / 5 minutes through interpolation.

5. A photovoltaic power generation prediction device, characterized in that, The device includes: The data processing module is used to acquire multi-source heterogeneous data corresponding to the prediction time point and preprocess the multi-source heterogeneous data; the multi-source heterogeneous data includes historical power generation data, NWP data, satellite remote sensing data, and cloud height data; the cloud height data is obtained by inversion calculation using the tbb_13 channel data of Himawari L1 Gridded data, including: selecting the highest value of the preprocessed data within the gridded area. As a ground temperature benchmark, the value of each pixel within the grid is calculated relative to the highest value. The difference According to the decreasing factor, through Calculate cloud height; where, Represents pixel coordinates The cloud height value corresponding to the grid point. It is a decreasing factor; A separate prediction module is used to input preprocessed multi-source heterogeneous data into deep learning sub-models and output future power prediction sequences for each sub-model. The deep learning sub-models include a historical power analysis model, a numerical weather prediction analysis model, a cloud region analysis model, and a cloud height analysis model. The historical power analysis model is used to predict the future power prediction sequence corresponding to the historical power generation data. The numerical weather prediction analysis model is used to predict the future power prediction sequence corresponding to the NWP data. The cloud region analysis model is used to predict the future power prediction sequence corresponding to the fusion of historical power generation data and satellite remote sensing data. The cloud height analysis model is used to predict the future power prediction sequence corresponding to the fusion of satellite remote sensing data and cloud height data. Specifically, the historical power analysis model is constructed based on the iTransformer framework, the numerical weather prediction analysis model is constructed based on the FA-MLP structure, the cloud region analysis model is constructed based on U-Net, TCN structure, and cross-attention, and the cloud height analysis model is constructed based on Prior_Attention and TCN structure. The fusion prediction module is used to fuse the future power prediction sequence through a learnable Fusion fusion layer to obtain the photovoltaic power prediction result. This includes: adaptively fusing the future power prediction sequences output by the historical power analysis model, the numerical forecast analysis model, the cloud region analysis model, and the cloud height analysis model through the Fusion fusion layer; the Fusion fusion layer uses a parameter matrix of shape [4, 16], where the number 4 represents the number of sub-models in the deep learning sub-model, and the number 16 represents the dimension of the time-series prediction value; the Fusion fusion layer automatically adjusts the weight allocation after training, and achieves adaptive accuracy contribution allocation based on data-driven methods to fuse the future power prediction sequence output by the deep learning sub-models.

6. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the photovoltaic power generation prediction method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the photovoltaic power generation prediction method according to any one of claims 1 to 4.

Images