CN-121997270-A - Multi-source data driven ultra-short-term photovoltaic power multi-step prediction method and system

CN121997270ACN 121997270 ACN121997270 ACN 121997270ACN-121997270-A

Abstract

The invention provides a multi-source data driven ultra-short-term photovoltaic power multi-step prediction method and system, which comprises four core modules, wherein an SDFM cloud image scale decoupling module is used for carrying out multi-scale feature extraction on a satellite cloud image and accurately capturing different cloud layer feature information, a DGAM feature fusion module is used for adaptively adjusting importance weights of the satellite cloud image, topography and meteorological data through a dynamic time feature gating mechanism, a transient attention network and a cross feature effect are combined, the problem of dynamic change of the multi-source data features along with time is solved, feature fusion is carried out, a DSCRM residual module is used for carrying out self-adaptive weighting on feature channels and efficiently extracting deep time sequence relations, and a CATFM prediction module is used for mapping the multi-source time sequence feature image through a space-time double weight matrix to complete ultra-short-term photovoltaic power multi-step prediction and provide reliable technical support for power system scheduling operation.

Inventors

TAN LING
HUA CHENYANG
XIA JINGMING

Assignees

南京信息工程大学

Dates

Publication Date: 20260508
Application Date: 20260131

Claims (6)

1. The multi-step prediction method for the ultra-short-term photovoltaic power driven by the multi-source data is characterized by comprising the following steps of: s1, acquiring a satellite cloud picture, topographic data, meteorological data and historical photovoltaic power data; s2, carrying out outlier screening and missing interpolation preprocessing operation on the topographic data, the meteorological data and the historical photovoltaic power data obtained in the step S1; S3, performing radiation calibration, geometric correction and reprojection on the satellite cloud image obtained in the step S1; s4, performing feature extraction on the satellite cloud picture processed in the step S3 through an SDFM cloud picture scale decoupling module to obtain multi-scale cloud picture features; S5, utilizing DGAM feature fusion modules to fuse the topographic data preprocessed in the step S2, meteorological data and the multiscale cloud image features extracted in the step S4 to obtain multisource fusion data; And S6, inputting the multi-source fusion data obtained in the step S5 into a DSCRM residual error module to extract time sequence information, obtaining a multi-source time sequence characteristic diagram, and then carrying out multi-step prediction on the ultra-short-term photovoltaic power through a CATFM prediction module.
2. The multi-source data driven ultra-short term photovoltaic power multi-step prediction method according to claim 1, wherein in step S4, the feature extraction is performed on the satellite cloud image processed in step S3 by the SDFM cloud image scale decoupling module, and multi-scale cloud image features are obtained by three paths, namely a fine-scale path, a middle-scale path and a large-scale path, and the method comprises the following steps: s21, capturing the local microscale cloud system structural features in the satellite cloud image through geometric self-adaptive variability convolution in a fine-scale path, wherein the geometric self-adaptive variability convolution formula is as follows: Wherein, the In position for outputting the feature map Is used as a reference to the value of (a), Representing the number of convolution kernel samples, A convolution weight parameter representing the nth sample point, Representing a fixed offset of the standard convolution kernel, Representing a learnable spatial offset, calculated by the offset prediction network, Representing a learnable modulation scalar, controlling the importance of each sample point, Representing an input feature value; S22, extracting cloud organization structure features such as cloud clusters, convection complexes and short-time rain belts between local details and a global system through a separable convolution kernel in the direction of a mesoscale path, wherein the method specifically comprises the following steps of: s221, performing bilinear interpolation rotation operation on the basic convolution kernel to generate a group of direction convolution kernels; S222, capturing characteristic patterns of a plurality of directions in a group of direction convolution kernels to reduce characteristic confusion caused by direction change, and calculating convolution output for each direction, wherein the formula is as follows: Wherein, the Representing the output characteristic tensor of the c-th channel, i-th row, j-th column, Representing the rotation of the convolution kernel, In order to output the channel index, In order to input the channel index, And The convolution kernel row and column offsets are represented respectively, The number of input channels is indicated and, The tensor of the input feature map is represented, Representing a point-by-point operation; S223, calculating the feature vector of each direction, wherein the formula is as follows: Wherein, the As a vector of the directional characteristic map, And For the height and width of the feature map, And A spatial row index and a column index are represented, Representing all channels; s224, calculating a direction attention score and an attention weight, wherein the formula is as follows: Wherein, the In order to be able to pay attention to the score, As an amount of the offset to be used, As the value of the direction of the light, As the weight coefficient of the light-emitting diode, As a function of the temperature parameter(s), In order to find the sum of the indices, Is a direction set; s225, weighting and summing all direction characteristics according to the attention weight; S23, extracting global features of the whole cloud picture background through cavity convolution of the large-scale path.
3. The multi-source data-driven ultra-short term photovoltaic power multi-step prediction method according to claim 1, wherein in step S5, the DGAM module is used to fuse the topographic data preprocessed in step S2, the meteorological data and the multi-scale cloud image features extracted in step S4 to obtain multi-source fusion data, and the method comprises the following steps: S31, aiming at the extracted multi-scale cloud image features, meteorological data and topographic data, a dynamic gating network is designed, a weight vector with dimensions matched with the original features is generated under the condition of time features, and a gating weight calculation formula is as follows: wherein, the 、 And The gating weights of cloud image characteristics, topographic data and meteorological data under the current time condition are respectively, Is characterized by cloud picture, Is characterized by the topography, Is characterized by the weather, In order to be a time-coded feature, The vector concatenation operation is represented by a vector, For a weight matrix of the gating network, As a result of the bias term, Is a Sigmoid activation function used to constrain the weight values to the [0,1] interval, For correcting the linear unit activation function for filtering negative information; S32, feature weighting is carried out through the weight vector, the feature weighting is mapped to a unified feature space by using the linear projection layer, and the dimension alignment is completed, wherein the formula is as follows: Wherein, the The characteristic dimension of the cloud image is represented, Representing the dimensions of the topographical features, Representing the dimensions of the weather feature, As the weight coefficient of the light-emitting diode, For the element-by-element multiplication, Representing the offset; s33, all original features are spliced to form a global feature representation, and then cross feature effect is obtained through linear transformation: Wherein, the In order for the splice feature to be useful, In order for the cross-characteristic effect to be a function of, Is a weight matrix coefficient; S34, information weight among fusion features is adjusted by using a cross gating effect, so that DGAM modules dynamically adjust internal cross relations according to time, wherein the formula is as follows: Wherein, the The cross-feature weights are represented as such, In order for the splice feature to be useful, Is a cross feature; s35, after the multi-scale cloud image characteristics, the gating characteristics and the crossing characteristics of meteorological data and topographic data are obtained, constructing unified characteristic representation through a characteristic fusion layer; s36, designing a transient attention mechanism to capture a dynamic dependency relationship in an input sequence, providing a more refined time feature expression for dynamic gating, and obtaining a final fusion feature, wherein the formula is as follows: Wherein, the Representing the final attention matrix, the first term For content relevance, second item For the relative positional offset to be a function of the position, The dimensions of the features are represented and, Representing the value.
4. The multi-source data-driven ultra-short term photovoltaic power multi-step prediction method according to claim 1, wherein in step S6, the obtained multi-source fusion data is input into a DSCRM residual module to extract time sequence information, and a multi-source time sequence feature diagram is obtained, and the method comprises the following steps: s41, reducing the channel dimension through a ReLU activation function after carrying out layer normalization and point-by-point convolution compression on the fusion features processed in the step S36, and obtaining a dimension reduction feature sequence, wherein the formula is as follows: Wherein, the The characteristic sequence of dimension reduction is represented, Is the input tensor of the residual sub-block, In order to convolve the kernel point by point, In order to convolve the offset vector point by point, Normalizing the representation layer; s42, converting the dimension-reduced feature sequence into a convolution calculation format And input into a sequence consisting of 3 DSCRM residual modules; S43, at each residual block, firstly, channel dimension compression is carried out, then, time sequence local receptive field modeling is realized by using depth convolution, and finally, the channel is restored to the original dimension, so as to obtain a multi-source time sequence characteristic diagram, and the residual structure formula in the block is as follows: wherein Y is the residual output, As a residual input, Representing regularization.
5. The multi-source data driven multi-step prediction method for ultra-short term photovoltaic power according to claim 1, wherein in step S6, the obtained multi-source timing sequence feature map is subjected to multi-step prediction for ultra-short term photovoltaic power by a CATFM prediction module, and the method comprises the following steps: S51, designing CATFM a prediction module, wherein the module comprises a time and space double-branch structure, and performing space-time processing on the multi-source time sequence characteristic diagram respectively; S52, carrying out global average pooling on the time dimension of the multi-source time sequence feature map through a time branch of CATFM prediction modules so as to compress time information, firstly, learning the nonlinear importance of each channel through two layers of fully-connected networks, and then outputting a time weight vector through a Sigmoid function; s53, performing causal filling depth convolution on the multi-source time sequence feature map through a spatial branch of the CATFM prediction module, then performing global average pooling in a channel dimension, retaining time dynamics, and finally normalizing along a time dimension through Softmax to obtain a spatial weight vector; s54, performing outer product operation on the space weight vector and the time weight vector to generate a space-time double weight matrix, and multiplying the space-time double weight matrix by a multi-source time sequence feature diagram element by element to realize self-adaptive regulation and control, and adjusting corresponding space-time weights according to different space-time features to obtain weighted features; And S55, carrying out point-by-point convolution on the weighted characteristics to integrate channel information, flattening, and directly mapping the flattened channel information into a multi-step prediction result of 15-minute intervals in the future 1h through a full-connection layer.
6. A multi-source data driven multi-step prediction system for ultra-short term photovoltaic power, suitable for use in the multi-source data driven multi-step prediction method of claim 15, comprising: The SDFM cloud picture scale decoupling module is used for extracting characteristics of the satellite cloud picture subjected to radiometric calibration, geometric correction and re-projection treatment, and obtaining multi-scale cloud picture characteristics through three paths of a fine-scale path, a middle-scale path and a large-scale path respectively; DGAM a feature fusion module, which is used for fusing the preprocessed topographic data, meteorological data and multi-scale cloud image features, designing a dynamic feature gating mechanism, adaptively adjusting the weight of multi-source data according to time conditions, capturing the local time sequence dependency relationship of photovoltaic power through a transient attention mechanism, and capturing the nonlinear interaction effect among different data sources by introducing a cross gating effect; DSCRM residual error module, which is used to extract the time sequence information in the multisource fusion data, compress the channel dimension through a plurality of residual error blocks, and then use the depth convolution to realize the modeling of the time sequence local receptive field, so as to obtain the multisource time sequence characteristic diagram; CATFM prediction module, which processes the multi-source time sequence feature map through the space-time double branch structure, and maps the extracted time sequence feature into power, thereby realizing multi-step prediction result of 15 minutes interval in future 1 h.

Description

Multi-source data driven ultra-short-term photovoltaic power multi-step prediction method and system Technical Field The invention belongs to the technical field of photovoltaic power prediction, and particularly relates to a multi-step prediction method and a multi-step prediction system for ultra-short-term photovoltaic power driven by multi-source data. Background The photovoltaic power generation is used as a clean renewable energy source, the power generation process is highly dependent on natural environment conditions, and is easily influenced by the combination of solar radiation intensity, cloud cover, environmental temperature, atmospheric state and other factors, so that the output power shows obvious randomness and fluctuation. This unstable nature presents a significant challenge for safe operation and dispatch control of the power system. Therefore, the development of high-precision photovoltaic power generation power prediction has important significance for improving the safety, reliability and overall economy of a photovoltaic power generation system. With the development of remote sensing technology and data acquisition means, photovoltaic power prediction methods which fuse satellite cloud images, meteorological data, topographic information and historical photovoltaic power data are gradually increased. The multi-source data can reflect the environmental change characteristics in the photovoltaic power generation process from different angles, and the representation capability of the prediction model on the photovoltaic power fluctuation rule is improved to a certain extent. However, due to the significant differences of different data sources in time scale, spatial resolution, data structure and the like, the existing multi-source photovoltaic power prediction method still has defects in the aspect of collaborative modeling. On one hand, most methods directly adopt a characteristic direct splicing or fixed weight fusion mode, so that dynamic relations among multi-source data are difficult to fully capture, and key information is easy to weaken or redundant characteristics interfere model learning. On the other hand, the existing cloud image feature extraction mode is single, the capturing capability of the cloud layer multi-scale space structure and the evolution characteristic thereof is limited, and the local cloud cluster change and the large-scale cloud system motion feature are difficult to reflect simultaneously. In an ultra-short term photovoltaic power multi-step prediction scenario, the non-linearity and timing dependence of the photovoltaic power time series make the prediction error easily accumulate as the prediction step increases. The existing method has certain limitation in the aspects of long-time dependent modeling and time sequence characteristic dynamic updating, and the prediction precision is easy to be reduced under a longer prediction step length. Meanwhile, part of high-precision models have complex structures and large calculation cost, and the simplified models have the problem of insufficient feature expression capability, so that effective balance between prediction precision and calculation efficiency is difficult to obtain. Disclosure of Invention The invention aims to solve the problems in the prior art, and provides a multi-step prediction method and system for multi-source data-driven ultra-short-term photovoltaic power. The invention provides a multi-source data driven ultra-short term photovoltaic power multi-step prediction method, which specifically comprises the following steps: s1, acquiring a satellite cloud picture, topographic data, meteorological data and historical photovoltaic power data; s2, carrying out outlier screening and missing interpolation preprocessing operation on the topographic data, the meteorological data and the historical photovoltaic power data obtained in the step S1; S3, performing radiation calibration, geometric correction and reprojection on the satellite cloud image obtained in the step S1; s4, performing feature extraction on the satellite cloud picture processed in the step S3 through an SDFM cloud picture scale decoupling module to obtain multi-scale cloud picture features; S5, utilizing DGAM feature fusion modules to fuse the topographic data preprocessed in the step S2, meteorological data and the multiscale cloud image features extracted in the step S4 to obtain multisource fusion data; And S6, inputting the multi-source fusion data obtained in the step S5 into a DSCRM residual error module to extract time sequence information, obtaining a multi-source time sequence characteristic diagram, and then carrying out multi-step prediction on the ultra-short-term photovoltaic power through a CATFM prediction module. Further, in step S2, the outlier screening and missing interpolation preprocessing operation is performed on the topographic data, the meteorological data and the historical photovoltaic power data obtained