Efficient edge computing method and related device for smart grid

Abstract

An efficient edge computing method and related device for a smart grid. The method comprises: collecting electric power series data of an electric meter by means of a terminal device, and preprocessing the electric power series data; uploading the preprocessed electric power series data to a terminal edge layer model, so that the terminal edge layer model performs series slicing on the electric power series data to obtain short series segment data; using an improved Transformer comprising a low-rank self-attention mechanism to perform modeling analysis on the short series segment data so as to complete model training; distilling knowledge obtained during model training to a multilayer perception unit, so that the multilayer perception unit performs inference analysis on the short series segment data; and deploying the multilayer perception unit obtained by means of knowledge distillation in a terminal edge computing platform for use in processing various electric power time series data modeling analysis tasks. Thus, the response speed and efficiency of a smart grid edge computing platform in processing a large amount of electric power time series data are improved.

Description

An intelligent power grid efficient edge computing method and related device

[0001] The present application claims priority to the Chinese patent application No. 202410638537.4, filed on May 22, 2024, and entitled "An intelligent power grid efficient edge computing method and related device", the whole content of which is incorporated herein by reference. TECHNICAL FIELD

[0002] The present application relates to the technical field of intelligent power grid, in particular to an intelligent power grid efficient edge computing method and related device. BACKGROUND

[0003] In recent years, the related technology of intelligent power grid cloud edge interaction scheme covers many aspects of applications such as power data collection, processing and providing intelligent analysis suggestions in power system scenarios. The edge computing technology greatly reduces the data transmission consumption, and effectively meets the basic needs of intelligent power grid in real-time business, application intelligence, security and privacy protection, etc. Benefiting from the rapid development of neural networks, the edge computing platform uses artificial intelligence models to process a large amount of complex power time series data, which has become a hot issue of research. Deep models can help terminal devices focus on mining the implicit dependencies in power time series data, and accordingly make intelligent suggestions for power grid management.

[0004] However, the model based on deep algorithm often causes a large amount of calculation time consumption, and the time delay generated when processing a large amount of power time series data affects the tasks with high timeliness requirements such as fault identification and load prediction in intelligent power grid system. SUMMARY

[0005] The present application provides an intelligent power grid efficient edge computing method and related device, which improves the response speed and efficiency of the intelligent power grid edge computing platform in processing a large amount of power time series data.

[0006] Therefore, the first aspect of the present application provides an intelligent power grid efficient edge computing method, which comprises:

[0007] S1, collecting power sequence data of an electric energy meter by a terminal device, and preprocessing the power sequence data;

[0008] S2, uploading the preprocessed power sequence data to a terminal edge layer model, so that the terminal edge layer model performs sequence slicing on the power sequence data to obtain short sequence segment data;

[0009] S3, modeling and analyzing the short sequence segment data by using a modified Transformer containing a low-rank self-attention mechanism to complete model training.

[0010] S4, distilling the knowledge obtained in the model training process in step S3 to a multi-layer perception unit, so that the multi-layer perception unit performs inference analysis on the short sequence segment data;

[0011] S5, deploying the multi-layer perception unit obtained by knowledge distillation in step S4 to a terminal edge computing platform for processing various power time series data modeling analysis tasks.

[0012] Optionally, step S2 further comprises: normalizing the short sequence segment data, thereby alleviating the difference in data distribution.

[0013] Optionally, the modeling and analysis of the short sequence segment data using the improved Transformer containing the low-rank self-attention mechanism, and completing the model training, comprises:

[0014] Generating a plurality of matrices based on the low-rank self-attention mechanism, and using low-rank factorization to approximately calculate the attention score; weighting the attention score and the short sequence segment data representation to obtain a complete sequence representation of the power data.

[0015] Optionally, the generating a plurality of matrices based on the low-rank self-attention mechanism, and using low-rank factorization to approximately calculate the attention score; weighting the original attention score and the short sequence segment data representation to obtain a complete sequence representation of the power data, comprises:

[0016] Inputting the short sequence segment data into the sequence encoding unit based on the Transformer in the edge layer model to calculate the second-order relationship information encoding between the short sequence segments;

[0017] Calculating the initial attention score based on the second-order relationship information encoding;

[0018] Using a feedforward network to capture nonlinear relationships and obtain the second-order relationship information encoding contained in the short sequence segment data to obtain the final attention score;

[0019] Weighting the attention score representation as the short sequence segment data representation to obtain a complete sequence representation of the power data.

[0020] Optionally, the distilling the knowledge obtained in the model training process in step S3 to a multi-layer perception unit comprises:

[0021] Extracting and injecting the knowledge obtained in the model training process in step S3, and jointly optimizing the time series modeling task and the knowledge distillation task using a joint training framework, thereby obtaining a multi-layer perception unit.

[0022] The second aspect of the application provides a smart grid efficient edge computing system, the system comprises:

[0023] A preprocessing unit is configured to collect power sequence data of an electric energy meter by a terminal device and pre-process the power sequence data;

[0024] A slicing unit is configured to upload the pre-processed power sequence data to a terminal edge layer model, so that the terminal edge layer model performs sequence slicing on the power sequence data to obtain short sequence segment data;

[0025] A training unit is configured to model and analyze the short sequence segment data by using a modified Transformer containing a low-rank self-attention mechanism, and complete model training;

[0026] A distillation unit is configured to distill knowledge obtained in the model training process of the training unit to a multi-layer perception unit, so that the multi-layer perception unit performs inference analysis on the short sequence segment data;

[0027] A deployment unit is configured to deploy the multi-layer perception unit obtained by the distillation unit to a terminal edge computing platform, for processing various power time series data modeling and analysis tasks.

[0028] Optionally, the system further comprises a standardization processing unit configured to perform standardization processing on the short sequence segment data, thereby alleviating data distribution differences.

[0029] Optionally, the training unit is specifically configured to:

[0030] Generate a plurality of matrices based on a low-rank self-attention mechanism, and approximate calculation of attention scores by using low-rank factorization; weight the attention scores and the short sequence segment data representation to obtain a complete power data sequence representation.

[0031] Optionally, the generating a plurality of matrices based on a low-rank self-attention mechanism and the approximate calculation of attention scores by using low-rank factorization; weighting the attention scores and the short sequence segment data representation to obtain a complete power data sequence representation specifically comprises:

[0032] Inputting the short sequence segment data into a sequence encoding unit based on a Transformer in an edge layer model to calculate second-order relationship information encoding between short sequence segments;

[0033] Calculating initial attention scores based on the second-order relationship information encoding;

[0034] Capturing nonlinear relationships and obtaining second-order relationship information encoding contained in the short sequence segment data by using a feedforward network to obtain final attention scores;

[0035] The weighted attention score is represented as the short sequence data representation, resulting in a complete sequence representation of the power data.

[0036] Optionally, the distillation unit is specifically used for:

[0037] The knowledge obtained during model training in step S3 is extracted and injected, and a joint training framework is used to jointly optimize the temporal modeling task and the knowledge distillation task, thereby obtaining a multilayer perceptual unit.

[0038] A third aspect of this application provides a transformer winding mechanical condition monitoring and evaluation device, the device comprising a processor and a memory:

[0039] The memory is used to store program code and transmit the program code to the processor;

[0040] The processor is used to execute the steps of the transformer winding mechanical condition monitoring and evaluation method as described in the first aspect above, according to the instructions in the program code.

[0041] A fourth aspect of this application provides a computer-readable storage medium for storing program code for executing the transformer winding mechanical condition monitoring and evaluation method described in the first aspect above.

[0042] As can be seen from the above technical solutions, this application has the following advantages:

[0043] This application provides an efficient edge computing method for smart grids. By designing a low-rank attention mechanism, it reduces the time complexity of sequence modeling algorithms and improves the response speed of edge computing platforms. Knowledge distillation technology is used to further improve the service speed of the method, greatly meeting the needs of edge computing platforms in real-world power scenarios. This method alleviates the computational pressure on edge computing platforms, enabling distributed analysis of power time-series data under different power usage scenarios to complete downstream tasks such as data classification and prediction. Compared with existing technologies, this application designs a low-rank self-attention mechanism for power data analysis. The low-rank self-attention mechanism generates multiple smaller matrices and uses low-rank factorization to calculate attention scores, greatly reducing the computational time complexity in the data modeling process. Attached Figure Description

[0044] Figure 1 is a flowchart illustrating an efficient edge computing method for smart grids provided in an embodiment of this application;

[0045] Figure 2 is a model structure diagram of an efficient edge computing method for smart grids based on time-series slicing algorithm provided in an embodiment of this application;

[0046] Fig. 3 is a structural schematic diagram of an intelligent power grid efficient edge computing system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0047] In order for those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0048] In order to comprehensively consider the knowledge of multiple power terminal data and take into account the efficiency and accuracy of the edge computing model, the present application combines sequence slicing technology and knowledge distillation technology to improve the applicability of the power time series data analysis model deployed on the edge computing platform.

[0049] In particular, the power sequence collected by the electric energy meter is acquired through the terminal device, and the data is uploaded to the terminal edge computing platform. Then, the original power data is sliced by the sequence slicing unit of the model on the edge computing platform, the long power sequence is cut into short sequence segments, and the short sequence segments are encoded and integrated by the sequence encoding unit of the model containing a low-rank attention mechanism. Finally, the knowledge learned in the model training process is distilled to a multi-layer perception unit and deployed to the edge computing platform, which is used for quickly and accurately classifying various power sequence data and obtaining target results. Based on this, the present application proposes an intelligent power grid efficient edge computing method and device based on a time series slicing algorithm, as follows:

[0050] Referring to Figs. 1 and 2, an intelligent power grid efficient edge computing method provided in an embodiment of the present application includes:

[0051] Step 101, collecting power sequence data of an electric energy meter through a terminal device, and preprocessing the power sequence data.

[0052] It should be noted that the power sequence data of the electric energy meter, such as load, voltage, current, etc., is collected through the terminal device, and each data is normalized and missing value 0 is filled in for preprocessing.

[0053] Step 102, uploading the preprocessed power sequence data to a terminal edge layer model, so that the terminal edge layer model performs sequence slicing on the power sequence data to obtain short sequence segment data.

[0054] Further, in one embodiment, step 102 further includes: performing standardization processing on the short sequence segment data, thereby alleviating the difference in data distribution.

[0055] It needs to be explained that the application proposes an intelligent power grid efficient edge computing model based on a time series slicing algorithm in the edge layer.

[0056] Specifically:

[0057] First, let represent a power data set collected by a terminal device, where x n ∈R T×F , T represents the length of the timestamp contained in the power data, and F represents the number of characteristic variables (such as positive active power, voltage, current, etc.) contained in the power data.

[0058] Next, let x i ∈X represent a power time series in the power data set. Before blocking, the method fills S repeated (that is, the last data item of the time series data with a total length of L) at the end of the original time series to preserve the edge information of the sequence. In particular, in order to alleviate the distribution offset effect between the training data and the test data, the method normalizes the time series data as follows: x i , that is, the power sequence after normalization.

[0059] Finally, assuming that each block length in the sequence is P and the blocking step is S, the pth short sequence segment of x i is represented as where represents the total number of blocks of x i sequence. Through the blocking operation, the token input in the next step of the method is reduced from L to L / S, which reduces the spatial complexity and computational complexity of subsequent attention score calculation by S times.

[0060] Step 103, modeling and analyzing the short sequence segment data by using a modified Transformer containing a low-rank self-attention mechanism, and completing model training.

[0061] In one embodiment, step 103 includes: generating a plurality of matrices based on the low-rank self-attention mechanism, approximating the calculation of attention scores by using low-rank factorization; and weighting the attention scores with the short sequence segment data representation to obtain a complete power data sequence representation.

[0062] Specifically, it includes:

[0063] inputting the short sequence segment data into the Transformer-based sequence encoding unit in the edge layer model to calculate the second-order relationship information encoding between the short sequence segments;

[0064] calculating the initial attention scores based on the second-order relationship information encoding;

[0065] The nonlinear relationship is captured by using a feedforward network, and the second-order relationship information encoding contained in the short sequence segment data is obtained, and the final attention score is obtained;

[0066] The weighted attention score is represented as a short sequence segment data representation, and a complete power data sequence representation is obtained.

[0067] It should be noted that the process of encoding the short sequence segment using the low-rank self-attention mechanism is as follows:

[0068] 1. According to the short sequence segment obtained in step 102, it is input into the edge layer model based on the sequence encoding unit of the Transformer, and the observed power time series data is encoded into a hidden layer representation. First, the method uses a learnable mapping matrix W p ∈R p×d , which maps the short sequence segment X P×N into the representation space R d×N . In order to enable the model to perceive the original time sequence information, the present application designs a learnable position encoding matrix W pos , which is added to the short sequence segment representation after mapping, that is, Wherein, represents the final learnable representation of the short sequence segment. Then, will be fed into the low-rank self-attention layer to obtain the second-order relationship information encoding contained in the short sequence segment

[0069] 2. The low-rank self-attention layer can aggregate information with approximately linear time complexity. Unlike the original self-attention mechanism, which uses scaled dot product to calculate the attention score, the low-rank self-attention mechanism generates multiple smaller matrices to approximate the attention score obtained by the original attention mechanism using low-rank factorization. First, the present application defines two trainable projection matrices and to perform low-rank embedding transformation. Here, C represents the low-rank scaling scale, represents the number of time series x i in the low-rank latent representation space. The attention score is calculated as follows:

[0070] Wherein, W Q , W K , W V are mapping matrices for learning query, key, and value vectors. E and F are used to convert the original Jxd-dimensional key and value vectors H·W K and H·W V into The low-rank vector of V is then computed, and the attention matrix is calculated. Through the low-rank mapping, the time complexity of the temporal encoder is reduced from O(JxJ) to where

[0071] 3. The sequence encoding unit utilizes a feedforward network to capture the nonlinear relationship therein and obtain the second-order relationship information encoding contained between short sequence segments That is, the final attention score: In this feedforward network, the method utilizes a layer of nonlinear activation ReLU(·) for nonlinear transformation, where W D and b D are the learnable parameter matrix and bias term.

[0072] 4. Weighted attention score and short sequence segment representation Obtain the sequence representation of the complete power data sequence. Taking the time series data classification as an example, the edge layer model can obtain the final prediction value τ θ (x) of the model according to the obtained sequence representation, and use the maximum likelihood loss to measure the difference between the prediction and the true label: L(θ) = ∑ (x，y) MSE(τ θ (x), y).

[0073] Step 104, distill the knowledge obtained in the model training process in step 103 to a multilayer perception unit, so that the multilayer perception unit performs inference analysis on short sequence segment data.

[0074] In one embodiment, step 104 includes extracting and injecting the knowledge obtained in the model training process in step 103, and jointly optimizing the time series modeling task and the knowledge distillation task using a joint training framework, thereby obtaining the multilayer perception unit.

[0075] It should be noted that the process of distilling the knowledge learned in the model training process to a multilayer perception unit is represented as follows:

[0076] 1. The above model (i.e., as a teacher model) is represented as The multilayer perception model (i.e., as a student model) is represented as where θ1= θ, and the principle of extracting and injecting knowledge is as follows: where, is the student model The maximum likelihood loss on a training data pair (x, y), is the teacher model and the student model The KL divergence loss between the predicted probability distribution, and λ is a hyperparameter used to balance the two loss terms. The final inference model is deployed on the edge computing platform.

[0077] 2. To jointly optimize the time series modeling task and the knowledge distillation task, the method uses a joint training framework to train the entire model according to L(θ) and L(θ2) to define the final loss function L: L = λ1·L(θ) + λ2L(θ2). Where λ1 and λ2 represent adjustable hyperparameters that balance the time series modeling learning loss and the knowledge distillation learning loss.

[0078] Step 105, deploying the multi-layer perception unit obtained by knowledge distillation in step 104 to the terminal edge computing platform for processing various power time series data modeling analysis tasks.

[0079] It should be noted that the multi-layer perception unit obtained by knowledge distillation, i.e. the student model is deployed to the smart grid edge computing platform to quickly and accurately process various power time series data modeling analysis tasks such as classification.

[0080] The smart grid efficient edge computing method provided by the embodiments of the present application reduces the time complexity of the sequence modeling algorithm by designing a low-rank attention mechanism, and improves the response speed of the edge computing platform. The service speed of the method is further improved by using the knowledge distillation technology, which greatly meets the needs of the edge computing platform in the actual power scene. The method of the present application is used to relieve the computing pressure of the edge computing platform, and realizes the distributed analysis of power time series data in different power use scenarios to complete downstream tasks such as data classification, prediction, etc. Compared with the prior art, the present application designs a low-rank self-attention mechanism for power data analysis. The low-rank self-attention mechanism generates multiple smaller matrices, and uses low-rank factorization to calculate attention scores, which greatly reduces the time complexity of the data modeling process.

[0081] The above is a smart grid efficient edge computing method provided in an embodiment of the present application, and the following is a smart grid efficient edge computing system provided in an embodiment of the present application.

[0082] Please refer to FIG. 3, a smart grid efficient edge computing system provided in an embodiment of the present application, comprising:

[0083] A preprocessing unit 201 is configured to collect power sequence data of an electric energy meter through a terminal device, and preprocess the power sequence data.

[0084] A slicing unit 202 is configured to upload the preprocessed power sequence data to a terminal edge layer model, so that the terminal edge layer model performs sequence slicing on the power sequence data to obtain short sequence segment data.

[0085] The training unit 203 is configured to model and analyze the short sequence data by using the improved Transformer comprising a low-rank self-attention mechanism, and complete model training.

[0086] The distillation unit 204 is configured to distill the knowledge obtained in the model training process in the training unit to a multi-layer perception unit, so that the multi-layer perception unit analyzes and reasons the short sequence data.

[0087] The deployment unit 205 is configured to deploy the multi-layer perception unit obtained by the distillation unit through knowledge distillation to a terminal edge computing platform, for processing various power time series data modeling and analysis tasks.

[0088] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described system and unit can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0089] The terms "first", "second", "third", "fourth" and the like in the specification of the present application and the above-described drawings, if any, are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device comprising a series of steps or units does not have to be limited to only those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0090] It should be understood that in the present application, "at least one" means one or more, and "multiple" means two or more. "And / or" is used to describe the association between the associated objects, which means that there can be three relationships, for example, "A and / or B" can represent three cases: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally represents an "or" relationship between the associated objects. "At least one of the following" or similar expressions means any combination of these items, including any combination of single or multiple items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0091] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other manners. For example, the described device embodiments are merely schematic. The division of the units is merely a logical function division. There can be another division manner for the actual implementation, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections can be indirect couplings or communication connections through some interfaces, devices or units, and can be in electrical, mechanical or other forms.

[0092] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0093] In addition, each functional unit in the various embodiments of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0094] The integrated unit, if realized in the form of a software functional unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the part of the prior art that makes a contribution or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (English full name: Read-Only Memory, English abbreviation: ROM), a random access memory (English full name: Random Access Memory, English abbreviation: RAM), a magnetic disk or an optical disk, and various program code storage media.

[0095] The above-described embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for part of the technical features; and these modifications 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 application.

Claims

1. A smart grid efficient edge computing method, characterized in that, The method comprises the following steps: S1, collecting power sequence data of an electric energy meter by a terminal device, and preprocessing the power sequence data; S2, uploading the preprocessed power sequence data to a terminal edge layer model, so that the terminal edge layer model performs sequence slicing on the power sequence data to obtain short sequence segment data; S3, modeling and analyzing the short sequence segment data by using a modified Transformer containing a low-rank self-attention mechanism, and completing model training; S4, distilling the knowledge obtained in the model training process in step S3 to a multi-layer perception unit, so that the multi-layer perception unit performs inference analysis on the short sequence segment data; S5, deploying the multi-layer perception unit obtained by knowledge distillation in step S4 to a terminal edge computing platform for processing various power time series data modeling and analysis tasks.

2. The smart grid efficient edge computing method of claim 1, wherein, Step S2 further comprises: standardizing the short sequence segment data.

3. The smart grid efficient edge computing method of claim 1, wherein, The modeling and analysis of the short sequence segment data by using the modified Transformer containing the low-rank self-attention mechanism to complete the model training comprises: Generating a plurality of matrices based on the low-rank self-attention mechanism, and using low-rank factorization to approximately calculate attention scores; weighting the attention scores and the short sequence segment data representation to obtain a complete power data sequence representation.

4. The smart grid efficient edge computing method of claim 3, wherein, The low-rank self-attention mechanism is used to generate a plurality of matrices, and low-rank factorization is used to approximately calculate attention scores. Weighting the original attention scores and the short sequence segment data representation to obtain a complete power data sequence representation comprises: Inputting the short sequence segment data into a sequence encoding unit based on the Transformer in the edge layer model to calculate the second-order relationship information encoding between the short sequence segments; Based on the second-order relationship information encoding, an initial attention score is calculated; Using a feedforward network to capture nonlinear relationships and obtain the second-order relationship information encoding contained in the short sequence segment data to obtain a final attention score; Weighting the attention score representation as the short sequence segment data representation to obtain a complete power data sequence representation.

5. The smart grid efficient edge computing method of claim 1, wherein, The knowledge obtained in the model training process in step S3 is distilled to a multi-layer perception unit, which comprises: Extracting and injecting the knowledge obtained in the model training process in step S3, and jointly optimizing the time series modeling task and the knowledge distillation task by using a joint training framework, so as to obtain a multi-layer perception unit.

6. A smart grid high efficient edge computing system, characterized in that, The method comprises the following steps: A preprocessing unit is configured to collect power sequence data of an electric energy meter by a terminal device, and preprocess the power sequence data; A slicing unit is configured to upload the preprocessed power sequence data to a terminal edge layer model, so that the terminal edge layer model performs sequence slicing on the power sequence data to obtain short sequence segment data; A training unit is configured to model and analyze the short sequence segment data by using a modified Transformer containing a low-rank self-attention mechanism, and complete model training; The distillation unit is used for distilling the knowledge obtained in the model training process in the training unit to a multilayer perception unit, so that the multilayer perception unit performs inference analysis on the short sequence segment data. The deployment unit is used for deploying the multilayer perception unit obtained by the distillation unit through knowledge distillation to a terminal edge computing platform, which is used for processing various power time series data modeling analysis tasks.

7. The smart grid efficient edge computing system of claim 6, wherein, Further comprising: The standardization processing unit is used for standardizing the short sequence segment data, thereby relieving the data distribution difference.

8. The smart grid efficient edge computing system of claim 6, wherein, The training unit is specifically used for: Based on the low-rank self-attention mechanism, a plurality of matrices are generated, and low-rank factor decomposition is used to approximately calculate the attention score; the attention score is weighted with the short sequence segment data representation to obtain a complete power data sequence representation.

9. A smart grid high efficiency edge computing device, characterized by, The device includes a processor and a memory: The memory is used to store program code and transmit the program code to the processor; The processor is used to execute the intelligent power grid efficient edge computing method according to the instructions in the program code.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium is used to store program code, and the program code is used to execute the intelligent power grid efficient edge computing method.

Images