Multi-source data driven ultra-short-term photovoltaic power multi-step prediction method and system

Abstract

The invention provides a multi-source data-driven ultra-short-term photovoltaic power multi-step prediction method and system, and the system comprises four core modules: an SDFM cloud picture scale decoupling module carries out the multi-scale feature extraction of a satellite cloud picture, and precisely captures the feature information of different cloud layers; the DGAM feature fusion module adaptively adjusts importance weights of satellite cloud pictures, terrains and meteorological data through a dynamic time feature gating mechanism, solves the problem that multi-source data features dynamically change along with time in combination with a transient attention network and a cross feature effect, and performs feature fusion; the DSCRM residual module adopts a lightweight residual structure, performs adaptive weighting on a feature channel, and efficiently extracts a deep time sequence relation; and the CATFM prediction module maps the multi-source time sequence feature map through a space-time double-weight matrix, so that ultra-short-term photovoltaic power multi-step prediction is completed, and reliable technical support is provided for dispatching operation of a power system.

Description

Technical Field

[0001] This invention belongs to the field of photovoltaic power prediction technology, and specifically designs a multi-source data-driven method and system for ultra-short-term photovoltaic power multi-step prediction. Background Technology

[0002] Photovoltaic power generation, as a clean and renewable energy source, is highly dependent on natural environmental conditions. It is easily affected by a combination of factors, including solar radiation intensity, cloud cover, ambient temperature, and atmospheric conditions, resulting in significant randomness and fluctuations in output power. This instability poses a considerable challenge to the safe operation and dispatch control of power systems. Therefore, conducting high-precision photovoltaic power generation forecasting is of great significance for improving the safety, reliability, and overall economic efficiency of photovoltaic power generation systems.

[0003] With the development of remote sensing technology and data acquisition methods, there is a growing number of photovoltaic power prediction methods that integrate satellite cloud images, meteorological data, topographic information, and historical photovoltaic power data. Multi-source data can reflect the environmental change characteristics during photovoltaic power generation from different perspectives, and to a certain extent, improve the ability of prediction models to represent the fluctuation patterns of photovoltaic power. However, due to significant differences in time scale, spatial resolution, and data structure among different data sources, existing multi-source photovoltaic power prediction methods still have shortcomings in collaborative modeling. On the one hand, most methods directly use feature splicing or fixed-weight fusion, which makes it difficult to fully capture the dynamic relationships between multi-source data, easily leading to the weakening of key information or redundant features interfering with model learning. On the other hand, existing cloud image feature extraction methods are relatively simple, with limited ability to capture the multi-scale spatial structure and evolution characteristics of clouds, making it difficult to simultaneously reflect local cloud cluster changes and large-scale cloud system movement characteristics.

[0004] In ultra-short-term photovoltaic (PV) power multi-step prediction scenarios, the nonlinearity and temporal dependence of PV power time series make prediction errors prone to accumulation as the prediction step size increases. Existing methods have certain limitations in long-term dependency modeling and dynamic updating of temporal features, and prediction accuracy tends to decrease with longer prediction step sizes. Meanwhile, some high-precision models have complex structures and high computational costs, while simplified models suffer from insufficient feature representation capabilities, making it difficult to achieve an effective balance between prediction accuracy and computational efficiency. Summary of the Invention

[0005] Purpose of the invention: In order to solve the problems existing in the prior art, the present invention provides a multi-source data-driven method and system for multi-step prediction of ultra-short-term photovoltaic power.

[0006] Technical solution: This invention provides a multi-source data-driven multi-step prediction method for ultra-short-term photovoltaic power, specifically including the following steps: S1 acquires satellite cloud images, terrain data, meteorological data, and historical photovoltaic power data; S2, perform outlier screening and missing value interpolation preprocessing on the terrain data, meteorological data and historical photovoltaic power data obtained in step S1; S3, Perform radiometric calibration, geometric correction, and reprojection operations on the satellite cloud image obtained in step S1; S4. The satellite cloud image processed in step S3 is subjected to feature extraction through the SDFM cloud image scale decoupling module to obtain multi-scale cloud image features. S5, use the DGAM feature fusion module to fuse the terrain data and meteorological data preprocessed in step S2 with the multi-scale cloud map features extracted in step S4 to obtain multi-source fused data; S6. Input the multi-source fusion data obtained in step S5 into the DSCRM residual module to extract time series information, obtain the multi-source time series feature map, and then perform multi-step prediction of ultra-short-term photovoltaic power through the CATFM prediction module.

[0007] Furthermore, the outlier screening and missing value interpolation preprocessing operation performed on the terrain data, meteorological data, and historical photovoltaic power data obtained in step S1, as described in step S2, includes the following steps: S21. Check if there are missing values ​​in the original terrain data, and use local neighborhood analysis to detect elevation anomalies. The formula for the local elevation standard deviation is as follows: in, Represents a cell Local standard deviation of elevation at the location, and These represent the row index and column index in the terrain raster, respectively. and This represents the neighborhood offset, with a value range of {-1, 0, 1}. Indicates the location of the line number , column number The original elevation value at that location, For pixels The arithmetic mean of the elevations of all valid pixels within a 3×3 neighborhood at a given location, where N represents the number of valid elevation pixels within the neighborhood; S22, reproject the terrain data to the target coordinate system (WGS84) and resample to 1 km resolution using bi-sex interpolation; S23, calculate the slope of the core terrain factor. The slope gradient is calculated using the third-order inverse distance squared weighted difference method to reduce the influence of noise. The formula is as follows: in, and These represent the rates of change of elevation in the horizontal and vertical directions, respectively. and Spatial resolution of the data in the horizontal and vertical directions, respectively. and These represent the row index and column index in the terrain raster, respectively. This represents the slope value. S24 calculates the core terrain factor slope aspect and returns the azimuth angle from the origin to the target point using the four-quadrant arctangent function, as shown in the following formula: in, The slope angle, It is the arctangent function in the four quadrants; S25 extracts multi-element time series data of the target area and surrounding grid points from the meteorological dataset, including temperature, relative humidity, air pressure, and 10-meter wind speed. Spatial and temporal matching is then performed using bilinear interpolation.

[0008] S26. Standardize and detect anomalies in historical photovoltaic power data to identify anomalies caused by equipment failure or data errors, and interpolate missing values. The formula for standardized residual anomaly detection is as follows:

[0009] in, For the true value, This is the theoretical value. For the installed capacity of the power station, For standardized residuals.

[0010] Furthermore, the operations performed in step S3 on the satellite cloud image obtained in step S1, such as radiometric calibration, geometric correction, and reprojection, include the following steps: S31, Perform radiometric calibration on the satellite cloud image obtained in step S1, calibrate the reflectance of the visible and near-infrared channels, and convert the observed count values ​​into surface reflectance using the following formula: in, Indicates reflectivity, This is the scaling gain coefficient. This is the original count value. This is the calibration offset; S32. Brightness temperature calibration is performed on the infrared channel in the cloud image. First, the original count values ​​are converted into the radiation altitude received by the satellite altitude sensor. Then, the brightness temperature is calculated using the inverse function of Planck's law, as shown in the following formula: in, Indicates spectral radiance. This is the scaling gain coefficient. This is the original count value. This is the calibration offset. Indicates brightness temperature. and Let represent the first radiation constant and the second radiation constant, respectively. Indicates the center wavelength of the channel; S33, Perform geometric correction on the cloud image data processed in S32, converting the original image coordinates into geographic coordinates; including the following sub-steps: S331, based on satellite attitude, orbital parameters and scanning geometric model, constructs a geometric model function from image coordinates to geographic latitude and longitude; S332, using the digital elevation data processed in step S22, performs fine correction on the geometric correction results of step S331, employing a quadratic polynomial for local fitting to minimize the positioning error. The formula is as follows:

[0011] in, and This indicates the row and column number of a pixel in the raw satellite image data. and Indicates longitude and latitude. , , , , and These are the coefficients of the longitude mapping polynomial. , , , , and These are the coefficients of the latitude mapping polynomial.

[0012] S33 resamples the data to a 1-kilometer resolution consistent with the grid of the photovoltaic power station area, maintaining spatial scale consistency.

[0013] Furthermore, step S4 involves extracting features from the satellite cloud image processed in step S3 using the SDFM cloud image scale decoupling module. Multi-scale cloud image features are obtained through three paths: fine-scale path, meso-scale path, and large-scale path. This includes the following steps: S41 captures local micro-scale cloud structure features in satellite cloud images through geometrically adaptive variable convolution in fine-scale paths. The formula for geometrically adaptive variable convolution is as follows: in, To output feature map at location The value, This indicates the number of sampling points in the convolution kernel. Indicates the first n Convolution weight parameters for each sampling point This represents the fixed offset of the standard convolution kernel. This represents the learnable spatial offset, calculated by the offset prediction network. This represents a learnable modulation scalar that controls the importance of each sampling point. Indicates the input feature value; S42 extracts cloud organization features, such as cloud clusters, convective complexes, and short-duration rainbands, from the mesoscale path through directional separable convolutional kernels. This process includes the following sub-steps: S421, apply a bilinear interpolation rotation operation to the basic convolution kernel to generate a set of directional convolution kernels; S422 captures feature patterns from multiple directions within a set of directional convolutional kernels to reduce feature confusion caused by directional changes. The convolutional output is calculated for each direction, as shown in the following formula: in, Indicates the first c The first channel, the... i Okay, number j The output feature tensor of the column, Indicates the rotation of the convolution kernel. For output channel index, For input channel index, and These represent the row offset and column offset of the convolution kernel, respectively. Indicates the number of input channels. This represents the input feature map tensor. This indicates point-by-point operation; S423, calculate the feature vector for each direction, using the following formula: in, For directional feature map vectors, and The height and width of the feature map, and Indicates the spatial row index and column index. Indicates all channels; S424, calculates the directional attention score and attention weight, using the following formula: in, For attention score, This is the offset. For direction value, These are the weighting coefficients. For temperature parameters, For the summation index, For the direction set; S425 sums all directional features according to attention weights; S43 extracts global features of the entire cloud map background through dilated convolution of large-scale paths.

[0014] Furthermore, step S5 involves using the DGAM module to fuse the preprocessed terrain data and meteorological data from step S2 with the multi-scale cloud image features extracted in step S4 to obtain multi-source fused data, including the following steps: S51, for the extracted multi-scale cloud image features, meteorological data, and topographic data, a dynamic gating network is designed. Using time features as a condition, a weight vector with dimensions matching the original features is generated. The gating weight calculation formula is as follows: in, , and These are the gating weights of cloud image features, terrain data, and meteorological data under the current time conditions. For cloud map features, For terrain features, For meteorological characteristics, For time-encoded features, This represents a vector concatenation operation. Here is the weight matrix of the gated network. For bias terms, The Sigmoid activation function is used to constrain the weight values ​​to the interval [0,1]. To modify the activation function of the linear unit, used to filter negative values; S52 performs feature weighting using weight vectors and then uses a linear projection layer to map them to a unified feature space, achieving dimension alignment. The formula is as follows: in, Indicates the feature dimension of the cloud map. Represents the dimensions of terrain features. Representing the dimensions of meteorological characteristics, These are the weighting coefficients. For element-wise multiplication, Indicates the offset; S53, concatenate all original features to form a global feature representation, and then obtain the cross-feature effect through linear transformation: in, For splicing features, This is a cross-characteristic effect. These are the coefficients of the weight matrix; S54 utilizes the cross-gating effect to adjust the information weights between fused features, enabling the DGAM module to dynamically adjust its internal cross-relationships over time; the formula is as follows: in, Represents the weights of cross features. For splicing features, It is a cross feature; S55, after obtaining the multi-scale cloud map features, the gating features of meteorological data and topographic data, and the cross features, constructs a unified feature representation through the feature fusion layer; S56. A transient attention mechanism is designed to capture the dynamic dependencies within the input sequence, providing a more refined temporal feature representation for dynamic gating, resulting in the final fused features, as shown in the following formula: in, Let the first term represent the final attention matrix. For content relevance, the second item For relative position offset, Representing feature dimension, Indicates the value.

[0015] Further, step S6 involves inputting the multi-source fusion data obtained in step S5 into the DSCRM residual module to extract time-series information, obtaining a multi-source time-series feature map, and then performing multi-step prediction of ultra-short-term photovoltaic power through the CATFM prediction module, including the following steps: S61, after the fused features processed in step S36 are compressed through layer normalization and pointwise convolution, the channel dimension is reduced by the ReLU activation function to obtain the dimensionality-reduced feature sequence, as shown in the following formula: in, Represents the dimensionality reduction feature sequence. For the input tensor of the residual sub-block, For pointwise convolution kernels, The pointwise convolution bias vector. Representation layer normalization; S62 converts the dimensionality-reduced feature sequence into a convolutional computation format. And input it into a sequence consisting of 3 DSCRM residual modules; S63, in each residual block, firstly, channel dimension compression is performed, then depthwise convolution is used to model the temporal local receptive field, and finally, the channels are restored to their original dimensions to obtain a multi-source temporal feature map. The residual structure formula within the block is as follows: Where Y is the residual output. For residual input, Indicates regularization; S64, Design the CATFM prediction module, which contains a temporal and spatial dual-branch structure to perform spatiotemporal processing on multi-source temporal feature maps respectively; S65 uses the time branch of the CATFM prediction module to perform global average pooling on the time dimension of the multi-source temporal feature map to compress temporal information. First, a two-layer fully connected network is used to learn the nonlinear importance of each channel, and then the time weight vector is output through the Sigmoid function. S66 applies causal-filled deep convolution to the multi-source temporal feature map through the spatial branch of the CATFM prediction module, then performs global average pooling in the channel dimension to preserve temporal dynamics, and finally normalizes it along the time dimension using Softmax to obtain the spatial weight vector. S67: First, perform an outer product operation between the spatial weight vector and the temporal weight vector to generate a spatiotemporal dual weight matrix. Then, multiply the spatiotemporal dual weight matrix element-wise with the multi-source temporal feature map to achieve adaptive control. Adjust the corresponding spatiotemporal weights according to different spatiotemporal features to obtain weighted features. S68 first performs pointwise convolution on the weighted features to integrate channel information, and then flattens them and directly maps them to multi-step prediction results with 15-minute intervals within the next hour through a fully connected layer.

[0016] A multi-source data-driven multi-step prediction system for ultra-short-term photovoltaic power, the prediction system comprising: The SDFM cloud image scale decoupling module is used to extract features from satellite cloud images that have undergone radiometric calibration, geometric correction, and reprojection processing. It obtains multi-scale cloud image features through three paths: fine-scale path, meso-scale path, and large-scale path. The DGAM feature fusion module is used to fuse preprocessed terrain data, meteorological data and multi-scale cloud map features. It designs a dynamic feature gating mechanism that can adaptively adjust the weights of multi-source data according to time conditions. It captures the local temporal dependence of photovoltaic power through a transient attention mechanism and introduces a cross-gating effect to capture the nonlinear interaction effect between different data sources. The DSCRM residual module is used to extract temporal information from multi-source fused data. It compresses the channel dimension through multiple residual blocks and then uses deep convolution to model the temporal local receptive field to obtain multi-source temporal feature maps. The CATFM prediction module processes multi-source time-series feature maps through a spatiotemporal dual-branch structure, mapping the extracted time-series features to power, thereby achieving multi-step prediction results at 15-minute intervals within the next hour.

[0017] Beneficial effects: First, the multi-source data-driven multi-step prediction method for ultra-short-term photovoltaic power proposed in this invention, after the SDFM cloud map scale decoupling module performs refined feature extraction from the cloud map, achieves multi-source heterogeneous data fusion through the dynamic gating mechanism and transient attention network in the DGAM feature fusion module. The DSCRM residual module is designed to extract deep-level temporal dependency information while reducing complexity, and finally, high-precision multi-step prediction is achieved through the CATFM prediction module.

[0018] Secondly, the DGAM feature fusion module proposed in this invention addresses the dynamic changes of multi-source heterogeneous data across spatiotemporal scales by adaptively adjusting the weights of satellite, terrain, and meteorological data based on temporal features. This module can optimize data fusion according to real-time meteorological conditions and time factors, significantly improving the feature fusion performance of the photovoltaic power prediction model. By introducing a transient attention network with relative position encoding, it can accurately capture local temporal dependencies, further enhancing the accuracy and robustness of multi-source data fusion.

[0019] Third, the DSCRM residual module and CATFM prediction module of this invention adopt a lightweight residual block structure, which effectively reduces the computational complexity and number of parameters of the model, ensuring efficient real-time prediction capabilities. By processing multi-source time-series feature maps through a spatiotemporal dual-branch structure, the extracted time-series features are mapped to power, achieving efficient multi-step prediction. Compared with traditional methods, the prediction model can provide more accurate ultra-short-term photovoltaic power predictions in a shorter time, providing more efficient technical support for power system scheduling and operation. Attached Figure Description

[0020] Figure 1 This is a flowchart of the multi-source data-driven multi-step prediction method for ultra-short-term photovoltaic power according to the present invention; Figure 2 This is a diagram showing the overall structure of the multi-source data-driven ultra-short-term photovoltaic power multi-step prediction model of the present invention. Figure 3 This is a diagram of the SDFM cloud map scale decoupling module of the present invention; Figure 4 This is a diagram of the DGAM feature fusion module of the present invention; Figure 5 This is a diagram of the DSCRM residual module of the present invention; Figure 6 This is a diagram of the CATFM prediction module of the present invention; Figure 7 and Figure 8 This is a diagram showing the results of the multi-source data-driven multi-step prediction of ultra-short-term photovoltaic power in this invention. Figure 9 This is a comparison chart of the prediction results of the present invention; Figure 10 This is a convergence experiment diagram of the present invention. Detailed Implementation

[0021] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0022] like Figure 1 As shown, this invention proposes a multi-source data-driven multi-step prediction method for ultra-short-term photovoltaic power; the prediction method includes the following steps: S1 acquires satellite cloud images, terrain data, meteorological data, and historical photovoltaic power data; S2, perform preprocessing operations such as outlier screening and missing value interpolation on the terrain data, meteorological data and historical photovoltaic power data obtained in step S1; S3, Perform radiometric calibration, geometric correction, reprojection, and other operations on the satellite cloud image obtained in step S1; S4. The satellite cloud image processed in step S3 is subjected to feature extraction through the SDFM cloud image scale decoupling module to obtain multi-scale cloud image features. S5, use the DGAM feature fusion module to fuse the terrain data and meteorological data preprocessed in step S2 with the multi-scale cloud map features extracted in step S4 to obtain multi-source fused data; S6. Input the multi-source fusion data obtained in step S5 into the DSCRM residual module to extract time series information, obtain the multi-source time series feature map, and then perform multi-step prediction of ultra-short-term photovoltaic power through the CATFM prediction module.

[0023] Figure 2 The following is a diagram illustrating the overall structure of a multi-source data-driven, ultra-short-term photovoltaic power multi-step prediction model. Figure 2 The prediction method of the present invention will be described in detail. Specifically, it includes the following steps:

[0024] S1 acquires satellite cloud images, terrain data, meteorological data, and historical photovoltaic power data.

[0025] S2 involves preprocessing the terrain data, meteorological data, and historical photovoltaic power data obtained in step S1, including outlier filtering and missing value interpolation. Specifically, this includes: S21. Check if there are missing values ​​in the original terrain data, and use local neighborhood analysis to detect elevation anomalies. The formula for the local elevation standard deviation is as follows: in, Represents a cell Local standard deviation of elevation at the location, and These represent the row index and column index in the terrain raster, respectively. and This represents the neighborhood offset, with a value range of {-1, 0, 1}. Indicates the location of the line number , column number The original elevation value at that location, For pixels The arithmetic mean of the elevations of all valid pixels within a 3×3 neighborhood at a given location, where N represents the number of valid elevation pixels within the neighborhood; S22, reproject the terrain data to the target coordinate system (WGS84) and resample to 1 km resolution using bi-sex interpolation; S23, calculate the slope of the core terrain factor. The slope gradient is calculated using the third-order inverse distance squared weighted difference method to reduce the influence of noise. The formula is as follows: in, and These represent the rates of change of elevation in the horizontal and vertical directions, respectively. and Spatial resolution of the data in the horizontal and vertical directions, respectively. and These represent the row index and column index in the terrain raster, respectively. This represents the slope value. S24 calculates the core terrain factor slope aspect and returns the azimuth angle from the origin to the target point using the four-quadrant arctangent function, as shown in the following formula: in, The slope angle, It is the arctangent function in the four quadrants; S25 extracts multi-element time series data of the target area and surrounding grid points from the meteorological dataset, including temperature, relative humidity, air pressure, and 10-meter wind speed. Spatial and temporal matching is then performed using bilinear interpolation.

[0026] S26. Standardize and detect anomalies in historical photovoltaic power data to identify anomalies caused by equipment failure or data errors, and interpolate missing values. The formula for standardized residual anomaly detection is as follows:

[0027] in, For the true value, This is the theoretical value. For the installed capacity of the power station, For standardized residuals.

[0028] S3, Perform radiometric calibration, geometric correction, and reprojection on the satellite cloud image obtained in step S1, including the following steps: S31, Perform radiometric calibration on the satellite cloud image obtained in step S1, calibrate the reflectance of the visible and near-infrared channels, and convert the observed count values ​​into surface reflectance using the following formula: in, Indicates reflectivity, This is the scaling gain coefficient. This is the original count value. This is the calibration offset; S32. Brightness temperature calibration is performed on the infrared channel in the cloud image. First, the original count values ​​are converted into the radiation altitude received by the satellite altitude sensor. Then, the brightness temperature is calculated using the inverse function of Planck's law, as shown in the following formula: in, Indicates spectral radiance. This is the scaling gain coefficient. This is the original count value. This is the calibration offset. Indicates brightness temperature. and Let represent the first radiation constant and the second radiation constant, respectively. Indicates the center wavelength of the channel; S33, Perform geometric correction on the cloud image data processed in S32, converting the original image coordinates into geographic coordinates; including the following sub-steps: S331, based on satellite attitude, orbital parameters and scanning geometric model, constructs a geometric model function from image coordinates to geographic latitude and longitude; S332, using the digital elevation data processed in step S22, performs fine correction on the geometric correction results of step S331, employing a quadratic polynomial for local fitting to minimize the positioning error. The formula is as follows:

[0029] in, and This indicates the row and column number of a pixel in the raw satellite image data. and Indicates longitude and latitude. , , , , and These are the coefficients of the longitude mapping polynomial. , , , , and These are the coefficients of the latitude mapping polynomial.

[0030] S33 resamples the data to a 1-kilometer resolution consistent with the grid of the photovoltaic power station area, maintaining spatial scale consistency.

[0031] S4, feature extraction is performed on the satellite cloud image processed in step S3 using the SDFM cloud image scale decoupling module. Multi-scale cloud image features are obtained through three paths: fine-scale path, medium-scale path, and large-scale path. The SDFM cloud image scale decoupling module diagram is shown below. Figure 3 As shown, it includes the following steps: S41 captures local micro-scale cloud structure features in satellite cloud images through geometrically adaptive variable convolution in fine-scale paths. The formula for geometrically adaptive variable convolution is as follows: in, To output feature map at location The value, This indicates the number of sampling points in the convolution kernel. Indicates the first n Convolution weight parameters for each sampling point This represents the fixed offset of the standard convolution kernel. This represents the learnable spatial offset, calculated by the offset prediction network. This represents a learnable modulation scalar that controls the importance of each sampling point. Indicates the input feature value; S42 extracts cloud organization features, such as cloud clusters, convective complexes, and short-duration rainbands, from the mesoscale path through directional separable convolutional kernels. This process includes the following sub-steps: S421, apply a bilinear interpolation rotation operation to the basic convolution kernel to generate a set of directional convolution kernels; S422 captures feature patterns from multiple directions within a set of directional convolutional kernels to reduce feature confusion caused by directional changes. The convolutional output is calculated for each direction, as shown in the following formula: in, Indicates the first c The first channel, the... i Okay, number j The output feature tensor of the column, Indicates the rotation of the convolution kernel. For output channel index, For input channel index, and These represent the row offset and column offset of the convolution kernel, respectively. Indicates the number of input channels. This represents the input feature map tensor. This indicates point-by-point operation; S423, calculate the feature vector for each direction, using the following formula: in, For directional feature map vectors, and The height and width of the feature map, and Indicates the spatial row index and column index. Indicates all channels; S424, calculates the directional attention score and attention weight, using the following formula: in, For attention score, This is the offset. For direction value, These are the weighting coefficients. For temperature parameters, For the summation index, For the direction set; S425 sums all directional features according to attention weights; S43 extracts global features of the entire cloud map background through dilated convolution of large-scale paths.

[0032] S5, using the DGAM module, the preprocessed terrain data and meteorological data from step S2 are fused with the multi-scale cloud image features extracted in step S4 to obtain multi-source fused data. The DGAM feature fusion module diagram is shown below. Figure 4 As shown, it includes the following steps: S51, for the extracted multi-scale cloud image features, meteorological data, and topographic data, a dynamic gating network is designed. Using time features as a condition, a weight vector with dimensions matching the original features is generated. The gating weight calculation formula is as follows: in, , and These are the gating weights of cloud image features, terrain data, and meteorological data under the current time conditions. For cloud map features, For terrain features, For meteorological characteristics, For time-encoded features, This represents a vector concatenation operation. Here is the weight matrix of the gated network. For bias terms, The Sigmoid activation function is used to constrain the weight values ​​to the interval [0,1]. To modify the activation function of the linear unit, used to filter negative values; S52 performs feature weighting using weight vectors and then uses a linear projection layer to map them to a unified feature space, achieving dimension alignment. The formula is as follows: in, Indicates the feature dimension of the cloud map. Represents the dimensions of terrain features. Representing the dimensions of meteorological characteristics, These are the weighting coefficients. For element-wise multiplication, Indicates the offset; S53, concatenate all original features to form a global feature representation, and then obtain the cross-feature effect through linear transformation: in, For splicing features, This is a cross-characteristic effect. These are the coefficients of the weight matrix; S54 utilizes the cross-gating effect to adjust the information weights between fused features, enabling the DGAM module to dynamically adjust its internal cross-relationships over time; the formula is as follows: in, Represents the weights of cross features. For splicing features, It is a cross feature; S55, after obtaining the multi-scale cloud map features, the gating features of meteorological data and topographic data, and the cross features, constructs a unified feature representation through the feature fusion layer; S56. A transient attention mechanism is designed to capture the dynamic dependencies within the input sequence, providing a more refined temporal feature representation for dynamic gating, resulting in the final fused features, as shown in the following formula: in, Let the first term represent the final attention matrix. For content relevance, the second item For relative position offset, Representing feature dimension, Indicates the value.

[0033] S6. Input the multi-source fusion data obtained in step S5 into the DSCRM residual module to extract time series information and obtain a multi-source time series feature map. Then, perform multi-step prediction of ultra-short-term photovoltaic power through the CATFM prediction module. The DSCRM residual module diagram is shown below. Figure 5 As shown in the diagram, the CATFM prediction module is as follows: Figure 6 As shown, it includes the following steps: S61, after the fused features processed in step S36 are compressed through layer normalization and pointwise convolution, the channel dimension is reduced by the ReLU activation function to obtain the dimensionality-reduced feature sequence, as shown in the following formula: in, Represents the dimensionality reduction feature sequence. For the input tensor of the residual sub-block, For pointwise convolution kernels, The pointwise convolution bias vector. Representation layer normalization; S62 converts the dimensionality-reduced feature sequence into a convolutional computation format. And input it into a sequence consisting of 3 DSCRM residual modules; S63, in each residual block, firstly, channel dimension compression is performed, then depthwise convolution is used to model the temporal local receptive field, and finally, the channels are restored to their original dimensions to obtain a multi-source temporal feature map. The residual structure formula within the block is as follows: Where Y is the residual output. For residual input, Indicates regularization; S64, Design the CATFM prediction module, which contains a temporal and spatial dual-branch structure to perform spatiotemporal processing on multi-source temporal feature maps respectively; S65 uses the time branch of the CATFM prediction module to perform global average pooling on the time dimension of the multi-source temporal feature map to compress temporal information. First, a two-layer fully connected network is used to learn the nonlinear importance of each channel, and then the time weight vector is output through the Sigmoid function. S66 applies causal-filled deep convolution to the multi-source temporal feature map through the spatial branch of the CATFM prediction module, then performs global average pooling in the channel dimension to preserve temporal dynamics, and finally normalizes it along the time dimension using Softmax to obtain the spatial weight vector. S67: First, perform an outer product operation between the spatial weight vector and the temporal weight vector to generate a spatiotemporal dual weight matrix. Then, multiply the spatiotemporal dual weight matrix element-wise with the multi-source temporal feature map to achieve adaptive control. Adjust the corresponding spatiotemporal weights according to different spatiotemporal features to obtain weighted features. S68 first performs pointwise convolution on the weighted features to integrate channel information, and then flattens them and directly maps them to multi-step prediction results with 15-minute intervals within the next hour through a fully connected layer.

[0034] A multi-source data-driven multi-step prediction system for ultra-short-term photovoltaic power, the prediction system comprising: The SDFM cloud image scale decoupling module is used to extract features from satellite cloud images that have undergone radiometric calibration, geometric correction, and reprojection processing. It obtains multi-scale cloud image features through three paths: fine-scale path, meso-scale path, and large-scale path. The DGAM feature fusion module is used to fuse preprocessed terrain data, meteorological data and multi-scale cloud map features. It designs a dynamic feature gating mechanism that can adaptively adjust the weights of multi-source data according to time conditions. It captures the local temporal dependence of photovoltaic power through a transient attention mechanism and introduces a cross-gating effect to capture the nonlinear interaction effect between different data sources. The DSCRM residual module is used to extract temporal information from multi-source fused data. It compresses the channel dimension through multiple residual blocks and then uses deep convolution to model the temporal local receptive field to obtain multi-source temporal feature maps. The CATFM prediction module processes multi-source time-series feature maps through a spatiotemporal dual-branch structure, mapping the extracted time-series features to power, thereby achieving multi-step prediction results at 15-minute intervals within the next hour.

[0035] Table 1 shows the average error results of different methods for multi-step prediction of ultra-short-term photovoltaic power driven by multi-source data. The bolded data are the optimal values. Figure 9 This paper presents a comparison of prediction results for different methods of multi-step prediction of ultra-short-term photovoltaic power driven by multi-source data. The present invention selected the latter 20% of the dataset for experiments with different models. Table 1 shows that the method of the present invention outperforms other methods, exhibiting the lowest MSE error (0.095 lower than the BiTransformer model, 0.122 lower than the ConvLSTM model, and 0.06 lower than the Vision Mamba model). The prediction error of the second-best BiMamba model is not significantly different from that of the present invention. Experimental results demonstrate that the present method achieves higher accuracy than other models in the multi-step prediction task of ultra-short-term photovoltaic power driven by multi-source data, exhibits the smallest decrease in accuracy with increasing prediction time, and demonstrates the most stable performance.

[0036] Table 1 Table 2 shows the ablation experiments; "×" indicates that the module was added, and "×" indicates that the module was not added. This invention constructs three comparative models: M1: no modules added, prediction is performed using fully connected layers; M2: prediction is performed using only the SDFM cloud map scale decoupling module; M3: prediction is performed using the SDFM cloud map scale decoupling module and the DGAM feature fusion module; M4: prediction is performed using the SDFM cloud map scale decoupling module, the DGAM feature fusion module, and the DSCRM residual module. Compared to the above four configurations, the method of this invention includes all modules, serving as the complete version.

[0037] Table 2 No single module can achieve optimal performance independently, and may even lead to a decrease in accuracy due to the limited number of features or increased model complexity. Only through the complete chain of SDFM, DGAM, DSCRM, and CATFM can a significant improvement in accuracy be achieved. Compared with the baseline model M1, the complete model reduces MSE by 32.1%, MAE by 26.2%, and RMSE by 20.7%, fully demonstrating the technical advancement and engineering practicality of the method of this invention.

[0038] Figure 7 and Figure 8 The prediction results of the multi-source data-driven ultra-short-term photovoltaic power multi-step prediction method are shown in the figure. Table 3 shows the error loss at different time lengths. Figure 7 It involves randomly selecting 100 samples for multi-step prediction. Figure 8 The method performs time series prediction on the remaining 20% ​​of the data. Experimental results show that the prediction accuracy gradually decreases as the step size increases, but overall, the prediction results of the method of this invention can achieve high prediction accuracy at all four time steps.

[0039] Table 3 Figure 10 The convergence performance of the model of this invention is demonstrated. The model exhibits good convergence during the training process. After about 50 epochs of training, the loss of the model on both the training set and the validation set has decreased to a low level and tends to stabilize, indicating that the model has fully learned the effective features in the data and has good generalization ability.

Claims

1. A multi-source data-driven multi-step prediction method for ultra-short-term photovoltaic power, characterized in that, Specifically, the steps include the following: S1 acquires satellite cloud images, terrain data, meteorological data, and historical photovoltaic power data; S2, perform outlier screening and missing value interpolation preprocessing on the terrain data, meteorological data and historical photovoltaic power data obtained in step S1; S3, Perform radiometric calibration, geometric correction, and reprojection operations on the satellite cloud image obtained in step S1; S4. The satellite cloud image processed in step S3 is subjected to feature extraction through the SDFM cloud image scale decoupling module to obtain multi-scale cloud image features. S5, use the DGAM feature fusion module to fuse the terrain data and meteorological data preprocessed in step S2 with the multi-scale cloud map features extracted in step S4 to obtain multi-source fused data; S6. Input the multi-source fusion data obtained in step S5 into the DSCRM residual module to extract time series information, obtain the multi-source time series feature map, and then perform multi-step prediction of ultra-short-term photovoltaic power through the CATFM prediction module.

2. The multi-source data-driven multi-step prediction method for ultra-short-term photovoltaic power according to claim 1, characterized in that, Step S4 describes extracting features from the satellite cloud image processed in step S3 using the SDFM cloud image scale decoupling module. Multi-scale cloud image features are obtained through three paths: fine-scale path, medium-scale path, and large-scale path. This includes the following steps: S21 captures local micro-scale cloud structure features in satellite cloud images through geometrically adaptive variable convolution in fine-scale paths. The formula for geometrically adaptive variable convolution is as follows: in, To output feature map at location The value, This indicates the number of sampling points in the convolution kernel. Indicates the first n Convolution weight parameters for each sampling point This represents the fixed offset of the standard convolution kernel. This represents the learnable spatial offset, calculated by the offset prediction network. This represents a learnable modulation scalar that controls the importance of each sampling point. Indicates the input feature value; S22, extracting cloud organization structure features, such as cloud clusters, convective complexes, and short-duration rainbands, from the perspective of mesoscale path-separable convolutional kernels, using these features. Specifically, this includes the following steps: S221, apply a bilinear interpolation rotation operation to the basic convolution kernel to generate a set of directional convolution kernels; S222 captures feature patterns from multiple directions within a set of directional convolutional kernels to reduce feature confusion caused by directional changes. The convolutional output is calculated for each direction, as shown in the following formula: in, Indicates the first c The first channel, the... i Okay, number j The output feature tensor of the column, Indicates the rotation of the convolution kernel. For output channel index, For input channel index, and These represent the row offset and column offset of the convolution kernel, respectively. Indicates the number of input channels. This represents the input feature map tensor. This indicates point-by-point operation; S223, calculate the eigenvector for each direction, using the following formula: in, For directional feature map vectors, and The height and width of the feature map, and Indicates the spatial row index and column index. Indicates all channels; S224, calculate the directional attention score and attention weight, using the following formula: in, For attention score, This is the offset. For direction value, These are the weighting coefficients. For temperature parameters, For the summation index, For the direction set; S225, sum all directional features according to attention weights; S23 extracts global features of the entire cloud map background through dilated convolution of large-scale paths.

3. The method for multi-source data-driven multi-step prediction of ultra-short-term photovoltaic power according to claim 1, characterized in that, Step S5 involves using the DGAM module to fuse the preprocessed terrain data and meteorological data from step S2 with the multi-scale cloud image features extracted in step S4 to obtain multi-source fused data. This includes the following steps: S31. For the extracted multi-scale cloud image features, meteorological data, and topographic data, a dynamic gating network is designed. Using time features as a condition, a weight vector with dimensions matching the original features is generated. The gating weight calculation formula is as follows: in, , and These are the gating weights of cloud image features, terrain data, and meteorological data under the current time conditions. For cloud map features, For terrain features, For meteorological characteristics, For time-encoded features, This represents a vector concatenation operation. Here is the weight matrix of the gated network. For bias terms, The Sigmoid activation function is used to constrain the weight values ​​to the interval [0,1]. To modify the activation function of the linear unit, used to filter negative values; S32 performs feature weighting using weight vectors and then uses a linear projection layer to map them to a unified feature space, achieving dimension alignment. The formula is as follows: in, Indicates the feature dimension of the cloud map. Represents the dimensions of terrain features. Representing the dimensions of meteorological characteristics, These are the weighting coefficients. For element-wise multiplication, Indicates the offset; S33: Concatenate all original features to form a global feature representation, and then obtain the cross-feature effect through linear transformation: in, For splicing features, This is a cross-characteristic effect. These are the coefficients of the weight matrix; S34 utilizes the cross-gating effect to adjust the information weights between fused features, enabling the DGAM module to dynamically adjust its internal cross-relationships over time; the formula is as follows: in, Represents the weights of cross features. For splicing features, It is a cross feature; S35, after obtaining the multi-scale cloud map features, the gating features of meteorological data and topographic data, and the cross features, a unified feature representation is constructed through the feature fusion layer; S36. A transient attention mechanism is designed to capture the dynamic dependencies within the input sequence, providing a more refined temporal feature representation for dynamic gating, resulting in the final fused features, as shown in the following formula: in, Let the first term represent the final attention matrix. For content relevance, the second item For relative position offset, Representing feature dimension, Indicates the value.

4. The multi-source data-driven multi-step prediction method for ultra-short-term photovoltaic power according to claim 1, characterized in that, Step S6 describes inputting the obtained multi-source fused data into the DSCRM residual module to extract time-series information and obtain a multi-source time-series feature map, including the following steps: S41, after the fused features processed in step S36 are compressed through layer normalization and pointwise convolution, the channel dimension is reduced by the ReLU activation function to obtain the dimensionality-reduced feature sequence, as shown in the following formula: in, Represents the dimensionality reduction feature sequence. For the input tensor of the residual sub-block, For pointwise convolution kernels, The pointwise convolution bias vector. Representation layer normalization; S42, convert the dimensionality-reduced feature sequence into a convolutional computation format. And input it into a sequence consisting of 3 DSCRM residual modules; S43, in each residual block, channel dimension compression is first performed, then depthwise convolution is used to model the temporal local receptive field, and finally the channels are restored to their original dimensions to obtain a multi-source temporal feature map. The residual structure formula within the block is as follows: Where Y is the residual output. For residual input, This indicates regularization.

5. The multi-source data-driven multi-step prediction method for ultra-short-term photovoltaic power according to claim 1, characterized in that, In step S6, the obtained multi-source time-series feature map is used to perform multi-step prediction of ultra-short-term photovoltaic power through the CATFM prediction module, including the following steps: S51, Design a CATFM prediction module. The module contains a temporal and spatial dual-branch structure to perform spatiotemporal processing on multi-source temporal feature maps respectively. S52 uses the time branch of the CATFM prediction module to perform global average pooling on the time dimension of the multi-source temporal feature map to compress temporal information. First, a two-layer fully connected network is used to learn the nonlinear importance of each channel, and then the time weight vector is output through the Sigmoid function. S53 applies causal-filled deep convolution to the multi-source temporal feature map through the spatial branch of the CATFM prediction module, then performs global average pooling in the channel dimension to preserve temporal dynamics, and finally normalizes it along the time dimension using Softmax to obtain the spatial weight vector. S54: First, the spatial weight vector and the temporal weight vector are multiplied together to generate a spatiotemporal dual weight matrix. Then, the spatiotemporal dual weight matrix is ​​multiplied element-wise with the multi-source temporal feature map to achieve adaptive control. According to different spatiotemporal features, the corresponding spatiotemporal weights are adjusted to obtain weighted features. S55 first performs point-by-point convolution on the weighted features to integrate channel information, and then flattens them and directly maps them to multi-step prediction results with 15-minute intervals within the next hour through a fully connected layer.

6. A multi-source data-driven ultra-short-term photovoltaic power multi-step prediction system, applicable to the multi-source data-driven ultra-short-term photovoltaic power multi-step prediction method according to any one of claims 15, characterized in that, Prediction systems include: The SDFM cloud image scale decoupling module is used to extract features from satellite cloud images that have undergone radiometric calibration, geometric correction, and reprojection processing. It obtains multi-scale cloud image features through three paths: fine-scale path, meso-scale path, and large-scale path. The DGAM feature fusion module is used to fuse preprocessed terrain data, meteorological data and multi-scale cloud map features. It designs a dynamic feature gating mechanism that can adaptively adjust the weights of multi-source data according to time conditions. It captures the local temporal dependence of photovoltaic power through a transient attention mechanism and introduces a cross-gating effect to capture the nonlinear interaction effect between different data sources. The DSCRM residual module is used to extract temporal information from multi-source fused data. It compresses the channel dimension through multiple residual blocks and then uses deep convolution to model the temporal local receptive field to obtain multi-source temporal feature maps. The CATFM prediction module processes multi-source time-series feature maps through a spatiotemporal dual-branch structure, mapping the extracted time-series features to power, thereby achieving multi-step prediction results at 15-minute intervals within the next hour.

Images