Power source load prediction method and device based on graph auto-encoder

Abstract

The invention relates to a power source load prediction method and device based on a graph auto-encoder. The method comprises the following steps: acquiring original power source load data associated with a source load prediction task from real-time power source load data according to prediction demand information of the source load prediction task; inputting the original power source load data into a pre-trained power source load prediction model, and processing the original power source load data based on a graph attention mechanism to obtain an original data feature graph; through a graph auto-encoder in a pre-trained power source load prediction model, adaptively extracting features from the original data feature graph and carrying out dimension reduction and simplification processing to obtain a coded feature graph; and inputting the coding feature graph into a graph convolution prediction network in a pre-trained power source load prediction model to obtain a power source load prediction result of the source load prediction task. By adopting the method, simple and efficient coding features can be adaptively extracted based on the graph auto-encoder, data support is provided for graph convolution prediction network processing, and the source load prediction effect can be effectively improved.

Description

Technical Field

[0001] The present application relates to the field of electric power technology, and in particular to a method, device, computer equipment, computer-readable storage medium and computer program product for predicting electric power source load based on a graph autoencoder. Background Art

[0002] Under the "dual carbon" goal and the construction of a new power system, the energy structure is transforming, and the penetration rate of renewable energy such as wind power and photovoltaics in the power system is gradually increasing. Among them, with the increase in the proportion of flexible loads (such as electric vehicles, smart homes, etc.), the multiple uncertainties of source and load have a great impact on the planning, prediction, analysis, decision-making, and control of the power system. The key to the operation of the power system is to meet the real-time supply and demand balance of electric energy. In-depth research on the randomness of source and load power, ensuring the high accuracy of source and load prediction, and ensuring the reliability of the unit output power generation plan are key factors in ensuring the stable operation of the power system.

[0003] In recent years, graph neural networks (GNN) have become a research hotspot in the source-load forecasting business. However, existing research focuses on obtaining available feature information for source-load forecasting tasks under different requirements. This is usually done through human experience model selection or screening based on correlation analysis criteria. There are problems such as complex feature screening process, high dependence on human experience, and difficulty in extracting highly correlated feature signals, resulting in poor power source-load forecasting results. Summary of the invention

[0004] Based on this, it is necessary to provide a method, device, computer equipment, computer-readable storage medium and computer program product based on a graph autoencoder that can improve the power source and load prediction effect in response to the above technical problems.

[0005] In a first aspect, the present application provides a method for predicting power source load based on a graph autoencoder, comprising:

[0006] According to the forecast demand information of the source load forecast task, original power source load data associated with the source load forecast task is acquired from the real-time power source load data;

[0007] Inputting the original power source and load data into a pre-trained power source and load prediction model, processing the original power source and load data based on a graph attention mechanism, and obtaining an original data feature graph;

[0008] Adaptively extract features from the original data feature graph through the graph autoencoder in the pre-trained power source and load prediction model, perform dimensionality reduction and simplification processing on the feature graph, and obtain an encoded feature graph;

[0009] The encoded feature graph is input into the graph convolution prediction network in the pre-trained power source and load prediction model to obtain the power source and load prediction result of the source and load prediction task; the power source and load prediction result is used to characterize the predicted power energy demand and load demand in the power system.

[0010] In one embodiment, the method further comprises:

[0011] Obtain sample power source and load data containing different sample prediction tasks and construct a training data set;

[0012] Constructing a corresponding data graph structure according to each of the sample power source and load data in the training data set, and processing based on a graph attention mechanism to obtain an original feature graph of each sample;

[0013] By constructing the graph autoencoder and the graph convolution prediction network, a power source load prediction model to be trained is obtained;

[0014] Based on the difference between the sample encoding features generated by the graph autoencoder and the sample original feature graph, and based on the difference between the sample prediction results generated by the graph convolutional prediction network and the sample real data, the power source and load prediction model to be trained is trained to obtain the pre-trained power source and load prediction model.

[0015] In one embodiment, the step of obtaining sample power source and load data including different sample prediction tasks and constructing a training data set includes:

[0016] According to the forecast demand information of each sample forecast task, sample power source and load data associated with each sample forecast task is obtained from historical power source and load data;

[0017] The training data set is obtained by performing data preprocessing on each of the sample power source and load data;

[0018] The data preprocessing includes any one or more of the following:

[0019] Outlier cleaning, interpolation filling, data transformation, normalization processing, standardization processing, and calculation of power-related statistical indicators.

[0020] In one embodiment, the corresponding data graph structure is constructed according to each of the sample power source load data in the training data set, and is processed based on a graph attention mechanism to obtain the original feature graph of each sample, including:

[0021] According to the prediction type determined from the prediction demand information of each sample prediction task, a data graph structure corresponding to each sample power source and load data is constructed; different prediction types correspond to different data graph structures;

[0022] The graph attention mechanism is used to adjust the feature expression of the nodes in each of the data graph structures to obtain the original feature graph of each of the samples.

[0023] In one embodiment, the method further comprises:

[0024] Encoding the original feature graph of the sample by an encoder in the graph autoencoder to generate the sample encoding feature based on low-dimensional feature representation;

[0025] The sample encoding features are decoded by a decoder in the graph autoencoder, and reversely transformed to obtain the original feature graph of the sample based on high-dimensional feature representation.

[0026] In one embodiment, the training of the power source load prediction model to be trained based on the difference between the sample encoding features generated by the graph autoencoder and the sample original feature graph, and the difference between the sample prediction results generated by the graph convolution prediction network and the sample real data, includes:

[0027] Using a first loss function for the encoder and the graph convolution prediction network and a second loss function for the decoder, based on the difference between the sample encoding feature and the sample original feature map, and the difference between the sample prediction result and the sample real data, the power source load prediction model to be trained is trained;

[0028] Among them, the first loss function is used to control the data difference between the sample prediction result and the sample real data to be minimum; the second loss function is used to control the feature difference between the sample encoding feature and the sample original feature map to be minimum.

[0029] In a second aspect, the present application also provides a power source load prediction device based on a graph autoencoder, comprising:

[0030] A real-time data acquisition module, used to acquire original power source and load data associated with the source and load prediction task from real-time power source and load data according to the prediction demand information of the source and load prediction task;

[0031] An original feature graph acquisition module is used to input the original power source and load data into a pre-trained power source and load prediction model, process the original power source and load data based on a graph attention mechanism, and obtain an original data feature graph;

[0032] A coding feature map acquisition module is used to adaptively extract features from the original data feature map and perform dimensionality reduction and simplification processing through the map autoencoder in the pre-trained power source and load prediction model to obtain a coding feature map;

[0033] The power source and load prediction result obtaining module is used to input the encoded feature graph into the graph convolution prediction network in the pre-trained power source and load prediction model to obtain the power source and load prediction result of the source and load prediction task; the power source and load prediction result is used to characterize the predicted power energy demand and load demand in the power system.

[0034] In a third aspect, the present application further provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the method for predicting power source load based on graph autoencoder as described in the first aspect are implemented.

[0035] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the method for predicting power source load based on a graph autoencoder as described in the first aspect are implemented.

[0036] In a fifth aspect, the present application also provides a computer program product, including a computer program, which, when executed by a processor, implements the steps of the method for predicting power source load based on graph autoencoder as described in the first aspect.

[0037] The above-mentioned method, device, computer equipment, computer-readable storage medium and computer program product for power source and load prediction based on graph autoencoder obtains the original power source and load data associated with the source and load prediction task from the real-time power source and load data according to the predicted demand information of the source and load prediction task, and then inputs the original power source and load data into the pre-trained power source and load prediction model, processes the original power source and load data based on the graph attention mechanism, obtains the original data feature map, and adaptively extracts features from the original data feature map through the graph autoencoder in the pre-trained power source and load prediction model to perform dimensionality reduction and simplification processing to obtain the encoding feature map, and then inputs the encoding feature map into the graph convolution prediction network in the pre-trained power source and load prediction model to obtain the power source and load prediction result of the source and load prediction task. The power source and load prediction result is used to characterize the predicted power energy demand and load demand in the power system, realizes the optimization of the power source and load prediction processing, and can adaptively extract concise and efficient encoding features based on the graph autoencoder according to the different needs of the source and load prediction task, provides data support for the graph convolution prediction network processing, and can effectively improve the source and load prediction effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the drawings required for use in the embodiments of the present application or related technical descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0039] FIG1 is a schematic flow chart of a method for predicting power source load based on a graph autoencoder in one embodiment;

[0040] FIG2 is a schematic diagram of a graph convolution power source load prediction process in combination with a graph autoencoder in one embodiment;

[0041] FIG3a is a schematic diagram of a data graph structure in one embodiment;

[0042] FIG3 b is a schematic diagram of a data graph structure in another embodiment;

[0043] FIG3c is a schematic diagram of a data graph structure in yet another embodiment;

[0044] FIG4 is a schematic diagram of a framework of an autoencoder according to an embodiment;

[0045] FIG5 is a flow chart of a method for predicting power source load based on a graph autoencoder in another embodiment;

[0046] FIG6 is a structural block diagram of a power source load prediction device based on a graph autoencoder in one embodiment;

[0047] FIG. 7 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0049] Graph neural network is a deep learning model based on graph structure. By integrating the high-dimensional features of the source-load prediction task into the attribute information of the node and mining the correlation between different nodes to supplement the attribute information of the edge, the feature information of neighboring nodes can be repeatedly aggregated through further feature transformation to obtain low-dimensional expressions of nodes and edges, making it have excellent performance in various prediction tasks. The data input into the prediction model is real, accurate, and meaningful, which is the basis for the graph neural network to build a streamlined and effective graph structure and achieve good prediction results. However, the parameters of the power grid operation state show the characteristics of high redundancy and low information density. The redundant feature information on the node not only makes it more difficult to construct an effective nonlinear mapping relationship for the model, but also further affects the attribute information between nodes (edges), making it difficult to mine effective information, reducing the prediction performance of the model. The traditional method is to select through artificial experience mode or screen based on correlation analysis criteria to obtain available feature information for source-load prediction tasks under different requirements. The feature screening process is complicated and highly dependent on artificial experience. It is difficult to extract feature signals with strong correlation.

[0050] In response to the above problems, this application proposes a graph convolutional power source and load prediction method that combines a graph autoencoder to realize adaptive extraction of coding features. By constructing a graph autoencoder, the original features and coding features of the source and load prediction task can be bidirectionally transformed, which can provide more efficient input and output for the prediction task, effectively ensure the accuracy of the prediction, and provide accurate data support for the stable operation of the power grid.

[0051] In an exemplary embodiment, as shown in FIG1 , a method for predicting power source load based on a graph autoencoder is provided. This embodiment uses the method applied to a terminal as an example. It can be understood that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps 101 to 104. Among them:

[0052] Step 101 : acquiring original power source and load data associated with the source and load prediction task from real-time power source and load data according to prediction demand information of the source and load prediction task.

[0053] In practical applications, since different source-load prediction tasks have different business scenarios, the prediction tasks of the same type of electrical objects also correspond to different data sources and prediction targets. For the source-load prediction task to be processed, the power source-load data related to the source-load prediction task can be obtained from the collected real-time power source-load data according to the prediction demand information of the source-load prediction task, as the original power source-load data corresponding to the source-load prediction task.

[0054] Step 102: input the original power source and load data into a pre-trained power source and load prediction model, process the original power source and load data based on a graph attention mechanism, and obtain an original data feature graph.

[0055] In a specific implementation, a pre-trained power source load prediction model can be used to perform power source load prediction. As shown in Figure 2, the original data of the source load prediction task (i.e., the original power source load data) can be input into the pre-trained power source load prediction model, and then the data can be processed based on the graph attention mechanism to obtain the original data Graph structure (i.e., the original data feature graph).

[0056] Step 103, using the graph autoencoder in the pre-trained power source and load prediction model, adaptively extracting features from the original data feature graph, performing dimensionality reduction and simplification processing, and obtaining an encoded feature graph.

[0057] After obtaining the original data feature graph, the original data feature graph (such as the original data Graph structure in Figure 2) can be input into the graph autoencoder. The graph autoencoder can be used to adaptively extract features from the original data feature graph and perform dimensionality reduction and simplification to obtain a coding feature graph based on low-dimensional graph representation (such as the coding feature Graph structure in Figure 2). Thus, by combining the graph autoencoder to extract more concise and efficient correlation features, the problem of insufficient feature information screening based on artificial experience patterns in traditional methods is solved, and a universal source-load prediction coding representation can be obtained, which helps to further guide the generation of more accurate prediction results and has strong adaptability and accuracy.

[0058] Step 104: input the encoded feature graph into the graph convolution prediction network in the pre-trained power source and load prediction model to obtain the power source and load prediction result of the source and load prediction task.

[0059] Among them, the power source and load forecast results can be used to characterize the predicted power energy demand and load demand in the power system.

[0060] For example, in the model application stage, based on the real-time raw data of the input source load prediction task, as shown in Figure 2, after being processed by the graph autoencoder and the graph convolutional neural network, the power source load prediction result of the source load prediction task can be output. Thus, by extracting the reduced and simplified features by combining the graph autoencoder before prediction, efficient feature data support is provided for the subsequent source load prediction processing, which can adapt to the needs of different source load prediction tasks, avoid the uncertainty of feature selection based on artificial experience models, and provide a common architecture for source load prediction business processing.

[0061] In one example, in the application stage, in view of the high redundancy and low information density of massive source-load prediction feature information, the method of artificial experience screening of relevant features is adopted, and the screening results have an uncertain impact on the performance of the source-load prediction model. This application is based on the power source-load prediction method of the graph autoencoder. By combining the advantages of graph representation learning in processing multi-source heterogeneous data, the massive source-load prediction electrical data is represented as graph structure data according to the different needs of the source-load prediction task. Then, based on the constructed graph autoencoder, the original features and the encoded features can be bidirectionally converted to use the streamlined and efficient encoded features as the composition of feature engineering to guide the high-precision prediction of the source-load.

[0062] Compared with traditional methods, it lacks the means of adaptively acquiring effective feature information, such as only considering information such as climate and natural disasters to select features suitable for new energy prediction tasks, or only considering features under the influence of time factors such as the same season and month in load prediction tasks. Although such feature extraction methods can avoid redundant information in massive electrical data, they will inevitably lead to the loss of some original feature information. The technical solution of this embodiment, by constructing a graph autoencoder with an adaptive graph structure, introduces a more flexible and adaptive feature extraction method, and realizes a universal coding representation for source-load prediction tasks under different needs. By effectively combining the coding representation with the power source-load prediction business, the efficient adaptive feature input can not only avoid the uncertainty of artificial screening of redundant information, but also automatically mine the feature information suitable for the source-load prediction task. At the same time, it can achieve effective streamlining on the basis of ensuring the integrity of the feature information, and provide support for subsequent high-precision prediction processing, effectively improving the prediction accuracy and ensuring the stable operation of the system.

[0063] In the above-mentioned power source and load prediction method based on graph autoencoder, the original power source and load data associated with the source load prediction task is obtained from the real-time power source and load data according to the prediction demand information of the source load prediction task, and then the original power source and load data is input into the pre-trained power source and load prediction model, and the original power source and load data is processed based on the graph attention mechanism to obtain the original data feature map, and the original data feature map is adaptively extracted through the graph autoencoder in the pre-trained power source and load prediction model to perform dimensionality reduction and simplification processing to obtain the encoding feature map, and then the encoding feature map is input into the graph convolution prediction network in the pre-trained power source and load prediction model to obtain the power source and load prediction result of the source load prediction task, thereby realizing the optimization of the power source and load prediction processing, and can adaptively extract concise and efficient encoding features based on the graph autoencoder according to the different needs of the source load prediction task, provide data support for the graph convolution prediction network processing, and effectively improve the source and load prediction effect.

[0064] In an exemplary embodiment, the following steps may also be included:

[0065] Obtain sample power source and load data containing different sample prediction tasks to construct a training data set; construct a corresponding data graph structure according to each sample power source and load data in the training data set, and process it based on the graph attention mechanism to obtain the original feature graph of each sample; obtain the power source and load prediction model to be trained by constructing the graph autoencoder and the graph convolutional prediction network; based on the difference between the sample encoding features generated by the graph autoencoder and the sample original feature graph, and based on the difference between the sample prediction results generated by the graph convolutional prediction network and the sample real data, train the power source and load prediction model to be trained to obtain the pre-trained power source and load prediction model.

[0066] In practical applications, through data set construction, data samples related to the source-load prediction task (i.e., sample power source-load data) can be obtained, and data on influencing factors under different source-load prediction needs can be collected, such as economic factors, seasonal months and other time factors, and training data sets can be obtained through data preprocessing; through Graph construction, the original multi-source heterogeneous data input can be converted into a Graph structure (i.e., sample original feature graph) based on the constructed training data set and the graph attention mechanism, so as to more efficiently represent the data and further utilize the powerful feature mining capabilities of graph neural networks.

[0067] In one example, through model construction, in terms of feature engineering, a graph autoencoder adapted to the Graph structure can be constructed. Based on the graph autoencoder, a representation model for bidirectional transformation of "original data Graph-encoded feature Graph" can be realized, which can effectively simplify feature information while ensuring integrity; in terms of prediction model, a graph convolution prediction network adapted to the encoded feature Graph can be constructed, and the powerful information mining capability of the graph convolutional neural network can be used to accurately construct the nonlinear mapping relationship between feature information and prediction results.

[0068] In another example, during the model training phase, Monte Carlo sampling can be used to collect training samples from the training data set and convert them into a graph data matrix corresponding to the graph structure. The original data graph structure (i.e., the sample original feature graph) is then used as the input of the graph autoencoder feature engineering to generate a reduced-dimensional and simplified encoded data graph structure (i.e., the sample encoding feature), and then the corresponding prediction result (i.e., the sample prediction result) can be generated through the graph convolutional neural network. During the training process, the encoder and graph convolutional neural network in the graph autoencoder can be updated based on the difference between the prediction result and the real sample (i.e., the difference between the sample prediction result and the sample real data); the decoder in the graph autoencoder can be updated based on the difference between the reverse-converted data graph structure and the original data graph structure (i.e., the difference between the sample encoding feature and the sample original feature graph).

[0069] In this embodiment, by obtaining data samples related to different source load prediction tasks, a training data set is constructed based on data preprocessing, and then based on the Graph construction method, a Graph that can be processed by the graph neural network can be constructed in combination with the screened training data. The graph attention mechanism can be used to mine the correlation between nodes and supplement the edge attribute information. Furthermore, by constructing a graph autoencoder, the two-way conversion of original features and encoded features can be achieved, and effective information simplification is completed on the basis of ensuring the integrity of feature information. The obtained encoded features can be used as a component of feature engineering to support the implementation of different prediction tasks. High-precision source load prediction can be achieved using a graph convolutional neural network.

[0070] In an exemplary embodiment, the step of obtaining sample power source and load data including different sample prediction tasks and constructing a training data set may include the following steps:

[0071] According to the prediction demand information of each of the sample prediction tasks, sample power source and load data associated with each of the sample prediction tasks are obtained from historical power source and load data; and the training data set is obtained by performing data preprocessing on each of the sample power source and load data.

[0072] The data preprocessing may include any one or more of the following:

[0073] Outlier cleaning, interpolation filling, data transformation, normalization processing, standardization processing, and calculation of power-related statistical indicators.

[0074] In the specific implementation, by obtaining real and accurate historical power source and load data, such as using power grid data acquisition devices to collect source and load power data and multi-source heterogeneous data such as meteorology and geography, and filtering out the electrical data related to the prediction task based on correlation analysis of the massive electrical data obtained, data samples related to the source and load prediction task can be obtained, and then the obtained data can be preprocessed by data cleaning, data supplementation, data transformation, etc., such as including but not limited to outlier cleaning, interpolation filling, data transformation, normalization, standardization, and calculation of relevant statistical indicators, so as to obtain data-driven construction of a real and accurate high-quality data set (i.e., a training data set).

[0075] Optionally, multiple factors such as meteorological data at the same time and historical time, geographic information, power data at historical time, and power data corresponding to the same time can be used as input and output pairs of training data sets and test data sets to provide data support for subsequent model training and model testing.

[0076] In an exemplary embodiment, the steps of constructing a corresponding data graph structure according to each of the sample power source load data in the training data set, and processing based on a graph attention mechanism to obtain each sample original feature graph may include the following steps:

[0077] According to the prediction type determined from the prediction demand information of each sample prediction task, a data graph structure corresponding to each sample power source and load data is constructed; different prediction types correspond to different data graph structures; the graph attention mechanism is used to adjust the feature expression of the nodes in each data graph structure to obtain the original feature graph of each sample.

[0078] In one example, during the graph construction process, the graph structure can be abstractly represented as a node feature matrix E and a modified adjacency matrix A at the data level. Based on the training data set, the number of graph nodes n can be set equal to the data dimension (type), all nodes can be fully connected, and the graph attention mechanism can be used to mine the correlation between features (such as wind speed and wind direction) to complete the construction of the original data graph structure G. 0 , that is, the original feature map of the sample.

[0079] Specifically, the Graph can be constructed in the following ways:

[0080] 1. Graph data matrix visualization: In the Graph, it contains two attributes: nodes and edges. G = {E, A} represents an undirected graph, where E is the node feature matrix and A is the modified adjacency matrix. |E| = n means that there are n nodes on the graph, each node represents a one-dimensional feature, and the features of all nodes can form an N×1-dimensional matrix E, which can represent D-dimensional features on each node; A is modified to represent the adjacency matrix. Since the nodes are fully connected, the dimension of the matrix A is N×N. In response to different requirements of source-load prediction tasks, as shown in Figures 3a, 3b, and 3c, for different prediction types, the number of features in the Graph (i.e., data graph structure) corresponding to each prediction type can be adaptively set according to the type of original data. For example, the photovoltaic prediction task Graph constructed in Figure 3a may include features such as temperature, historical data, irradiation intensity, air pressure, and relative humidity; the wind power prediction task Graph constructed in Figure 3b may include features such as wind speed, historical data, wind direction, air pressure, relative humidity, and air pressure; the load prediction task Graph constructed in Figure 3c may include relative humidity, historical data, daily cycle features, temperature, date code features, air pressure, etc.

[0081] 2. Graph Attention Mechanism: Graph Attention Network (GAT) can define aggregation functions through attention mechanism. In the graph attention network, the calculation of the node association weight in the modified adjacency matrix depends on the feature expression of the node. For example, the graph attention network takes the feature expression of nodes i and j as input, calculates the attention weight between i and j and normalizes it, and further aggregates the expression of the surrounding nodes to itself in the form of weighted sum. The calculation formulas of attention weight and expression update are as follows:

[0082] (1)

[0083] (2)

[0084] Among them, the parameter W can be used to complete the feature dimension transformation of each node, the parameter a can be used to calculate the attention weight between nodes, and || represents vector concatenation. Represents the calculated weight between nodes i and j, represents a non-linear activation function.

[0085] In an exemplary embodiment, the following steps may also be included:

[0086] The sample original feature map is encoded by the encoder in the graph autoencoder to generate the sample encoding features based on low-dimensional feature representation; the sample encoding features are decoded by the decoder in the graph autoencoder, and reversely transformed to obtain the sample original feature map based on high-dimensional feature representation.

[0087] In practical applications, the graph autoencoder is constructed with the original data graph structure as input. By constructing an encoder, the high-dimensional input G 0 (i.e., the original feature map of the sample based on the high-dimensional feature representation) is encoded into a low-dimensional graph representation G 1 (i.e., sample encoding features based on low-dimensional feature representation) can control the neural network to learn the features that are most relevant to the prediction results. By constructing a decoder, the low-dimensional G 1 Reverse transformation into the original input G 0 , thereby ensuring that the generated low-dimensional feature representation guarantees the integrity of information. The graph autoencoder based on bidirectional encoding transformation provides an effective technical means for adaptive high-precision prediction.

[0088] Specifically, the graph autoencoder can be constructed in the following way:

[0089] In the graph autoencoder (the framework of the graph autoencoder as shown in Figure 4), two neural network models can be included to process the grid operation feature graph, which are divided into an encoder and a decoder. Among them, the structure of the encoder can be a two-layer graph convolutional neural network, and the structure of the decoder can also be a two-layer graph convolutional neural network; the encoder encodes the original data feature graph to generate a lower-dimensional encoding feature, and the decoder decodes the encoding feature; for the loss function of the graph autoencoder is the feature difference between the input and output, the decoder can restore the data through the high- and low-dimensional data features, and the low-dimensional data can be regarded as feature information that is more correlated with the prediction results, thereby realizing streamlined and efficient feature engineering.

[0090] In an optional embodiment, the graph convolution prediction network can be constructed in the following manner:

[0091] The idea of ​​graph convolutional neural network is to aggregate the information of neighboring nodes and obtain more powerful feature expressions to deeply explore the potential distribution of power grid operation scenario data. The following calculation formula can be used:

[0092] (3)

[0093] Among them, A'=A+I N , A' is the adjacency matrix with self-connection, I N is the identity matrix. Dii=∑j Aij', D is the degree matrix of A', is the trainable parameter in the convolutional layer of the graph convolutional neural network, Represents the input feature of the lth layer. After the matrix multiplication in the above formula, it is forward propagated through the activation function RELU.

[0094] For the training phase and the testing phase, the original data set can be divided into a training set and a test set respectively to train the graph autoencoder and the graph convolutional neural prediction network, and the weights of the internal parameters of the neural network model can be updated through the gradient descent algorithm.

[0095] In this embodiment, by cleaning and screening data according to different requirements of the source load prediction task, a graph structure containing source load feature information is initially constructed; then, a graph autoencoder adapted to the graph structure can be introduced into the source load prediction task, and based on the graph autoencoder including an encoder and a decoder, a dual conversion of original features and encoded features can be realized, wherein the encoder can realize the conversion of high-dimensional, redundant and nonlinear original features on the graph structure to low-dimensional and streamlined encoded features, and can adaptively provide efficient node feature information for the source load prediction task, and the decoder can realize the reverse conversion of encoded features to original features, ensuring that the encoded features guarantee the integrity of the original feature information; and then, the graph attention mechanism can be combined to mine the correlation between nodes (edges), obtain the graph representation of the source load prediction task, and combine the graph convolutional neural network to perform high-precision source load prediction. Thus, it is possible to adaptively mine key data in massive feature information, and complete the effective simplification of feature information on the basis of ensuring integrity, so as to serve the accurate and reliable prediction of power source load.

[0096] In an exemplary embodiment, the training of the power source load prediction model to be trained based on the difference between the sample encoding features generated by the graph autoencoder and the sample original feature graph, and based on the difference between the sample prediction results generated by the graph convolutional prediction network and the sample real data, may include the following steps:

[0097] The power source load prediction model to be trained is trained based on the difference between the sample encoding feature and the sample original feature map, and the difference between the sample prediction result and the sample real data, using a first loss function for the encoder and the graph convolution prediction network and a second loss function for the decoder.

[0098] Among them, the first loss function can be used to control the data difference between the sample prediction result and the sample real data to be minimized; the second loss function can be used to control the feature difference between the sample encoding feature and the sample original feature map to be minimized.

[0099] In one example, the power source load prediction model may include two sets of neural networks, the first set is a graph autoencoder, where the input of the encoder is a high-dimensional raw data graph G 0 , the output is a low-dimensional graph representation G 1 ; The input and output of the decoder are the opposite; The second group of neural networks is a graph convolutional neural network, whose input is a low-dimensional graph representation G 1 , the output is the prediction result.

[0100] In another example, the loss function of the encoder and the graph convolutional neural network (i.e., the first loss function) aims to minimize the distance between the output source load prediction result and the real data (i.e., the data difference between the sample prediction result and the sample real data is minimized); the loss function of the decoder (i.e., the second loss function) aims to minimize the distance between the original data graph G and the real data. 0’ With G 0 The distance between the sample encoding feature and the sample original feature map is the smallest (that is, the feature difference between the sample encoding feature and the sample original feature map is the smallest). The neural network in the power source load prediction model can reversely update the parameters inside the graph neural network through gradient descent.

[0101] In an exemplary embodiment, as shown in FIG5 , a flow chart of another method for predicting power source load based on a graph autoencoder is provided. In this embodiment, the method includes the following steps:

[0102] In step 501, according to the prediction demand information of each sample prediction task, the sample power source and load data associated with each sample prediction task is obtained from the historical power source and load data, and the training data set is obtained by preprocessing the data of each sample power source and load data. In step 502, according to the prediction type determined from the prediction demand information of each sample prediction task, the data graph structure corresponding to each sample power source and load data is constructed, and the feature expression of the nodes in each data graph structure is adjusted by using the graph attention mechanism to obtain the original feature graph of each sample. In step 503, the power source and load prediction model to be trained is obtained by constructing a graph autoencoder and a graph convolution prediction network. In step 504, the sample original feature graph is encoded by the encoder in the graph autoencoder to generate a sample encoding feature based on a low-dimensional feature representation, and the sample encoding feature is decoded by the decoder in the graph autoencoder, and the reverse conversion is obtained to obtain the sample original feature graph based on a high-dimensional feature representation. In step 505, the first loss function for the encoder and the graph convolution prediction network and the second loss function for the decoder are used to train the power source load prediction model to be trained based on the difference between the sample encoding feature and the original feature map of the sample, and the difference between the sample prediction result and the real data of the sample, to obtain a pre-trained power source load prediction model. In step 506, according to the prediction demand information of the source load prediction task, the original power source load data associated with the source load prediction task is obtained from the real-time power source load data, the original power source load data is input into the pre-trained power source load prediction model, and the original power source load data is processed based on the graph attention mechanism to obtain the original data feature map. In step 507, the original data feature map is adaptively extracted through the graph autoencoder in the pre-trained power source load prediction model to perform dimensionality reduction and simplification processing to obtain the encoding feature map, and the encoding feature map is input into the graph convolution prediction network in the pre-trained power source load prediction model to obtain the power source load prediction result of the source load prediction task. It should be noted that the specific limitations of the above steps can be found in the above specific limitations of a method for predicting power source load based on a graph autoencoder, which will not be repeated here.

[0103] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0104] Based on the same inventive concept, the embodiment of the present application also provides a graph autoencoder-based power source load prediction device for implementing the graph autoencoder-based power source load prediction method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more embodiments of the graph autoencoder-based power source load prediction device provided below can refer to the above limitations on the graph autoencoder-based power source load prediction method, which will not be repeated here.

[0105] In an exemplary embodiment, as shown in FIG6 , a power source load prediction device based on a graph autoencoder is provided, comprising:

[0106] The real-time data acquisition module 601 is used to acquire original power source and load data associated with the source and load prediction task from the real-time power source and load data according to the prediction demand information of the source and load prediction task;

[0107] The original feature graph acquisition module 602 is used to input the original power source and load data into the pre-trained power source and load prediction model, process the original power source and load data based on the graph attention mechanism, and obtain the original data feature graph;

[0108] The coding feature graph acquisition module 603 is used to adaptively extract features from the original data feature graph and perform dimensionality reduction and simplification processing through the graph autoencoder in the pre-trained power source and load prediction model to obtain a coding feature graph;

[0109] The power source and load prediction result obtaining module 604 is used to input the encoded feature graph into the graph convolution prediction network in the pre-trained power source and load prediction model to obtain the power source and load prediction result of the source and load prediction task; the power source and load prediction result is used to characterize the predicted power energy demand and load demand in the power system.

[0110] In one embodiment, the apparatus further comprises:

[0111] A training data set construction module is used to obtain sample power source and load data containing different sample prediction tasks and construct a training data set;

[0112] A sample original feature map acquisition module is used to construct a corresponding data graph structure according to each of the sample power source and load data in the training data set, and process based on the graph attention mechanism to obtain each sample original feature map;

[0113] A model construction module, used to obtain a power source load prediction model to be trained by constructing the graph autoencoder and the graph convolution prediction network;

[0114] A model training module is used to train the power source and load prediction model to be trained based on the difference between the sample encoding features generated by the graph autoencoder and the sample original feature graph, and the difference between the sample prediction results generated by the graph convolutional prediction network and the sample real data, so as to obtain the pre-trained power source and load prediction model.

[0115] In one embodiment, the training data set construction module includes:

[0116] A sample data acquisition submodule, used to acquire sample power source and load data associated with each of the sample prediction tasks from historical power source and load data according to the prediction demand information of each of the sample prediction tasks;

[0117] A data preprocessing submodule, used for obtaining the training data set by performing data preprocessing on each of the sample power source and load data;

[0118] The data preprocessing includes any one or more of the following:

[0119] Outlier cleaning, interpolation filling, data transformation, normalization processing, standardization processing, and calculation of power-related statistical indicators.

[0120] In one embodiment, the sample original feature map obtaining module includes:

[0121] A data graph structure construction submodule, used to construct a data graph structure corresponding to each of the sample power source and load data according to the prediction type determined from the prediction demand information of each of the sample prediction tasks; different prediction types correspond to different data graph structures;

[0122] The attention mechanism processing submodule is used to adopt the graph attention mechanism to adjust the feature expression of the nodes in each of the data graph structures to obtain the original feature graph of each of the samples.

[0123] In one embodiment, the apparatus further comprises:

[0124] A low-dimensional feature generation module, used to encode the sample original feature map through the encoder in the graph autoencoder to generate the sample encoding feature based on the low-dimensional feature representation;

[0125] The high-dimensional feature recovery module is used to decode the sample encoding features through the decoder in the image autoencoder, and reversely transform to obtain the original feature map of the sample based on the high-dimensional feature representation.

[0126] In one embodiment, the model training module includes:

[0127] A loss function processing submodule, configured to train the power source load prediction model to be trained based on the difference between the sample encoding feature and the sample original feature map, and the difference between the sample prediction result and the sample real data, using a first loss function for the encoder and the graph convolution prediction network and a second loss function for the decoder;

[0128] Among them, the first loss function is used to control the data difference between the sample prediction result and the sample real data to be minimum; the second loss function is used to control the feature difference between the sample encoding feature and the sample original feature map to be minimum.

[0129] Each module in the above-mentioned power source load prediction device based on graph autoencoder can be implemented in whole or in part by software, hardware and their combination. Each of the above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the corresponding operations of each of the above modules.

[0130] In an exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be shown in FIG7. The computer device includes a processor, a memory, an input / output interface, a communication interface, a display unit, and an input device. The processor, the memory, and the input / output interface are connected via a system bus, and the communication interface, the display unit, and the input device are connected to the system bus via the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be implemented through WIFI, a mobile cellular network, a near field communication (NFC) or other technologies. When the computer program is executed by the processor, a method for predicting power source load based on a graph self-encoder is implemented.

[0131] Those skilled in the art will understand that the structure shown in FIG. 7 is merely a block diagram of a partial structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different arrangement of components.

[0132] In an exemplary embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0133] According to the forecast demand information of the source load forecast task, original power source load data associated with the source load forecast task is acquired from the real-time power source load data;

[0134] Inputting the original power source and load data into a pre-trained power source and load prediction model, processing the original power source and load data based on a graph attention mechanism, and obtaining an original data feature graph;

[0135] Adaptively extract features from the original data feature graph through the graph autoencoder in the pre-trained power source and load prediction model, perform dimensionality reduction and simplification processing on the feature graph, and obtain an encoded feature graph;

[0136] The encoded feature graph is input into the graph convolution prediction network in the pre-trained power source and load prediction model to obtain the power source and load prediction result of the source and load prediction task; the power source and load prediction result is used to characterize the predicted power energy demand and load demand in the power system.

[0137] In one embodiment, when the processor executes the computer program, it also implements the steps of the power source load prediction method based on the graph autoencoder in the other embodiments described above.

[0138] In one embodiment, a computer readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:

[0139] According to the forecast demand information of the source load forecast task, original power source load data associated with the source load forecast task is acquired from the real-time power source load data;

[0140] Inputting the original power source and load data into a pre-trained power source and load prediction model, processing the original power source and load data based on a graph attention mechanism, and obtaining an original data feature graph;

[0141] Adaptively extract features from the original data feature graph through the graph autoencoder in the pre-trained power source and load prediction model, perform dimensionality reduction and simplification processing on the feature graph, and obtain an encoded feature graph;

[0142] The encoded feature graph is input into the graph convolution prediction network in the pre-trained power source and load prediction model to obtain the power source and load prediction result of the source and load prediction task; the power source and load prediction result is used to characterize the predicted power energy demand and load demand in the power system.

[0143] In one embodiment, when the computer program is executed by a processor, the steps of the power source load prediction method based on the graph autoencoder in the other embodiments described above are also implemented.

[0144] In one embodiment, a computer program product is provided, comprising a computer program, which, when executed by a processor, implements the following steps:

[0145] According to the forecast demand information of the source load forecast task, original power source load data associated with the source load forecast task is acquired from the real-time power source load data;

[0146] Inputting the original power source and load data into a pre-trained power source and load prediction model, processing the original power source and load data based on a graph attention mechanism, and obtaining an original data feature graph;

[0147] Adaptively extract features from the original data feature graph through the graph autoencoder in the pre-trained power source and load prediction model, perform dimensionality reduction and simplification processing on the feature graph, and obtain an encoded feature graph;

[0148] The encoded feature graph is input into the graph convolution prediction network in the pre-trained power source and load prediction model to obtain the power source and load prediction result of the source and load prediction task; the power source and load prediction result is used to characterize the predicted power energy demand and load demand in the power system.

[0149] In one embodiment, when the computer program is executed by a processor, the steps of the power source load prediction method based on the graph autoencoder in the other embodiments described above are also implemented.

[0150] It should be noted that the collection, use and processing of the relevant data involved in this application must comply with relevant regulations.

[0151] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, an artificial intelligence (AI) processor, etc., but are not limited to this.

[0152] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0153] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for predicting power source load based on graph autoencoder, characterized in that: The method comprises: According to the forecast demand information of the source load forecast task, original power source load data associated with the source load forecast task is acquired from the real-time power source load data; Inputting the original power source and load data into a pre-trained power source and load prediction model, processing the original power source and load data based on a graph attention mechanism, and obtaining an original data feature graph; Adaptively extract features from the original data feature graph through the graph autoencoder in the pre-trained power source and load prediction model, perform dimensionality reduction and simplification processing on the feature graph, and obtain an encoded feature graph; The encoded feature graph is input into the graph convolution prediction network in the pre-trained power source and load prediction model to obtain the power source and load prediction result of the source and load prediction task; the power source and load prediction result is used to characterize the predicted power energy demand and load demand in the power system.

2. The method according to claim 1, characterized in that The method further comprises: Obtain sample power source and load data containing different sample prediction tasks and construct a training data set; Constructing a corresponding data graph structure according to each of the sample power source and load data in the training data set, and processing based on a graph attention mechanism to obtain an original feature graph of each sample; By constructing the graph autoencoder and the graph convolution prediction network, a power source load prediction model to be trained is obtained; Based on the difference between the sample encoding features generated by the graph autoencoder and the sample original feature graph, and based on the difference between the sample prediction results generated by the graph convolutional prediction network and the sample real data, the power source and load prediction model to be trained is trained to obtain the pre-trained power source and load prediction model.

3. The method according to claim 2, characterized in that The obtaining of sample power source and load data including different sample prediction tasks and constructing a training data set includes: According to the forecast demand information of each sample forecast task, sample power source and load data associated with each sample forecast task is obtained from historical power source and load data; The training data set is obtained by performing data preprocessing on each of the sample power source and load data; The data preprocessing includes any one or more of the following: Outlier cleaning, interpolation filling, data transformation, normalization processing, standardization processing, and calculation of power-related statistical indicators.

4. The method according to claim 2, characterized in that: The corresponding data graph structure is constructed according to each of the sample power source and load data in the training data set, and processed based on the graph attention mechanism to obtain the original feature graph of each sample, including: According to the prediction type determined from the prediction demand information of each sample prediction task, a data graph structure corresponding to each sample power source and load data is constructed; different prediction types correspond to different data graph structures; The graph attention mechanism is used to adjust the feature expression of the nodes in each of the data graph structures to obtain the original feature graph of each of the samples.

5. The method according to claim 2, characterized in that: The method further comprises: Encoding the original feature graph of the sample by an encoder in the graph autoencoder to generate the sample encoding feature based on low-dimensional feature representation; The sample encoding features are decoded by a decoder in the graph autoencoder, and reversely transformed to obtain the original feature graph of the sample based on high-dimensional feature representation.

6. The method according to claim 5, characterized in that The training of the power source load prediction model to be trained based on the difference between the sample encoding feature generated by the graph autoencoder and the sample original feature graph, and the difference between the sample prediction result generated by the graph convolution prediction network and the sample real data, includes: Using a first loss function for the encoder and the graph convolution prediction network and a second loss function for the decoder, based on the difference between the sample encoding feature and the sample original feature map, and the difference between the sample prediction result and the sample real data, the power source load prediction model to be trained is trained; Among them, the first loss function is used to control the data difference between the sample prediction result and the sample real data to be minimum; the second loss function is used to control the feature difference between the sample encoding feature and the sample original feature map to be minimum.

7. A power source load prediction device based on graph autoencoder, characterized in that: The device comprises: A real-time data acquisition module, used to acquire original power source and load data associated with the source and load prediction task from real-time power source and load data according to the prediction demand information of the source and load prediction task; An original feature graph acquisition module is used to input the original power source and load data into a pre-trained power source and load prediction model, process the original power source and load data based on a graph attention mechanism, and obtain an original data feature graph; A coding feature map acquisition module is used to adaptively extract features from the original data feature map and perform dimensionality reduction and simplification processing through the map autoencoder in the pre-trained power source and load prediction model to obtain a coding feature map; The power source and load prediction result obtaining module is used to input the encoded feature graph into the graph convolution prediction network in the pre-trained power source and load prediction model to obtain the power source and load prediction result of the source and load prediction task; the power source and load prediction result is used to characterize the predicted power energy demand and load demand in the power system.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images