Power load prediction method and device based on multi-head attention mechanism fusion graph convolutional network and time sequence

Abstract

The embodiment of the invention provides a power load prediction method and device based on a multi-head attention mechanism fusion graph convolutional network and a time series, and the method comprises the steps: constructing a power system topology model according to the topological relation between equipment in a power system; acquiring node features of each node and node features of neighbor nodes in the power system topology model, and performing feature aggregation splicing of multiple attention heads on the node features of each node and the node features of the neighbor nodes in the power system topology model by using a multi-head attention mechanism fusion graph convolutional network, the updated node features of each node are obtained; acquiring historical time sequence features of each node, and performing fusion processing on the historical time sequence features of each node and the updated node features to obtain fusion feature data; and inputting the fusion feature data into a trained deep neural network prediction model to obtain a power load prediction result of each node.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a power load prediction method and device based on a multi-head attention mechanism, a graph convolution network and a time series. BACKGROUND

[0002] In modern power systems, there are many methods for current power load prediction. Traditional prediction methods such as regression analysis, time series method and grey model have certain effects and stability when dealing with simple and linear load changes. However, these methods often have difficulty in accurately capturing the complex nonlinear characteristics and uncertainty factors of power load. Intelligent prediction methods such as artificial neural networks, support vector machines and fuzzy logic have certain advantages in dealing with complex relationships, but also have some limitations, such as neural networks being prone to local optimal solutions, resulting in inaccurate prediction results. In addition, existing technologies only rely on historical load data and historical device power data, but do not comprehensively depict the overall operation system of the factory, and the correlation between them is missing, which affects the prediction accuracy. SUMMARY

[0003] Therefore, the embodiments of the present application provide a power load prediction method based on a multi-head attention mechanism, a graph convolution network and a time series. One or more embodiments of the present application also relate to a power load prediction device based on a multi-head attention mechanism, a graph convolution network and a time series, a computing device, a computer readable storage medium and a computer program to solve the technical defects in the prior art.

[0004] According to a first aspect of the embodiments of the present application, a power load prediction method based on a multi-head attention mechanism, a graph convolution network and a time series is provided, comprising: obtaining a topological relationship between devices in a power system, and constructing a power system topological model according to the topological relationship between devices in the power system; obtaining node features of each node in the power system topological model and node features of neighbor nodes thereof, and using a multi-head attention mechanism to fuse graph convolution networks to perform feature aggregation and splicing of a plurality of attention heads on the node features of each node in the power system topological model and the node features of neighbor nodes thereof, to obtain updated node features of each node; obtaining historical time series features of each node in the power system topological model, and obtaining fusion feature data of each node by fusing the historical time series features of each node with the updated node features; obtaining a power load prediction result of each node by inputting the fusion feature data of each node into a trained deep neural network prediction model.

[0005] Preferably, the constructing a power system topology model according to the topological relationship between devices in the power system comprises: According to the topological relationship between devices in the power system, each device node in the power system is modeled as a node, the current flow relationship or load dependency relationship between device nodes is modeled as an edge, and the device node having a current flow relationship or load dependency relationship with the device node is modeled as a neighbor node; According to all the nodes, edges and neighbor nodes in the power system, a power system topology model is constructed.

[0006] Preferably, the multi-head attention mechanism fusion graph convolution network performs feature aggregation and splicing of the node features of each node in the power system topology model and the node features of its neighbor nodes through multiple attention heads, to obtain the updated node features of each node, comprising: By respectively performing linear transformation of the node features of each node and the node features of its neighbor nodes through the kth attention head, the linear transformation features of each node through the kth attention head and the linear transformation features of its neighbor nodes through the kth attention head are obtained; Using the linear transformation features of each node through the kth attention head and the linear transformation features of its neighbor nodes through the kth attention head, the attention coefficients between each node and its neighbor nodes through the kth attention head are calculated; Using the attention coefficients between each node and its neighbor nodes through the kth attention head and the linear transformation features of its neighbor nodes through the kth attention head, the aggregation features of each node through the kth attention head are calculated; By splicing the aggregation features of all heads of each node, the updated node features of each node are obtained.

[0007] Preferably, the splicing the aggregation features of all heads of each node to obtain the updated node features of each node comprises:

[0008] wherein, denotes a node updated node features; denotes a node aggregation features of the kth attention head; || denotes a splicing operation; k denotes the kth attention head; K is the number of multi-head attention mechanisms.

[0009] Preferably, the historical time series features of each node are historical load time series features or historical weather data time series features.

[0010] Preferably, the step of obtaining the power load prediction result of each node by inputting the fusion feature data of each node into the trained deep neural network prediction model comprises:

[0011] wherein: is the power load prediction result of each node; Z is the fusion feature data of each node; is a weight matrix of an output layer of the deep neural network prediction model; is a bias term.

[0012] Preferably, the method further comprises: obtaining a topological relationship between devices in a plurality of power systems and a real result of power load of each device, and constructing a power system topological model according to the topological relationship between devices in each power system; obtaining node features of each node in the power system topological model and node features of neighbor nodes thereof, and performing feature aggregation and splicing of a plurality of attention heads on the node features of each node in the power system topological model and the node features of neighbor nodes thereof by using a multi-head attention mechanism fusion graph convolution network to obtain updated node features of each node; obtaining historical time series features of each node in the power system topological model, and obtaining fusion feature data of each node by fusing the historical time series features of each node with the updated node features; training a deep neural network prediction model based on the fusion feature data of each node and the real result of power load of each device to obtain the trained deep neural network prediction model.

[0013] According to a second aspect of the embodiments of the present application, a power load prediction device based on a multi-head attention mechanism fusion graph convolution network and time series is provided, comprising: an obtaining module configured to obtain a topological relationship between devices in a power system, and construct a power system topological model according to the topological relationship between devices in the power system; an aggregation and splicing module configured to obtain node features of each node in the power system topological model and node features of neighbor nodes thereof, and perform feature aggregation and splicing of a plurality of attention heads on the node features of each node in the power system topological model and the node features of neighbor nodes thereof by using a multi-head attention mechanism fusion graph convolution network to obtain updated node features of each node; The fusion processing module is configured to acquire historical time sequence features of each node in the power system topology model, and obtain fusion feature data of each node by fusing the historical time sequence features of each node with the updated node features. The prediction module is configured to input the fusion feature data of each node into a trained deep neural network prediction model to obtain a power load prediction result of each node.

[0014] According to a third aspect of the embodiments of the present application, a computing device is provided, comprising: a memory and a processor; The memory is configured to store computer executable instructions, and the processor is configured to execute the computer executable instructions to implement the steps of any one of the power load prediction methods based on the multi-head attention mechanism fusion graph convolution network and time sequence.

[0015] According to a fourth aspect of the embodiments of the present application, a computer readable storage medium is provided, which stores computer executable instructions, and the instructions are executed by a processor to implement the steps of any one of the power load prediction methods based on the multi-head attention mechanism fusion graph convolution network and time sequence.

[0016] According to a fifth aspect of the embodiments of the present application, a computer program is provided, and when the computer program is executed in a computer, the computer is caused to perform the steps of the power load prediction method based on the multi-head attention mechanism fusion graph convolution network and time sequence.

[0017] The power load prediction scheme based on the multi-head attention mechanism fusion graph convolution network and time sequence provided by the embodiment of the application, by acquiring the topological relationship between devices in a power system, and constructing a power system topological model according to the topological relationship between devices in the power system, acquiring the node features of each node in the power system topological model and the node features of its neighbor nodes, and using the multi-head attention mechanism fusion graph convolution network to perform feature aggregation and splicing of a plurality of attention heads on the node features of each node in the power system topological model and the node features of its neighbor nodes, obtaining the updated node features of each node, acquiring the historical time sequence features of each node in the power system topological model, and obtaining the fusion feature data of each node by fusing the historical time sequence features of each node with the updated node features, and by inputting the fusion feature data of each node into a trained deep neural network prediction model, obtaining the power load prediction result of each node. The embodiment of the application can more accurately capture the complex relationship between nodes by dynamically selecting the importance of neighbor nodes, and improve the understanding ability of the model to the structural features of the power network. By introducing the multi-head attention mechanism, the expression ability and robustness of the model are enhanced, the influence of multiple neighbor nodes can be considered comprehensively, and information loss and single attention deviation can be avoided. By combining the advantages of the graph convolution network and the time sequence data, the topological relationship between devices and the historical load change rule are fully utilized, and high accuracy and stability of load prediction are realized. The method is suitable for complex power demand scenarios, provides more scientific load prediction and decision support for energy storage enterprises, and promotes efficient operation and optimized management of the power system. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 is a flowchart of a power load prediction method based on a multi-head attention mechanism fusion graph convolution network and time sequence provided by an embodiment of the application; Figure 2 is a flowchart of a deep neural network prediction model training method provided by an embodiment of the application; Figure 3 is a schematic diagram of a power load prediction device based on a multi-head attention mechanism fusion graph convolution network and time sequence provided by an embodiment of the application; Figure 4 is a structural block diagram of a computing device provided by an embodiment of the application. DETAILED DESCRIPTION

[0019] In the following description, many specific details are set forth in order to provide a thorough understanding of the application. However, the application can be practiced in many different ways beyond the specific details disclosed herein, and with modifications apparent to those skilled in the art, and therefore the application is not limited to the specific implementations disclosed herein.

[0020] The terminology used in this disclosure of one or more embodiments is for the purpose of describing particular embodiments only and is not intended to be limiting of one or more embodiments of the disclosure. As used in this disclosure and the appended claims, the singular forms "a," "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will also be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0021] It is to be understood that the singular forms "a," "an," and "the" include plural referents unless the context clearly dictates otherwise. It is to be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0022] In the present application, a power load prediction method based on a multi-head attention mechanism fusion graph convolution network and time series is provided. The present application also relates to a power load prediction device based on a multi-head attention mechanism fusion graph convolution network and time series, a computing device, a computer-readable storage medium, and a computer program, which are described in detail in the following embodiments.

[0023] Figure 1 A flowchart of a power load prediction method based on a multi-head attention mechanism fusion graph convolution network and time series according to an embodiment of the present application is shown, which specifically includes the following steps.

[0024] Step S101: Obtain the topological relationship between devices in a power system, and construct a power system topological model according to the topological relationship between devices in the power system. In an optional implementation, constructing a power system topological model according to the topological relationship between devices in the power system includes: modeling each device node in the power system as a node according to the topological relationship between devices in the power system, modeling the current flow relationship or load dependency relationship between device nodes as an edge, and modeling a device node having a current flow relationship or load dependency relationship with a device node as a neighbor node; and constructing a power system topological model according to all the nodes, edges, and neighbor nodes in the power system.

[0025] Step S102: Obtain the node features of each node in the power system topology model and the node features of its neighbor nodes, and use the multi-head attention mechanism fusion graph convolution network to perform feature aggregation and splicing of the node features of each node in the power system topology model and the node features of its neighbor nodes through multiple attention heads, to obtain updated node features of each node; In an optional implementation, the multi-head attention mechanism fusion graph convolution network performing feature aggregation and splicing of the node features of each node in the power system topology model and the node features of its neighbor nodes through multiple attention heads to obtain updated node features of each node includes: performing linear transformation of the node features of each node and the node features of its neighbor nodes through the kth attention head respectively to obtain linear transformation features of the kth attention head of each node and the linear transformation features of the kth attention head of its neighbor nodes; calculating the attention coefficients of the kth attention head between each node and its neighbor nodes by using the linear transformation features of the kth attention head of each node and the linear transformation features of the kth attention head of its neighbor nodes; calculating the aggregation features of the kth attention head of each node by using the attention coefficients of the kth attention head between each node and its neighbor nodes and the linear transformation features of the kth attention head of its neighbor nodes; and obtaining updated node features of each node by splicing the aggregation features of all heads of each node.

[0026] In an optional implementation, the obtaining of the updated node features of each node by splicing the aggregation features of all heads of each node includes:

[0027] wherein, denotes a node updated node feature; denotes a node aggregation features of the kth attention head; || denotes a splicing operation; k denotes the kth attention head; and K is the number of multi-head attention mechanisms.

[0028] Step S103: Obtain the historical time sequence features of each node in the power system topology model, and obtain the fusion feature data of each node by fusing the historical time sequence features of each node with the updated node features; In an optional implementation, the historical time sequence features of each node are historical load time sequence features or historical weather data time sequence features.

[0029] Step S104: obtaining the power load prediction result of each node by inputting the fusion feature data of each node into the trained deep neural network prediction model.

[0030] In an optional implementation, the obtaining the power load prediction result of each node by inputting the fusion feature data of each node into the trained deep neural network prediction model comprises:

[0031] Wherein: is the power load prediction result of each node; Z is the fusion feature data of each node; is the weight matrix of the output layer of the deep neural network prediction model; is the bias term.

[0032] In an optional implementation, the method further comprises: obtaining the topological relationship between devices in a plurality of power systems and the real result of the power load of each device, and constructing a power system topological model according to the topological relationship between devices in each power system; obtaining the node feature of each node in the power system topological model and the node feature of its neighbor node, and using the multi-head attention mechanism fusion graph convolution network to perform feature aggregation and splicing of a plurality of attention heads on the node feature of each node in the power system topological model and the node feature of its neighbor node, to obtain the updated node feature of each node; obtaining the historical time series feature of each node in the power system topological model, and obtaining the fusion feature data of each node by fusing the historical time series feature of each node and the updated node feature; training a deep neural network prediction model based on the fusion feature data of each node and the real result of the power load of each device, to obtain the trained deep neural network prediction model.

[0033] The power load prediction scheme based on the multi-head attention mechanism fusion graph convolution network and time series provided by the embodiments of the present application aims to dynamically select the importance of neighbor nodes, more accurately capture the complex relationship between nodes, and improve the understanding ability of the model to the structural features of the power network. By introducing the multi-head attention mechanism, the expression ability and robustness of the model are enhanced, the influence of multiple neighbor nodes can be considered comprehensively, and information loss and single attention deviation can be avoided. By combining the advantages of the graph convolution network and the time series data, the topological relationship between devices and the historical load change rule are fully utilized, and high accuracy and stability of the load prediction are realized. The method is suitable for complex power demand scenarios, provides more scientific load prediction and decision support for energy storage enterprises, and promotes the efficient operation and optimized management of the power system.

[0034] The technical solutions of the present application will be further described below Definition of nodes and edges, each device (such as machines, sensors, etc.) will be a node The edge represents the current flow relationship or load-dependent relationship between devices and neighboring devices . The weight of the edge can be based on the current flow between devices, power consumption or other related metrics.

[0035] Example: In a production system, it contains energy storage devices , production equipment , and other non-electricity devices (such as trucks, forklifts, etc. auxiliary equipment). The connection weight between these devices can be determined according to the comprehensive proportion of their electricity consumption and power consumption, and combined with their actual importance in the production process to make empirical adjustments.

[0036] The graph can be represented as , where is the node set, is the edge set. Each node has a feature vector , which can be represented as: [load, current, voltage, device state,...].

[0037] The graph attention network learns the features of the nodes by introducing an attention mechanism, so it can dynamically adjust the feature aggregation process according to the influence of the neighbor nodes. Calculate the attention coefficient: for each node The neighbor nodes , calculate the attention coefficient : Where: is the feature vector of node ; denotes the feature vector of the neighbor node ; is the weight matrix of the linear transformation, which is configured according to random assignment or historical experience parameters; is the weight vector used to calculate attention; II represents the feature splicing operation; softmax is the normalization function, and LeakyReLU is the activation function; The feature of each node is updated as: Where denotes the neighbor set of node ; denotes the updated node feature of node ; σ represents the activation function.

[0038] In power grid load forecasting, the load changes of some devices have greater impact on the target device, while some have less impact. The attention mechanism is introduced to enable the model to automatically learn and identify the importance of different neighbor nodes to the current node. The attention coefficient is a dynamic weight learned from node features, representing the actual contribution of neighbor node information to the target node under the current model. The attention mechanism can be seen as giving each edge a dynamic, data-driven weight based on the existing edge (current flow relationship). This weight not only considers the existence of the edge, but also considers the current features of the connected devices, reflecting their different importance to load forecasting. That is, the attention coefficient is the attention degree of the model to the neighbor device (dynamic edge weight) calculated based on node features, enhancing the flexibility of the model in identifying the impact of different neighbors. Through this mechanism, the load impact of neighbor devices can be more accurately combined to improve the overall load forecasting effect To enhance the expressiveness of the model, different attention weights are multiplied to different nodes, as follows: where is the number of heads, U represents the sequential concatenation of the outputs of multiple attention heads, is the neighbor node set of node . is the attention coefficient between node and its neighbor . It is usually calculated as follows: first, use linear transformation to obtain node features , then for each pair of nodes , calculate the attention coefficient between node and its neighbor , where is a learnable weight vector used to calculate attention, and II represents the concatenation operation. Finally, feature aggregation, each attention head weights and aggregates the features of neighbor nodes according to the calculated attention coefficient, including: for each attention head, the features of neighbor nodes are first linearly transformed into , then weighted and summed according to the calculated attention coefficient , processed by the activation function σ to obtain the output feature of node of this attention head. Finally, the outputs of all heads are concatenated to obtain the updated feature of node .

[0039] Specifically, the following steps are included: 1) Linear transformation: on the current node and its neighbor nodes feature vectors , Linear transformation of the k-th attention head:

[0040] 2) Calculate attention energy (unnormalized attention coefficient) for the current node and its neighbor nodes Concatenate the transformed features, and calculate:

[0041] 3) Normalize the attention coefficient Do softmax for all neighbor nodes

[0042] 4) Weighted aggregation of neighbor node features Feature aggregation is the weighted sum, and an activation function is added:

[0043] 5) Multi-attention head concatenation Concatenate the aggregation outputs of all attention heads:

[0044] The neighbor node set N(i) is defined based on the edge set in the graph structure G=(V,E). Specifically: Further, for the current node , its neighbor node set N(i) is the set of all nodes directly connected to node through edge . In the power network scenario, nodes represent devices, and edges represent the current flow relationship or load dependency relationship between devices. Therefore, the neighbor node is the device node that has a direct connection with device in terms of power flow or dependency. This neighbor relationship can be quickly indexed through the adjacency matrix or adjacency list of the graph. For example, if the element Aij in the adjacency matrix A is not equal to 0, then node is a neighbor of node .

[0045] Time series data processing: Time series data (such as historical load data) can be represented as a matrix , where each row corresponds to a time step feature (e.g., past 24 hours of load, weather data, etc.).

[0046] ​Feature fusion: combine the node features aggregated by the above model with the time series features: where is the node feature of the last layer, i.e., when is L-1. Wherein, the symbol represents the splicing operation of the feature matrix.

[0047] Prediction layer, the combined features are input into a fully connected layer for load prediction: where: is the predicted load value. is the weight matrix of the output layer. is the bias term. Use mean square error (MSE) as the loss function for model training: where is the number of samples, is the true load value.

[0048] The deep neural network prediction model training method adopted by the present application is based on the back propagation algorithm and gradient descent optimization technology, which adjusts the network parameters through iteration to improve the fitting and prediction ability of the model to the input data, as shown in Figure 2 The specific steps include: Step S201: data preprocessing: normalize and standardize the input model aggregated node feature data and time series load data to ensure the stability of the data distribution and improve the training efficiency.

[0049] Step S202: forward propagation: input the preprocessed data into the deep neural network prediction model, and calculate the output prediction value through linear transformation and nonlinear activation function of each layer of neurons.

[0050] Step S203: error calculation: use the defined loss function (such as mean square error MSE) to calculate the difference between the predicted output and the true load value, and quantify the prediction error of the model.

[0051] Step S204: back propagation: based on the error value, use the chain rule to calculate the gradient of each layer parameter, and pass the error information forward layer by layer to guide the parameter update direction.

[0052] Step S205: parameter update: use optimization algorithms (such as Adam optimizer or stochastic gradient descent SGD) to adjust the network weights and biases according to the calculated gradient, and gradually reduce the prediction error.

[0053] Step S206: training iteration: repeat the steps of forward propagation, error calculation, back propagation and parameter update until the preset stopping condition (such as reaching the maximum training round or error convergence threshold) is met.

[0054] Step S207: Model verification and tuning: Use the independent validation set to evaluate the performance of the trained model, adjust the hyperparameters (such as learning rate, batch size, network layer number, etc.) to optimize the model effect, and effectively prevent overfitting phenomenon.

[0055] The application can fully utilize the relationship between devices and historical load patterns by combining the graph convolution network and the load prediction algorithm of time series data, improve the accuracy and robustness of load prediction. The model designed in this way can adapt to complex power demand scenarios and provide better decision support for energy storage enterprises.

[0056] The application also provides a power load prediction device based on a multi-head attention mechanism fusing a graph convolution network and time series, Figure 3 An embodiment of the application provides a structure diagram of a power load prediction device based on a multi-head attention mechanism fusing a graph convolution network and time series. As shown in the figure, Figure 3 The device comprises: an acquisition module configured to acquire the topological relationship between devices in a power system, and construct a power system topological model according to the topological relationship between devices in the power system; an aggregation and splicing module configured to acquire the node features of each node in the power system topological model and the node features of its neighbor nodes, and use a multi-head attention mechanism to fuse a graph convolution network to perform feature aggregation and splicing of multiple attention heads on the node features of each node in the power system topological model and the node features of its neighbor nodes, to obtain updated node features of each node; a fusion processing module configured to acquire historical time series features of each node in the power system topological model, and obtain fusion feature data of each node by fusing the historical time series features of each node with the updated node features; and a prediction module configured to input the fusion feature data of each node into a trained deep neural network prediction model to obtain a power load prediction result of each node.

[0057] The above is a schematic scheme of the power load prediction device based on a multi-head attention mechanism fusing a graph convolution network and time series. It should be noted that the technical scheme of the power load prediction device based on a multi-head attention mechanism fusing a graph convolution network and time series is the same as the technical scheme of the power load prediction method based on a multi-head attention mechanism fusing a graph convolution network and time series described above. The technical scheme of the power load prediction device based on a multi-head attention mechanism fusing a graph convolution network and time series, which is not described in detail, can be referred to the description of the technical scheme of the power load prediction method based on a multi-head attention mechanism fusing a graph convolution network and time series.

[0058] Figure 4A structural block diagram of a computing device 400 is shown, according to one embodiment of the present application. The components of the computing device 400 include, but are not limited to, a memory 410 and a processor 420. The processor 420 is connected with the memory 410 through a bus 430, and a database 450 is used to save data.

[0059] The computing device 400 also includes an access device 440 that enables the computing device 400 to communicate via one or more networks 460. Examples of these networks include a public switched telephone network (PSTN), a local area network (LAN), a wide area network (WAN), a personal area network (PAN), or a combination of networks such as the Internet. The access device 440 can include one or more of any type of network interface (for example, a network interface card (NIC)), wired or wireless, such as an IEEE 802.11 wireless local area network (WLAN) wireless interface, a Worldwide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, a near-field communication (NFC) interface, and the like.

[0060] In one embodiment of the present application, the above-mentioned components of the computing device 400 and other components not shown in the above-mentioned components can be connected with each other, for example, through a bus. It should be understood that, Figure 4 Figure 4 The structural block diagram of the computing device shown is only for the purpose of example, and is not a limitation on the scope of the present application. Other components can be added or replaced as needed by those skilled in the art.

[0061] The computing device 400 can be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (for example, a tablet computer, a personal digital assistant, a laptop computer, a notebook computer, a netbook, and the like), a mobile phone (for example, a smartphone), a wearable computing device (for example, a smart watch, smart glasses, and the like), or other types of mobile devices, or a stationary computing device such as a desktop computer or PC. The computing device 400 can also be a mobile or stationary server.

[0062] The processor 420 is configured to execute computer-executable instructions, which, when executed by the processor, implement the steps of the above-mentioned power load prediction method based on the multi-head attention mechanism fusion graph convolutional network and time series.

[0063] ​The above is a schematic scheme of the computing device of the embodiment. It should be noted that the technical scheme of the computing device and the technical scheme of the power load prediction method based on the multi-head attention mechanism fusion graph convolution network and time sequence belong to the same concept, and the details of the technical scheme of the computing device that are not described in detail can be referred to the description of the technical scheme of the power load prediction method based on the multi-head attention mechanism fusion graph convolution network and time sequence.

[0064] An embodiment of the present application further provides a computer readable storage medium, which stores computer executable instructions, and the computer executable instructions are executed by a processor to implement the steps of the power load prediction method based on the multi-head attention mechanism fusion graph convolution network and time sequence.

[0065] The above is a schematic scheme of the computer readable storage medium of the embodiment. It should be noted that the technical scheme of the storage medium and the technical scheme of the power load prediction method based on the multi-head attention mechanism fusion graph convolution network and time sequence belong to the same concept, and the details of the technical scheme of the storage medium that are not described in detail can be referred to the description of the technical scheme of the power load prediction method based on the multi-head attention mechanism fusion graph convolution network and time sequence.

[0066] An embodiment of the present application further provides a computer program, and when the computer program is executed in a computer, the computer program causes the computer to execute the steps of the power load prediction method based on the multi-head attention mechanism fusion graph convolution network and time sequence.

[0067] The above is a schematic scheme of the computer program of the embodiment. It should be noted that the technical scheme of the computer program and the technical scheme of the power load prediction method based on the multi-head attention mechanism fusion graph convolution network and time sequence belong to the same concept, and the details of the technical scheme of the computer program that are not described in detail can be referred to the description of the technical scheme of the power load prediction method based on the multi-head attention mechanism fusion graph convolution network and time sequence.

[0068] The specific embodiments of the present application are described above. Other embodiments are within the scope of the appended claims. In some cases, acts or steps recited in the claims can be performed in an order other than that in which they are recited in the embodiments, and still achieve desirable results. Also, the process depicted in the figures can not necessarily require the particular order shown, or sequential order to achieve the desired results. In certain implementations, multitasking and parallel processing can be advantageous.

[0069] The computer readable medium can include any entity or apparatus capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, software distribution medium, etc. It should be noted that the computer readable medium can include appropriate additions or subtractions according to the requirements of legislation and patent practice in the jurisdiction, for example, in some jurisdictions, according to legislation and patent practice, the computer readable medium does not include electrical carrier signals and telecommunication signals.

[0070] It should be noted that, for the foregoing method embodiments, in order to facilitate description, they are all described as a combination of a series of actions, but those skilled in the art should know that the embodiments of the present application are not limited to the order of the actions described, because according to the embodiments of the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification all belong to preferred embodiments, and the actions and modules involved are not necessarily necessary for the embodiments of the present application.

[0071] In the above embodiments, the description of each embodiment has its own emphasis, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0072] The preferred embodiments of the present application disclosed above are only used to help explain the present application. The alternative embodiments do not describe all the details and limit the application to the specific embodiments described. Obviously, according to the content of the embodiments of the present application, many modifications and changes can be made. The present application selects and describes these embodiments in order to better explain the principles and practical applications of the embodiments of the present application, so that those skilled in the art can well understand and use the present application. The present application is limited by the claims and their entire scope and equivalents.

Claims

1. A power load forecasting method based on a multi-head attention mechanism fusion graph convolution network and time series, characterized in that, The method comprises: acquiring the topological relationship between devices in the power system, and constructing a power system topological model according to the topological relationship between devices in the power system; acquiring the node features of each node in the power system topological model and the node features of its neighbor nodes, and using a multi-head attention mechanism fusion graph convolution network to perform feature aggregation and splicing of the node features of each node in the power system topological model and the node features of its neighbor nodes through multiple attention heads, to obtain updated node features of each node; acquiring historical time sequence features of each node in the power system topological model, and obtaining fusion feature data of each node by fusing the historical time sequence features of each node with the updated node features; inputting the fusion feature data of each node into a trained deep neural network prediction model to obtain power load prediction results of each node.

2. The method of claim 1, wherein, The method comprises: According to the topological relationship between devices in the power system, modeling each device node in the power system as a node, modeling the current flow relationship or load dependency relationship between device nodes as an edge, and modeling the device nodes having a current flow relationship or load dependency relationship with the device nodes as neighbor nodes; According to all the nodes, edges and neighbor nodes in the power system, a power system topological model is constructed.

3. The method of claim 2, wherein, The multi-head attention mechanism fusion graph convolution network performs feature aggregation and splicing of the node features of each node in the power system topological model and the node features of its neighbor nodes through multiple attention heads, to obtain updated node features of each node, which comprises: performing linear transformation of the node features of each node and the node features of its neighbor nodes through the kth attention head respectively, to obtain linear transformation features of the kth attention head of each node and its neighbor nodes; using the linear transformation features of the kth attention head of each node and its neighbor nodes to calculate the attention coefficients of the kth attention head between each node and its neighbor nodes; using the attention coefficients of the kth attention head between each node and its neighbor nodes and the linear transformation features of the kth attention head of its neighbor nodes to calculate the aggregation features of the kth attention head of each node; performing splicing processing on the aggregation features of all heads of each node to obtain updated node features of each node.

4. The method of claim 3, wherein, The method comprises: ； wherein, representing a node updated node features; representing a node aggregated features of the k-th attention head; || represents a concatenation operation; k represents the k-th attention head; K is the number of multi-head attention mechanisms.

5. The method of claim 1, wherein, The historical time sequence features of each node are historical load time sequence features or historical weather data time sequence features.

6. The method of claim 1, wherein, The method comprises: ； wherein: is the power load prediction result of each node; Z is the fusion feature data of each node; is the weight matrix of the output layer of the deep neural network prediction model; is the bias term.

7. The method of claim 1, wherein, Further comprising: Obtain the topological relationship between devices in a plurality of power systems and the actual power load of each device, and construct a power system topological model according to the topological relationship between devices in each power system; Obtain the node features of each node in the power system topological model and the node features of its neighbor nodes, and use a multi-head attention mechanism fusion graph convolution network to aggregate and splice the node features of each node in the power system topological model and the node features of its neighbor nodes through multiple attention heads, to obtain updated node features of each node; Obtain the historical time series features of each node in the power system topological model, and obtain the fusion feature data of each node by fusing the historical time series features of each node with the updated node features; Train a deep neural network prediction model based on the fusion feature data of each node and the actual power load of each device, to obtain the trained deep neural network prediction model.

8. A power load prediction device based on a multi-head attention mechanism fusion graph convolution network and time series, characterized in that, Comprise: An acquisition module configured to obtain the topological relationship between devices in a power system, and construct a power system topological model according to the topological relationship between devices in the power system; An aggregation and splicing module configured to obtain the node features of each node in the power system topological model and the node features of its neighbor nodes, and use a multi-head attention mechanism fusion graph convolution network to aggregate and splice the node features of each node in the power system topological model and the node features of its neighbor nodes through multiple attention heads, to obtain updated node features of each node; A fusion processing module configured to obtain the historical time series features of each node in the power system topological model, and obtain the fusion feature data of each node by fusing the historical time series features of each node with the updated node features; A prediction module configured to input the fusion feature data of each node into the trained deep neural network prediction model, to obtain the power load prediction result of each node.

9. A computing device comprising: a memory and a processor; the memory is configured to store computer executable instructions, and the processor is configured to execute the computer executable instructions, which, when executed by the processor, implement the steps of the power load prediction method based on the multi-head attention mechanism fusion graph convolution network and time series according to any one of claims 1 to 7.

10. A computer readable storage medium storing computer executable instructions, which, when executed by a processor, implement the steps of the power load prediction method based on the multi-head attention mechanism fusion graph convolution network and time series according to any one of claims 1 to 7.

Images