A method for predicting multi-type task requirements for a computing power network

Abstract

This invention relates to computing network prediction and proposes a method for predicting multi-type task requirements in computing networks. This method analyzes and extracts features from multi-type task requirements in computing network scenarios, uses a deep neural network to analyze and process these requirements, outputs classification results, and then processes the classified task requirements through a combination of frequency attention mechanism and non-stationary factors. This not only automatically extracts deeper-level correlation features from global data but also automatically identifies abrupt changes in task requirements, thereby improving the prediction accuracy of multi-type task requirements in computing network scenarios. Compared with traditional methods, this method characterizes computing network task requirement data from a combined time and frequency domain perspective and introduces non-stationary factors, effectively identifying abrupt changes in task requirements and exhibiting higher accuracy and applicability.

Description

Technical Field

[0001] This invention relates to computing power network prediction. Background Technology

[0002] With the comprehensive development of the computing power network center industry, diverse computing power task demands are constantly emerging. In order to improve the utilization rate of computing power resources and reduce operating costs, it is necessary to establish a practical and effective demand forecasting system.

[0003] Demand forecasting in computing networks is closely related to time information and serves as the basis for resource scheduling decisions throughout the network. In practical applications, as tasks progress, a mismatch between computing power demand and resources arises. Manually analyzing and predicting resource usage for these massive, diverse task demands is virtually impossible. Current research on time series analysis has been developed and applied in various fields, such as energy consumption, traffic planning, stock analysis, disease transmission, and weather forecasting. However, most time series algorithms focus on uncovering periodic and trend patterns in data, neglecting the non-stationary characteristics of data in real-world applications. For example, the time distribution of task demands in computing networks is not a smooth curve; it includes short-term surges in demand due to unexpected tasks. Simply focusing on periodic and trend patterns in demand within the computing network, while ignoring sudden changes in task demands, is ineffective for demand forecasting in computing network scenarios. There is an urgent need to integrate sudden changes in task demands to further improve the accuracy of demand forecasting.

[0004] The above research reveals that a major challenge in demand forecasting for computing networks is analyzing the non-stationary demand characteristics present in complex task requirements. Traditional forecasting methods smooth data, making it difficult to capture the temporal correlation of sudden demands, thus lacking effective forecasting capabilities. General time-series-based demand forecasting methods cannot handle sudden or highly volatile task demands and have limited applicability in computing network scenarios.

[0005] There are two existing implementation schemes:

[0006] 1) Frequency attention mechanism prediction models can convert time-domain data into frequency-domain data. Through frequency domain analysis, a global spectrum of the sequence can be obtained, allowing for a direct observation of the signal's frequency distribution. Then, a self-attention mechanism is used to calculate attention values, thereby capturing the dependencies between different time steps in the time series. However, because this approach converts the sequence data from the time domain to the frequency domain, it can only capture the global frequency distribution of the sequence, while local non-stationary data within the sequence may be obscured or ignored.

[0007] 2) The non-stationary attention mechanism model, by improving upon the self-attention mechanism, can capture the dynamic features of non-stationary data that changes over time in time series. Traditional self-attention mechanisms use fixed parameters to calculate similarity when calculating attention weights, which cannot adapt to the changes in non-stationary data in time series. The non-stationary attention mechanism model, however, introduces time-dimensional information, allowing the model to adjust the calculation method of attention weights according to changes over time, thus better capturing the non-stationarity in time series. Because it considers non-stationary feature extraction, this approach can better handle locally non-stationary data, but it may have certain limitations in predicting overall trends and global volatility. Summary of the Invention

[0008] The purpose of this invention is to propose a multi-type task demand prediction method for computing power networks, enhancing the ability to predict sudden demands in computing power network scenarios and improving the applicability of time series-based demand prediction methods under irregular conditions. This invention not only automatically extracts deeper correlation features from global data but also automatically identifies non-stationary data in the sequence data. This improvement enables the model to more accurately predict the changing trends and volatility in time series, thereby enhancing the model's predictive power and application scope.

[0009] Technical solution of the present invention:

[0010] A method for predicting the demand of multiple types of tasks in a computing network includes the following steps:

[0011] Step S1: Obtain initial task requirement data based on the original task requirement data, perform data standardization processing based on the initial task requirement data to obtain a standardized sequence, and divide the sequence into training set, test set and validation set;

[0012] Step S2: Based on the standardized sequence in step S1, the classification labels corresponding to different task requirements are obtained by the output of the deep neural network classifier;

[0013] Step S3: Based on the classification labels obtained in step S2 and the training set divided in step S1, perform sequence feature engineering and calculate the non-stationary factor of the sequence, and re-encode it through an encoder based on a non-stationary frequency attention mechanism.

[0014] Step S4: Based on the encoding result of the non-stationary frequency attention mechanism encoder in Step S3, input it into the decoder to generate the prediction result and save the current training parameters; then, take the validation set divided in Step S1 as input, use the currently saved training parameters to generate the prediction value, compare it with the actual value of the validation set, optimize the loss function, and thus obtain the optimal training parameters.

[0015] Step S5: Based on the optimal training parameters obtained in step S4, perform prediction calculations using the test set divided in step S1 to evaluate the prediction performance and output the prediction results.

[0016] Further, step S1 includes:

[0017] S11: Select feature columns closely related to the task requirements from the original task requirement data and recombine them to obtain a feature sequence;

[0018] S12: Based on the feature sequence constructed by S11, the mean and standard deviation are calculated in each dimension. The mean represents the average value in each dimension, while the standard deviation reflects the range and dispersion of variation in each dimension.

[0019] S13: Based on the mean and standard deviation calculated in S12, scale the feature sequence constructed in S11. By subtracting the mean from the value of the feature sequence and then dividing by the standard deviation, a standardized feature sequence with zero mean and unit standard deviation is obtained.

[0020] S14: Divide the standardized feature sequences obtained in S13 into training set, test set and validation set;

[0021] Preferably, the set is divided into a training set, a test set, and a validation set based on the standardized feature sequence, including:

[0022] Set the training set time window, the test set time window, and the validation set time window;

[0023] The training set is divided into the standardized feature sequence based on the training set time window;

[0024] A test set is divided into the standardized feature sequence based on the test set time window;

[0025] A validation set is defined in the standardized feature sequence based on the validation set time window;

[0026] The training set time window, the test set time window, and the validation set time window are consecutive time windows, and the training set time window precedes the test set time window and the validation set time window in time.

[0027] Further, step S2 specifically includes:

[0028] The deep neural network classifier includes multiple neural network modules, pooling layers, fully connected layers, and a Softmax activation function;

[0029] The standardized feature sequence is provided as input to the deep neural network classifier. The deep neural network classifier extracts key information from the input features through a multi-layer neural network module. The pooling layer reduces the data dimensionality while retaining important features. The fully connected layer is responsible for expanding the features extracted by the previous layer into a one-dimensional vector. Finally, the output of the previous layer is converted into a probability distribution through the Softmax activation function, so that the output value of each category represents the probability that the input belongs to that category, thereby realizing the multi-class classification task.

[0030] Further, step S3 includes:

[0031] S31) After fusing the classification labels and the training set divided in step S1, embedding encoding is performed;

[0032] S32) Using the standardized sequence and the mean and standard deviation calculated based on the sequence as input, a multilayer perceptron is used to infer the non-stationary factor of the output sequence;

[0033] S33) The embedding encoding result from S31 and the non-stationary factor obtained from S32 are calculated using the non-stationary frequency attention mechanism module to obtain a real-valued vector in the time domain. The non-stationary frequency attention mechanism is a variant of the attention mechanism that focuses on processing non-stationary sequence data, highlighting important change points and frequency components in the sequence;

[0034] The non-stationary frequency attention mechanism module includes three modules: a non-stationary frequency attention mechanism module, a sequence decomposition module, and a feedforward network. The steps for obtaining real-valued vectors in the time domain include:

[0035] Based on the embedding encoding vector, generate three vector representations: query, key, and value. Then, use the fast Fourier transform to obtain a complex vector in the frequency domain.

[0036] The enhanced complex vector after transformation is obtained by weighting and summing the complex vector using a random weight matrix.

[0037] Based on the enhanced complex vector and the non-stationary factor, a new vector is obtained by calculation after transformation through the non-stationary frequency attention mechanism.

[0038] The real vector in the time domain is restored by inverse fast Fourier transform based on the transformed vector.

[0039] S34) Based on the output of the non-steady-state frequency attention mechanism module in step S33, input it into the sequence decomposition module to decompose and obtain periodic feature vectors;

[0040] The sequence decomposition module performs the following steps:

[0041] First, the output of the non-steady-state frequency attention mechanism module is padded with front and back real number vectors in the time domain;

[0042] Then, by setting the sliding window and step size, the average value is calculated on the filled sequence according to the sliding window size to obtain the trend feature vector;

[0043] Finally, the periodic feature vector is obtained by subtracting the trend feature vector from the original sequence vector, where the original sequence vector is the sequence vector before decomposition.

[0044] S35) Based on the periodic feature vector obtained from step S34, it is input into the feedforward network, and the final re-encoded result is output. The feedforward network includes multiple convolutional layers, multiple dropout layers, and activation functions, which are cascaded together to further extract and encode the periodic feature vector.

[0045] Furthermore, in step S4, the decoder includes three modules: a cross-attention mechanism module, a sequence decomposition module, and a feedforward network.

[0046] The steps in step S4 include:

[0047] S41) Generate a query guided by task requirements based on the training set divided in step 1. Specifically, by processing and analyzing the training set, feature representations related to task requirements are obtained, and the query is constructed using these feature representations.

[0048] S42) Based on the encoding result output by the non-steady-state frequency attention mechanism encoder in step S3, a set of keys and values ​​associated with the task requirements is generated. These keys and values ​​are used to associate and describe the task requirements, and contain key features and information of the input data. The key represents the focus or feature of the task requirements, and the value is the corresponding information or weight.

[0049] S43) The query generated by S41 based on the training set and the key and value generated by S42 based on the encoder output are used as inputs to the cross-attention mechanism module to calculate a new vector representation. This process can capture the correlation between the encoder and decoder and generate a vector that integrates the features of both.

[0050] S44) The vector transformed by the cross-attention mechanism module is further decomposed by the sequence decomposition module to obtain the periodic feature vector and trend feature vector of the sequence.

[0051] S45) The periodic feature vectors obtained by the sequence decomposition module are re-encoded and output through the feedforward network. The output result is used as the periodic vector part of the final prediction result. The feedforward network consists of multiple convolutional layers, multiple dropout layers and activation functions cascaded together, which can further extract and encode the periodic feature vectors.

[0052] S46) The output of the feedforward network is added to the trend feature vector obtained by the sequence decomposition module to obtain the final prediction result, and the current training parameters are saved; the final prediction result includes a periodic feature vector part and a trend feature vector part.

[0053] S47) Using the validation set divided in step S1 as input, the training parameters saved in S46 are used for validation and prediction values ​​are generated. The prediction values ​​are compared with the actual values ​​of the validation set, the loss function is calculated, and the training parameters are optimized through iterative training to obtain the optimal training parameters.

[0054] Further, step S5 includes:

[0055] Based on the optimal training parameters obtained from the training, the unused test set divided in step S1 is used as input to calculate the error index between the prediction result and its true value, so as to evaluate the prediction performance and output the prediction result.

[0056] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0057] The multi-type task requirement prediction method for computing power networks provided by this invention analyzes and extracts features from multi-type task requirements in computing power network scenarios. It then uses a deep neural network to analyze and process these requirements, outputting classification results. The classified task requirements are further processed using a combination of frequency attention mechanism and non-stationary factors. This approach not only automatically extracts deeper-level correlation features from global data but also automatically identifies abrupt changes in task requirements, thereby improving the prediction accuracy of multi-type task requirements in computing power network scenarios. Compared to traditional methods, this method characterizes computing power network task requirement data from a combined time and frequency domain perspective and introduces non-stationary factors, effectively identifying abrupt changes in task requirements and exhibiting higher accuracy and applicability. Attached Figure Description

[0058] Figure 1 A flowchart illustrating the multi-type task demand prediction method for computing power networks of the present invention is shown.

[0059] Figure 2 A schematic diagram of a deep neural network structure is shown;

[0060] Figure 3 A schematic diagram of a multilayer perceptron is shown.

[0061] Figure 4 A schematic diagram of the encoder structure with a nonsteady-state frequency attention mechanism is shown.

[0062] Figure 5 A schematic diagram of the decoder structure is shown; Detailed Implementation

[0063] The key innovation of this invention lies in combining frequency attention and non-stationary attention to compensate for the shortcomings of their respective models, thereby achieving better performance in time series forecasting. Frequency attention captures the global trend and volatility of a series, while non-stationary attention captures its local non-stationary features. By combining these two attention mechanisms, this invention can simultaneously consider both the overall distribution characteristics and local non-stationary features of a series, thus providing a more comprehensive understanding and prediction of time series behavior. This combination provides more accurate prediction results and demonstrates higher predictive performance and robustness in practical applications. By fully leveraging the advantages of frequency and non-stationary attention, this invention enables a better understanding and analysis of time series data, providing a more reliable basis for decision-making and forecasting.

[0064] The technical solutions provided in this application will be further described below with reference to specific embodiments and accompanying drawings. The advantages and features of this application will become clearer from the following description.

[0065] Terminology Explanation:

[0066] The attention mechanism originates from research on human vision. In cognitive science, due to the bottleneck of information processing, humans selectively focus on a portion of all information while ignoring other visible information. This mechanism is commonly referred to as the attention mechanism.

[0067] A multilayer perceptron (MLP) is a basic feedforward neural network model consisting of multiple layers of neurons, each fully connected to the previous one. Its basic building block is the neuron (or node). Each neuron receives input from neurons in the previous layer, performs a non-linear transformation on the input using an activation function, and then passes the transformed result to the next layer's neurons. Each neuron has a set of learnable weights and biases used to adjust the input weights and biases to better fit the training data.

[0068] Example

[0069] A method for predicting the demand for multiple types of tasks in a computing power network includes the following steps:

[0070] Step S1: Obtain initial task requirement data based on the original task requirement data, perform data standardization processing based on the initial task requirement data to obtain a standardized sequence, and divide the sequence into training set, test set and validation set.

[0071] Specifically, the original task requirement data is obtained, and feature columns closely related to the task requirements are selected from the original task requirement data to construct the initial task requirement data. Where x i This represents the demand characteristics at time i; then the mean is calculated. and standard deviation The initial task requirement data x is scaled. Obtain the standardized sequence

[0072] Based on the standardized sequences, we set training set time windows, test set time windows, and validation set time windows to partition the training set x′ from the standardized sequences. train Test set x′ test , Validation set x′ valid The training set time window, test set time window, and validation set time window are all continuous time windows.

[0073] Preferably, the sizes of the training set time window, the test set time window, and the validation set time window are set in a ratio of 7:2:1, and the training set time window, the test set time window, and the validation set time window are consecutive time periods.

[0074] Step S2 involves using the standardized sequence from step S1 as the basis for obtaining classification labels corresponding to different task requirements through the output of a deep neural network classifier. This process delves into the hidden patterns and features within the sequence.

[0075] The deep neural network classifier (such as...) Figure 2 It consists of multiple neural network modules, pooling layers, fully connected layers, and the Softmax activation function.

[0076] The standardized feature sequence is fed into the deep neural network classifier to extract key information from the input features. The pooling layer reduces the data dimensionality while retaining important features. Then, the fully connected layer is responsible for expanding the features extracted by the previous layer into a one-dimensional vector. Finally, the softmax activation function is used to convert the output of the previous layer into a probability distribution, so that the output value of each class represents the probability that the input belongs to that class, thereby realizing the multi-class classification task.

[0077] The deep neural network classifier learns abstract representations of input sequences through a combination of multiple neural network layers and classifies the sequences based on these representations. Standardized feature sequences are provided as input to the deep neural network classifier, which learns and iteratively optimizes through the cascading of multiple neural network modules, pooling layers, fully connected layers, and the Softmax activation function, enabling it to distinguish and classify the input data.

[0078] Specifically:

[0079] First, the input feature sequence undergoes feature extraction through multiple neural network modules. The extracted results are then passed through a pooling layer to further refine key information. This pooling layer employs an average pooling strategy, averaging the features across local regions to effectively reduce dimensionality while retaining the most representative features. This process not only reduces the complexity of subsequent computations but also improves robustness to input variations.

[0080] Then, based on the feature vector output by the pooling layer, it is flattened by a fully connected layer, that is, mapped from a high-dimensional space to one dimension, so that it can be directly applied to the classification task. The fully connected layer maps the features to the class-related latent space through a series of weight matrix transformations.

[0081] Finally, the output of the fully connected layer is normalized using the Softmax activation function and converted into a probability distribution form, thereby assigning a probability value to each predefined category and realizing the classification label prediction of the sequence.

[0082] Step S3: Based on the classification labels obtained in step S2 and the training set divided in step S1, sequence feature engineering is performed and the non-stationary factor of the sequence is calculated. The sequence is then re-encoded using an encoder based on a non-stationary frequency attention mechanism.

[0083] First, based on the training set x′ partitioned in step S1 train And the classification label c obtained in step S2, are used for fusion feature embedding encoding, x emb_out =embedding(x′) train ,c);

[0084] Next, the standardized sequence x′ obtained in step S1 and the average value μ calculated based on this sequence are combined. x and standard deviation σ x As input, using a multilayer perceptron ( Figure 3 To infer the non-stationary factors τ and Δ of the output sequence, logτ = MLP(σ) x ,x′) and Δ=MLP(μ x ,x′);

[0085] The above describes the result x of the embedding encoding. emb_out Both the non-stationary factors τ and Δ will be used as inputs to the encoder based on the non-stationary frequency attention mechanism.

[0086] Figure 4 The structure of the non-stationary frequency attention mechanism encoder is shown, which includes three modules: a non-stationary frequency attention mechanism module, a sequence decomposition module, and a feedforward network.

[0087] The non-steady-state frequency attention mechanism module:

[0088] Based on the result x of the embedding encoding emb_out Generate query q, key k, and value v, and obtain three complex vectors in the frequency domain through Fast Fourier Transform. Applying a random weight matrix W to these three complex vectors Q W K W V The three enhanced complex vectors are obtained by weighted summation. The three enhanced complex vectors and the non-stationary factors (τ and Δ) are used as inputs to the encoding function of the non-stationary frequency attention mechanism, i.e., the calculation is performed. Obtain the vector transformed by the frequency attention mechanism In the formula, Softmax(·) represents the normalization function, and d k The dimension representing the key k.

[0089] Vector Z transformed based on frequency attention mechanism out Through inverse Fast Fourier Transform (FFT) -1 Restore the real vector x′ in the time domain en =FFT -1 (Z out ).

[0090] The sequence decomposition module:

[0091] Further extracting periodic features, the output x′ of the non-steady-state frequency attention mechanism module is... en Sequence decomposition is performed, specifically: first, for x′ en Perform vector padding operations (padding(x′)) en Then, a moving average operation (AvgPool) is performed to obtain a series of averages. These averages are then concatenated to obtain the trend characteristics (x′) of the filled sequence. en_trend The periodic eigenvector can be obtained by subtracting the trend eigenvector from the original sequence vector, where the original sequence vector is the sequence vector x′ before decomposition. en That is, the periodic eigenvector (x′) en_period) = original sequence vector (x′) en Trend eigenvector (x′) en_trebd ).

[0092] More specifically, the sliding window size w and step size s = 1 are set. To ensure that the length of the vector after moving average is consistent with the length of the original vector, at x′ en Vector supplement There are x′0, in x′ en Vectors are supplemented later x′ L-1 After padding, an averaging operation with a sliding window size of w is performed on the sequence of vector length L+w-1. The resulting vector composed of moving averages is denoted as the trend feature vector x′. en_trend Thus, the periodic eigenvector x′ is obtained. en_period , represented as x′ en_period =x′ en -x′ en_trend .

[0093] The feedforward network:

[0094] Able to perform periodic eigenvectors x′ en_period Further feature extraction and encoding are then performed. Specifically:

[0095] The feedforward network comprises multiple convolutional layers, multiple dropout layers, and a ReLU activation function, and performs the following process:

[0096] First, the input vector x′ is extracted through a convolutional layer. en_period Local features x′ con =conv(x′) en_period );

[0097] Secondly, the nonlinear transformation x′ of the data is completed using the ReLU activation function. relu =ReLU(x′) conv1 );

[0098] To prevent overfitting, x′ is then output through a dropout layer. dropout =dropout(x′) relu ,0.5);

[0099] Then, the local features x′ of the input vector are further extracted through convolutional layers. conv =conv(x′) dropout ),

[0100] Finally, the final result x′ is output through the dropout layer. en_out =dropout(x′)relu ,0.5).

[0101] Step S4: Based on the encoding result of the non-stationary frequency attention mechanism encoder in step S3, input it into the decoder to generate the prediction result and save the current training parameters; then, take the validation set divided in step S1 as input, use the currently saved training parameters to generate the prediction value, compare it with the actual value of the validation set, optimize the loss function, and thus obtain the optimal training parameters.

[0102] Figure 5 The decoder structure is shown, which includes three modules: a cross-attention mechanism module, a sequence decomposition module, and a feedforward network.

[0103] The encoding result x′ output in step S3 en_out It is passed as input to the decoder for computation.

[0104] Specifically, step S4 includes:

[0105] S41: Based on the training set x′ partitioned in step S1 train Generate a query Q containing task requirements;

[0106] S42: Based on the encoding result x′ output by the encoder using the non-steady-state frequency attention mechanism in step S3. en_out Generate a key K containing the task requirements association. en Sum V en ;

[0107] S43: The cross-attention mechanism module uses the non-stationary frequency attention mechanism coding function to calculate the cross vector X of the encoder and decoder. res , represented as In the formula, Softmax(·) represents the normalization function, and d k The dimension representing the key k;

[0108] S44: Then, X obtained through the cross-attention mechanism module... res The periodic feature vector X is obtained by using the sequence decomposition module. res_period and trend eigenvector X res_trend The design of the sequence decomposition module here refers to the sequence decomposition module described in step S3;

[0109] S45: The periodic eigenvector X res_period The periodic vector part X of the final prediction result is re-encoded by a feedforward network. out_period The design of the feedforward network here refers to the feedforward network module described in S3.

[0110] S46: The output X of the feedforward networkout_period With trend eigenvector X res_trend Add them together to get the final prediction result X out_predict And save the current training parameters;

[0111] S47: Divide the validation set x′ in step S1 valid As input, the training parameters stored in S46 are used for validation and to generate the predicted value X. out_predict X out_predict and the actual value x′ in the validation set valid_true Compare and calculate the loss function loss = MSE(x′). valid_true ,X out_predict The loss value is continuously reduced through iterative training. When the loss value is at its minimum, the current training parameters are saved as the optimal training parameters.

[0112] Step S5: Based on the optimal training parameters obtained in step S4, perform prediction calculations using the test set divided in step S1 to evaluate the prediction performance and output the prediction results.

[0113] Specifically, based on the optimal training parameters obtained in step S4, the test set x′ partitioned in step S1 is obtained. test As input, the predicted result X is calculated. out and X out Its true value x′ test_true Calculate the loss function: loss = MSE(x′) test_true ,X out This is used to evaluate the predictive performance.

[0114] This invention overcomes the limitations of traditional demand forecasting methods in predicting the accuracy of various task demands in computing power network scenarios. It increases the demand forecasting system's ability to predict various demand tasks in computing power network scenarios by introducing non-stationary factors and frequency attention mechanisms, which effectively improves the accuracy of demand forecasting for complex tasks, as well as enhances the utilization rate of computing power resources and reduces operating costs.

[0115] The above description is merely a description of preferred embodiments of this application and is not intended to limit the scope of this application in any way. Any changes or modifications made by those skilled in the art based on the above-disclosed technical content should be considered as equivalent and valid embodiments and fall within the scope of protection of the technical solution of this application.

Claims

1. A method for predicting the demand for multiple types of tasks in a computing power network, characterized in that, Includes the following steps: Step S1: Obtain initial task requirement data based on the original task requirement data, perform data standardization processing based on the initial task requirement data to obtain a standardized sequence, and divide the sequence into training set, test set and validation set; Step S2: Based on the standardized sequence in step S1, the classification labels corresponding to different task requirements are obtained by the output of the deep neural network classifier; Step S3: Based on the classification labels obtained in step S2 and the training set divided in step S1, perform sequence feature engineering and calculate the non-stationary factor of the sequence, and re-encode it through a non-stationary frequency attention mechanism encoder; Step S4: Based on the encoding result of the non-stationary frequency attention mechanism encoder in step S3, input it into the decoder to generate the prediction result, and save the current training parameters; Subsequently, the validation set partitioned in step S1 is used as input, and the currently saved training parameters are used to generate predicted values. These predicted values ​​are then compared with the actual values ​​of the validation set to optimize the loss function and obtain the optimal training parameters. Step S5: Based on the optimal training parameters obtained in step S4, perform prediction calculations using the test set divided in step S1 to evaluate the prediction performance and output the prediction results.

2. The method for predicting multi-type task requirements for computing power networks as described in claim 1, characterized in that, Step S1 includes: S11: Select feature columns closely related to the task requirements from the original task requirement data and recombine them to obtain a feature sequence; S12: Based on the feature sequence constructed by S11, the mean and standard deviation are calculated in each dimension. The mean represents the average value in each dimension, while the standard deviation reflects the range and dispersion of variation in each dimension. S13: Based on the mean and standard deviation calculated in S12, scale the feature sequence constructed in S11. By subtracting the mean from the value of the feature sequence and then dividing by the standard deviation, a standardized feature sequence with zero mean and unit standard deviation is obtained. S14: Divide the standardized feature sequences obtained in S13 into training set, test set and validation set.

3. The method for predicting multi-type task requirements for computing power networks as described in claim 2, characterized in that, Based on standardized feature sequences, the dataset is divided into training, test, and validation sets, including: Set the training set time window, the test set time window, and the validation set time window; The training set is divided into the standardized feature sequence based on the training set time window; A test set is divided into the standardized feature sequence based on the test set time window; A validation set is defined in the standardized feature sequence based on the validation set time window; The training set time window, the test set time window, and the validation set time window are consecutive time windows, and the training set time window precedes the test set time window and the validation set time window in time.

4. The method for predicting multi-type task requirements for computing power networks as described in claim 1, characterized in that, Step S2: The deep neural network classifier includes multiple neural network modules, pooling layers, fully connected layers, and a Softmax activation function; The standardized feature sequence is provided as input to the deep neural network classifier. The deep neural network classifier extracts key information from the input features through a multi-layer neural network module. The pooling layer reduces the data dimensionality while retaining important features. The fully connected layer is responsible for expanding the features extracted by the previous layer into a one-dimensional vector. Finally, the output of the previous layer is converted into a probability distribution through the Softmax activation function, so that the output value of each category represents the probability that the input belongs to that category, thereby realizing the multi-class classification task.

5. The method for predicting multi-type task requirements for computing power networks as described in claim 1, characterized in that, Step S3 includes: S31) After fusing the classification labels and the training set divided in step S1, embedding encoding is performed; S32) Using the standardized sequence and the mean and standard deviation calculated based on the sequence as input, a multilayer perceptron is used to infer the non-stationary factor of the output sequence; S33) The embedding encoding result in S31, i.e. the obtained embedding encoding vector, and the non-stationary factor obtained in S32 are calculated by the non-stationary frequency attention mechanism module to obtain a real vector in the time domain. S34) Based on the output of the non-steady-state frequency attention mechanism module in step S33, input it into the sequence decomposition module to decompose and obtain periodic feature vectors; S35) Based on the periodic feature vector obtained by decomposition in step S34, it is input into the feedforward network and the final re-encoded result is output; the feedforward network includes multiple convolutional layers, multiple dropout layers and activation functions, which are cascaded together to further extract and encode the periodic feature vector.

6. The method for predicting multi-type task requirements for computing power networks as described in claim 5, characterized in that, Non-stationary frequency attention mechanism is a variant of attention mechanism that focuses on processing non-stationary sequence data, highlighting important points of change and frequency components in the sequence; The non-stationary frequency attention mechanism module includes three modules: a non-stationary frequency attention mechanism module, a sequence decomposition module, and a feedforward network. The steps for obtaining real-valued vectors in the time domain include: Based on the embedding encoding vector, generate three vector representations: query, key, and value. Then, use the fast Fourier transform to obtain a complex vector in the frequency domain. The enhanced complex vector after transformation is obtained by weighting and summing the complex vector using a random weight matrix. Based on the enhanced complex vector and the non-stationary factor, a new vector is obtained by calculation after transformation through the non-stationary frequency attention mechanism. The real vector in the time domain is restored by inverse fast Fourier transform based on the transformed vector.

7. A method for predicting the demand for multiple types of tasks in a computing network as described in claim 5. Its features are, The sequence decomposition module performs the following steps: First, the output of the non-steady-state frequency attention mechanism module is padded with front and back real number vectors in the time domain; Then, by setting the sliding window and step size, the average value is calculated on the filled sequence according to the sliding window size to obtain the trend feature vector; Finally, the periodic feature vector is obtained by subtracting the trend feature vector from the original sequence vector, where the original sequence vector is the sequence vector before decomposition.

8. The method for predicting multi-type task requirements for computing power networks as described in claim 1, characterized in that, Step S4: The decoder includes three modules: a cross-attention mechanism module, a sequence decomposition module, and a feedforward network. The steps in step S4 include: S41) Generate a query with task requirements based on the training set divided in step 1. Specifically, by processing and analyzing the training set, obtain feature representations related to task requirements, and construct the query using these feature representations. S42) Based on the encoding result output by the non-steady-state frequency attention mechanism encoder in step S3, a set of keys and values ​​associated with the task requirements are generated. These keys and values ​​are used to associate and describe the task requirements, which include the key features and information of the input data; the key represents the focus or feature of the task requirements, and the value is the corresponding information or weight. S43) The query generated by S41 based on the training set and the key and value generated by S42 based on the encoder output are used as inputs to the cross-attention mechanism module to calculate a new vector representation. This process can capture the correlation between the encoder and decoder and generate a vector that integrates the features of both. S44) The vector transformed by the cross-attention mechanism module is further decomposed by the sequence decomposition module to obtain the periodic feature vector and trend feature vector of the sequence; S45) The periodic feature vectors obtained by the sequence decomposition module are re-encoded and output through the feedforward network. The output result is used as the periodic vector part of the final prediction result. The feedforward network consists of multiple convolutional layers, multiple dropout layers and activation functions cascaded together, which can further extract and encode the periodic feature vectors. S46) The output of the feedforward network is added to the trend feature vector obtained by the sequence decomposition module to obtain the final prediction result, and the current training parameters are saved; the final prediction result includes a periodic feature vector part and a trend feature vector part. S47) Using the validation set divided in step S1 as input, the training parameters saved in S46 are used for validation and prediction values ​​are generated. The prediction values ​​are compared with the actual values ​​of the validation set, the loss function is calculated, and the training parameters are optimized through iterative training to obtain the optimal training parameters.

Images