Training methods for load forecasting models, load forecasting methods, devices and equipment

Abstract

This application provides a training method, apparatus, and device for a load prediction model, relating to the field of digital twin technology. The method includes: acquiring a sample load sequence of computing nodes and the actual load value at the sample prediction time; extracting features from the sample load sequence using an initial convolutional neural network and an initial gated recurrent neural network, respectively, to obtain a first sample feature vector and a second sample feature vector; performing load prediction through an initial fully connected layer based on the first and second sample feature vectors to obtain the predicted sample load value; and adjusting the parameters of the initial convolutional neural network, the initial gated recurrent neural network, and the initial fully connected layer based on the predicted sample load value and the target sample load value to obtain the load prediction model. This application can improve experimental efficiency by predicting the load of different computing nodes at future times, allowing for the selection of computing nodes with lower loads for data computation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of digital twinning, in particular to a load prediction model training method, a load prediction method, a device and equipment. BACKGROUND

[0002] With the continuous development of aircraft equipment, single-area flight tests have been unable to meet the increasingly complex flight test requirements, therefore, joint tests implemented through digital twinning logical range systems have gradually become the development trend of flight tests.

[0003] In the process of flight tests using digital twinning logical range systems, multiple ranges distributed in different areas and numerous devices in the ranges will be involved, which will generate a large amount of raw data, in order to ensure the real-time performance of flight tests, the large amount of raw data needs to be calculated and processed in time.

[0004] During the peak period of flight tests, due to the limited computing resources and storage resources of the digital twinning logical range system, a large amount of tasks and data will flow into the system, causing rapid increase of system load, resulting in delay in processing of computing and storage tasks, which cannot meet the real-time requirements of flight tests and reduces the test efficiency. SUMMARY

[0005] The present application aims at the deficiencies in the prior art, and provides a load prediction model training method, a load prediction method, a device and equipment, so as to predict the load of different computing nodes in the digital twinning logical range at future time, select computing nodes with lower load for data calculation, reduce task processing delay, and improve test efficiency.

[0006] To achieve the above-mentioned purpose, the technical solutions adopted by the embodiments of the present application are as follows:

[0007] In a first aspect, the embodiments of the present application provide a load prediction model training method, which comprises:

[0008] obtaining a sample load sequence of any computing node of a digital twinning logical range and a real load value at a sample prediction time after a plurality of continuous sample times, the sample load sequence comprising sample load values of the any computing node at the plurality of continuous sample times;

[0009] performing feature extraction on the sample load sequence by using an initial convolutional neural network to obtain a first sample feature vector;

[0010] performing feature extraction on the sample load sequence by using an initial gated recurrent neural network to obtain a second sample feature vector;

[0011] According to the first sample feature vector and the second sample feature vector, load prediction is performed through an initial full connection layer to obtain a sample prediction load value of the computing node at the sample prediction moment;

[0012] According to the sample prediction load value and the real load value, parameters of the initial convolutional neural network, the initial gated recurrent neural network and the initial full connection layer are adjusted to obtain a load prediction model.

[0013] Optionally, the method further comprises:

[0014] The first sample load value of the any computing node at the continuous multiple sample moments and the second sample load value of the test resource of the digital twin logical target range at the continuous multiple sample moments are obtained.

[0015] The sample load sequence is calculated according to the first sample load value at the continuous multiple sample moments and the second sample load value at the continuous multiple sample moments.

[0016] Optionally, before the feature extraction of the sample load sequence by the initial convolutional neural network to obtain the first sample feature vector, the method further comprises:

[0017] The sample load sequence is two-dimensionally converted to generate a static time sequence diagram corresponding to the sample load sequence.

[0018] The feature extraction of the sample load sequence by the initial convolutional neural network to obtain the first sample feature vector comprises:

[0019] The static time sequence diagram is feature-extracted by the initial convolutional neural network to obtain the first sample feature vector.

[0020] Optionally, the feature extraction of the static time sequence diagram by the initial convolutional neural network to obtain the first sample feature vector comprises:

[0021] The static time sequence diagram is convoluted to obtain a sample feature map, and the sample feature map comprises multiple channel feature maps.

[0022] The multiple channel feature maps are attention-extracted by an attention mechanism to determine channel weights of the multiple channel feature maps.

[0023] The first sample feature vector is calculated according to the multiple channel feature maps and the channel weights.

[0024] Optionally, the calculating the first sample feature vector according to the plurality of channel feature maps and the channel weights comprises:

[0025] selecting a target channel feature map from the plurality of channel feature maps according to the channel weights of the plurality of channel feature maps;

[0026] calculating the first sample feature vector according to the target channel feature map and the channel weight corresponding to the target channel feature map.

[0027] Optionally, the performing load prediction on the first sample feature vector and the second sample feature vector through an initial full connection layer to obtain a sample predicted load value of the computing node at the sample prediction moment comprises:

[0028] weighting the first sample feature vector and the second sample feature vector according to an initial first weight and an initial second weight to obtain a weighted sample feature vector;

[0029] performing load prediction on the weighted sample feature vector by using the initial full connection layer to obtain the sample predicted load value.

[0030] In a second aspect, an embodiment of the present application further provides a load prediction method, and the method comprises:

[0031] obtaining a load sequence of a plurality of computing nodes of a digital twin logical target range, wherein the load sequence of each computing node comprises load values of the digital twin logical target range at a plurality of continuous moments;

[0032] performing feature extraction on the load sequence of each computing node by using a convolutional neural network in a load prediction model to obtain a first feature vector;

[0033] performing feature extraction on the load sequence of each computing node by using a gated recurrent neural network in the load prediction model to obtain a second feature vector;

[0034] performing load prediction on the first feature vector and the second feature vector through a full connection layer in the load prediction model to obtain a predicted load value of each computing node at the prediction moment, wherein the load prediction model is obtained by training the load prediction model by using the training method of the load prediction model according to any one of the first aspect;

[0035] determining a target computing node according to the predicted load values of the plurality of computing nodes, wherein the target computing node is used for data calculation at the prediction moment.

[0036] In a third aspect, an embodiment of the present application further provides a training device of a load prediction model, and the training device comprises:

[0037] a sample data acquisition module, configured to acquire sample load sequences of any computing node of a digital twin logical range and real load values at a sample prediction moment after a plurality of continuous sample moments, wherein the sample load sequences comprise sample load values of the any computing node at the plurality of continuous sample moments;

[0038] a first sample feature extraction module, configured to perform feature extraction on the sample load sequences by using an initial convolutional neural network to obtain a first sample feature vector;

[0039] a second sample feature extraction module, configured to perform feature extraction on the sample load sequences by using an initial gated recurrent neural network to obtain a second sample feature vector;

[0040] a sample load prediction module, configured to perform load prediction on the first sample feature vector and the second sample feature vector by using an initial full connection layer to obtain a sample predicted load value of the computing node at the sample prediction moment;

[0041] a model parameter adjustment module, configured to perform parameter adjustment on the initial convolutional neural network, the initial gated recurrent neural network and the initial full connection layer according to the sample predicted load value and the real load value to obtain a load prediction model.

[0042] Optionally, the sample data acquisition module comprises:

[0043] a sample load value acquisition unit, configured to acquire first sample load values of the any computing node at the plurality of continuous sample moments and second sample load values of test resources of the digital twin logical range at the plurality of continuous sample moments;

[0044] a sample load sequence generation unit, configured to calculate the sample load sequences according to the first sample load values at the plurality of continuous sample moments and the second sample load values at the plurality of continuous sample moments.

[0045] Optionally, before the first sample feature extraction module, the apparatus further comprises:

[0046] a two-dimensional conversion module, configured to perform two-dimensional conversion on the sample load sequences to generate a static timing diagram corresponding to the sample load sequences;

[0047] The first sample feature extraction module is specifically configured to perform feature extraction on the static timing diagram by using the initial convolutional neural network to obtain the first sample feature vector.

[0048] Optionally, the first sample feature extraction module comprises:

[0049] a convolution processing unit, configured to obtain a sample feature map by performing convolution processing on the static timing diagram, the sample feature map comprising a plurality of channel feature maps;

[0050] an attention extraction unit, configured to perform attention extraction on the plurality of channel feature maps by using an attention mechanism to determine channel weights of the plurality of channel feature maps;

[0051] a first sample feature extraction unit, configured to calculate the first sample feature vector according to the plurality of channel feature maps and the channel weights.

[0052] Optionally, the first sample feature extraction unit comprises:

[0053] a channel selection subunit, configured to select a target channel feature map from the plurality of channel feature maps according to the channel weights of the plurality of channel feature maps;

[0054] a first sample feature extraction subunit, configured to calculate the first sample feature vector according to the target channel feature map and the channel weight corresponding to the target channel feature map.

[0055] Optionally, the sample load prediction module is specifically configured to weight the first sample feature vector and the second sample feature vector according to an initial first weight and an initial second weight to obtain a weighted sample feature vector, and perform load prediction on the weighted sample feature vector by using the initial full connection layer to obtain the sample predicted load value.

[0056] In a fourth aspect, an embodiment of the present application further provides a load prediction device, the device comprising:

[0057] a load sequence acquisition module, configured to acquire a load sequence of a plurality of computing nodes of a digital twin logical range, the load sequence of each computing node comprising: load values of the digital twin logical range at a plurality of continuous time points;

[0058] a first feature extraction module, configured to perform feature extraction on the load sequence of each computing node by using a convolutional neural network in a load prediction model to obtain a first feature vector;

[0059] a second feature extraction module, configured to perform feature extraction on the load sequence of each computing node by using a gated recurrent neural network in the load prediction model to obtain a second feature vector;

[0060] a load prediction module, configured to perform load prediction through a full connection layer in the load prediction model according to the first feature vector and the second feature vector, to obtain a predicted load value of each computing node at the predicted time, wherein the load prediction model is obtained by training using the training method of the load prediction model according to any one of the first aspect;

[0061] a node selection module, configured to determine a target computing node according to the predicted load values of the plurality of computing nodes, the target computing node being used for data calculation at the predicted time.

[0062] In a fifth aspect, an embodiment of the present application further provides an electronic device, including a processor, a storage medium and a bus, the storage medium stores program instructions executable by the processor, when the electronic device is running, the processor and the storage medium communicate through the bus, and the processor executes the program instructions to perform the steps of the training method of the load prediction model according to any one of the first aspect, or the steps of the load prediction method according to the second aspect.

[0063] In a sixth aspect, an embodiment of the present application further provides a computer readable storage medium, the storage medium stores a computer program, when the computer program is run by a processor, the steps of the training method of the load prediction model according to any one of the first aspect, or the steps of the load prediction method according to the second aspect are performed.

[0064] The present application has the following beneficial effects:

[0065] The training method of the load prediction model, the load prediction method, the device and the electronic device provided by the present application respectively use an initial convolutional neural network and an initial gated recurrent neural network to extract features of a sample load sequence, to perform load prediction through an initial full connection layer according to the extracted first sample feature vector and the second sample feature vector, to obtain a sample predicted load value, to adjust network parameters according to the sample predicted load value and a real load value, to obtain a load prediction model including a convolutional neural network and a gated recurrent neural network, to use the convolutional neural network and the gated recurrent neural network in combination to extract spatial features and long-time sequence features of data, and to perform load prediction based on the spatial features and the long-time sequence features, so that the prediction accuracy can be improved. BRIEF DESCRIPTION OF DRAWINGS

[0066] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation to the scope. For those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0067] Figure 1 Architecture diagram of the digital twin logic range system provided in this application embodiment;

[0068] Figure 2 A flowchart illustrating the training method for the load prediction model provided in this application embodiment. Figure 1 ;

[0069] Figure 3 This is a schematic diagram of the structure of a gated recurrent neural network provided in an embodiment of this application;

[0070] Figure 4 A flowchart illustrating the training method for the load prediction model provided in this application embodiment. Figure 2 ;

[0071] Figure 5 A flowchart illustrating the training method for the load prediction model provided in this application embodiment. Figure 3 ;

[0072] Figure 6 A flowchart illustrating the training method for the load prediction model provided in this application embodiment. Figure 4 ;

[0073] Figure 7 A flowchart illustrating the training method for the load prediction model provided in this application embodiment. Figure 5 ;

[0074] Figure 8 A flowchart illustrating the training method for the load prediction model provided in this application embodiment. Figure 6 ;

[0075] Figure 9 A flowchart illustrating the load prediction method provided in this application embodiment;

[0076] Figure 10 Network architecture diagram for the load prediction model;

[0077] Figure 11 A schematic diagram of the structure of a training device for a load prediction model provided in an embodiment of this application;

[0078] Figure 12 This is a schematic diagram of the structure of the load prediction device provided in the embodiments of this application;

[0079] Figure 13 A schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0080] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present application in conjunction with 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 of the present application.

[0081] Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0082] In addition, the terms "first", "second", and the like in the specification and claims of the present application and the above drawings 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 those 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 including 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.

[0083] It should be noted that the features in the embodiments of the present application can be combined with each other without conflict.

[0084] Before introducing the training method of the load prediction model, the load prediction method, the device and the equipment provided by the present application, in order to better understand the scheme of the present application, the digital twin logical target system applied by the present application is introduced first.

[0085] Please refer to Figure 1 The architecture diagram of the digital twin logical target system provided by the embodiments of the present application is shown as Figure 1 The digital twin logical target system includes a logical target test platform, at least one physical test resource and a plurality of computing nodes.

[0086] The logical target test platform can run on a computer device. Through the logical target test platform, a plurality of logical target fields can be created according to test or training tasks. Each logical target field includes at least one virtual test resource. The test or training task is executed through the virtual test resource in each logical target field.

[0087] The physical test resource is an entity test resource existing in a real target range, including a physical test device and a simulation test device. Each virtual test resource corresponds to a physical test resource. By binding the virtual test resource and the physical test resource, a test task is sent to each logical target range in the logical target range test platform, so that the virtual test resource in each logical target range sends a test instruction to the physical test resource bound to the virtual test resource based on the test task, so that the physical test resource generates test data according to the test instruction, and sends the test data to one of the plurality of computing nodes, so that the computing node calculates according to the test data, and drives the rendering and state update of the virtual test resource according to the calculation result.

[0088] The computer device where the logical target range test platform is located is deployed with a pre-trained load prediction model. The computer device can obtain the load of each computing node in a period of time, and predict the load of each computing node at a future time by using the load prediction model according to the load in the period of time, so as to select the computing node with the minimum load at the future time from the plurality of computing nodes for executing the computing task at the future time.

[0089] In some embodiments, the load of each computing node at the future time is determined by the computing load of the computing node in a period of time and the amount of data to be received by the computing node at the future time. The amount of data to be received by the computing node at the future time can be determined according to the task load of the physical test resource corresponding to each virtual test resource in the logical target range in a period of time, that is, the computing load of each computing node and the task load of the physical test resource corresponding to each virtual test resource in any logical target range can be obtained to predict the load of each computing node at the future time according to the computing load and the task load, so as to select a target computing node from the plurality of computing nodes, and the target computing node performs data calculation according to the test data returned by the physical test resource corresponding to each virtual test resource in any logical target range, to obtain the test result corresponding to the test task.

[0090] Based on the above digital twin logical target range system, the training method of the load prediction model provided by the embodiments of the present application, the load prediction method, the device and the equipment are described below.

[0091] Please refer to Figure 2 The flowchart of the training method of the load prediction model provided by the embodiments of the present application is shown in Figure 1 As Figure 2 shown, the training method can include:

[0092] S11: Obtain a sample load sequence of any computing node of the digital twin logical range and a real load value at a sample prediction moment after a plurality of continuous sample moments, the sample load sequence comprising sample load values of the any computing node at the plurality of continuous sample moments.

[0093] In the embodiment, the digital twin logical range system comprises a plurality of computing nodes. For a test task of each logical range, test instructions are sent to physical test resources corresponding to virtual test resources in the logical range through the virtual test resources, the physical test resources perform tests based on the test instructions and generate test data, and each computing node obtains a test result corresponding to the test task by performing calculation on the test data.

[0094] In a high incidence period of flight tests, a plurality of logical ranges in a logical range test platform need to perform a large number of test tasks. For each test task, a physical test resource corresponding to a virtual test resource of a logical range generates a group of test data, and a computing node needs to perform calculation on the test data. In order to ensure the real-time performance of the computing node and improve the test efficiency, the test resource needs to be calculated by the computing node with the lowest load.

[0095] When performing model training, in order to predict the load value of the computing node at the prediction moment, the sample load values of the computing node at a plurality of continuous sample moments and the real load value at the sample prediction moment need to be obtained, the sample load values at the plurality of continuous sample moments constitute a sample load sequence, and the sample load sequence and the real load value at the prediction moment are used as inputs of the model.

[0096] In some embodiments, before performing model training, the original data set S is split into a training set, a validation set and a test set at a preset ratio, the training set is used to train the model, the validation set is used to select and adjust the parameters of the model, and the test set is used to test the performance of the trained model. For example, the preset ratio can be 7:1:2.

[0097] Before splitting the original data set S, the data in the original data set is checked and the missing values are filled, and the filling method can be mean filling.

[0098] When constructing the sample load sequence, the test set is usually cut by a sliding window to form a time sequence with a fixed length. Assuming that the original load time sequence L={b1, b2, …, bn} in the test set with a length of n, the sliding window size is w, and the prediction step is d, after data cutting by the sliding window, n-w-d+1 groups of sample load sequences can be constructed.

[0099] For example, the first row in the following formula is the sample load sequence in the input model, and the second row is the label of the sample load sequence, that is, the real load value at the prediction time w+d.

[0100]

[0101] In one possible implementation, the data in the training set and the validation set are normalized before being input into the model. The formula for normalization can be represented as:

[0102]

[0103] where x' is the normalized sample load value, x is the original sample load value, min(x) is the minimum sample load value in the training set or the validation set, and max(x) is the maximum sample load value in the training set or the validation set.

[0104] S12: The initial convolutional neural network is used to extract features of the sample load sequence to obtain a first sample feature vector.

[0105] In this embodiment, the convolutional neural network (CNN) is used to extract spatial features of the sample load sequence. The initial convolutional neural network is composed of a convolutional layer, a pooling layer, and a fully connected layer. The convolutional layer is used to convolve the sample load sequence to extract features to obtain feature values. The pooling layer is used to reduce the size of the convolution kernel in the convolutional layer, reduce parameters to prevent overfitting, and perform dimension reduction processing. The pooling layer has multiple pooling methods such as average pooling, maximum pooling, and random pooling, and the pooling method can be selected as needed. The fully connected layer mainly maps the data after convolution and pooling in the hidden feature space of the convolutional neural network, performs dimension conversion, and outputs a first sample feature vector V1 = {a1, a2,..., an} of dimension n. n .

[0106] S13: The initial gated recurrent unit is used to extract features of the sample load sequence to obtain a second sample feature vector.

[0107] In the embodiment, the gated recurrent unit (GRU) is used to extract long-time features of the sample load sequence. For example, please refer to Figure 3 , the structural diagram of the gated recurrent unit provided by the embodiment of the present application, as shown in Figure 3 , the GRU is composed of a reset gate (rt), an update gate (z t ), a hidden layer state parameter h t at the current time, and a memory parameter .

[0108] First, the update gate (z t ) is calculated, and the calculation formula can be represented as:

[0109] z t = σ(W (z) x t + U (z) h t-1 )

[0110] Second, the reset gate (r t ) is calculated, which is used to calculate the information forgotten during transmission, and the calculation formula can be represented as:

[0111] r t = σ(W (r) x t + U (r) h t-1 )

[0112] According to the reset gate, the new memory parameter is recalculated , and the calculation formula can be represented as:

[0113]

[0114] According to the memory parameter , the hidden layer state parameter h t at the current moment is calculated, and the calculation formula can be represented as:

[0115]

[0116] Then the output at the current moment can be represented as: y t = σ(W0·h t ).

[0117] Where x t is the sample load value input at the current moment t, h t-1 is the hidden layer state parameter at the last moment, and y t is the sample feature vector representing the load change between adjacent moments t and t-1 based on the sample load value input at the current moment t and the hidden layer state parameter at the last moment t-1.

[0118] The initial gated recurrent neural network learns the change state of the sample load value at the continuous multiple sample moments based on the sample load value at the continuous multiple sample moments in the sample load sequence, to determine the second sample feature vector V2 = {b1, b2, …, b n} of the load transformation at the continuous multiple sample moments. Wherein the dimension of the second sample feature vector is the same as that of the first sample feature vector.

[0119] S14: performing load prediction by the initial fully connected layer according to the first sample feature vector and the second sample feature vector, to obtain a sample prediction load value of the computing node at the sample prediction moment.

[0120] In the embodiment, the first sample feature vector V1 is used to represent the feature of the sample load values at the continuous multiple sample moments in the spatial dimension in the sample load sequence, the second sample feature vector V2 is used to represent the feature of the sample load values at the continuous multiple sample moments in the long time sequence dimension in the sample load sequence, and the fused feature vector V is obtained by fusing the first sample feature vector V1 and the second sample feature vector V2. The fused feature vector V=(V1, V2)=(a1, a2, …, a n , b1, b2, …, b n ), and the sample prediction load value is obtained by performing dimension transformation on the fused feature vector V by the initial fully connected layer.

[0121] S15: performing parameter adjustment on the initial convolutional neural network, the initial gated recurrent neural network and the initial fully connected layer according to the sample prediction load value and the real load value, to obtain a load prediction model.

[0122] In the embodiment, the loss value is calculated according to the sample prediction load value and the real load value, and the network parameters in the initial convolutional neural network, the initial gated recurrent neural network and the initial fully connected layer are adjusted by the gradient descent algorithm according to the loss value. Through multiple rounds of training, the load prediction model composed of the target convolutional neural network, the target gated recurrent neural network and the target fully connected layer is obtained.

[0123] In a possible implementation, after multiple rounds of training, there can be multiple groups of network parameters that make the load prediction model converge. In order to determine the optimal load prediction model, the evaluation index of the multiple groups of network parameters corresponding to the multiple candidate load prediction models can be calculated by using the test set, so as to select the optimal load prediction model from the multiple groups of candidate load prediction models according to the evaluation index.

[0124] For example, the evaluation index can include at least one of the mean absolute error (Mean Absolute Error, MAE), the mean squared error (Mean Squared Error, MSE) and the determination coefficient (R 2 ) of the sample prediction load value and the real load value.

[0125]

[0126]

[0127]

[0128] The training method of the load prediction model provided in the above embodiment extracts features of the sample load sequence by using an initial convolutional neural network and an initial gated recurrent neural network respectively, performs load prediction by an initial full connection layer according to the extracted first sample feature vector and second sample feature vector, obtains a sample predicted load value, adjusts network parameters according to the sample predicted load value and the real load value, and obtains a load prediction model including a convolutional neural network and a gated recurrent neural network. The combination of the convolutional neural network and the gated recurrent neural network is used to extract spatial features and long-time sequence features of data, and load prediction is performed based on the spatial features and the long-time sequence features, which can improve the prediction accuracy.

[0129] The above possible implementation of obtaining the sample load sequence is described below in combination with an embodiment.

[0130] Please refer to Figure 4 The flowchart of the training method of the load prediction model provided in the embodiment of the present application is shown in Figure 2 As shown in Figure 4 The process of S11 of obtaining the sample load sequence of any computing node of the digital twin logical target range can include:

[0131] S21: Obtain a first sample load value of any computing node at a plurality of consecutive sample time points, and a second sample load value of a test resource of the digital twin logical target range at the plurality of consecutive sample time points.

[0132] S22: Calculate a sample load sequence according to the first sample load value at the plurality of consecutive sample time points and the second sample load value at the plurality of consecutive sample time points.

[0133] In this embodiment, the first sample load value is the computing load of the computing node at each sample time point, i.e., the host load of the computing node, and the second sample load value is the task load of the physical test resource corresponding to the virtual test resource in the logical target range at each sample time point, i.e., the number of instructions sent to the physical test resource corresponding to each logical target range. The sample load value at each sample time point is calculated according to the host load at each sample time point and the task load at each sample time point, and the sample load sequence is generated according to the sample load values at the plurality of consecutive sample time points. The host load can include CPU load, memory load, and disk IO usage rate load of the computing node.

[0134] For example, it is assumed that the logical target test platform includes three logical targets, each of which corresponds to different numbers of physical test devices and simulation test devices. Then, the sample load value of any computing node at any time is obtained by combining the host load of the computing node and the task load of the physical test devices and simulation test devices according to a preset proportion.

[0135] For example, the host load includes CPU load and memory load, and the host load is formed according to a preset ratio. If the CPU load is 56 and the memory load is 47, the host load formed according to a ratio of 1:1 is 51.5. If the number of instructions received by the physical test device and the simulation test device is 320, the sample load value obtained by combining the host load and the task load according to a ratio of 1:1 is 185.75, and the sample load value at the current sample moment is obtained. Sampling is performed according to a fixed time interval to form a sample load sequence.

[0136] The training method of the load prediction model provided in the above embodiment can obtain the first sample load value of the computing node at the continuous multiple sample moments and the second sample load value of the test resource at the continuous multiple sample moments, and jointly form a sample load sequence, so as to realize the prediction of the load value of the computing node at the prediction moment based on the load condition of the computing node and the load condition of the test resource, and improve the accuracy of load prediction.

[0137] In a possible implementation, refer to Figure 5 The flowchart of the training method of the load prediction model provided in the embodiment of the present application is shown in Figure 3 As shown in Figure 5 Before the above S12 uses the initial convolutional neural network to perform feature extraction on the sample load sequence to obtain the first sample feature vector, the method can further include:

[0138] S31: performing two-dimensional conversion on the sample load sequence to generate a static time sequence diagram corresponding to the sample load sequence.

[0139] The process of S12 in the above, which uses the initial convolutional neural network to perform feature extraction on the sample load sequence to obtain the first sample feature vector, can include:

[0140] S32: using the initial convolutional neural network to perform feature extraction on the static time sequence diagram to obtain the first sample feature vector.

[0141] In the present embodiment, since CNN needs to perform convolution on images to extract feature values, before using the initial convolutional neural network to perform feature extraction on the sample load sequence, the sample load sequence needs to be converted into a two-dimensional static time sequence diagram first. The conversion method is to generate a static time sequence diagram based on the sample load values in the sample load sequence and the sample moments corresponding to each sample load value. After generating the static time sequence diagram, the initial convolutional neural network performs feature extraction on the static time sequence diagram to obtain the first sample feature vector.

[0142] In an example, a Gram angular field (GAF) can be used to convert the sample load sequence into a two-dimensional static time series graph, and the process of the Gram angular field algorithm can be as follows:

[0143] First, for the sample load sequence X = {x1, x2, …, x n}, the sample load sequence X is scaled to the interval [0, 1] using a Min-Max scaler, and the scaling formula can be expressed as:

[0144]

[0145] After scaling, the scaled sample load sequence is converted to a polar coordinate system to obtain the angle value Φ i and the radius value r i corresponding to each sample load value in the sample load sequence, and the calculation formula can be expressed as:

[0146]

[0147] The Gram angular sum field (GASF) is calculated according to the angle value Φ i corresponding to each sample load value, to generate a static time series graph according to the Gram angular sum field, and the calculation formula of the Gram angular sum field can be expressed as:

[0148]

[0149] The training method of the load prediction model provided in the above embodiment converts the sample load sequence into a static time series graph, and extracts features of the static time series graph by the initial convolutional neural network, so that spatial features of the sample load sequence can be extracted, thereby enabling the load prediction to be performed together with long time sequence features of the sample load sequence, and improving the accuracy and precision of model prediction.

[0150] The following describes one possible implementation of the above process of extracting features of the static time series graph by the initial convolutional neural network to obtain a first sample feature vector, in combination with embodiments.

[0151] Please refer to Figure 6 , the flowchart of the training method of the load prediction model provided in the embodiments of the present application Figure 4 , as shown in Figure 6 , the process of extracting features of the static time series graph by the initial convolutional neural network to obtain a first sample feature vector in the above S32 can include:

[0152] S41: obtaining a sample feature map by performing convolution processing on the static time series graph, and the sample feature map includes a plurality of channel feature maps.

[0153] S42: An attention mechanism is used to extract attention from multiple channel feature maps and determine the channel weights of the multiple channel feature maps.

[0154] S43: Calculate the feature vector of the first sample based on the feature maps of multiple channels and the channel weights.

[0155] In this embodiment, multiple convolutional kernels in the convolutional layer are used to convolve the static time series map to obtain multiple channel feature maps. An attention mechanism is used to focus on the relationship between the channels during the feature extraction process, thereby enhancing the CNN's ability to extract features.

[0156] The attention mechanism (Squeeze-and-Excitation Networks, SENet) extracts the importance of feature maps from multiple channels, determines the channel weights corresponding to each channel feature map, and then weights the multiple channel feature maps according to the channel weights to obtain the first sample feature vector. Specifically, the channel weights are used to enhance the features of channel feature maps with higher contributions and suppress the features of channel feature maps with lower contributions.

[0157] For example, suppose the feature map of the samples after convolution is D = {d1, d2, ..., dn}. c}, D∈R H*W*C Where H and W are the height and width of the sample feature map, and C is the number of channels.

[0158] The attention mechanism process includes compression, activation, and scaling operations.

[0159] Among them, compression operation (F sq The features in each channel dimension of the sample feature map are compressed into a single value using global average pooling. The compressed features then yield global statistical values ​​along the channel dimension. The formula for calculating the global statistical value of the c-th channel can be expressed as:

[0160]

[0161] Where Dc represents the channel feature of the c-th channel in the sample feature map.

[0162] Excitation operation (F) ex) two fully connected layers (FC) structures are adopted to fuse the nonlinear relationship between channel features, wherein the first fully connected layer performs full connection operation on the feature vector obtained by compression operation, so as to convert it into a vector of C / r dimensions (r is the compression ratio), and then uses a Relu activation function to perform nonlinear transformation; the second fully connected layer restores the vector of C / r dimensions obtained in the previous step to a vector of C dimensions; finally, the C-dimensional vector obtained is normalized by using a sigmoid function, to obtain a channel weight vector s distributed between 0 and 1, representing the importance of each channel, that is, the channel weight. The calculation formula of the excitation operation can be expressed as:

[0163] s=F ex (z,W)=σ(g(z,W))=σ(W2δ(W1z))

[0164] Wherein, σ (*) is a sigmiod function, δ (*) is a Relu function, z is the output of the compression operation, W1 and W2 are the weight matrices of the first fully connected layer and the second fully connected layer respectively.

[0165] Scaling operation (F sc ) uses multiplication operation to weight the weight vector s obtained by the excitation operation with the sample feature map D channel by channel, to obtain a weighted feature map, and performs dimension conversion on the weighted feature map to obtain a first sample feature vector.

[0166] For example, the calculation formula of the scaling operation can be expressed as:

[0167]

[0168] In a possible implementation, please refer to Figure 7 The flowchart of the training method of the load prediction model provided by the embodiment of the application Figure 5 As Figure 7 indicated, the process of calculating the first sample feature vector according to the plurality of channel feature maps and the channel weight in S43 can include:

[0169] S51: selecting a target channel feature map from the plurality of channel feature maps according to the channel weight of the plurality of channel feature maps.

[0170] S52: calculating a first sample feature vector according to the target channel feature map and the channel weight corresponding to the target channel feature map.

[0171] In this embodiment, in order to enhance the quality of the features extracted by the CNN, the plurality of channel feature maps can be sorted according to the channel weights of the plurality of channel feature maps, and the target channel feature maps with a contribution degree greater than a preset contribution degree can be selected according to the sorted plurality of channel feature maps. For example, one-half of the channel feature maps with larger channel weights can be selected as the target channel feature maps.

[0172] The first sample feature vector can be obtained by performing weighted calculation on the target channel feature maps according to the channel weights of the target channel feature maps.

[0173] The training method of the load prediction model provided in the above embodiments can calculate the channel weights of the respective channel feature maps through the attention mechanism, and can perform weighting on the plurality of channel feature maps according to the channel weights, so as to enhance the channel features with greater contribution degrees and suppress the channel features with smaller contribution degrees, thereby improving the quality of the features extracted by the convolutional neural network and improving the prediction accuracy of the load prediction model.

[0174] In a possible implementation manner, refer to Figure 8 The flowchart of the training method of the load prediction model provided in the embodiments of the present application is shown in Figure 6 As shown in Figure 8 The process of S14 performing load prediction on the first sample feature vector and the second sample feature vector through the initial full connection layer to obtain the sample prediction load value of the computing node at the sample prediction moment can include:

[0175] S61: weighting the first sample feature vector and the second sample feature vector according to the initial first weight and the initial second weight to obtain a weighted sample feature vector.

[0176] S62: performing load prediction on the weighted sample feature vector through the initial full connection layer to obtain the sample prediction load value.

[0177] In this embodiment, when performing feature fusion on the first sample feature vector and the second sample feature vector, the initial first weight and the initial second weight are used to weight the first sample feature vector and the second sample feature vector to obtain a weighted sample feature vector, i.e., a fused feature vector, and the initial full connection layer is used to perform dimension transformation on the fused feature vector V to obtain the sample prediction load value.

[0178] In the training process of the load prediction model, the initial first weight and the initial second weight can be trained by using the attention mechanism, so that the feature vectors with greater contribution degrees to the prediction accuracy have greater weights, and the feature vectors with smaller contribution degrees to the prediction accuracy have smaller weights, and the final first weight and the second weight are obtained.

[0179] The training method of the load prediction model provided in the above embodiment weights the first sample feature vector and the second sample feature vector through the initial first weight and the initial second weight, so that a feature vector with a large contribution to prediction accuracy has a larger weight, and a feature vector with a small contribution to prediction accuracy has a smaller weight, thereby improving the prediction accuracy of the load prediction model.

[0180] For example, refer to Table 1 for evaluation indexes of various load prediction models. As shown in Table 1, the sliding window is set to 32, and the prediction step d is set to 1 (that is, the first 32 load sequence data are used to predict the 33rd load data) during the experiment. The load prediction model of the embodiment of the application is composed of SE-GAF-CNN-GRU. Compared with the traditional single prediction model, each evaluation index is better than the evaluation index of the single prediction model.

[0181] Prediction model R2 MAE MSE CNN 0.7377 6.2184 60.8958 LSTM 0.8074 5.3099 43.6376 GRU 0.7854 5.6388 50.8448 SE-GAF-CNN-GRU 0.8526 4.4333 34.9264

[0182] The load prediction method using the above load prediction model is described below in combination with embodiments.

[0183] Please refer to Figure 9 The flowchart of the load prediction method provided in the embodiment of the application is shown in Figure 9 The load prediction method includes the following steps.

[0184] S71: Obtain a load sequence of a plurality of computing nodes of a digital twin logical target range, and the load sequence of each computing node includes load values of the digital twin logical target range at a plurality of continuous time points.

[0185] S72: Perform feature extraction on the load sequence of each computing node by using a convolutional neural network in the load prediction model to obtain a first feature vector.

[0186] S73: Perform feature extraction on the load sequence of each computing node by using a gated recurrent neural network in the load prediction model to obtain a second feature vector.

[0187] S74: Perform load prediction by using a full connection layer in the load prediction model according to the first feature vector and the second feature vector to obtain a predicted load value of each computing node at a prediction time point.

[0188] The load prediction model is trained by using the training method of the load prediction model.

[0189] S75: Determine a target computing node according to the predicted load values of the plurality of computing nodes, and the target computing node is used for data calculation at the prediction time point.

[0190] In this embodiment, please refer to Figure 10A network architecture diagram of a load prediction model is shown in FIG. 1. Figure 10 As shown in FIG. 1, a plurality of load sequences of a plurality of computing nodes are obtained, and each load sequence of each computing node is represented as {X1, X2, …, Xn}. The load sequence of each computing node is transformed into a static timing diagram by using a Gram angular field, a CNN in the load prediction model includes two convolutional layers ConV1 and ConV2 and two pooling layers Pool, which are used for feature extraction, an activation function tanh is used between each convolutional layer and each pooling layer, and a channel attention mechanism SENet is introduced after the feature extraction, the contribution of each channel in the extracted features is extracted, and the features are weighted according to the contribution to obtain a first sample feature vector V1.

[0191] The CRU in the load prediction model includes at least two GRUs and a fully connected layer FC, the at least two GRUs can extract more long-time sequence features, and the fully connected layer FC is used for dimension transformation of the long-time sequence features to obtain a second sample feature vector V2.

[0192] The first weight and the second weight determined by using the attention mechanism are used for feature fusion of the first sample feature vector V1 and the second sample feature vector V2 to obtain a feature vector V, and a predicted load value Y at a prediction time is output by the fully connected layer FC.

[0193] According to the predicted load values of the plurality of computing nodes at the prediction time, a target computing node is determined, so that the target computing node performs calculation according to test data of a physical test resource to obtain a test result.

[0194] In some embodiments, the computing node is a computer device deployed in a target range, which can also be referred to as a data center host. The predicted load of the computer device of each target range can be calculated according to the host load of the computer device corresponding to each target range and the task load of all physical test devices and simulation test devices in each target range, so as to determine a target computer device for calculating the test data of the physical test devices and the simulation test devices of each target range at a prediction time. After the target computer device calculates a test result according to the test data of the physical test devices and the simulation test devices of each target range, the test result is sent to a logical target range test platform, so that the logical target range test platform renders and updates the state of the virtual test resources corresponding to the physical test devices and the simulation test devices according to the test result.

[0195] In some possible implementations, if the physical test devices and the simulation test devices of each target range generate a large amount of test data, a plurality of target computer devices with the lowest predicted load can be selected according to the predicted load of the computer device of each target range, the large amount of test data is divided into a plurality of groups of test data according to the sizes of the predicted loads of the plurality of target computer devices, and the plurality of groups of test data are respectively sent to the plurality of target computer devices.

[0196] The physical test equipment and the simulation test equipment of each target range are in communication connection with the computer equipment of the target range, and the computer equipment of different target ranges can be in communication connection through the logical target range test platform, so as to transmit the test data of the physical test equipment and the simulation test equipment between the computer equipment according to the predicted load of each computer equipment through the logical target range test platform.

[0197] The load prediction method provided by the above embodiment calculates the load value at the prediction moment according to the load values in a period of time by using the load prediction model, so as to select the computing node with the lowest load value from multiple computing nodes to perform data calculation at the prediction moment, thereby improving the calculation efficiency, reducing the delay of flight test task processing, and improving the flight test efficiency.

[0198] Please refer to Figure 11 The structure schematic diagram of the training device of the load prediction model provided by the embodiment of the present application is shown in Figure 11 The training device comprises:

[0199] The sample data acquisition module 11 is configured to acquire a sample load sequence of any computing node of the digital twin logical target range and a real load value at a sample prediction moment after a plurality of continuous sample moments, and the sample load sequence comprises sample load values of any computing node at the plurality of continuous sample moments.

[0200] The first sample feature extraction module 12 is configured to extract features of the sample load sequence by using an initial convolutional neural network to obtain a first sample feature vector.

[0201] The second sample feature extraction module 13 is configured to extract features of the sample load sequence by using an initial gated recurrent neural network to obtain a second sample feature vector.

[0202] The sample load prediction module 14 is configured to perform load prediction by using an initial full connection layer according to the first sample feature vector and the second sample feature vector to obtain a sample predicted load value of the computing node at the sample prediction moment.

[0203] The model parameter adjustment module 15 is configured to perform parameter adjustment on the initial convolutional neural network, the initial gated recurrent neural network and the initial full connection layer according to the sample predicted load value and the real load value to obtain the load prediction model.

[0204] Optionally, the sample data acquisition module 11 comprises:

[0205] The sample load value acquisition unit is configured to acquire a first sample load value of any computing node at a plurality of continuous sample moments and a second sample load value of a test resource of the digital twin logical target range at the plurality of continuous sample moments.

[0206] The sample load sequence generation unit is configured to calculate a sample load sequence according to the first sample load values at the continuous multiple sample moments and the second sample load values at the continuous multiple sample moments.

[0207] Optionally, before the first sample feature extraction module 12, the device further comprises:

[0208] The two-dimensional conversion module is configured to perform two-dimensional conversion on the sample load sequence to generate a static timing diagram corresponding to the sample load sequence.

[0209] The first sample feature extraction module 12 is specifically configured to perform feature extraction on the static timing diagram by using an initial convolutional neural network to obtain a first sample feature vector.

[0210] Optionally, the first sample feature extraction module 12 comprises:

[0211] The convolution processing unit is configured to obtain a sample feature map by performing convolution processing on the static timing diagram, the sample feature map comprising a plurality of channel feature maps.

[0212] The attention extraction unit is configured to perform attention extraction on the plurality of channel feature maps by using an attention mechanism to determine channel weights of the plurality of channel feature maps.

[0213] The first sample feature extraction unit is configured to calculate the first sample feature vector according to the plurality of channel feature maps and the channel weights.

[0214] Optionally, the first sample feature extraction unit comprises:

[0215] The channel selection subunit is configured to select a target channel feature map from the plurality of channel feature maps according to the channel weights of the plurality of channel feature maps.

[0216] The first sample feature extraction subunit is configured to calculate the first sample feature vector according to the target channel feature map and the channel weight corresponding to the target channel feature map.

[0217] Optionally, the sample load prediction module 14 is specifically configured to weight the first sample feature vector and the second sample feature vector according to the initial first weight and the initial second weight to obtain a weighted sample feature vector, and perform load prediction on the weighted sample feature vector by using an initial fully connected layer to obtain a sample predicted load value.

[0218] Please refer to Figure 12 The structure diagram of the load prediction device provided by the embodiment of the present application is shown in FIG. 1, which can comprise: Figure 12

[0219] ​The load sequence acquisition module 71 is configured to acquire a load sequence of each of a plurality of computing nodes of the digital twin logical target range, and the load sequence of each computing node includes load values of the digital twin logical target range at a plurality of continuous time points.

[0220] The first feature extraction module 72 is configured to perform feature extraction on the load sequence of each computing node by using a convolutional neural network in the load prediction model, to obtain a first feature vector.

[0221] The second feature extraction module 73 is configured to perform feature extraction on the load sequence of each computing node by using a gated recurrent neural network in the load prediction model, to obtain a second feature vector.

[0222] The load prediction module 74 is configured to perform load prediction on the first feature vector and the second feature vector by using a full connection layer in the load prediction model, to obtain a predicted load value of each computing node at a prediction time point, wherein the load prediction model is obtained by training the load prediction model by using the training method.

[0223] The node selection module 75 is configured to determine a target computing node according to the predicted load values of the plurality of computing nodes, and the target computing node is used for data calculation at the prediction time point.

[0224] The above apparatus is configured to execute the method provided in the foregoing embodiments, and has similar implementation principles and technical effects, which will not be described here in detail.

[0225] The above modules can be one or more integrated circuits configured to implement the above method, for example, one or more application specific integrated circuits (ASICs), or one or more microprocessors, or one or more field programmable gate arrays (FPGAs), etc. For another example, when the above modules are implemented in the form of a processing element scheduling code, the processing element can be a general-purpose processor, such as a central processing unit (CPU) or other processor that can invoke program code. For another example, the modules can be integrated together in the form of a system on a chip (SOC).

[0226] Please refer to Figure 13 , the schematic diagram of an electronic device provided in the embodiments of the present application is as follows Figure 13As shown, the electronic device 100 includes a processor 101, a storage medium 102, and a bus. The storage medium 102 stores program instructions executable by the processor 101. When the electronic device 100 is running, the processor 101 communicates with the storage medium 102 through the bus. The processor 101 executes the program instructions to perform the above method embodiments. The specific implementation and technical effects are similar, and will not be repeated here.

[0227] Optionally, the present application also provides a computer readable storage medium, and the storage medium stores a computer program. The computer program is executed by a processor to perform the above method embodiments.

[0228] In several embodiments provided by the present application, it should be understood that the disclosed apparatus and method can be implemented in other manners. For example, the described embodiments of the apparatus are merely schematic, and 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 between the units can be indirect couplings or communication connections through some interfaces, devices or units, and can be electrical, mechanical or in other forms.

[0229] 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, that is, can be located in one place, or can be distributed on multiple network units. According to actual needs, some or all of the units can be selected to achieve the purpose of the embodiments.

[0230] In addition, each functional unit in the various embodiments of the present application can be integrated into a processing unit, or each unit can exist physically, or two or more units can be integrated into one unit. The integrated unit can be implemented in the form of hardware, or in the form of hardware plus software function units.

[0231] The integrated unit in the form of software function unit can be stored in a computer readable storage medium. The software function unit is stored in a storage medium, and includes a plurality of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute part of steps of the method according to various embodiments of the present application. The storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various storage media capable of storing program codes.

[0232] The above merely describes specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A training method for a load prediction model, characterized in that, The application is applied to a digital twin logical target range system, and the digital twin logical target range system comprises a logical target range test platform, at least one physical test resource and a plurality of computing nodes, the logical target range test platform is used for creating a logical target range, each virtual test resource of each logical target range corresponds to one physical test resource, and the training method comprises: Obtaining a sample load sequence of any computing node of a digital twin logical target range and a real load value at a sample prediction moment after a plurality of continuous sample moments, the sample load sequence comprising sample load values of the any computing node at the plurality of continuous sample moments, the sample load values comprising weighted values of host load of the any computing node and task load of a physical test resource corresponding to a virtual test resource of the logical target range; Performing feature extraction on the sample load sequence by using an initial convolutional neural network to obtain a first sample feature vector; Performing feature extraction on the sample load sequence by using an initial gated recurrent neural network to obtain a second sample feature vector; Performing load prediction by using an initial full connection layer according to the first sample feature vector and the second sample feature vector to obtain a sample predicted load value of the computing node at the sample prediction moment; Performing parameter adjustment on the initial convolutional neural network, the initial gated recurrent neural network and the initial full connection layer according to the sample predicted load value and the real load value to obtain a load prediction model; Before the step of performing feature extraction on the sample load sequence by using an initial convolutional neural network to obtain a first sample feature vector, the method further comprises: Performing two-dimensional conversion on the sample load sequence to generate a static timing diagram corresponding to the sample load sequence; The step of performing feature extraction on the sample load sequence by using an initial convolutional neural network to obtain a first sample feature vector comprises: Performing feature extraction on the static timing diagram by using the initial convolutional neural network to obtain the first sample feature vector; The step of performing feature extraction on the static timing diagram by using the initial convolutional neural network to obtain the first sample feature vector comprises: Obtaining a sample feature map by performing convolution processing on the static timing diagram, the sample feature map comprising a plurality of channel feature maps; Performing attention extraction on the plurality of channel feature maps by using an attention mechanism to determine channel weights of the plurality of channel feature maps; Calculating the first sample feature vector according to the plurality of channel feature maps and the channel weights.

2. The training method of claim 1, wherein, The step of obtaining a sample load sequence of any computing node of a digital twin logical target range comprises: Obtaining first sample load values of the any computing node at the plurality of continuous sample moments and second sample load values of test resources of the digital twin logical target range at the plurality of continuous sample moments; Calculating the sample load sequence according to the first sample load values at the plurality of continuous sample moments and the second sample load values at the plurality of continuous sample moments.

3. The training method of claim 1, wherein The step of calculating the first sample feature vector according to the plurality of channel feature maps and the channel weights comprises: select a target channel feature map from the plurality of channel feature maps according to channel weights of the plurality of channel feature maps; calculate the first sample feature vector according to the target channel feature map and the channel weight corresponding to the target channel feature map.

4. The training method of claim 1, wherein, The load prediction is performed on the first sample feature vector and the second sample feature vector through an initial full connection layer to obtain a sample prediction load value of the computing node at the sample prediction moment according to the first sample feature vector and the second sample feature vector. The first sample feature vector and the second sample feature vector are weighted according to an initial first weight and an initial second weight to obtain a weighted sample feature vector. The load prediction is performed on the weighted sample feature vector through the initial full connection layer to obtain the sample prediction load value.

5. A load prediction method characterized by, The method comprises: obtaining a load sequence of a plurality of computing nodes of a digital twin logical range, the load sequence of each computing node comprising: load values of the digital twin logical range at a plurality of continuous moments; extracting features of the load sequence of each computing node through a convolutional neural network in a load prediction model to obtain a first feature vector; extracting features of the load sequence of each computing node through a gated recurrent neural network in the load prediction model to obtain a second feature vector; performing load prediction through a full connection layer in the load prediction model according to the first feature vector and the second feature vector to obtain a predicted load value of each computing node at the prediction moment, wherein the load prediction model is obtained by training the load prediction model through the training method according to any one of claims 1-4; determining a target computing node according to the predicted load values of the plurality of computing nodes, the target computing node being used for data calculation at the prediction moment. 6.A training apparatus of a load prediction model, comprising: The digital twin logical range system comprises: a logical range test platform, at least one physical test resource, and a plurality of computing nodes, the logical range test platform being used for creating logical ranges, each virtual test resource of each logical range corresponding to one physical test resource, and the training device comprising: a sample data acquisition module configured to obtain a sample load sequence of any computing node of a digital twin logical range and a real load value at a sample prediction moment after a plurality of continuous sample moments, the sample load sequence comprising: sample load values of the any computing node at the plurality of continuous sample moments, the sample load values comprising weighted values of host load of the any computing node and task load of a physical test resource corresponding to a virtual test resource of the logical range; a first sample feature extraction module configured to extract features of the sample load sequence through an initial convolutional neural network to obtain a first sample feature vector; a second sample feature extraction module configured to extract features of the sample load sequence through an initial gated recurrent neural network to obtain a second sample feature vector; and a load prediction module configured to perform load prediction on the first sample feature vector and the second sample feature vector through an initial full connection layer to obtain a sample prediction load value of the any computing node at the sample prediction moment. a sample load prediction module, configured to perform load prediction on the first sample feature vector and the second sample feature vector through an initial full connection layer to obtain a sample predicted load value of the computing node at a sample prediction moment; a model parameter adjustment module, configured to perform parameter adjustment on the initial convolutional neural network, the initial gated recurrent neural network and the initial full connection layer according to the sample predicted load value and the real load value, to obtain a load prediction model; the device further comprises: a two-dimensional conversion module, configured to perform two-dimensional conversion on the sample load sequence to generate a static time sequence diagram corresponding to the sample load sequence; the first sample feature extraction module is specifically configured to perform feature extraction on the static time sequence diagram by using the initial convolutional neural network to obtain the first sample feature vector; the first sample feature extraction module comprises: a convolution processing unit, configured to perform convolution processing on the static time sequence diagram to obtain a sample feature map, the sample feature map comprising a plurality of channel feature maps; an attention extraction unit, configured to perform attention extraction on the plurality of channel feature maps by using an attention mechanism to determine channel weights of the plurality of channel feature maps; a first sample feature extraction unit, configured to calculate the first sample feature vector according to the plurality of channel feature maps and the channel weights.

7. A load prediction device characterized by comprising: the device comprises: a load sequence acquisition module, configured to acquire load sequences of a plurality of computing nodes of a digital twin logical target range, the load sequence of each computing node comprising load values of the digital twin logical target range at consecutive moments; a first feature extraction module, configured to perform feature extraction on the load sequence of each computing node by using a convolutional neural network in a load prediction model to obtain a first feature vector; a second feature extraction module, configured to perform feature extraction on the load sequence of each computing node by using a gated recurrent neural network in the load prediction model to obtain a second feature vector; a load prediction module, configured to perform load prediction on the first feature vector and the second feature vector through a full connection layer in the load prediction model to obtain a predicted load value of each computing node at a prediction moment, wherein the load prediction model is obtained by training the load prediction model according to the training method of the load prediction model in any one of claims 1-4; a node selection module, configured to determine a target computing node according to the predicted load values of the plurality of computing nodes, the target computing node being used for data calculation at the prediction moment.

8. An electronic device, comprising: comprises: a processor, a storage medium and a bus, the storage medium storing program instructions executable by the processor, when the electronic device is running, the processor and the storage medium communicate through the bus, the processor executes the program instructions to perform the steps of the training method of the load prediction model in any one of claims 1-4, or the steps of the load prediction method in claim 5.

Images