Mass distributed photovoltaic data acquisition point planning and power ultra-short-term prediction method

Abstract

The invention discloses a mass distributed photovoltaic data acquisition point planning and power ultra-short-term prediction method, relates to the field of new energy prediction, and constructs a two-stage method including distributed photovoltaic real-time monitoring point selection and mass photovoltaic ultra-short-term prediction by considering complex spatial-temporal correlation of output of a large-scale photovoltaic unit. A distributed photovoltaic sparse set selection strategy is determined based on the geographic position and output characteristics of distributed photovoltaic, and an ultra-short-term prediction model considering photovoltaic output spatial-temporal correlation is established according to distributed photovoltaic real-time monitoring data and in combination with a gated recurrent neural network and a graph convolutional network model. And efficient and accurate prediction of mass distributed photovoltaic is realized.

Description

Technical Field

[0001] The present invention relates to the field of new energy prediction, and in particular to a method for planning massive distributed photovoltaic data collection points and ultra-short-term power prediction. Background Art

[0002] In recent years, with the gradual increase in the penetration rate of photovoltaics in the power system, the intermittency and uncertainty of distributed resources have affected the reliability of the power grid. Ultra-short-term photovoltaic prediction requires the power generation system to be equipped with corresponding sensing and communication infrastructure to complete remote monitoring. The ultra-short-term prediction of distributed photovoltaic power generation helps the power grid to monitor, optimize, control and protect distributed power sources in real time, and thus has attracted widespread attention. Therefore, studying the ultra-short-term prediction method of massive photovoltaics is of great significance to maintaining the safe and stable operation of the power grid and ensuring people's production and life.

[0003] Traditional photovoltaic ultra-short-term prediction mainly focuses on the power prediction of a single distributed photovoltaic. Since photovoltaic output has a strong spatiotemporal correlation, better prediction results can be obtained by distributing multiple photovoltaic power stations through spatiotemporal modeling and establishing an overall prediction model. In recent years, graph neural networks (GNNs) have been widely used to solve the ultra-short-term power prediction problem of large power systems for massive photovoltaic generators, such as graph attention networks (GATs), spatiotemporal GNN networks (STGNNs), optimal graph construction, and graph spatiotemporal attention neural networks (GSTANNs). However, the above studies are only applicable to application scenarios where real-time measurement data of all photovoltaic sites are known; some studies use low-rank tensor learning to deal with the problem of missing data of distributed photovoltaic sites, but it assumes that only a small number of photovoltaic power stations have missing data, which may not be consistent with the actual situation; another algorithm uses the K-means method for photovoltaic clustering to obtain the average position of the photovoltaic cluster, but the centroid represents a virtual photovoltaic power station and does not correspond to a specific photovoltaic unit, so it is impossible to perform actual remote monitoring infrastructure installation. Summary of the invention

[0004] The present invention takes into account the complex spatiotemporal correlation of the output of large-scale photovoltaic units, constructs a two-stage method including the selection of distributed photovoltaic real-time monitoring points and ultra-short-term prediction of massive photovoltaics, determines the distributed photovoltaic sparse set selection strategy based on the geographical location and output characteristics of distributed photovoltaics, and based on the distributed photovoltaic real-time monitoring data, combines the gated recurrent neural network and the graph convolutional network model to establish an ultra-short-term prediction model that takes into account the spatiotemporal correlation of photovoltaic output, thereby achieving efficient and accurate prediction of massive distributed photovoltaics and solving the problems raised in the above-mentioned background technology.

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

[0006] The method for planning massive distributed photovoltaic data collection points and ultra-short-term power prediction includes the following steps:

[0007] ① Considering the geographical location and output characteristics of massive distributed photovoltaics, photovoltaic data collection points are selected according to the clustering algorithm: the distributed photovoltaic data collection points should be a set of representative photovoltaic units that can capture the output power characteristics of the entire photovoltaic generator set. In order to achieve this goal, K-medoids clustering analysis is used to divide all photovoltaic units into groups based on the geographical distribution of distributed photovoltaics. In each group, the distributed photovoltaic data collection point is one of the most representative units, and the sum of the distances between the collection point and all units of the same category is the smallest. K-medoids is an unsupervised algorithm that divides distributed photovoltaics into k categories using the Euclidean distance and the similarity measure between the point and the centroid. In the present invention, the geographical coordinates of all distributed photovoltaic units, i.e., longitude and latitude, are used as the input variables of the K-medoids clustering algorithm. Suppose there are n distributed photovoltaic units in total, and L = {l i}, 1≤i≤n represents the unit coordinates of distributed photovoltaic units, C 1 ,…,C k Represents k clusters among n distributed photovoltaic units. Selected distributed photovoltaic data collection point s 1 ,…,s k It can be obtained by solving the following optimization problem:

[0008]

[0009] Since the center point of each cluster is selected as the distributed photovoltaic data collection point, the number of distributed photovoltaic data collection points is the same as the number of clusters in the K-medoids algorithm, and the number of clusters is determined by economic indicators. After selecting the distributed photovoltaic data collection point, the installation location of the remote monitoring infrastructure can be determined, so that the distributed photovoltaic cluster can perform ultra-short-term photovoltaic forecasts based on actual photovoltaic output data and meteorological measurement data.

[0010] K-medoids algorithm is a classic clustering algorithm based on partitioning. In 1990, Rousseeuw et al. proposed PAM (Partition Around Medoids) and CLARA (Clustering Large Application) algorithms. Most domestic and foreign scholars focus their research on the initialization of cluster centers and the determination of the k value of clustering data, but there is no unified definition method for the initialization of cluster centers and the determination of the number of categories k. K-medoids algorithm is also a commonly used data mining method. It is an unsupervised algorithm that can divide data points with the same category and common attributes into the same class. In the same class, the similarity of cluster objects is very high. Conversely, in different classes, the similarity of cluster objects is very low. Through this partitioning method, we can get more accurate categories than the k-means method. K-medoids algorithm is often used for text clustering in large data sets. The basic principle of the algorithm is to extract the centers of k initial classes from the initial data set, and then divide the remaining points into corresponding classes according to the size of the distance. After obtaining the initial category, the center point is repeatedly replaced by non-center points to improve the clustering quality. K-medoids algorithm uses clustering cost function To determine the clustering quality criterion, when updating the cluster center, a randomly selected non-center point replaces one of the original center points, and then the cost function is recalculated. If the clustering effect is improved, the replacement is successful, otherwise it is restored to the center point before the replacement. After repeated iterations, the clustering ends until the clustering result no longer changes. This enhances the robustness of the algorithm and works well for processing small data sets. In order to achieve the best clustering effect, the sum of square errors must be as small as possible. At this time, it is necessary to compare the difference between the sum of square errors generated by applying the new center point and the initial cluster center to determine whether the newly replaced center point is better than the center point before replacement. If the sum of squares of the intra-class distance after replacement is less than the sum of squares of the intra-class distance before replacement, then the new center point is better than the previous center point. In the repeated replacement process, the cost function continues to decrease until the difference reaches a certain threshold or no longer changes, and the optimal clustering result is obtained. The input of K-medoid is the data set X = {x 1 ,x 2 ,…,x n}, the output is k clusters that meet the minimum variance standard of each cluster center object. The specific algorithm steps are as follows: Step 1: arbitrarily select k objects from n data objects as the initial cluster center representatives; Step 2: Calculate the distance between each object and the center object, and re-divide each object according to the minimum distance; Step 3: arbitrarily select a non-center object x random , calculate x random and the center point pj The change of the total difference function after the exchange; Step 4: If the total difference function is negative, the replacement is accepted, x random The remaining center points that have not been replaced become the k centers of the next clustering; Step 5: Repeat steps 2 to 4 until the clustering result does not change. The basic formula of the K-medoids algorithm is shown below. The formula for the sum of squares of the distances between any sample point and the cluster center point in space is defined as follows:

[0011]

[0012] Among them, X is the training set, that is, the clustering sample, p j It is C j ,j=1,2,…,k. d(X,C) represents the sum of the squares of the distances from all objects in the class to their respective center points. The replacement process of the center point is as follows:

[0013]

[0014] The ultimate goal of clustering is to make the gap between classes as large as possible, while the gap between objects within a class is as small as possible. The larger the gap between categories, the more obvious the distinction between categories; and the smaller the gap between objects within a class, the greater the similarity between objects within the class, thus realizing the basic principle of clustering. When the K-medoids algorithm selects a new center point each time, it calculates the difference between the sum of the squares of the distances from all data points in the original category to the class center and the sum of the squares of the distances from all data points after replacement to the new center point, and takes the center point that makes the total difference function negative as the new class center, avoiding the influence of the algorithm on outliers and having strong robustness. At the same time, the clustering results are independent of the input order of the data object points, so the clustering method also has the invariance of data object translation and orthogonal transformation.

[0015] ② Combine the gated recurrent neural network and graph convolutional network technology to establish a massive photovoltaic ultra-short-term prediction framework that considers spatiotemporal correlation: The time correlation of photovoltaic ultra-short-term output power is established through a gated recurrent neural network (GRU). The gated recurrent neural network adds a gating mechanism to the simple recurrent neural network to control the transmission of information in the neural network. It can better capture the dependencies with larger step sizes in the time series and solve the gradient decay or explosion problems in long-term memory and back propagation. Compared with the LSTM network, the GRU network has the advantages of fewer parameters, simple structure, and higher computational efficiency, and is more suitable for building larger networks. At present, the GRU network has been proven to have achieved very good results in certain application scenarios. The GRU network is mainly composed of two gated units. The introduction of a reset gate helps capture short-term dependencies in the time series, and the introduction of an update gate helps capture the time series. The inputs of the reset gate and update gate of the gated unit are both the current time step input X t With the previous time step hidden state H t-1 , the output is calculated by the fully connected layer with sigmoid activation function. Assuming the number of hidden units is h, given a small batch input X at time step t t ∈i n×h (the sample length is n, the number of input feature vectors is d) and the hidden state H of the previous time step t-1 ∈i n×h . Reset Gate R t ∈i n×h and update gate Z t ∈i n×h The calculation of is as follows:

[0016] R t =σ gru (X t W xr +H t-1 W hr +b r )

[0017] Z t =σ gru (X t W xz +H t-1 W hz +b z )

[0018] Among them, W xr , Wxz∈i d×h and W hr , W hz ∈i h×h is the network weight parameter; b r , b z is the network bias parameter; σ gruis the sigmoid activation function, which mainly converts the value range of each element in the reset gate and the update gate to between 0 and 1. t * ∈i n×h It is obtained through a series of operations on the output of the reset gate at the current time step and the hidden state at the previous time step. When the element value in the reset gate is close to 0, it means that the hidden state of the previous time step will be discarded. When the element value is close to 1, it means that the hidden state of the previous time step will be retained. The calculation method is as follows:

[0019]

[0020] Among them, W xh ∈i d×h and W hh ∈i h×h is the network weight parameter; b h is the network bias parameter; the Tanh activation function is to convert all element values ​​to between [-1,1]. In the LSTM network, the input gate and the forget gate are complementary and have a certain degree of redundancy. Unlike the LSTM network, the GRU network directly uses an update gate to control the balance between input and forgetting. The hidden state H at time step t t ∈i n×h is the update gate Z through the current time step t The hidden state H of the previous time step t-1 and the candidate hidden state H at the current time step t * Combining operations to get:

[0021]

[0022] From the above formula, we can see that when Z t =0,R t =1, the GRU network will degenerate into a simple recurrent neural network; when Z t =0,R t = 0, the current state is H t Only with the current input X t Related to the historical state H t-1 It doesn't matter. t =1, the current state is H t It is equal to the hidden state H at the previous moment t-1 .

[0023] The spatial correlation of photovoltaic output is completed through graph convolutional networks. Graph structured data is a kind of non-Euclidean data, which consists of nodes and edges and can represent various complex relationships, such as social networks, knowledge graphs, molecular structures, etc. Convolutional neural networks have been proven to have the ability to extract multi-scale local spatial features and construct highly expressive features. They can process conventional Euclidean data such as images and texts. Their local connections, weight sharing, and multi-level structures are also important for graph problems. However, the distribution and structure of graph data often change dynamically, which makes it difficult to define the CNN convolution kernel, and the parameters need to be constantly updated, making it difficult to perform stable training. On this basis, the graph convolutional network extends the convolution operation from traditional grid data (such as images) to graph data. Graph convolutional neural network is a new type of image processing method first proposed by Kipf's team. Similar to the function of CNN, GCN can also quickly and effectively extract graph characteristics from graph data, and has important application value in graph node recognition, graph edge prediction, and graph embedded representation. Assume that there are N nodes in a graph data, each node has its own features, and these features form an N×D-dimensional matrix X. Then the relationship between each node will form an N×N-dimensional matrix A, also known as the adjacency matrix. X and A are the inputs of GCN. For each node, extract the feature information of the node from each of its neighboring nodes, including the features of the node itself; select an appropriate activation function and perform the same operation on all nodes; input these calculated feature quantities into the neural network to obtain valid numerical values ​​for the feature quantities. GCN can effectively utilize the topological structure and node features of the graph to extract high-dimensional semantic representations of the nodes, thereby improving the performance of graph-related tasks. There are two main types of GCN, namely spectral domain graph convolution and spatial domain graph convolution. Spectral domain-based graph convolution first processes the graph data through the Laplacian matrix. According to spectral graph theory, the adjacency matrix is ​​represented as A. A simple Laplacian matrix is ​​defined as L=DA∈i n×n . D=diag(d(v 1 ), ..., d(v N ))∈i n×n is the diagonal degree matrix, where d(·) is the number of nodes v i The normalized Laplacian matrix is ​​then defined as Where I is the identity matrix. Obviously, the Laplacian matrix L is a real symmetric matrix. Given a joint point v i The corresponding input vector f is the output vector h obtained by calculating the product of the Laplace matrix L and f. Its physical meaning can be explained by the following formula:

[0024] h=Lf=(DA)f=Df-Af

[0025] Among them, the output vector h represents the joint point vi Its adjacent joint point v j At the same time, the Laplace matrix is ​​also a semi-positive matrix, which can be proved by the following formula. The quadratic form of the Laplace formula L is as follows:

[0026]

[0027] As shown in the above formula, the quadratic form of the Laplacian matrix L is the sum of the squares of the differences between each joint point in the graph and its neighbors. From two perspectives, the physical meaning of the Laplacian matrix is ​​that it is a measure of the difference between each node in the graph and its neighbors. This shows that the adjacency matrix provides the connection strength of the edges between nodes. Next, the spectral domain graph convolution operation is performed. The crucial Laplacian matrix L is the basic content of the spectral domain graph convolution operation. Convolution in the joint point domain cannot be roughly expressed as a meaningful operator. However, according to the graph convolution theorem, the convolution operator* G In the spectral domain it is defined as:

[0028] w* G h=U((U T w)e(U T ,h))

[0029] Among them, U=[u 0 ,u 1 , ..., u n-1 ]∈i n×n is the Fourier basis. In the Fourier basis, the spectral convolution of the signal w is defined as w = U T w. Fourier transform U of the convolution of two signals (w,h) T is equal to the dot product of their Fourier transforms. From this, we can get the spectral domain graph convolution formula:

[0030]

[0031] in, Δ=UΛU T ∈i n×n ; Λ=diag([λ 0 ,λ 1 , …, λ n-1 ])∈i n×n , where λ i represents the eigenvalue of the Laplacian matrix. w(Λ) is the filter to be learned in the spectral domain graph convolution operation.

[0032] In the present invention, the goal of photovoltaic prediction is to predict the photovoltaic output in a certain period of time in the future based on the historical information of the neighboring photovoltaic information collection points. The historical information includes historical photovoltaic output information and meteorological measurement data, and is defined as follows: Definition 1: Photovoltaic network G. An unweighted photovoltaic network G = (V, E) is used to describe the geographical location structure of photovoltaics, where V represents the node set, V = {v 1 ,v 2 ,…,v N}, N is the number of nodes, E is the set of edges. The adjacency matrix A represents the connection relationship of photovoltaics, A∈R N×N , the adjacency matrix contains only 0 and 1 elements, 0 means no connection. Definition 2: Feature matrix X N×P Where P represents the number of node attribute features, using X t ∈R N×i represents the power of each photovoltaic base at time i. Therefore, the spatiotemporal prediction problem of photovoltaic can be transformed into training the mapping function f based on the photovoltaic network G and the feature matrix X, and predicting the output in the next period:

[0033] [X t+1 ,…,X t+T ] = f(G; (X t-n ,…,X t-1 ,X t ))

[0034] Where n is the length of the historical sequence; f is the length of the predicted time series. Obtaining the spatial correlation of photovoltaic units is a key issue in photovoltaic forecasting. Traditional convolutional neural networks can obtain local spatial features, but can only be used in Euclidean space, such as images, regular grids, etc. The geographic location is in graphical form, which means that the CNN model cannot reflect the complex correlation of photovoltaic geography and therefore cannot accurately capture spatial dependencies. Graph convolutional networks (GCNs) can process arbitrary graph-structured data. GCN models have been successfully used in many applications, including document classification, image classification, and unsupervised learning. Given an adjacency matrix A and a feature matrix X, the GCN model constructs a filter in the Fourier domain and acts on the graph nodes, mining the spatial features between nodes through their first-order neighbors, and superimposing multiple convolutional layers to establish a GCN model, which can be expressed as:

[0035]

[0036] Among them, A=A+I N Represents the adjacency matrix with added self-connection, I N is the identity matrix, D is the degree matrix, H (l) represents the output of the lth layer; θ (l)represents the parameters of the first layer; σ(·) represents the nonlinear sigmoid function model. In the present invention, a 2-layer GCN model is selected to obtain the spatial dependency relationship, which is expressed as:

[0037] f(X,A)=σ(AReLU(AXW 0 )W 1 )

[0038] in, is the preprocessing step; W 0 ∈R P×H represents the weight matrix layer from the input layer to the hidden layer; P is the length of the feature matrix; H is the number of neurons in the hidden layer; W 1 ∈R H×T represents the weight matrix from the hidden layer to the output layer; f(X,A)∈R N×T Represents an output of length T; ReLU() represents a rectified linear unit, which is a commonly used activation function in deep neural networks. In summary, the present invention uses the GCN model to learn the spatial characteristics of photovoltaics. As shown in the figure below, assuming that node 1 is a photovoltaic information collection point, the GCN model can obtain the geographical topological relationship of adjacent photovoltaic units, and then obtain spatial dependence to predict photovoltaic output. Time correlation is another key issue in photovoltaic output prediction. At present, the most widely used neural network model is the recurrent neural network (RNN). Due to defects such as gradient disappearance and gradient explosion, traditional recurrent neural networks are not suitable for long-term series prediction problems. The LSTM model and the GRU model are variants of the recurrent neural network and have been proven to solve the above problems. The basic principle of LSTM is roughly the same as that of GRU, that is, a gating mechanism is used to memorize as much long-term information as possible. However, due to the complex structure of LSTM, its training time is relatively long. Compared with the LSTM model, the GRU model has a relatively simple structure, uses fewer parameters, and has faster training capabilities. Therefore, the present invention selects the GRU model to train time correlation. The input variables of GRU include the photovoltaic features at time t and the hidden state at time t-1. While obtaining the features at the current moment, the model still retains the trend of changes in historical photovoltaic output and time correlation features. During the training process, the prediction goal is to reduce the error between the actual power generation of the photovoltaic unit and the predicted value. t and Y t Represents the actual power and predicted power. The loss function of the T-GCN model is as follows:

[0039] loss=||Y t -Y t ||+λL reg

[0040] Among them, the first term is used to minimize the error between the actual output and the predicted output; the second term is a regularization term, which helps to avoid overfitting problems; λ is a hyperparameter. Through spatial correlation and temporal correlation modeling, the present invention adopts a temporal graph convolutional network model (Temporal Graph Convolutional Network, T-GCN) based on graph convolutional network and gated recurrent unit. The specific calculation process is as follows:

[0041] u t =σ(W u [f(A,X t ),h t-1 ]+b u )

[0042] r t =σ(W r [f(A,X t ),h t-1 ]+b r )

[0043] c t =tanh(W c [f(A,X t ),(r t *h t-1 )]+b c )

[0044] h t =u t *h t-1 +(1-u t )*c t

[0045] Among them, f(A,X) represents the convolution process of the graph, which is defined in equation (3.15); W and b represent the weights and biases in training. In summary, the T-GCN model can handle complex problems with spatial dependence and temporal dynamics: on the one hand, the graph convolutional network is used to capture the geographical location relationship of photovoltaic units and obtain the spatial correlation of the output of adjacent photovoltaic units; on the other hand, the gated recurrent unit captures the dynamic changes of photovoltaic output and realizes short-term power prediction in photovoltaic units with time correlation.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention takes into account the complex spatiotemporal correlation of the output of large-scale photovoltaic units, constructs a two-stage method including the selection of distributed photovoltaic real-time monitoring points and ultra-short-term prediction of massive photovoltaics, determines the distributed photovoltaic sparse set selection strategy based on the geographical location and output characteristics of distributed photovoltaics, and based on the distributed photovoltaic real-time monitoring data, combines the gated recurrent neural network and the graph convolutional network model to establish an ultra-short-term prediction model that takes into account the spatiotemporal correlation of photovoltaic output, thereby achieving efficient and accurate prediction of massive distributed photovoltaics. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 The figure is a flow chart of the method of the present invention.

[0048] Figure 2 Schematic diagram of the T-GCN prediction framework implemented in the present invention.

[0049] Figure 3 Schematic diagram of determining the number of collection points using the elbow method in an embodiment of the present invention.

[0050] Figure 4 The figure is a schematic diagram of the prediction effect of the algorithm proposed in the embodiment of the present invention on a certain distributed photovoltaic.

[0051] Figure 5 It is a schematic diagram of the prediction effect of the algorithm proposed in the embodiment of the present invention in a certain cluster. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] See also Figure 1 In the embodiments of the present invention, ① the geographical location and output characteristics of massive distributed photovoltaics are considered, and photovoltaic data collection points are selected according to the clustering algorithm; ② the gated recurrent neural network and graph convolutional network technology are combined to establish a massive photovoltaic ultra-short-term prediction framework that considers spatiotemporal correlation.

[0054] As a further embodiment of the present invention, please refer to Figure 2 ,This embodiment involves the power output data of 13 distributed photovoltaic power plants using smart inverters in the Boston metropolitan area, and the data comes from the public access dataset SOLRENVIEW. The distributed photovoltaic units include various types, including residential, commercial and public utility photovoltaic systems, and their geographic coordinates are shown in Table 1:

[0055] Table 1 Distributed photovoltaic geographic information

[0056]

[0057]

[0058] There are 3 distributed photovoltaic power generation data collection points, and the clustering is shown in Table 2:

[0059] Table 2 Distributed photovoltaic clustering and collection point settings

[0060]

[0061] The present invention uses the power output data of 15 distributed photovoltaic power stations using smart inverters in the Boston metropolitan area. The data set includes photovoltaic output data every 10 minutes from March 5, 2023 to May 28, 2023, that is, 144 data points per day. Taking into account the power generation window between sunrise and sunset, data points from 6 am to 19 pm are selected. The prediction model proposed in the present invention predicts massive distributed photovoltaics based on the real-time measurement data of distributed photovoltaic power generation data collection points. Affected by system size, efficiency, pollution, shading, aging and other conditions, different photovoltaic units have different output powers under the same weather conditions. In order to avoid the impact of data scale on the prediction effect, each distributed photovoltaic output data is normalized:

[0062]

[0063] The present invention divides the data into a training set and a test set, the first 80% of the data points are used for training, and the last 10% are used for testing, and two prediction models, WEIGHTED AVG and TOP M LSTM, are used as comparison methods. WEIGHTED AVG: This method is the most direct solution when there is no other high-precision prediction method available. The main idea of ​​this method is that the power output of neighboring photovoltaic units is similar, so the calculation method for the per-unit output value of the photovoltaic units that are not selected as data collection points is: the average value of the normalized output power of the neighboring distributed photovoltaic data collection points, and the calculation formula is as follows:

[0064]

[0065] in, is the predicted output power of the ith PV unit at time t+1; M i Represents distributed photovoltaic data collection point; d i,j = Dist(l i ,l j) represents the distance between two photovoltaic units. Photovoltaic prediction of distributed photovoltaic data collection points is performed through the GRU model. TOP M GRU: Similar to the WEIGHTED AVG algorithm, TOP M GRU uses the measurement data of distributed photovoltaic data collection points for power prediction. The difference between the two algorithms is: WEIGHTED AVG does not use historical data sequences, but only uses the power data of distributed photovoltaic data collection points at the current moment; in order to overcome this defect, TOP MGRU adopts the standard GRU and uses the historical output sequence of distributed photovoltaic data collection points as the input feature of the power prediction of ordinary photovoltaic units. The present invention sets the time step of GRU to 12 data points (i.e., 2 hours of photovoltaic output data), adopts the Adam optimizer, the learning rate is 0.001, the number of iterations is 200, and the number of batches is 32.

[0066] See also Figure 2 , the T-GCN prediction framework in the embodiment of the present invention: the left side is the spatiotemporal photovoltaic output prediction, and the right side is the specific structure of the T-GCN unit: h t-1 is the output at time t-1; GC represents the graph convolution process; u t , R t is the update gate and reset gate at time t; h t is the output at time t, and the specific calculation process is:

[0067] u t =σ(W u [f(A,X t ),h t-1 ]+b u )

[0068] r t =σ(W r [f(A,X t ),h t-1 ]+b r )

[0069] c t =tanh(W c [f(A,X t ),(r t *h t-1 )]+b c )

[0070] h t =u t *h t-1 +(1-u t )*c t

[0071] Among them, f(A,X) represents the convolution process of the graph; W and b represent the weights and biases in training.

[0072] See also Figure 3 ,Selecting the appropriate number of cluster centers is a key step in the K-medoids clustering algorithm.,The present invention uses the elbow method to determine the cluster center k value.,The steps of the elbow method are to draw a mapping curve between the objective function and the number of clusters, and then find the "elbow" node in the curve. Figure 3 The impact of the number of photovoltaic data collection points on prediction accuracy is evaluated: the number of clusters k ranges from 1 to 14. When the k value increases from 1 to 4, the objective function value K-medoids drops sharply, but the rate of decline gradually slows down; when the k value is greater than 6, the objective function value tends to remain unchanged; therefore, selecting a small number (3 photovoltaic data collection points are used in this invention) of photovoltaic units to install remote monitoring infrastructure is sufficient to achieve accurate prediction of massive distributed photovoltaics. When the number of photovoltaic data collection points continues to increase, the incremental benefit will drop rapidly. This phenomenon also verifies the idea and significance of sparse measurement.

[0073] Table 3 shows the prediction effect of a single photovoltaic, using the root mean square error RMSE as the prediction accuracy measurement standard. The prediction results show that among the 13 distributed photovoltaics, the T-GCN model has the highest prediction accuracy and performs better. Figure 4 This is a schematic diagram of the forecast of distributed photovoltaics on a sample day.

[0074] Table 3 RMSE (pu) of single distributed photovoltaic prediction

[0075]

[0076] Since the power grid may be affected by the aggregation behavior of distributed photovoltaics, Figure 5 The prediction effects of different models on the overall photovoltaic cluster are shown. Table 4 shows the RMSE of the overall prediction of each photovoltaic cluster and all distributed photovoltaics. Since the fluctuation amplitude of the photovoltaic cluster is smaller than that of a single photovoltaic unit, the prediction accuracy of T-GCN is improved after photovoltaic aggregation. From the prediction results, it can be seen that the T-GCN model is suitable for real-time monitoring of large-scale distributed photovoltaic power generation units, and has the advantages of small prediction error and high stability.

[0077] Table 4 Photovoltaic cluster prediction RMSE (pu)

[0078]

[0079]

[0080] Among the above algorithms, what is original to the present invention, has never been disclosed, and its working method is different from any existing literature records: the present invention takes into account the complex spatiotemporal correlation of the output of large-scale photovoltaic units, constructs a two-stage method including distributed photovoltaic real-time monitoring point selection and massive photovoltaic ultra-short-term prediction, determines the distributed photovoltaic sparse set selection strategy based on the geographical location and output characteristics of distributed photovoltaics, and establishes an ultra-short-term prediction model considering the spatiotemporal correlation of photovoltaic output based on the distributed photovoltaic real-time monitoring data, combining the gated recurrent neural network and the graph convolutional network model, to achieve efficient and accurate prediction of massive distributed photovoltaics.

[0081] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

[0082] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. Massive distributed photovoltaic data collection point planning and power ultra-short-term prediction method, characterized by: The steps include: ① Considering the geographical location and output characteristics of massive distributed photovoltaics, photovoltaic data collection points are selected according to the clustering algorithm; Using K-medoids cluster analysis, all photovoltaic units are divided into groups based on the geographical distribution of distributed photovoltaics. In each group, the sum of the distances between the distributed photovoltaic data collection point and all units of the same category is the smallest. K-medoid is an unsupervised algorithm that divides distributed photovoltaics into k categories using the Euclidean distance and the similarity measure between the point and the centroid. The input of K-medoid is the data set X = {x1, x2, …, x n }, the output is k clusters that meet the minimum variance standard of each cluster center object. The specific algorithm steps are as follows: Step 1: arbitrarily select k objects from n data objects as the initial cluster center representatives; Step 2: Calculate the distance between each object and the center object, and re-divide each object according to the minimum distance; Step 3: arbitrarily select a non-center object x random , calculate x random and the center point p j The change of the total difference function after the exchange; Step 4: If the total difference function is negative, the replacement is accepted, x random The remaining center points that have not been replaced become the k centers of the next clustering; Step 5: Repeat steps 2 to 4 until the clustering result does not change; The formula for the sum of squares of distances between any sample point and the cluster center point in space is defined as follows: Among them, X is the training set, that is, the clustering sample, p j It is C j ,j=1,2,…,k, d(X,C) represents the sum of the squares of the distances from all objects in the class to their respective center points. The replacement process of the center point is as follows: Each time the K-medoids algorithm selects a new center point, it calculates the difference between the sum of the squares of the distances from all data points in the original category to the class center and the sum of the squares of the distances from all data points after replacement to the new center point, and takes the center point that makes the total difference function negative as the new class center. The geographical coordinates of all distributed photovoltaic units, i.e., longitude and latitude, are used as the input variables of the K-medoids clustering algorithm. Assuming that there are n distributed photovoltaic units in total, L = {l i }, 1≤i≤n represents the unit coordinates of distributed photovoltaic units, C1,…,C k represents k clusters among n distributed photovoltaic units, and the selected distributed photovoltaic data collection points s1,…,s k It is obtained by solving the following optimization problem: Since the center point of each cluster is selected as the distributed photovoltaic data collection point, the number of distributed photovoltaic data collection points is the same as the number of clusters in the K-medoids algorithm. The number of clusters is determined by economic indicators. After the distributed photovoltaic data collection points are selected, the installation location of the remote monitoring infrastructure is determined, so that the distributed photovoltaic cluster can perform photovoltaic ultra-short-term prediction based on the actual photovoltaic output data and meteorological measurement data; ② Combining gated recurrent neural network and graph convolutional network technology, a massive photovoltaic ultra-short-term prediction framework that considers temporal and spatial correlation is established, specifically considering the temporal correlation and spatial correlation of photovoltaic output; The goal of photovoltaic prediction is to predict the photovoltaic output in a certain period of time in the future based on the historical information of the neighboring photovoltaic information collection points. The historical information includes historical photovoltaic output information and meteorological measurement data, and is defined as follows: Definition 1: Photovoltaic network G, an unweighted photovoltaic network G = (V, E) is used to describe the geographical location structure of photovoltaics, V represents the node set, V = {v1, v2, ..., v N }, N is the number of nodes, E represents the set of edges, and the adjacency matrix A represents the connection relationship of photovoltaics, A∈R N×N , the adjacency matrix contains only 0 and 1 elements, 0 means no connection; Definition 2: The feature matrix X N×P , where P represents the number of node attribute features, using X t ∈R N×i Represents the power of each photovoltaic base at time i. The spatiotemporal prediction problem of photovoltaics is transformed into training the mapping function f based on the photovoltaic network G and the feature matrix X, and predicting the output in the next period: [X t+1 ,…,X t+T ]=f(G;(X t-n ,…,X t-1 ,X t )) Among them, n is the length of the historical sequence; f is the length of the predicted time series; during the training process, the prediction goal is to reduce the error between the actual power generation of the photovoltaic unit and the predicted value, using Y t and Y t Representing the actual power and predicted power, the loss function of the T-GCN model is as follows: loss=||Y t -Y t ||+λL reg Among them, the first term is used to minimize the error between the actual output and the predicted output; the second term is the regularization term; λ is a hyperparameter; given an adjacency matrix A and a feature matrix X, the GCN model constructs a filter in the Fourier domain and acts on the graph nodes, mining the spatial features between nodes through their first-order neighbors, and superimposing multiple convolutional layers to establish the GCN model, which can be expressed as: Among them, A=A+I N Represents the adjacency matrix with added self-connection, I N is the identity matrix, D is the degree matrix, H (l) represents the output of the lth layer; θ (l) represents the parameters of the lth layer; σ(·) represents the nonlinear sigmoid function model. A 2-layer GCN model is selected to obtain the spatial dependency, which is expressed as: f(X,A)=σ(AReLU(AXW0)W1) in, is the preprocessing step; W0∈R P×H represents the weight matrix layer from the input layer to the hidden layer; P is the length of the feature matrix; H is the number of neurons in the hidden layer; W1∈R H×T represents the weight matrix from the hidden layer to the output layer; f(X,A)∈R N×T Represents the output of length T; ReLU() represents the rectified linear unit, which is a commonly used activation function in deep neural networks; the GCN model is used to learn the spatial characteristics of photovoltaics. Assuming that node 1 is the photovoltaic information collection point, the GCN model obtains the geographical topological relationship of adjacent photovoltaic units, and then obtains spatial dependence to predict photovoltaic output; the GRU model is selected to train time correlation, h t-1 represents the hidden layer state at time t-1; X t is the photovoltaic characteristic information at the tth moment; r t is the reset gate, which is used to control the degree of ignoring the state information of the previous moment; u t is the update gate, which is used to control the degree of influence of the previous state on the current state; C t The storage content at the tth moment, that is, the output state, the input variables of GRU include the photovoltaic characteristics at the tth moment and the hidden state at the t-1th moment. While obtaining the characteristics of the current moment, the model retains the changing trend and time correlation characteristics of the historical photovoltaic output, and is modeled through spatial correlation and time correlation.

Images