Power time sequence prediction method based on improved dynamic and static graph attention and related device

Abstract

The invention discloses an electric power time sequence prediction method based on improved dynamic and static graph attention and a related device. The method comprises the following steps: acquiring electric power time sequence data; inputting the power time sequence data into a time sequence prediction model, and outputting a prediction result; wherein the time sequence prediction model comprises a multi-scale time convolution module for extracting time sequence correlation by using a gating mechanism; the graph attention module is used for aggregating information of dynamic and static matrixes by using two independent graph attention networks so as to optimize time sequence characteristics; and the dynamic and static graph learning module generates a mapping function of a dynamic graph and an adjacent matrix of a static graph through adaptive learning of a dynamic and static graph learning layer, and captures a hidden relationship between different data. The objective of the invention is to solve the problem of insufficient prediction accuracy caused by the fact that a traditional prediction model often neglects time-space characteristics of power grid data.

Description

Technical Field

[0001] The present invention belongs to the field of power data security, and in particular relates to a power time series prediction method and a related device based on improving dynamic and static graph attention. Background Art

[0002] In the power industry, ensuring the safe and stable operation of the power grid and the reliability of power supply is one of the core tasks. Since the power system is affected by a variety of potential threats, including extreme weather conditions, equipment aging and failure, system failures, and possible malicious attacks, these factors may cause power outages or reduced efficiency, causing serious impacts on economic and social activities. Therefore, timely prediction and effective response to these potential threats are crucial to maintaining the safety and reliability of the power grid.

[0003] With the advancement of science and technology, especially the rapid development of artificial intelligence (AI) and machine learning (ML) technologies, data-driven prediction models are increasingly used in the power industry. These models can reveal hidden operating rules and potential risks by mining and analyzing massive amounts of power data, providing a scientific basis for the management and decision-making of power systems. Among many machine learning technologies, graph neural networks (GNNs) stand out for their unique ability to process graph-structured data. As a complex network structure, the power system consists of nodes (such as generators, transformers, transmission lines, etc.) and edges (such as physical connections or electrical connections) that form a typical graph structure. By capturing the correlation between nodes and the overall structural information of the graph, GNNs have achieved remarkable results in the fields of power system state estimation, load forecasting, and fault detection.

[0004] Although GNN has shown great potential in power systems, it still has certain limitations in dealing with the dynamics and complexity of power systems. The operating state of the power system is constantly changing, including load fluctuations, equipment switching, fault occurrence and recovery, etc. These dynamic changes put forward higher requirements on the accuracy and real-time performance of the prediction model. At the same time, the topological structure of the power system is complex, with many nodes, and the interaction relationship between nodes is complex and changeable, which further increases the difficulty of prediction. Summary of the invention

[0005] In view of the problems existing in the prior art, the present invention provides a power time series prediction method and related devices based on improving the attention of dynamic and static graphs, which aims to solve the problem that traditional prediction models often ignore the temporal and spatial characteristics of power grid data, resulting in insufficient prediction accuracy.

[0006] In order to solve the above technical problems, the present invention is implemented by the following technical solutions:

[0007] A power time series prediction method based on improved dynamic and static graph attention, comprising:

[0008] Obtain power time series data;

[0009] The power time series data is input into a time series prediction model, and a prediction result is output; wherein the time series prediction model includes:

[0010] Multi-scale temporal convolution module, which uses gating mechanism to extract temporal correlation;

[0011] The graph attention module uses two independent graph attention networks to aggregate the information of dynamic and static matrices to optimize temporal features;

[0012] The dynamic and static graph learning module adaptively generates the mapping function of the dynamic graph and the adjacency matrix of the static graph through the dynamic and static graph learning layer, and captures the hidden relationship between different data.

[0013] Furthermore, the extraction of timing correlation using a gating mechanism is specifically as follows:

[0014] A multi-scale convolutional neural network layer with a gating mechanism is used to process power time series data and extract time series correlation.

[0015] Furthermore, two independent graph attention networks are used to summarize the information of dynamic and static matrices to optimize the timing features, specifically:

[0016] Two independent graph attention networks are integrated to summarize the information of dynamic and static matrices respectively and optimize the temporal features.

[0017] Furthermore, the time series prediction model also includes:

[0018] Combined with the Transformer-based multi-head attention idea, a dynamic matrix is ​​constructed, and the dynamic graph is obtained through linear layer projection and summation operations.

[0019] Furthermore, the time series prediction model also includes:

[0020] Skip connection and output module,The skip connection layer is set to normalize the aggregated information, and the channel size is converted to the required output size through a convolutional neural network composed of two filters.

[0021] A power timing prediction device based on improved dynamic and static graph attention, comprising:

[0022] An acquisition module, used for acquiring power time series data;

[0023] An output module is used to input the power time series data into a time series prediction model and output a prediction result; wherein the time series prediction model includes:

[0024] Multi-scale temporal convolution module, which uses gating mechanism to extract temporal correlation;

[0025] The graph attention module uses two independent graph attention networks to aggregate the information of dynamic and static matrices to optimize temporal features;

[0026] The dynamic and static graph learning module adaptively generates the mapping function of the dynamic graph and the adjacency matrix of the static graph through the dynamic and static graph learning layer, and captures the hidden relationship between different data.

[0027] A device comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of a power timing prediction method based on improving the attention of dynamic and static graphs are implemented.

[0028] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of a power timing prediction method based on improving the attention of dynamic and static graphs are implemented.

[0029] Compared with the prior art, the present invention has at least the following beneficial effects:

[0030] The present invention provides a method for power time series prediction based on improved dynamic and static graph attention, which combines the advantages of dynamic graph attention network and static graph attention network, and can not only capture the static correlation between nodes within the power system, but also adapt to the dynamic characteristics of the system that change over time. Specifically, the dynamic graph attention network can pay attention to data changes at important time points in real time, improving the model's response speed to emergencies; while the static graph attention network can deeply analyze the topological structure of the power grid, identify key nodes and potential weak links, thereby enhancing the accuracy and robustness of the model in power time series prediction and threat identification tasks. This prediction model that combines dynamic and static information provides new technical means and solutions for the safe and stable operation of power systems.

[0031] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solutions in the specific implementation modes of the present invention, the drawings required for use in the description of the specific implementation modes will be briefly introduced below. Obviously, the drawings described below are some implementation modes of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0033] Figure 1 The present invention is a flowchart of an embodiment of a method for predicting power timing based on improving attention of dynamic and static graphs.

[0034] Figure 2 This is the specific architecture diagram of the TVDNN model.

[0035] Figure 3 This is a long short-term memory network diagram.

[0036] Figure 4 This is the specific architecture diagram of the DS-GAT model.

[0037] Figure 5 Learn layer structure diagrams for dynamic graphs.

[0038] Figure 6 Graph of the graph attention module.

[0039] Figure 7 To improve the overall framework diagram of the DSGAT model.

[0040] Figure 8 This is the setting diagram of the experimental simulation environment and parameters.

[0041] Fig. 9 This is a comparison chart of the convergence speed.

[0042] Fig.10 This is a graph showing the change in the root mean square error (RMSE) between the two data sets.

[0043] Fig.11 This is a graph of CORR values ​​(linear relationship between true values ​​and predicted values).

[0044] Fig.12 Comparison of the accuracy of threat prediction by different methods.

[0045] Fig.13 This is a comparison chart of the RSE values ​​of different models on two data sets when the parameter Layer changes. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are 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.

[0047] Reference Figure 1 and Figure 2 As shown, the present embodiment provides a method for power time series prediction based on improved dynamic and static graph attention, which specifically includes the following steps:

[0048] Step 1: Use the power time series data missing processing module in the TVDNN model to correct and fill in the relevant data of active power missing, so as to ensure the integrity of the power time series data. The size of the date window of historical data is set to T h In addition, the missing data can be filled by using the daily moving average method using the relevant data at the same time of the historical date in the date window, thereby fitting the fluctuation trend of the current missing data and the historical data at some times.

[0049] Specifically, the specific calculation formula for fitting the fluctuation trend of the current missing data and historical data at some moments is as follows:

[0050]

[0051] Among them, val f Indicates the fill value; val(h) indicates the historical data at the same time the day before the current date.

[0052] Step 2: For the active power time series data, at time T, the power data value has a strong dependence on the data at the previous time. The horizontal short-term year-on-year anomaly detection method can be used to obtain the periodic characteristics followed by the data in the most recent time window T. In the experiment, T is set to n, and the fill value at the previous moment is compared with the value of the past n-1 moments. If it is greater than the threshold, count is increased by 1; otherwise, count is set to nums, and the data point can be considered as an abnormal point.

[0053] Specifically, the comparison formula is as follows:

[0054]

[0055] In the formula, H represents the short-term month-on-month horizontal anomaly detection result (0 represents no anomaly; 1 represents anomaly); val i represents the data at time i (i∈1,…,n-1); t d Indicates a dynamic threshold.

[0056] Step 3: After the missing active power is supplemented, due to the different power loads generated by different industries, the experiment performs weighted summing and averaging of the active power of different industries at the substation level, city level, and province level, and calculates the proportion of induction motors in a step-by-step manner.

[0057] Specifically, the specific calculation formula for the proportion of induction motors that are aggregated from electricity consumption of different industries to substations, from different substations to various prefecture-level cities, and from various prefecture-level cities to various provinces is as follows:

[0058]

[0059] In the formula, km s represents the proportion of induction motors in the substation; n represents the type of electricity used by the industry under the substation; k n Indicates the corresponding industry weight; P n Indicates the active power of the corresponding industry; km c Indicates the proportion of induction motors in prefecture-level cities; km p Indicates the proportion of induction motors in a province.

[0060] Step 4: Because the induction motor ratio data may have trends and periodicities and be affected by holidays and potential characteristics of long-term historical data, the experiment also designed an induction motor ratio prediction module based on a dual-view deep neural network model.

[0061] Specifically, the specific calculation formula of the predicted value of the induction motor ratio from the perspective of prediction results and shallow time series feature decomposition is as follows:

[0062]

[0063] Where t represents the time; P(t) and L(t) represent the predicted values; w1 and w2 represent the weights; Y(t) represents the final prediction result; g(t) represents the logistic regression growth module; s(t) represents the periodic module; h(t) represents the irregular fluctuation module.

[0064] Step 5: In order to discover the nonlinear growth trend changes hidden in the power time series, the experiment uses a piecewise logistic regression growth algorithm.

[0065] Specifically, the specific calculation formula of the piecewise logistic regression growth algorithm is as follows:

[0066]

[0067] In the formula, g(t) represents trend change, s(t) represents periodic change, and C(t) represents the saturation value of the curve; (k+a(t) T ) represents the curve growth rate; δ = δ j Indicates t j Changes in time; m+a(t) T represents the midpoint of the curve; γ represents s j The offset of the moment smoothing; s j represents the jth change point; P represents the period; a n ,b n represents the smoothing parameter; N represents the Fourier order.

[0068] Step 6: Reference Figure 3As shown in the figure, the experiment uses a long short-term memory neural network (LSTM) to realize the deep temporal feature mining perspective, and adds a forget gate, input gate, and output gate to the traditional structure to select information to retain very important information. Thus, an improved long short-term memory network structure is obtained.

[0069] Step 7: Reference Figure 4 In order to mine the long-term and short-term correlation patterns implied between power time series metrics, the DSGAT model uses two graph learning layers to learn the dynamic and static adjacency matrices respectively. The temporal convolution module uses a gating mechanism to extract time series correlations, which is mainly composed of two parallel multi-scale convolutional neural network layers to process power time series data. In the graph attention module, DSGAT uses two independent graph attention networks to summarize the information of dynamic and static matrices to optimize time series features. When the DS-GAT model is applied to reality, time series data not only contains long-term features but also has constantly changing short-term characteristics. In view of this, the experiment introduces a dynamic and static graph learning module, and adaptively learns to generate the mapping function of the dynamic graph and the adjacency matrix of the static graph through the dynamic and static graph learning layer to capture the hidden relationship between different data.

[0070] Step 8. Reference Figure 5 ,In order to mine more information between different variables, the experiment combines the idea of ​​multi-head attention based on Transformer to construct the dynamic matrix.

[0071] Specifically, the specific calculation formula of the attention matrix is ​​as follows:

[0072]

[0073] The specific calculation formula for summing the multi-head attention matrix is ​​as follows:

[0074]

[0075] In the formula, Attention(Q,K) represents the attention matrix; the dropout layer is used to improve the generalization performance of the model. In addition, a linear layer is used to project the head i , and sum the multi-head attention matrix to get MultiHead∈R N×N .W h , W i Q With W i K They all represent parameter matrices.

[0076] Then use Dot products are performed as residual connections, and multi-layer perceptrons are used for feature projection to obtain dynamic graphs.

[0077] Specifically, the specific calculation formula for using a multi-layer perceptron to perform feature projection to obtain a dynamic graph is as follows:

[0078]

[0079] In the formula, E r =W r X, LN both represent normalization with dimension d; W r ,W1,W2 both represent parameter matrices; b1,b2 represent bias parameters.

[0080] Step 9. In order to make the static graph have better convergence and interpretability, the experiment designed an information fusion layer based on the gated recurrent unit.

[0081] Specifically, the specific calculation formula of the information fusion process is as follows:

[0082]

[0083] Where E∈R N×d represents a given node embedding; X represents the power time series input data; X T Indicates dimension; W, U both represent weight matrices; Represents a new node embedding.

[0084] Then the static dependency matrix SA between different nodes is calculated. The specific calculation formula of the static dependency matrix SA between different nodes is as follows:

[0085]

[0086] In the formula, ReLU and Softmax both represent activation functions; the former is responsible for filtering operations, and the latter is responsible for normalizing the adaptive static matrix.

[0087] Then, in order to mine the periodic characteristics of power time series data variables, the experiment was improved based on the time convolution network, and a multi-scale time convolution module was designed based on it. The specific calculation formula of the multi-scale time convolution module is as follows:

[0088]

[0089] The tanh activation function is used as a feature filter, and the sigmoid activation function is used as a filter to control the amount of information passed to the next module in the process, S∈R T represents one-dimensional time series input; f 1×2 ∈R 2 ,f 1×3 ∈R 3 ,f1×6 ∈R 6 ,f 1×7 ∈R 7 All represent filters of different scales. The purpose of the graph attention module is to weightedly fuse the information of different nodes with the information of neighboring nodes.

[0090] Step 10: Reference Figure 6 In view of the problem that different information propagation steps will cause the graph attention network to be over-smoothed, the experiment retains the original state of some nodes during the propagation process to achieve multi-faceted properties. Then, in order to retain some lost node information to prevent the information from being affected by noise in the extreme case that there is no space, an information selection step is introduced to filter out important information.

[0091] Specifically, the specific calculation formula for the information propagation step is as follows:

[0092]

[0093] The specific calculation formula of the information selection step is as follows:

[0094]

[0095] In the formula, β represents a hyperparameter; K represents the propagation depth; W (k) represents the parameter matrix. In the experiment, the parameter matrix is ​​regarded as a feature selector; H in Represents the hidden state of the previous layer output; H out Represents the hidden state of the current convolutional layer output. In order to model the interaction between the long-term and short-term modes, the experiment uses two graph attention modules to aggregate the node information on the dynamic graph and the static graph. The specific calculation formula for the aggregated output is as follows:

[0096] H f =H static +H dynamic

[0097] Step 11: Reference Figure 7 , and finally there is a skip connection layer, which is used to standardize the information aggregated by the graph attention module to obtain the same sequence length and pass it to the output module. The output module consists of a convolutional neural network with two filters, which converts the input channel size into the required output size. The experiment uses multivariate power time series data to carry out single-step prediction and obtain the final improved DSGAT model.

[0098] Example

[0099] refer to Figure 8In order to verify the actual performance of the electric power time series prediction model based on improved dynamic and static graph attention (IDS-GAT) constructed by the present invention, the experiment selected a wind power energy prediction model based on graph attention network and bidirectional deep learning long short-term memory (Graph Attention Network and Bidirectional Deep Learning Long Short Term Memory, GAN-LSTM), a wind power effective prediction model based on improved EMD (IEMD), and an ocean environment information prediction model based on big data and time series (IClustering-ANN) for comparison with the method constructed by the research institute. At the same time, in order to avoid accidental errors in the data during the experiment, it is necessary to ensure that all experimental parameters are set the same. For the setting of experimental simulation environment and parameters, see Figure 8 .

[0100] refer to Fig. 9 The experiment selected two public benchmark datasets related to the power system, the Solar-Energy dataset and the Electricity dataset, for the power time series prediction task. The higher the fitness value, the faster the model converges to a stable state. Therefore, the experiment first compares the fitness of the four algorithms running on different datasets. For details, see Fig. 9 . Fig. 9 In the figure, (a) shows the convergence speed changes of the four methods on the Solar-Energy dataset. It can be found that when the system iterates to the 30th time, the fitness value of the IDS-GAT model reaches a stable state and has the highest value, up to 99.01. When the system iterates to 104 times, 58 times and 42 times respectively, the fitness values ​​of the GAN-LSTM model, IEMD model and IClustering-ANN model reach the maximum values, which are 93.58, 85.97 and 79.01 respectively. Fig. 9 In the figure, (b) shows the change in convergence speed of different models on the Electricity dataset. When the system iterates to the 16th time, the fitness value of the IDS-GAT model begins to remain stable, with a value as high as 99.88; while the fitness values ​​of the other three algorithms are still changing. When the number of iterations of the system is about 60 times, the fitness values ​​of the GAN-LSTM model, the IEMD model, and the IClustering-ANN model begin to stabilize. By comparison, the IDS-GAT model has a faster convergence speed and can reach a stable operation state more quickly when monitoring and predicting power time series data.

[0101] refer to Fig.10 Then, the average relative errors of different algorithms on the two data sets are compared. The specific results are shown in Fig.10 . Fig.10 In the figure, (a) shows the change of the root mean square difference of the four methods on the Solar-Energy dataset. As time goes on, the root mean square difference of all algorithms increases to varying degrees. When the time is 100s, the RMSE value of the IDS-GAT model is 2.14% and begins to stabilize. However, during the entire process of time, the RMSE values ​​of the GAN-LSTM model, the IEMD model, and the IClustering-ANN model did not show a stable value. In addition, they have been in a state of continuous increase and change, and the RMSE value is getting larger and larger. Fig.10 In the figure, (b) shows the change of the root mean square difference of different models on the Electricity dataset. The RMSE value of the IDS-GAT model is 4.66%, which has remained the minimum value. The RMSE values ​​of the other three algorithms have been increasing, far greater than the IDS-GAT model. In summary, the performance of the IDS-GAT model on the two datasets is constantly being optimized, and the prediction accuracy is gradually improving.

[0102] refer to Fig.11 Then, the changes of CORR values ​​on the two data sets are compared. Fig.11 . Fig.11 In the figure, (a) and (b) are the changes in the CORR values ​​of the four methods on the Solar-Energy dataset and the Electricity dataset, respectively. It can be found that with the increase in system iterations, the CORR values ​​of the four algorithms began to increase continuously. Among them, on the Solar-Energy dataset, when the system iterated to 30 times, the CORR value of the IDS-GAT model was the largest, up to 0.982; on the Electricity dataset, when the system iterated to 36 times, the CORR value of the IDS-GAT model was as high as 0.756. The CORR values ​​of the IDS-GAT model on the two datasets have been significantly higher and are much larger than the other three algorithms. This indicates that the IDS-GAT model is more accurate in predicting power timing and has a more stable operating state during the prediction process.

[0103] refer to Fig.12 Finally, the constructed model is applied to the power system of a company in the United States, and the actual threats suffered by the power sequence are compared with the threat predictions obtained by the model. Fig.12 . Fig.12It can be seen that as time goes by, the prediction accuracy of the threats to the power system by the four methods shows a state of continuous fluctuation. When the time is 200s, the prediction accuracy of the IEMD algorithm for threats has a minimum value, approaching 94%, and the fluctuation range of the prediction accuracy during the entire change process is 94% to 99%. At the same time, the fluctuation range of the accuracy of the GAN-LSTM model and the IClustering-ANN model for threat prediction is also between 94% and 99%, but the range of accuracy fluctuation is smaller. The prediction accuracy of the IDS-GAT model for threats is more intensive, and the fluctuation range is concentrated around 98%, and then approaches 99%. By comparison, it can be seen that the IDS-GAT model has a higher prediction accuracy for the threats suffered by the power system, and can be used as the main method for threat prediction.

[0104] refer to Fig.13 Then, the RSE values ​​of the four models for predicting the threat to the power system are compared when the parameter Layer changes. Fig.13 . Fig.13 In the example, the IDS-GAT method is set to A, the GAN-LSTM method is set to B, the IEMD method is set to C, and the IClustering-ANN method is set to D. Fig.13 In the figure, (a) shows the RSE value changes of the four methods on the Solar-Energy dataset. It can be found that with the increase of the parameter Layer, the RSE values ​​of the four methods show the same trend of fluctuation. However, the RSE value of the IDS-GAT method has always been the smallest, with a numerical fluctuation range of 0.021 to 0.035. The RSE values ​​of the GAN-LSTM method, the IEMD method, and the IClustering-ANN method are significantly greater than the IDS-GAT method. Among them, the RSE value of the IClustering-ANN method has always been the largest, up to 0.385. Fig.13 In the figure, (b) shows the RSE value changes of different models on the Electricity dataset. It can be seen that the RSE value of the IDS-GAT method has been fluctuating between (0.050 and 0.10) and is significantly smaller than the other three algorithms. The above results show that the IDS-GAT method has the smallest RSE value for predicting threats to the power system, has higher accuracy, and can achieve accurate prediction of threats to the power system.

[0105] This method aims to improve the accuracy and robustness of prediction by using a power time series prediction model based on improved dynamic and static graph attention, which often ignores the spatiotemporal characteristics of power grid data in traditional prediction models, and leads to the problem of insufficient prediction accuracy. The results show that on the Solar-Energy dataset and the Electricity dataset, when the system iterates to the 30th and 16th times respectively, the fitness values ​​of the present invention are as high as 99.01 and 99.88 respectively. At the same time, on the Solar-Energy dataset, when the running time is 100s, the RMSE value of the present invention is 2.14%; and when the system reaches 30 times, the CORR value of the present invention is as high as 0.982. In practical applications, with the increase of the parameter Layer, the RSE value of the present invention has been at the minimum, and the numerical fluctuation range is between 0.021 and 0.035. The above results show that the present invention has significant advantages in power time series and threat prediction tasks, and has a high prediction accuracy; the experiment provides a useful reference for the dispatching, operation and management of the power system, and has important practical significance for improving the safe, stable and economic operation of the power system.

[0106] The method has significant advantages in power time series and threat prediction tasks, with high prediction accuracy, providing a useful reference for the dispatching, operation and management of power systems, and has important practical significance for improving the safe, stable and economic operation of power systems. By combining dynamic and static graph attention mechanisms, the model can capture the complex spatiotemporal dependencies of power grid data, thereby significantly improving the accuracy of power time series prediction. The high fitness values ​​on the Solar-Energy and Electricity datasets prove this, providing a more reliable basis for the dispatching and decision-making of power systems. Accurate power time series prediction helps the power grid dispatching center to discover potential supply and demand imbalances and operational risks in advance, so as to take corresponding measures for prevention and adjustment. This helps to enhance the stability and reliability of the power grid and reduce operational failures and power outages caused by inaccurate predictions. By predicting changes in power load, the power grid can arrange power generation plans and resource allocation more accurately. This helps to reduce energy waste, improve energy utilization efficiency, reduce operating costs, and achieve economical and efficient operation of the power system. With the large-scale access of renewable energy, the importance of power time series prediction is becoming increasingly prominent. The prediction model of the present invention can better adapt to the volatility and uncertainty of renewable energy, and provides strong support for the integration and optimized operation of renewable energy. The present invention is an important part of smart grid construction. By improving the accuracy and robustness of power time series prediction, it helps to promote the development of smart grids and realize intelligent, automated and efficient operation of power grids.

[0107] In one embodiment, a power timing prediction device based on improving dynamic and static graph attention is also provided, comprising:

[0108] The acquisition module is used to acquire power time series data.

[0109] An output module is used to input the power time series data into a time series prediction model and output a prediction result. The time series prediction model includes:

[0110] Multi-scale temporal convolution module, which uses gating mechanism to extract temporal correlation;

[0111] The graph attention module uses two independent graph attention networks to aggregate the information of dynamic and static matrices to optimize temporal features;

[0112] The dynamic and static graph learning module adaptively generates the mapping function of the dynamic graph and the adjacency matrix of the static graph through the dynamic and static graph learning layer, and captures the hidden relationship between different data.

[0113] In one embodiment of the present invention, a computer device is provided, the computer device comprising a processor and a memory, the memory being used to store a computer program, the computer program comprising program instructions, and the processor being used to execute the program instructions stored in the computer storage medium. The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc., which are the computing core and control core of the terminal, which are suitable for implementing one or more instructions, and are specifically suitable for loading and executing one or more instructions to implement the corresponding method flow or corresponding functions; the processor described in the embodiment of the present invention can be used to implement an operation of a power timing prediction method based on improved dynamic and static graph attention.

[0114] In one embodiment of the present invention, a method for predicting power timing based on improved attention of dynamic and static graphs can be stored in a computer-readable storage medium if it is implemented in the form of a software functional unit and sold or used as an independent product. Based on this understanding, the present invention implements all or part of the process in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned method embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. Computer-readable storage media include permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules or other data.

[0115] The computer storage medium can be any available medium or data storage device that can be accessed by a computer, including but not limited to magnetic storage (such as floppy disks, hard disks, magnetic tapes, magneto-optical disks (MO), etc.), optical storage (such as CD, DVD, BD, HVD, etc.), and semiconductor storage (such as ROM, EPROM, EEPROM, non-volatile memory (NANDFLASH), solid-state drive (SSD)), etc.

[0116] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0117] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products of the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0118] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0119] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0120] In the present invention, the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" etc. mean that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the described specific features, structures, materials or characteristics may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0121] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A power time series prediction method based on improved dynamic and static graph attention, characterized in that: include: Obtain power time series data; The power time series data is input into a time series prediction model, and a prediction result is output; wherein the time series prediction model includes: Multi-scale temporal convolution module, which uses gating mechanism to extract temporal correlation; The graph attention module uses two independent graph attention networks to aggregate the information of dynamic and static matrices to optimize temporal features; The dynamic and static graph learning module adaptively generates the mapping function of the dynamic graph and the adjacency matrix of the static graph through the dynamic and static graph learning layer, and captures the hidden relationship between different data.

2. According to claim 1, a method for predicting power time series based on improved dynamic and static graph attention is characterized in that: The gating mechanism is used to extract the timing correlation, specifically: A multi-scale convolutional neural network layer with a gating mechanism is used to process power time series data and extract time series correlation.

3. The power time series prediction method based on improved dynamic and static graph attention according to claim 1 is characterized in that: The two independent graph attention networks are used to summarize the information of dynamic and static matrices to optimize the timing features, specifically: Two independent graph attention networks are integrated to summarize the information of dynamic and static matrices respectively and optimize the temporal features.

4. The power time series prediction method based on improved dynamic and static graph attention according to claim 1 is characterized in that: The time series prediction model also includes: Combined with the Transformer-based multi-head attention idea, a dynamic matrix is ​​constructed, and the dynamic graph is obtained through linear layer projection and summation operations.

5. The power time series prediction method based on improved dynamic and static graph attention according to claim 1 is characterized in that: The time series prediction model also includes: Skip connection and output module,The skip connection layer is set to normalize the aggregated information, and the channel size is converted to the required output size through a convolutional neural network composed of two filters.

6. A power timing prediction device based on improved dynamic and static graph attention, characterized in that: include: An acquisition module, used for acquiring power time series data; An output module is used to input the power time series data into a time series prediction model and output a prediction result; wherein the time series prediction model includes: Multi-scale temporal convolution module, which uses gating mechanism to extract temporal correlation; The graph attention module uses two independent graph attention networks to aggregate the information of dynamic and static matrices to optimize temporal features; The dynamic and static graph learning module adaptively generates the mapping function of the dynamic graph and the adjacency matrix of the static graph through the dynamic and static graph learning layer, and captures the hidden relationship between different data.

7. A device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of a method for power timing prediction based on improving dynamic and static graph attention as described in any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of a method for power timing prediction based on improved dynamic and static graph attention as described in any one of claims 1 to 5 are implemented.

Images