Post-spectroscopic photovoltaic decomposition method based on spatiotemporal feature adaptive extraction and deep fusion

Abstract

The application discloses a kind of post-meter photovoltaic decomposition methods based on spatiotemporal feature adaptive extraction and deep fusion, comprising: collecting the net load data of user and post-meter photovoltaic output label data and pre-processing, according to proportion division into training set, verification set and test set;Dynamic representation user node space correlation is constructed by adaptive adjacency matrix, and dynamic spatiotemporal graph is generated;Post-meter photovoltaic decomposition model including time series feature extraction module, spatial feature extraction module and spatiotemporal state module is constructed, and adaptive adjacency matrix and model parameters are iteratively updated based on training set;In test phase, based on optimal adaptive adjacency matrix and model, carry out post-meter photovoltaic decomposition.The application realizes the high-precision decomposition of the power generation of household distributed photovoltaic device installed after intelligent electric meter, which helps to realize massive household distributed photovoltaic monitoring at low cost, and improves the carrying capacity and control ability of distribution network to post-meter photovoltaic.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of post-meter photovoltaic decomposition, and particularly relates to a post-meter photovoltaic decomposition method based on adaptive extraction and deep fusion of space-time features. BACKGROUND

[0002] In order to reduce long-distance transmission line losses, optimize power distribution network investment, and promote local consumption of renewable energy, more and more distributed photovoltaics are installed and consumed on the low-voltage user side. However, most low-voltage household photovoltaics are installed after the smart meter, and the smart meter only measures the net load data, so the power grid metering system cannot obtain photovoltaic power generation information and load information. Household photovoltaics have caused problems such as difficulty in voltage control of the distribution network, decreased load prediction accuracy, and ineffective adjustment of roof photovoltaic consumption trading strategies. In order to meet the requirements of "observation, measurement, adjustment, and control" for distributed photovoltaic power generation, installing independent metering devices on existing large-scale post-meter household photovoltaic power generation systems is a direct solution, but it requires a lot of manpower and material resources during deployment. Therefore, researching a post-meter photovoltaic decomposition method based on only net load data can help to achieve low-cost monitoring of a large number of household photovoltaics and is the basis for improving the carrying and regulation capacity of the distribution network for post-meter photovoltaics.

[0003] At present, data-driven post-meter photovoltaic decomposition has the following difficulties:

[0004] 1. There are user behavior heterogeneity disturbances in user net load data, such as differences in appliance usage patterns and randomness of device start-stop, which makes it difficult to represent the spatial correlation of post-meter photovoltaic output that is strongly related to the geographical location distribution of post-meter photovoltaic power generation devices using only user net load data.

[0005] 2. There are obvious period differences between the output of post-meter photovoltaics and the periodic variation process of household appliances. In the time domain, multiple time series with different periods are coupled with each other, and the net load data presents complex periodicity, making it difficult to extract the periodic time series features of post-meter photovoltaics.

[0006] 3. The change of post-meter photovoltaic output is affected by geographical location distribution, showing wide-area spatial correlation, and is affected by the movement of atmospheric clouds, showing local time-lag correlation. These two correlations present a complex coupling relationship, making it difficult to extract the space-time correlation of post-meter photovoltaic output from net load data.

[0007] 4. There is a dynamic coupling characteristic between the space-time features of post-meter photovoltaics, and there is a lack of explicit modeling means for the space-time features of post-meter photovoltaics, making it difficult to effectively utilize the space-time correlation between users and affecting the depth of space-time feature fusion of post-meter photovoltaics.

[0008] The above problems increase the difficulty of post-meter photovoltaic decomposition to some extent, reduce the accuracy of post-meter photovoltaic decomposition, and limit the application of traditional post-meter photovoltaic decomposition methods in new energy penetration rate increasing new power systems.

[0009] Therefore, a post-meter photovoltaic decomposition method based on adaptive extraction and deep fusion of space-time characteristics is provided to realize high-precision and high-robustness post-meter photovoltaic decomposition. As a part of distribution network operation and planning, it will help promote the application of artificial intelligence technology in new power systems and improve the perception of the operation state of the distribution network, providing accurate data support for source-grid-load collaborative optimization, safe and stable operation, and efficient consumption of new energy. SUMMARY

[0010] The purpose of the present application is to overcome the shortcomings and deficiencies of the prior art, and to provide a post-meter photovoltaic decomposition method based on adaptive extraction and deep fusion of space-time characteristics, to solve the problem of insufficient extraction and deep fusion of periodic time series features and spatial correlation features in traditional post-meter photovoltaic decomposition methods, and to help improve the carrying and regulation capacity of renewable energy in distribution networks.

[0011] To achieve the above purpose, the technical scheme provided by the present application is: a post-meter photovoltaic decomposition method based on adaptive extraction and deep fusion of space-time characteristics, comprising the following steps:

[0012] 1) Collecting the net load data and post-meter photovoltaic output label data of users and preprocessing them, dividing the preprocessed net load data and post-meter photovoltaic output label data into training set, validation set and test set;

[0013] 2) Initializing the adaptive adjacency matrix according to the mutual information value between the net load data, constructing an adaptive adjacency matrix dynamically updated by a decomposition error backpropagation mechanism, dynamically representing the spatial correlation relationship of user nodes, and constructing a dynamic space-time graph containing the space-time correlation relationship between users;

[0014] 3) Calculating the decomposition error based on the post-meter photovoltaic output label data in the training set, using the backpropagation mechanism to iteratively update the adaptive adjacency matrix and the learnable parameters in the constructed post-meter photovoltaic decomposition model, and in all iteration rounds, selecting the model and adaptive adjacency matrix with the best decomposition performance for the user net load data in the validation set as the optimal model and optimal adaptive adjacency matrix for saving;

[0015] 4) In the test phase, based on the optimal model and the optimal adaptive adjacency matrix, the post-meter photovoltaic decomposition of the user net load data in the test set is carried out, and the post-meter photovoltaic output of the user in the test set is obtained.

[0016] Further, in step 1), the preprocessing includes outlier removal, missing value filling, and normalization. The photovoltaic after the meter is a low-voltage household photovoltaic power generation device installed after the smart meter. The user net load data is the net value of the user's actual electricity load minus the output of its photovoltaic after the meter.

[0017] Furthermore, the specific operation steps of step 2) are as follows:

[0018] 2.1) Traverse all possible user combinations in the user set, and calculate the user load based on the net load data obtained in the training set in step 1). and users Mutual information values ​​between net load sequences As shown in the following formula:

[0019] ;

[0020] In the formula, and Representing users respectively Net load sequence and users At sequence point Net load power; and They represent and The marginal distribution; express and The joint distribution;

[0021] 2.2) The calculated mutual information value With a given threshold To make a comparison, and The value assigned to and Determine the initial adaptive adjacency matrix The Middle Line 1 Column values As shown in the following formula:

[0022] ;

[0023] 2.3) For the initial adaptive adjacency matrix Perform singular value decomposition to calculate the node embedding vector. and As shown in the following formula:

[0024] ;

[0025] ;

[0026] ;

[0027] wherein, 、 and denote the left singular matrix, the diagonal matrix of singular values and the right diagonal matrix, respectively, and denote the first columns of the matrix and the first columns of the matrix and transpose, respectively, denotes the diagonal matrix that retains the first singular values and takes the power operation;

[0028] 2.4) embedding the node vectors and as learnable parameter vectors, constructing the adaptive adjacency matrix that is dynamically updated by the decomposition error backpropagation mechanism as follows:

[0029] ;

[0030] wherein, and denote the normalized exponential and the linear rectified activation function, respectively;

[0031] 2.5) regarding each user in the user set as a node, constructing the dynamic spatio-temporal graph that contains the spatio-temporal correlation between users as follows:

[0032] ;

[0033] wherein, 、 and denote the user net load sequence matrix, the user node set and the set of edges connecting the user nodes on the graph, respectively.

[0034] Further, in step 3), the post-table photovoltaic decomposition model is composed of a time series feature extraction module, a spatial feature extraction module and a spatio-temporal state module.

[0035] Further, the process of extracting the user node time series feature representation on the dynamic spatio-temporal graph from the user net load sequence matrix by the time series feature extraction module is as follows:

[0036] a, performing numerical embedding operation on the user net load sequence matrix as follows:

[0037] ;

[0038] wherein, represents the output result of the numerical embedding operation, represents a one-dimensional convolution operation;

[0039] Then, the user load sequence matrix is subjected to a position embedding operation as shown in the following formula:

[0040] ;

[0041] wherein, represents the position encoding of the sequence point , represents the time series feature dimension, represents the scaling factor; b, the output results of the two-part embedding operation of the sum of the numerical embedding and the position embedding, are added to generate the original time series features of the load data

[0042] as shown in the following formula:

[0043] ;

[0044] wherein, represents the output result of the position embedding operation;

[0045] c, the original time series features of the load data are converted from the time domain representation to the frequency domain representation using the fast Fourier transform method as shown in the following formula:

[0046] ;

[0047] wherein, represents the sequence length; represents the frequency domain index; and are the time domain representation corresponding to the load data at the sequence point and the frequency domain representation corresponding to , respectively; represents the rotation factor, represents the complex number symbol;

[0048] Then, the modulus of each component is calculated, and the average value is taken in the time series feature dimension as shown in the following formula:

[0049] ;

[0050] wherein, and represent the modulus of and , respectively​the average value of the time sequence dimension;

[0051] d. Representing the importance of the modulus size, taking the first main components that dominate the periodic time sequence changes of the net load data, as shown in the following formula:

[0052] ;

[0053] In the formula, represent the first components corresponding to the frequency value of the maximum modulus; , and represent the frequencies of the first, second and periodic dominant components, respectively;

[0054] Then, calculate the periodic information of the net load data, as shown in the following formula:

[0055] ;

[0056] In the formula, and represent the period and frequency of the first component in the first periodic dominant component with the largest amplitude;

[0057] e. Based on the period and frequency information of the first periodic dominant components in the frequency domain of the net load data, the original time sequence characteristics of the net load data are divided in turn, and aligned in the time dimension, and finally rearranged into period alignment feature blocks , wherein represents the real number field;

[0058] f. Construct a periodic feature extraction model, and use a convolution kernel with a size of to extract the periodic features of the periodic alignment feature blocks in turn, as shown in the following formula:

[0059] ;

[0060] In the formula, represents the feature map obtained by the first component in the first periodic dominant component in the first convolution branch; represents the two-dimensional convolution operation of the first convolution branch; represents the convolution kernel with a size of ; ​​

[0061] Then, the periodic features are stacked along the feature dimension to generate the output, as shown in the following equation:

[0062] ;

[0063] In the formula, Indicates the preceding The first of the periodic dominant components The periodic characteristics of each component; This indicates a feature map stacking operation; , , , , and In order to represent the previous The first of the periodic dominant components The feature maps obtained by each component in the first to sixth convolutional branches;

[0064] g. According to the previous... The weighting coefficients for the magnitudes of the periodic dominant components are calculated as follows:

[0065] ;

[0066] In the formula, and They represent the first The amplitude of each component before and after activation;

[0067] Then the periodic features One by one, restore them to the same shape as the original temporal characteristics. Then, weighted aggregation The periodic features corresponding to each periodic dominant component generate the temporal feature representation of user nodes on the dynamic spatiotemporal graph. As shown in the following formula:

[0068] .

[0069] Furthermore, the spatial feature extraction module updates the temporal feature representation of user nodes on the dynamic spatiotemporal graph to the spatiotemporal feature representation of user nodes and generates the dynamic spatiotemporal graph embedding process as follows:

[0070] a. Extend the adaptive adjacency matrix sequentially from first order using matrix exponentiation. Step, use Order-adaptive adjacency matrix As The probability transition matrix of the jump diffusion process, then... The probability transition matrices of each graph diffusion process are arranged into a graph diffusion tensor. As shown in the following formula:

[0071] ;

[0072] In the formula, , and respectively represent the probability transition matrix of one-hop, two-hop and three-hop diffusion process on dynamic spatio-temporal graph;

[0073] b. The wide-area user node spatial correlation relationship is represented by the graph diffusion tensor, the multi-hop neighbor feature information is aggregated to update the user node spatio-temporal feature representation, and the dynamic spatio-temporal graph embedding is generated, as shown in the following formula:

[0074] ;

[0075] ;

[0076] In the formula, represents the graph diffusion convolution kernel; , and respectively represent the learnable weight parameter matrix of one-hop, two-hop and three-hop diffusion process; represents the learnable weight parameter matrix of one-hop, two-hop and three-hop diffusion process; represents the dynamic spatio-temporal graph embedding; represents the dimension size of the user node spatio-temporal feature representation; represents the element-wise multiplication operation; represents the total number of users.

[0077] Further, the process of the spatio-temporal state module based on the dynamic spatio-temporal graph embedding deep fusion user node spatio-temporal feature representation to generate the photovoltaic decomposition result is as follows:

[0078] a. The extracted dynamic spatio-temporal graph embedding is projected into a spatio-temporal state space representation , as shown in the following formula:

[0079] ;

[0080] In the formula, represents a linear projection function;

[0081] b. The spatio-temporal state space representation is split into a main component and a gating component , as shown in the following formula:

[0082] ;

[0083] In the formula, represents a gating split operation;

[0084] c. extracting the principal component using a one-dimensional convolutional neural network and performing activation as follows:

[0085] ;

[0086] wherein, represents the locally enhanced principal component; represents an activation function;

[0087] d. generating four selective matrices based on dynamic spatio-temporal graph embedding , , and as follows:

[0088] ;

[0089] ;

[0090] ;

[0091] ;

[0092] wherein, represents an enhanced temporal scale parameter matrix; represents a matrix multiplication operation according to Einstein summation convention; represents a broadcast operation function of feature dimension; is an activation function; represents a learnable parameter matrix with the same shape as the broadcast operation output result; and represent a randomly initialized state transition matrix and an enhanced state transition matrix, respectively; represents the power function of the base of the natural logarithm; and represent an enhanced input control matrix and an identity matrix, respectively; represents an output control matrix;

[0093] e. performing a selective spatio-temporal state transition process on the locally enhanced principal component based on the four generated selective matrices , , and , and deeply fusing the locally enhanced principal component to output an enhanced spatio-temporal feature representation for each discrete sequence point The selective spatiotemporal state transition process of the selection is as follows:

[0094] ;

[0095] In the formula, and respectively represent the local enhanced main component and the enhanced spatiotemporal feature representation of the discrete sequence point ; and respectively represent the hidden state of the discrete sequence point and ;

[0096] f, the activated gating component is multiplied with the enhanced spatiotemporal feature representation , and a linear projection generates the final post-metering photovoltaic decomposition result , as shown in the following formula:

[0097] .

[0098] Compared with the prior art, the present application has the following advantages and beneficial effects:

[0099] 1. The present application proposes a periodic decoupling-based time series feature extraction module, which first decouples the user net load sequence by frequency domain analysis, and then extracts the multi-scale periodic features of the post-metering photovoltaic output, effectively improving the periodic feature extraction capability of the photovoltaic decomposition model, and effectively utilizing the period difference between the post-metering photovoltaic output and the periodic variation process of the household appliance electricity to improve the model decomposition performance.

[0100] 2. The present application proposes a spatial feature extraction module based on dynamic spatial correlation representation, which can adaptively capture the spatial correlation of user nodes in a data-driven manner, and further globally model the user nodes to represent the spatial correlation of wide-area user nodes, dynamically update the spatiotemporal feature representation of the user nodes, and solve the problem of insufficient representation of the spatial correlation of user nodes in traditional methods, and can efficiently mine the spatiotemporal coupling rules of post-metering photovoltaic output.

[0101] 3. The present application proposes a dynamic spatiotemporal graph embedding-based spatiotemporal state module, which explicitly models the spatiotemporal correlation between users through a selective spatiotemporal state transition process based on dynamic spatiotemporal graph embedding, deeply fuses the spatiotemporal feature representation of the user nodes, and outputs the post-metering photovoltaic decomposition result, solving the problem of insufficient spatiotemporal feature fusion capability of traditional methods, and fully utilizing the complex coupling relationship existing in the spatiotemporal feature representation of the user nodes to improve the model decomposition performance and robustness.

[0102] In conclusion, the present application provides a new solution for how to efficiently utilize net load data to realize massive post-meter photovoltaic monitoring, has outstanding engineering application value, and has good popularization prospects. BRIEF DESCRIPTION OF DRAWINGS

[0103] Figure 1 is a schematic diagram of a periodic decoupling process.

[0104] Figure 2 is a schematic diagram of multi-scale periodic feature extraction.

[0105] Figure 3 is a schematic diagram of selective spatiotemporal state transition.

[0106] Figure 4 is a post-meter photovoltaic output feature extracted by different models and a real post-meter photovoltaic spatiotemporal distribution diagram. DETAILED DESCRIPTION

[0107] The present application will be described in further detail below with reference to the embodiments and drawings, but the embodiments of the present application are not limited thereto.

[0108] Embodiment 1

[0109] The present embodiment discloses a post-meter photovoltaic decomposition method based on spatiotemporal feature adaptive extraction and deep fusion, and the specific circumstances are as follows:

[0110] 1) Collect user net load data and post-meter photovoltaic output label data and perform preprocessing, including outlier rejection, missing value filling and normalization, divide the preprocessed user net load data and post-meter photovoltaic output label data into training set, validation set and test set according to the proportion, and the specific steps are as follows:

[0111] 1.1) Collect the net load data and post-meter photovoltaic output label data saved in the form of daily load sequence of each user, form a daily load sequence pair , as shown in the following formula:

[0112] ;

[0113] In the formula, , and respectively represent the daily load sequence pair, the net load sequence and the post-meter photovoltaic output label sequence of the user ;

[0114] 1.2) Traverse each daily load sequence pair formed in step 1.1), judge the data missing condition of the net load sequence and the post-meter photovoltaic output label sequence, if the data missing rate of any one sequence in the daily load sequence pair is higher than 25%, the daily load sequence pair is discarded;

[0115] 1.3) Traverse each sequence point in the retained sequence. If the difference between the sequence point and the sequence mean is greater than 3 times the sequence standard deviation, it is regarded as outlier data and removed. After removal, the sequence point is treated as a missing value.

[0116] 1.4) Determine if there are adjacent sequence points to the missing value. If both adjacent sequence points are missing values, assign the missing value a value of 0. If the sequence point corresponding to the missing value is the first or last sequence point in the sequence, fill the missing value with the unique adjacent sequence point. If both adjacent sequence points to the missing value are normal values, fill the missing value with the average of the adjacent sequence points. If the adjacent sequence points to the missing value are one normal value and one missing value, fill the missing value with the normal value.

[0117] 1.5) The daily load series is normalized using the Z-Score normalization method, as shown in the following formula:

[0118] ;

[0119] ;

[0120] In the formula, , , and Representing users respectively At sequence point Net load power, net load series mean, net load series standard deviation, and normalized series points Net load power, , , and Representing users respectively At sequence point The table includes the photovoltaic power output, the average value of the photovoltaic power output label sequence, the standard deviation of the photovoltaic power output label sequence, and the normalized sequence points. Photovoltaic power output after the meter;

[0121] 1.6) All preprocessed daily load sequences are then sorted according to... The proportions are divided into training set, validation set, and test set, as shown in the example. , and The values ​​are 7, 1, and 2.

[0122] 2) Initialize the adaptive adjacency matrix based on the mutual information values ​​between user net load data, construct an adaptive adjacency matrix dynamically updated by the decomposition error backpropagation mechanism, dynamically represent the spatial correlation between user nodes, and then construct a dynamic spatiotemporal graph containing the spatiotemporal correlation between users. The specific steps are as follows:

[0123] 2.1) Traverse all possible user combinations in the user set, and calculate the mutual information value between the net load sequence of user and user according to the net load data obtained in step 1) in the training set, as shown in the following formula:

[0124] ;

[0125] In the formula, and respectively represent the net load sequence of user and the net load power of user at sequence point ; and respectively represent the marginal distribution of and ; represents the joint distribution of and .

[0126] 2.2) Compare the calculated mutual information value with a given threshold value , and assign the values of and to and , to determine the value of the row and the column in the initial adaptive adjacency matrix , as shown in the following formula:

[0127] ;

[0128] 2.3) Singular value decomposition is performed on the initial adaptive adjacency matrix , to calculate the node embedding vectors and , as shown in the following formula:

[0129] ;

[0130] ;

[0131] ;

[0132] In the formula, , and respectively represent the left singular matrix, the singular value diagonal matrix and the right diagonal matrix, and​ denote the first columns of the matrix and the first columns of the matrix and transpose, denote the diagonal matrix that retains the first singular values and takes power operation. In the embodiment, the value of

[0133] 2.4) Take the node embedding vectors and as the learnable parameter vectors, construct the adaptive adjacency matrix updated dynamically by the decomposition error backpropagation mechanism as follows:

[0134] ;

[0135] In the formula, and denote the normalized exponential and linear rectified activation function respectively.

[0136] 2.5) Take each user in the user set as a node, construct a dynamic spatio-temporal graph containing the spatio-temporal correlation between users as follows:

[0137] ;

[0138] In the formula, , and denote the user net load sequence matrix, the user node set and the set of edges connecting the user nodes on the graph respectively.

[0139] 3) Based on the post-metering photovoltaic output label data in the training set, calculate the decomposition error, and use the backpropagation mechanism to iteratively update the adaptive adjacency matrix and the learnable parameters in the constructed post-metering photovoltaic decomposition model. In all iteration rounds, select the model and adaptive adjacency matrix with the best decomposition performance for the user net load data in the validation set as the optimal model and optimal adaptive adjacency matrix to save, and the specific steps are as follows:

[0140] 3.1) Based on the post-metering photovoltaic output label data in the training set obtained in step 1), calculate the decomposition error as follows:

[0141] ;

[0142] In the formula, denotes the decomposition error; and denote the user at the sequence point post-table photovoltaic decomposition results and post-table photovoltaic output label values of the table;

[0143] 3.2) updating the adaptive adjacency matrix and the learnable parameters in the post-table photovoltaic decomposition model by using a back propagation method to minimize the decomposition error;

[0144] 3.3) iteration step 3.2) times, in each iteration process, calculating the mean absolute error based on the post-table photovoltaic output label data in the validation set obtained in step 1) as a decomposition performance evaluation index, as shown in the following formula:

[0145] ;

[0146] Then, the iteration round corresponding to the minimum is selected The post-table photovoltaic decomposition model and the adaptive adjacency matrix corresponding to the minimum iteration round are saved as the best model and the optimal adaptive adjacency matrix. In the embodiment, The value of is 500.

[0147] The post-table photovoltaic decomposition model is composed of a time sequence feature extraction module, a spatial feature extraction module and a space-time state module. The time sequence feature extraction module analyzes user net load data to obtain frequency domain information by using a fast Fourier transform method, decouples the net load data periodically in combination with time domain and frequency domain information, extracts and weightedly aggregates post-table photovoltaic output multi-scale periodic features, and generates user node time sequence feature representation on a dynamic space-time graph. The spatial feature extraction module takes the user node time sequence feature representation on the dynamic space-time graph as input, performs global spatial modeling based on the adaptive adjacency matrix of the current iteration round, further represents wide-area user node spatial correlation, updates the user node time sequence feature representation on the dynamic space-time graph into user node space-time feature representation through a multi-hop aggregation mechanism, and generates dynamic space-time graph embedding. The space-time state module performs an explicit modeling on the space-time correlation between users based on the dynamic space-time graph embedding, executes a selective space-time state transition process, and deeply fuses the user node space-time feature representation to generate post-table photovoltaic decomposition results.

[0148] Specifically, the process of extracting the user node time sequence feature representation on the dynamic space-time graph from the user net load sequence matrix by the time sequence feature extraction module is as follows:

[0149] a, performing a numerical embedding operation on the user net load sequence matrix as shown in the following formula:

[0150] ;

[0151] In the formula, represents the output result of the numerical embedding operation, represents a one-dimensional convolution operation.

[0152] Then, the user load sequence matrix is subjected to a position embedding operation as follows:

[0153] ;

[0154] In the formula, represents the position encoding of the sequence point , represents the time sequence feature dimension, represents the scaling factor. In the embodiment, and are respectively 128 and 0.4.

[0155] b. The output results of the two-part embedding operation of the sum value embedding and the position embedding are combined to generate the original time sequence feature of the load data as follows:

[0156] ;

[0157] In the formula, represents the output result of the position embedding operation.

[0158] c. The original time sequence feature of the load data is converted from the time domain representation to the frequency domain representation using the fast Fourier transform method as follows:

[0159] ;

[0160] In the formula, represents the sequence length; represents the frequency domain index; and are respectively the time domain representation corresponding to the load data at the sequence point and the frequency domain representation corresponding to the load data at ; represents the rotation factor, represents the complex number symbol;

[0161] Then, the modulus of each component is calculated, and the average value is taken in the time sequence feature dimension as follows:

[0162] ;

[0163] In the formula, and respectively represent the modulus of and ​​The average value in the temporal feature dimension;

[0164] d. Importance is indicated by the magnitude of the modulus; the first value is taken. The component that plays a dominant role in the periodic temporal variation of net load data is shown in the following formula:

[0165] ;

[0166] In the formula, Indicates taking The largest median value The frequency values ​​corresponding to each component; , and They represent the 1st, 2nd and 3rd respectively. The frequency of the dominant periodic component.

[0167] Then, the net load data period information is calculated as shown in the following formula:

[0168] ;

[0169] In the formula, and These represent the front with the largest amplitude. The first of the periodic dominant components The period and frequency of each component. Example: and The values ​​are 5 and 96 respectively.

[0170] e. Based on the net load data in the frequency domain The periodicity and frequency information of each periodic dominant component are used to sequentially analyze the original time-series characteristics of the net load data. Divide the data, align it along the time dimension, and finally rearrange it into... Each periodic aligned feature block ,in Represents the real number field. (Appendix) Figure 1 The generation process of the first two periodic feature alignment blocks is shown. This invention integrates time domain and frequency domain information, decouples and reconstructs the time domain representation of the original time series features into periodic alignment feature blocks, so that it can characterize the time series variation law of photovoltaic power output from two different dimensions: intra-cycle variation and inter-cycle variation.

[0171] f. Construct a periodic feature extraction model, using successively a size of... The convolutional kernel extracts the periodic features of the periodically aligned feature blocks, as shown in the following formula:

[0172] ;

[0173] In the formula, Indicates the preceding The first of the periodic dominant components The component in the first Feature maps obtained from convolutional branches; Indicates the first Two-dimensional convolution operation with multiple convolution branches; Indicates size is The convolution kernel;

[0174] Then, the periodic features are stacked along the feature dimension to generate the output, as shown in the following equation:

[0175] ;

[0176] In the formula, Indicates the preceding The first of the periodic dominant components The periodic characteristics of each component; This indicates a feature map stacking operation; , , , , and In order to represent the previous The first of the periodic dominant components Feature maps of each component obtained in the first to sixth convolutional branches; Appendix Figure 2 This invention demonstrates the process of multi-scale periodic feature extraction. The invention extracts periodic features of multiple time scales in periodically aligned feature blocks through a series of convolutional kernels of different sizes, and stacks periodic features from different receptive fields in the feature dimension to generate output.

[0177] g. According to the previous... The weighting coefficients for the magnitudes of the periodic dominant components are calculated as follows:

[0178] ;

[0179] In the formula, and They represent the first The amplitude of each component before and after activation;

[0180] Then the periodic features One by one, restore them to the same shape as the original temporal characteristics. Then, weighted aggregation The periodic features corresponding to each periodic dominant component generate the temporal feature representation of user nodes on the dynamic spatiotemporal graph. As shown in the following formula:

[0181] .

[0182] Specifically, the process of updating the user node time sequence feature representation on the dynamic spatio-temporal graph to the user node spatio-temporal feature representation and generating the dynamic spatio-temporal graph embedding by the spatial feature extraction module is as follows:

[0183] a. The adaptive adjacency matrix is sequentially expanded from the first order to the order using the order adaptive adjacency matrix as the probability transition matrix of the jump diffusion process, and then the probability transition matrices of graph diffusion processes are arranged into a graph diffusion tensor , as shown in the following formula:

[0184] ;

[0185] In the formula, , and respectively represent the probability transition matrices of one-hop, two-hop and three-hop diffusion processes on the dynamic spatio-temporal graph. In the embodiment, the value of is 8.

[0186] b. The wide-area user node spatial correlation relationship represented by the graph diffusion tensor is used to aggregate multi-hop neighbor feature information to update the user node spatio-temporal feature representation and generate the dynamic spatio-temporal graph embedding, as shown in the following formula:

[0187] ;

[0188] ;

[0189] In the formula, represents a graph diffusion convolution kernel; , and respectively represent the learnable weight parameter matrices of one-hop, two-hop and three-hop diffusion processes; represents the learnable weight parameter matrix of hop diffusion process; represents the dynamic spatio-temporal graph embedding; represents the dimension size of the user node spatio-temporal feature representation; represents an element-wise multiplication operation; represents the total number of users. In the embodiment, the value of is 196; the value of is 268.

[0190] Specifically, the process of generating the photovoltaic decomposition result by the spatio-temporal state module based on the deep fusion of the user node spatio-temporal feature representation and the dynamic spatio-temporal graph embedding is as follows:

[0191] a. embedding the extracted dynamic spatio-temporal graph projecting as spatio-temporal state space representation as follows:

[0192] ;

[0193] wherein, denotes a linear projection function.

[0194] b. splitting the spatio-temporal state space representation into principal component and gating component as follows:

[0195] ;

[0196] wherein, denotes a gating splitting operation.

[0197] c. extracting local dependency of principal component using one-dimensional convolutional neural network and performing activation as follows:

[0198] ;

[0199] wherein, denotes the locally enhanced principal component; denotes an activation function.

[0200] d. generating four selective matrices , , and based on the dynamic spatio-temporal graph embedding as follows:

[0201] ;

[0202] ;

[0203] ;

[0204] ;

[0205] wherein, denotes an enhanced temporal scale parameter matrix; denotes a matrix multiplication operation following Einstein summation convention; denotes a broadcast operation function of feature dimension; is an activation function; denotes a learnable parameter matrix with the same shape as the broadcast operation output result; and respectively represent the state transition matrix and the enhanced state transition matrix initialized randomly; denotes the power function of the base of natural logarithm ; and respectively represent the enhanced input control matrix and the unit matrix; denotes the output control matrix.

[0206] e、the generated four selective matrices , , and , perform a selective spatio-temporal state transition process on the locally enhanced main components, and deeply integrate the locally enhanced main components output enhanced spatio-temporal feature representation , the selective spatio-temporal state transition process of each discrete sequence point is as follows:

[0207] ;

[0208] In the formula, and respectively represent the locally enhanced main components and the enhanced spatio-temporal feature representation of the discrete sequence point ; and respectively represent the hidden states of the discrete sequence points and . The Figure 2 demonstrates the selective spatio-temporal state transition process, and the selective matrices , , and perform the selective spatio-temporal state transition process on the user nodes on the dynamic spatio-temporal graph in parallel by passing spatial information, and realize deep integration of spatio-temporal features.

[0209] f、after element-wise multiplication of the activated gating component and the enhanced spatio-temporal feature representation , linear projection generates the final post-metering photovoltaic decomposition result , as shown in the following formula:

[0210] .

[0211] 4) In the test phase, based on the optimal model and the optimal adaptive adjacency matrix, the post-metering photovoltaic decomposition of the user net load data in the test set is carried out, and the post-metering photovoltaic output of the user in the test set is obtained.

[0212] Embodiment 2

[0213] Reference is made to Tables 1, 2, 3 and the attached figures Figure 4 As a second embodiment of the present application, on the basis of the first embodiment, the experimental effect is provided to verify the beneficial effect.

[0214] Preferably, to further illustrate the effectiveness of the application, this embodiment uses the user daily load sequence from 2012 to 2013 in the Ausgrid public data set to construct an experimental data set and test it. First, 32 user daily load sequences with abnormal output or directly controlled load in the experimental data set are removed. After the remaining 268 user daily load sequences are subjected to the abnormal value removal described in step 1.3), the missing value filling described in step 1.4) and the normalization operation described in step 1.5) of the present application, they are divided into training set, validation set and test set according to the ratio of 7:1:2. The net load data and post-table photovoltaic output label data of each user in the Ausgrid public data set are saved in the form of daily load sequence, as shown in Table 1.

[0215] Table 1: Example of user daily load sequence

[0216]

[0217] Further, in this embodiment, to verify the superiority of the present application, the sequence-to-sequence convolutional neural network model (Sequence2Sequence-Convolutional neural network, S2S-CNN), convolutional neural network-bidirectional long short term memory neural network model (Convolutional Neural Network-Bidirectional Long Short Term Memory Neural Network, CNN-BiLSTM), graph attention recurrent neural network-extreme learning machine model (Graph Attention Recurrent Neural Network-Extreme Learning Machine, GARNN-ELM) and sparse attention gated recurrent unit-dictionary learning model (Sparse Attention Gated Recurrent Unit-Dictionary Learning, SAGGRU-DL) with advanced post-table photovoltaic decomposition performance in existing research are selected for comparison with the model constructed in the present application.

[0218] Further, in terms of test indicators, the mean absolute error (MAE), mean absolute percentage error (MAPE) and root-mean-square error (RMSE) are selected as the model decomposition performance evaluation indicators. The calculation formula is as shown in the following formula:

[0219] ;

[0220] ;

[0221] ;

[0222] Further, Table 2 shows the post-meter photovoltaic decomposition performance of the proposed model. The experimental results show that the present application is superior to other models in the three decomposition performance evaluation indicators. Compared with the two spatio-temporal models (SAGNN-DL, GARNN-ELM) that only use mutual information to represent the spatial correlation of user space and perform feature fusion based on dictionary learning, the present application improves by 61.2% and 58.8%, 51.2% and 57.5%, and 61.6% and 59.1% in the three decomposition performance evaluation indicators, respectively. The results show that the adaptive adjacency matrix constructed by the present application can more accurately represent the spatial correlation of user nodes hidden in the net load data, and the selective state transition based on dynamic spatio-temporal graph embedding enhances the spatio-temporal feature fusion process, which can obtain more accurate decomposition results, verifying the superiority of the present application.

[0223] Table 2 Comparison of post-meter photovoltaic decomposition performance of different models

[0224]

[0225] Further, in this embodiment, in order to verify the decomposition robustness of the present application in the data missing abnormal scene, post-meter photovoltaic decomposition performance comparison experiments with data missing rates of 5%, 10% and 15% are carried out. The specific method is as follows: the number of non-zero data in the input of the test set net load data is counted , according to the given data missing rate , randomly select non-zero data to set to 0, and then input each model to simulate the data missing scene of daily load sequence, and calculate and compare the average , and of the post-meter photovoltaic decomposition results of all user nodes of each decomposition model.

[0226] Further, Table 3 shows the post-meter photovoltaic decomposition performance of different models in the data loss abnormal scene. The experimental results show that as the data loss rate increases, the decomposition error of different models increases to a certain extent, and the decomposition performance decreases. When the data loss rate continues to rise, most of the user load data is lost, and the abnormal values in the spatial and temporal features of the post-meter photovoltaic output extracted by the model are inevitable. The SAGNN-DL and GARNN-ELM based on the dictionary learning method cannot select abnormal spatial and temporal features in the spatial and temporal feature fusion stage, resulting in a significant decrease in decomposition accuracy of 94% and 101%. The present application selectively fuses the spatial and temporal features through the selective spatial and temporal state transition process, only focuses on the key information that is conducive to the post-meter photovoltaic decomposition task, thereby reducing the adverse effects of abnormal spatial and temporal features. In the high data loss rate scene of 15%, the decomposition performance still guarantees a decrease of less than 87%, which shows strong decomposition robustness and verifies the superiority of the present application.

[0227] Table 3 Comparison of post-meter photovoltaic decomposition performance of different models under different data loss rates

[0228]

[0229] Further, in the present embodiment, in order to verify the extraction ability of the present application to the post-meter photovoltaic output feature, the principal component analysis (PCA) method is used to reduce the dimension of the post-meter photovoltaic output feature extracted from the net load sequence by each model and the real post-meter photovoltaic output. By comparing the feature space distribution of different models and the spatial and temporal distribution of the real post-meter photovoltaic output, the feature extraction ability of different models can be visually evaluated.

[0230] Further, the present application Figure 4 shows the post-meter photovoltaic output feature extracted by different models and the spatial and temporal distribution of the real post-meter photovoltaic output. From the Figure 4It can be seen that the shape of the dimensionality reduction result of the real spatiotemporal distribution of the post-table photovoltaic output presents a "spindle" type as a whole along the first dimension characteristic direction, and the characteristic distribution of each sequence point is relatively uniform without obvious boundaries. Since S2S-CNN and CNN-BiLSTM do not consider the spatial correlation of users, only the time characteristics of the post-table photovoltaic output are extracted, which leads to insufficient feature extraction, and there is a large blank in the center of the distribution shape. Since SAGGRU-DL and GARNN-ELM only use the mutual information of the net load data to represent the spatial correlation of the user nodes, the spatiotemporal feature extraction result is affected by the different user electricity behavior habits, and the distribution shape presents a clear outline, and the characteristic distribution is more dense, which has a large difference with the real spatiotemporal distribution of the post-table photovoltaic output, and is not conducive to the reconstruction of the post-table photovoltaic output. Among them, SAGGRU-DL introduces more sparse constraints in the feature extraction process, which leads to distortion of the characteristic distribution, and has an adverse effect on the subsequent spatiotemporal feature fusion. The spatiotemporal characteristics of the post-table photovoltaic output extracted by the application are highly similar to the spatiotemporal distribution shape of the real photovoltaic output, and can reflect the spatiotemporal coupling relationship of the post-table photovoltaic output. Therefore, the experimental results verify the superiority of the time sequence feature extraction module and the spatial feature extraction module in spatiotemporal feature extraction, and also verify the effectiveness of the time-space state module from the side.

[0231] The above embodiments are preferred embodiments of the present application, but the embodiments of the present application are not limited by the above embodiments, and any changes, modifications, substitutions, combinations, simplifications made without departing from the spirit and principles of the present application should be equivalent replacement methods, which are all included in the protection scope of the present application.

Claims

1. A post-spectroscopic photovoltaic decomposition method based on spatiotemporal feature adaptive extraction and deep fusion, characterized in that, Includes the following steps: 1) Collect users' net load data and post-meter photovoltaic output label data and preprocess them. Divide the preprocessed net load data and post-meter photovoltaic output label data into training set, validation set and test set; 2) Initialize the adaptive adjacency matrix based on the mutual information value between the net load data, construct the adaptive adjacency matrix that is dynamically updated by the backpropagation mechanism of decomposition error, and after dynamically representing the spatial relationship between user nodes, construct a dynamic spatiotemporal graph containing the spatiotemporal relationship between users. 3) Calculate the decomposition error based on the photovoltaic output label data after the table in the training set, and use the backpropagation mechanism to iteratively update the adaptive adjacency matrix and the learnable parameters in the constructed photovoltaic decomposition model after the table. In all iteration rounds, select the model and adaptive adjacency matrix with the best decomposition performance for the user net load data in the validation set as the optimal model and optimal adaptive adjacency matrix and save them. The photovoltaic decomposition model after the table consists of a time-series feature extraction module, a spatial feature extraction module, and a spatiotemporal state module. The time-series feature extraction module uses the fast Fourier transform method to analyze the net load data to obtain frequency domain information. After periodically decoupling the net load data by combining the time domain and frequency domain information, it extracts and weights the multi-scale periodic features of the photovoltaic output after the table to generate the time-series feature representation of user nodes on the dynamic spatiotemporal diagram. The spatial feature extraction module takes the temporal feature representation of user nodes on the dynamic spatiotemporal graph as input. It first performs global spatial modeling based on the adaptive adjacency matrix of the current iteration round, further characterizing the spatial association relationship of wide-area user nodes. Through a multi-hop aggregation mechanism, it updates the temporal feature representation of user nodes on the dynamic spatiotemporal graph to the spatiotemporal feature representation of user nodes and generates a dynamic spatiotemporal graph embedding. The spatiotemporal state module explicitly models the spatiotemporal association relationship between users based on the dynamic spatiotemporal graph embedding, and then performs a selective spatiotemporal state transition process to deeply fuse the spatiotemporal feature representation of user nodes to generate the photovoltaic decomposition results in the table. 4) During the testing phase, the photovoltaic output of users in the test set is decomposed based on the optimal model and the optimal adaptive adjacency matrix. 2.The post-table photovoltaic decomposition method based on spatio-temporal features adaptive extraction and deep fusion according to claim 1, characterized in that, In step 1), the preprocessing includes outlier removal, missing value filling, and normalization. The photovoltaic system after the meter is a low-voltage household photovoltaic power generation device installed after the smart meter. The user net load data is the net value of the user's actual electricity load minus the output of its photovoltaic system after the meter. 3.The post-table photovoltaic decomposition method based on spatio-temporal features adaptive extraction and deep fusion according to claim 2, characterized in that, The specific steps for step 1) are as follows: 1.1) Collecting the net load data and the post-metered photovoltaic output data of each user in the form of daily load sequences, forming pairs of daily load sequences as shown in the following formula: ； wherein, , and denote the user's day load sequence pair, net load sequence and post-meter photovoltaic output label sequence, respectively; 1.2) Traverse each daily load sequence pair formed in step 1.1) and determine the data missing status of the net load sequence and the photovoltaic output tag sequence after the table. If the data missing rate of any sequence in the daily load sequence pair is higher than 25%, then the daily load sequence pair is discarded. 1.3) Traverse each sequence point in the retained sequence. If the difference between the sequence point and the sequence mean is greater than 3 times the sequence standard deviation, it is regarded as outlier data and removed. After removal, the sequence point is treated as a missing value. 1.4) Determine if there are adjacent sequence points to the missing value. If both adjacent sequence points are missing values, assign the missing value a value of 0. If the sequence point corresponding to the missing value is the first or last sequence point in the sequence, fill the missing value with the unique adjacent sequence point. If both adjacent sequence points to the missing value are normal values, fill the missing value with the average of the adjacent sequence points. If the adjacent sequence points to the missing value are one normal value and one missing value, fill the missing value with the normal value. 1.5) The daily load series is normalized using the Z-Score normalization method, as shown in the following formula: ； ； In the formula, , , and Representing users respectively At sequence point Net load power, net load series mean, net load series standard deviation, and normalized series points Net load power, , , and Representing users respectively At sequence point The table includes the photovoltaic power output, the average value of the photovoltaic power output label sequence, the standard deviation of the photovoltaic power output label sequence, and the normalized sequence points. Photovoltaic power output after the meter; 1.6) Divide all preprocessed daily load sequences into training set, validation set and test set according to a preset ratio.

4. The photovoltaic decomposition method based on spatiotemporal feature adaptive extraction and deep fusion as described in claim 3, characterized in that, In step 2), the spatial association relationship of user nodes is the coupling of the spatial association relationship of photovoltaic power output after the table of neighboring geographical locations and the spatial association relationship of user electricity consumption behavior. The spatiotemporal association relationship between users is the coupling relationship of user net load data in the time dimension and spatial dimension.

5. The photovoltaic decomposition method based on spatiotemporal feature adaptive extraction and deep fusion as described in claim 4, characterized in that, The specific steps for step 2) are as follows: 2.1) Traverse all possible user combinations in the user set, and calculate the user load based on the net load data obtained in the training set in step 1). and users Mutual information values ​​between net load sequences As shown in the following formula: ； In the formula, and Representing users respectively Net load sequence and users At sequence point Net load power; and They represent and The marginal distribution; express and The joint distribution; 2.2) The calculated mutual information value With a given threshold To make a comparison, and The value assigned to and Determine the initial adaptive adjacency matrix The Middle Line 1 Column values As shown in the following formula: ； 2.3) For the initial adaptive adjacency matrix Perform singular value decomposition to calculate the node embedding vector. and As shown in the following formula: ； ； ； In the formula, , and Let these represent the left singular matrix, the singular value diagonal matrix, and the right diagonal matrix, respectively. and They represent taking the matrix respectively The former Column and matrix The former Columns merge and transpose Indicates to retain the previous text A diagonal matrix with singular values ​​and perform... Exponentiation; 2.4) Embedding nodes into vectors and As a learnable parameter vector, an adaptive adjacency matrix is ​​constructed and dynamically updated by the backpropagation mechanism of decomposition error. As shown in the following formula: ； In the formula, and These represent the normalization exponent and the linear rectified activation function, respectively. 2.5) Treat each user in the user set as a node and construct a dynamic spatiotemporal graph that includes the spatiotemporal relationships between users. As shown in the following formula: ； In the formula, , and These represent the user net load sequence matrix, the set of user nodes, and the set of edges connecting user nodes in the graph, respectively.

6. The photovoltaic decomposition method based on spatiotemporal feature adaptive extraction and deep fusion as described in claim 5, characterized in that, The process by which the temporal feature extraction module extracts the temporal feature representation of user nodes on the dynamic spatiotemporal graph from the user net load sequence matrix is ​​as follows: a. User net load sequence matrix Perform numerical embedding operations as shown in the following equation: ； In the formula, This represents the output of the numerical embedding operation. This represents a one-dimensional convolution operation; Then, the user net load sequence matrix The position embedding operation is performed as shown in the following equation: ； In the formula, Represents sequence point Location encoding, Represents the temporal feature dimension. Indicates the scaling factor; b. The outputs of the summation of numerical embedding and positional embedding operations generate the original time-series characteristics of the net payload data. As shown in the following formula: ； In the formula, This indicates the output of the position embedding operation; c. Use the Fast Fourier Transform method to extract the original time-series characteristics of the net load data. Converting from time-domain representation to frequency-domain representation As shown in the following formula: ； In the formula, Indicates the sequence length; Indicates frequency domain index; and Net load data at sequence points The corresponding time-domain representation and in The corresponding frequency domain representation; Indicates the rotation factor. Represents the complex number symbol; Then, calculate The magnitude of each component is calculated and averaged over the temporal feature dimension, as shown in the following formula: ； In the formula, and They represent model and The average value in the temporal feature dimension; d. Importance is indicated by the magnitude of the modulus; the first value is taken. The component that plays a dominant role in the periodic temporal variation of net load data is shown in the following formula: ； In the formula, Indicates taking The largest median value The frequency values ​​corresponding to each component; , and They represent the 1st, 2nd and 3rd respectively. The frequency of the dominant periodic component; Then, the net load data period information is calculated as shown in the following formula: ； In the formula, and These represent the front with the largest amplitude. The first of the periodic dominant components The period and frequency of each component; e. Based on the net load data in the frequency domain The periodicity and frequency information of each periodic dominant component are used to sequentially analyze the original time-series characteristics of the net load data. Divide the data, align it along the time dimension, and finally rearrange it into... Each periodic aligned feature block ,in Represents the real number field; f. Construct a periodic feature extraction model, using successively a size of... The convolutional kernel extracts the periodic features of the periodically aligned feature blocks, as shown in the following formula: ； In the formula, Indicates the preceding The first of the periodic dominant components The component in the first Feature maps obtained from convolutional branches; Indicates the first Two-dimensional convolution operation with multiple convolution branches; Indicates size is The convolution kernel; Then, the periodic features are stacked along the feature dimension to generate the output, as shown in the following equation: ； In the formula, Indicates the preceding The first of the periodic dominant components The periodic characteristics of each component; This indicates a feature map stacking operation; , , , , and In order to represent the previous The first of the periodic dominant components The feature maps obtained by each component in the first to sixth convolutional branches; g. According to the previous... The weighting coefficients for the magnitudes of the periodic dominant components are calculated as follows: ； In the formula, and They represent the first The amplitude of each component before and after activation; Then the periodic features One by one, restore them to the same shape as the original temporal characteristics. Then, weighted aggregation The periodic features corresponding to each periodic dominant component generate the temporal feature representation of user nodes on the dynamic spatiotemporal graph. As shown in the following formula: 。 7. The photovoltaic decomposition method based on spatiotemporal feature adaptive extraction and deep fusion as described in claim 6, characterized in that, The process by which the spatial feature extraction module updates the temporal feature representation of user nodes on the dynamic spatiotemporal graph to the spatiotemporal feature representation of user nodes and generates the dynamic spatiotemporal graph embedding is as follows: a. Extend the adaptive adjacency matrix sequentially from first order using matrix exponentiation. Step, use Order-adaptive adjacency matrix As The probability transition matrix of the jump diffusion process, then... The probability transition matrices of each graph diffusion process are arranged into a graph diffusion tensor. As shown in the following formula: ； In the formula, , and These represent the probability transition matrices for the one-hop, two-hop, and three-hop diffusion processes on the dynamic spatiotemporal graph, respectively. b. Utilizing the spatial relationships of wide-area user nodes represented by the graph diffusion tensor, multi-hop neighbor feature information is aggregated to update the spatiotemporal feature representation of user nodes, generating a dynamic spatiotemporal graph embedding, as shown in the following equation: ； ； In the formula, Represents the graph-diffused convolution kernel; , and These represent the learnable weight parameter matrices for the one-hop, two-hop, and three-hop diffusion processes, respectively. express The learnable weight parameter matrix for the jump diffusion process; This indicates the embedding of a dynamic spatiotemporal graph; This indicates the dimensionality of the spatiotemporal features represented by the user node; This represents the element-wise multiplication operation; This represents the total number of users.

8. The photovoltaic decomposition method based on spatiotemporal feature adaptive extraction and deep fusion as described in claim 7, characterized in that, The process by which the spatiotemporal state module generates the photovoltaic decomposition results after embedding the spatiotemporal feature representation of the user node into a table based on a dynamic spatiotemporal graph is as follows: a. Embed the extracted dynamic spatiotemporal graph Projection as a spatiotemporal state space representation As shown in the following formula: ； In the formula, Represents a linear projection function; b. Representing the spatiotemporal state space Break down into main components and gating components As shown in the following formula: ； In the formula, Indicates a gating split operation; c. Extracting principal components using a one-dimensional convolutional neural network The local dependencies are identified and activated, as shown in the following equation: ； In the formula, Indicates the main components after local enhancement; Indicates the activation function; d. Generate four selective matrices based on dynamic spatiotemporal graph embedding. , , and As shown in the following formula: ； ； ； ； In the formula, Represents the enhanced time-scale parameter matrix; This indicates matrix multiplication according to Einstein's summation convention; The broadcast operation function represents the feature dimension; For activation functions; This represents a learnable parameter matrix with the same shape as the output of the broadcast operation; and These represent the randomly initialized state transition matrix and the enhanced state transition matrix, respectively. The base of the natural logarithm The power function; and These represent the enhanced input control matrix and the identity matrix, respectively. Indicates the output control matrix; e. Based on the four selectively generated matrices , , and A selective spatiotemporal state transition process is performed on the locally enhanced principal components to achieve deep fusion of the locally enhanced principal components. Output Enhanced Spatiotemporal Feature Representation Each discrete sequence point The selective spatiotemporal state transition process is shown in the following equation: ； In the formula, and Representing discrete sequence points respectively The main components of local enhancement and the spatiotemporal features of enhancement are represented; and Representing discrete sequence points respectively and The hidden state; f. The activated gating component Enhanced spatiotemporal feature representation After element-wise multiplication, linear projection generates the final table of photovoltaic decomposition results. As shown in the following formula: 。 9. The photovoltaic decomposition method based on spatiotemporal feature adaptive extraction and deep fusion as described in claim 8, characterized in that, The specific steps for step 3) are as follows: 3.1) Based on the photovoltaic output label data obtained in step 1) in the training set, calculate the decomposition error as shown in the following formula: ； In the formula, Indicates the decomposition error; and Representing users respectively At sequence point The table shows the photovoltaic decomposition results and the photovoltaic output label values. 3.2) Minimize the decomposition error using the backpropagation method to update the adaptive adjacency matrix and the learnable parameters in the post-table photovoltaic decomposition model; 3.3) Iteration steps 3.2) In each iteration, the mean absolute error is calculated based on the photovoltaic output tag data in the validation set obtained in step 1). As a performance evaluation index for decomposition, it is shown in the following formula: ； Then, select The photovoltaic decomposition model and adaptive adjacency matrix corresponding to the minimum iteration round are saved as the optimal model and optimal adaptive adjacency matrix.

Images