Calculation center load prediction method, device, equipment and medium
By using a load forecasting model with a GRU-MHSA-GRU three-tier architecture and combining multiple data factors, the problem of low load forecasting accuracy in computing centers has been solved, achieving more efficient and stable forecasting results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-26
- Publication Date
- 2026-03-13
AI Technical Summary
Existing methods for predicting the load of computing centers do not consider all factors comprehensively enough, and traditional neural network models are limited in their performance in deep feature mining of data, resulting in low prediction accuracy.
A load forecasting model with a three-tiered architecture of GRU-MHSA-GRU is adopted. It combines historical load, temperature, irradiance, humidity and PUE data. The model uses GRU for primary time-series encoding, middle layer global multi-head self-attention to analyze the global dependencies between features, and final layer GRU to fuse global and local features. Finally, linear regression is performed through a fully connected layer.
It improved the accuracy of computing center load prediction. By standardizing data of different dimensions, it enhanced the efficiency and stability of model training and significantly improved prediction performance.
Smart Images
Figure CN121658332A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computing center load prediction technology, specifically relating to a computing center load prediction method, device, equipment and medium. Background Technology
[0002] With the rapid development of next-generation artificial intelligence technologies, especially the widespread application of open-source large models in various industries, the demand for computing power is growing rapidly. The continuous increase in computing power scale is driving up computing power energy consumption. Accurate prediction of the load on computing centers is crucial for achieving a dynamic balance between the power supply system and load demand.
[0003] The load of computing centers is affected by various factors such as weather, weekdays, and historical load, exhibiting rapid fluctuations and strong nonlinear characteristics, making existing load forecasting methods difficult to apply: on the one hand, the methods do not consider all factors, often only taking into account weather and historical load information; on the other hand, most methods are based on traditional neural network models, which are limited in performance in areas such as deep feature mining of data, resulting in low prediction accuracy. Summary of the Invention
[0004] The purpose of this invention is to provide a method, apparatus, device and medium for predicting the load of computing centers, thereby solving the problem of low accuracy in predicting the load of computing centers in the prior art.
[0005] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for predicting the load of a computing center, comprising the following specific steps: Collect relevant data from the computing center; The relevant data is input into a pre-built load forecasting model, which outputs a load forecast value. The load forecasting model adopts a GRU-MHSA-GRU three-tier architecture, including: The first-layer gated recurrent unit (GRU) is used to perform primary temporal encoding on the relevant data, capture local temporal features, and obtain feature vectors. The intermediate-layer global multi-head self-attention (MHSA) is used to parse the global dependencies between the feature vectors output by the first-layer GRU, and obtain global features and local temporal features. The final gated recurrent unit (GRU) is used to fuse the global features and local temporal features and output the final hidden state. The fully connected layer is used to perform linear regression on the final hidden state to generate load forecasts.
[0006] Preferably, in the step of collecting relevant data from the computing center, the relevant data includes historical load data, temperature data, irradiance data, humidity data, weekday data, and PUE data.
[0007] Preferably, in the step of processing the relevant data into relevant data, the historical load data, temperature, irradiance, humidity and PUE data of the computing center are standardized, and the weekday type data are one-hot encoded to obtain the raw data, which is then input into the load prediction model.
[0008] Preferably, the first-layer gated recurrent unit (GRU), used for performing primary temporal encoding on relevant data, capturing local temporal features, and obtaining feature vectors, includes the following steps: The relevant data is input into the first-layer gated recurrent unit (GRU), and the first-layer gated recurrent unit (GRU) outputs the feature vector. The feature vector is mapped to a three-dimensional space through a linear transformation to generate a query vector Q, a key vector K, and a value vector V. Calculate the attention score between the query vector Q and the key vector K; scale the attention score to obtain the scaled dot product attention score. The scaled dot product attention score is calculated using the Softmax function and then converted into a probability distribution. The final attention is obtained by weighting the probability distribution using attention weights. The final attention is concatenated and subjected to a linear transformation to obtain the final output.
[0009] Preferably, the calculation query vector Q and key vector K The attention score between the two points; the scaling step to obtain the scaled dot product attention score includes: Calculate the query vector Q and key vector K Attention score between: ; in, Represents the query vector Q and key vector K Perform dot product calculation between them; Represents the key vector K Transpose of; Divide the attention score by a constant The scaled dot product attention score is obtained as follows: ; in, It is the dimension of the key vector.
[0010] Preferably, the step of calculating the scaled dot product attention score using the Softmax function and converting the scaled dot product attention score into a probability distribution includes: .
[0011] Preferably, the step of using attention weights to perform a weighted summation of the probability distribution to obtain the final attention step includes: ; The final attention is concatenated and subjected to a linear transformation to obtain the final output. : ; ; in, , , This represents the appropriate parameters learned by the model during training. It is the final output linear transformation matrix. h For the number of attention heads.
[0012] The multi-head self-attention layer is calculated As input to the second GRU, it is combined with the fully connected layer to output the load prediction results of the computing center.
[0013] Preferably, in the step of collecting relevant data from the computing center, the dataset of relevant data is divided into a training set, a validation set, and a test set according to a preset ratio. The training set is used for training the load prediction model, the validation set is used for debugging hyperparameters during the training process, and the test set is used for final testing of the performance of the load prediction model.
[0014] In a second aspect, the present invention provides a computing center load prediction device, comprising: The data acquisition module is used to collect relevant data from the computing center. The processing module is used to process the relevant data into relevant data. The prediction module is used to input the relevant data into a pre-built load prediction model, and the load prediction model outputs load prediction values. The load forecasting model adopts a GRU-MHSA-GRU three-tier architecture, including: The first-layer gated recurrent unit (GRU) is used to perform primary temporal encoding on the relevant data, capture local temporal features, and obtain feature vectors. The intermediate-layer global multi-head self-attention (MHSA) is used to parse the global dependencies between the feature vectors output by the first-layer GRU, and obtain global features and local temporal features. The final gated recurrent unit (GRU) is used to fuse the global features and local temporal features and output the final hidden state. The fully connected layer is used to perform linear regression on the final hidden state to generate load forecasts.
[0015] In a third aspect, the present invention provides an electronic device including a processor and a memory, the processor being configured to execute a computer program stored in the memory to implement the computing center load prediction method described above.
[0016] In a fourth aspect, the present invention provides a computer-readable storage medium storing at least one instruction that, when executed by a processor, implements the computing center load prediction method.
[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: To address the challenge of feature mining caused by the complex correlations among computing center load-related data, this invention combines MHSA with GRU. GRU can dynamically adjust the information flow using its update and reset gates, accurately capturing the temporal dynamic evolution of signals and achieving temporal memory through hidden state propagation, ensuring the model's sensitivity to local temporal patterns. MHSA can jointly model the global dependencies of input features from different subspaces, significantly enhancing the model's ability to capture key features across time steps in the data and improving the accuracy of computing center load prediction.
[0018] This invention, in terms of load-related data acquisition, not only includes historical load data but also temperature data, irradiance data, humidity data, weekday data, and PUE data, providing a more comprehensive consideration of internal and external factors affecting the load of computing centers. Different standardization methods are employed for load data of different orders of magnitude and units: standard deviation standardization is used for load data; range standardization is used for temperature, irradiance, humidity, and PUE data; and one-hot encoding is used for weekday data. By standardizing data of different dimensions, the efficiency and stability of network model training are greatly improved. Attached Figure Description
[0019] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart of a computing center load prediction method according to Embodiment 1 of the present invention; Figure 2 This is a method for predicting the load of a computing center, as described in Embodiment 1 of the present invention. Figure 3 This is a method for predicting the load of a computing center, as described in Embodiment 1 of the present invention. Figure 4 This is a structural block diagram of a computing center load prediction device according to Embodiment 2 of the present invention; Figure 5 This is a structural block diagram of an electronic device according to Embodiment 3 of the present invention. Detailed Implementation
[0020] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.
[0021] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.
[0022] GRU: Gated Recurrent Unit; MHSA: Multi-Head Self-Attention; PUE: Power Usage Effectiveness, the ratio of all energy consumed by the data center to the energy consumed by the IT load.
[0023] Example 1 A method for predicting the load of a computing center includes the following specific steps: S1. Collect relevant data from the computing center.
[0024] The data related to the computing center includes historical load data, temperature data, irradiance data, humidity data, weekday data, and PUE data, as shown in Table 1.
[0025] Table 1. Data related to the load of the computing center
[0026] To ensure the effectiveness and reliability of training subsequent load prediction models and to reduce the impact of differences in the magnitude of different features, the computing power-related data are standardized.
[0027] The historical load data of the computing center is standardized, and the mean of the data sample set is calculated. and standard deviation The formula is as follows: (1) (2) In the formula, n is the sample size. This represents the load value in a single sample. Based on this, the load data samples are standardized as follows: (3) In the formula, This is the original load sample data. This is the standardized sample data.
[0028] The temperature, irradiance, humidity, and PUE data are standardized using a minimization method, as shown in the following formula: (4) In the formula, yi represents the original data such as temperature. The values are standardized data, y represents the original data vector such as temperature, and max() and min() represent the functions for finding the maximum and minimum values, respectively.
[0029] Whether it is a working day or not is one-hot encoded, as shown in Table 2 below.
[0030] Table 2 Monday to Sunday Codes
[0031] The standardized sample data and the encoded workdays are merged to obtain a merged feature vector; a sliding window sequence is constructed based on the merged feature sequence; sampling is performed based on the sliding window sequence to construct a tensor as the original data, which is then used as input to the load prediction model.
[0032] S2. Input the relevant data into the pre-built load forecasting model, and the load forecasting model outputs the load forecast value.
[0033] Specifically, the load forecasting model adopts a three-tiered architecture of GRU-MHSA-GRU to construct a progressive feature extraction process from local time-series awareness to global dependency modeling.
[0034] The first-layer GRU performs primary temporal encoding on the raw data through a gating mechanism to capture short-term fluctuation features and local temporal dependencies. The middle-layer MHSA analyzes the global association patterns across time steps and uses parallelized attention weights to quantify the intensity of long-distance feature interactions. The final-layer GRU adaptively integrates local detailed features with global context information through a dynamic gating fusion mechanism, outputting a high-order representation with spatiotemporal consistency.
[0035] like Figure 2 As shown, the prediction steps of the load forecasting model are as follows: The relevant data from the computing center is sent to the first-layer GRU for primary temporal encoding to capture local temporal features and obtain feature vectors; The middle-level MHSA analyzes the global dependencies of feature vectors to obtain global features and local temporal features; The obtained global features and local temporal features are input into the last-layer gating unit (GRU). The last-layer gating unit (GRU) fuses the global features and local temporal features and outputs the final hidden state. The final hidden state is input into the fully connected layer, and the load forecast value is obtained after calculation by a linear regression function.
[0036] The MHSA-GRU network model in this invention has a simple structure, can effectively utilize the correlation between important features, shortens the training time, and makes the prediction model converge faster, with more stable prediction performance and higher accuracy.
[0037] Specifically, regarding the feature vector output by the first-layer gated recurrent unit (GRU) X , to feature vector X By mapping to a three-dimensional space through linear transformation, three new vector representations are generated: query vector (Q), key vector (K), and value vector (V). (5) In the formula, , , This represents the appropriate parameters learned by the prediction model during training.
[0038] Query vector Q and key vector K, and calculate the attention score between each pair of elements using the dot product: (6) To prevent the dot product value from becoming too large and causing the gradient vanishing problem, the dot product score is scaled by dividing it by a constant. ,in It is the dimension of the key vector, that is: (7) Applying the Softmax function to the calculated scores transforms them into a probability distribution, representing the weight of each value, i.e.: (8) The values are weighted and summed using attention weights to obtain the final attention: (9) in, It is the number of columns in matrix Q and K, i.e., the vector dimension.
[0039] Multi-head attention mechanisms combine multiple scaled dot-product attention modules. While maintaining the same total number of parameters, the query (Q), key (K), and value (V) parameters are partitioned multiple times. Attention weights are calculated for each partitioned parameter in a different high-dimensional subspace, thus focusing on different parts of the input vector. By processing the computation results of multiple subspaces in parallel, the attention information from all subspaces is finally aggregated. (10) (11) In the formula, , , This represents the appropriate parameters learned by the model during training. It is the linear transformation matrix of the final output, and the calculated final output is... As input to the final layer GRU.
[0040] The final output of the MHSA is processed by the gated recurrent unit (GRU). It captures the temporal correlation between features and outputs the final hidden state.
[0041] The fully connected layer is used to perform linear regression on the final hidden state to generate load forecasts.
[0042] As a specific example of the above embodiments, in the load prediction model, the Units of the first-layer gated recurrent unit (GRU) is set to A, the Head of the middle-layer MHSA is set to B, the key_dim is set to C, the layer weights are initialized using Xavier uniform initialization, the dropout rate is set to D, and the Units of the last-layer gated recurrent unit (GRU) are still set to A. However, the last-layer gated recurrent unit (GRU) only needs to output the final hidden state, which is finally output by the Dense layer through linear regression. The parameter settings are shown in Table 3.
[0043] Table 3 Parameter Settings for GRU-MHSA-GRU Load Forecasting Model
[0044] As a preferred example of the above embodiment, the dataset of relevant data obtained in S1 is divided into a training set, a validation set, and a test set in a ratio of 6:2:2. The training set is used for training the load prediction model, the validation set is used for debugging hyperparameters during the training process, and the test set is used for finally testing the performance of the model.
[0045] During training, the load prediction model is trained using training set data. The hyperparameters in the load prediction model have a significant impact on the final prediction accuracy, therefore continuous adjustment and optimization are necessary. For computing center load prediction, the number of neurons output by the final gated recurrent unit (GRU) is... N ,in N This represents the final number of consecutive predicted load moments. The last fully connected layer uses a linear regression function as the activation function to normalize the output of each neuron by normalizing the likelihood variable. To represent the output of each neuron: (12) in, These are the output features of the final fully connected layer. Then, the classification cross-entropy function is used as the loss function, which can be expressed as: (13) In the formula, For the first i The actual output of each neuron, Predict the output for the target. N To predict the number of consecutive time points.
[0046] For the load forecasting model, adjust the following key hyperparameters: number of gated loop units, number of gated loop unit layers, number of heads in the multi-head attention mechanism, number of layers in the multi-head attention mechanism, Dropout ratio, and selection of different optimizers (such as Adam, RMSprop, etc.).
[0047] As a preferred example of the above embodiments, overfitting is prevented by employing early stopping or other methods during the training process.
[0048] The load prediction models trained with different hyperparameter combinations were evaluated for their prediction accuracy on the validation and test sets. After multiple comparative experiments, the hyperparameter combination with the highest prediction accuracy was selected and the corresponding optimal model was saved. A final test was conducted on the optimal model to obtain its final prediction accuracy, which was then used as the final evaluation metric for this experiment.
[0049] The evaluation metrics include Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percent Error (MAPE). The smaller the corresponding error, the higher the prediction accuracy of the method. Their definitions are as follows: (14) In the formula, NThe number of data samples in the test set, for i The actual load value of the computing center at any given time. for i Forecast value of the computing center load at any time.
[0050] Example 2 like Figure 4 As shown, based on the same inventive concept as the above embodiments, the present invention also provides a computing center load prediction device, comprising: The data acquisition module is used to collect relevant data from the computing center. The processing module is used to process the relevant data into relevant data. The prediction module is used to input the relevant data into a pre-built load prediction model, and the load prediction model outputs load prediction values. The load forecasting model adopts a GRU-MHSA-GRU three-tier architecture, including: The first-layer gated recurrent unit (GRU) is used to perform primary temporal encoding on the relevant data, capture local temporal features, and obtain feature vectors. The intermediate-layer global multi-head self-attention (MHSA) is used to parse the global dependencies between the feature vectors output by the first-layer GRU, and obtain global features and local temporal features. The final gated recurrent unit (GRU) is used to fuse the global features and local temporal features and output the final hidden state. The fully connected layer is used to perform linear regression on the final hidden state to generate load forecasts.
[0051] Example 3 like Figure 5 As shown, the present invention also provides an electronic device 100 for implementing a computing center load prediction method; The electronic device 100 includes a memory 101, at least one processor 102, a computer program 103 stored in the memory 101 and executable on at least one processor 102, and at least one communication bus 104.
[0052] The memory 101 can be used to store the computer program 103. The processor 102 implements the steps of the computing center load prediction method of Embodiment 1 by running or executing the computer program stored in the memory 101 and calling the data stored in the memory 101.
[0053] The memory 101 may primarily include a program storage area and a data storage area. The program storage area may store the operating system, application programs required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created based on the use of the electronic device 100 (such as audio data), etc. In addition, the memory 101 may include non-volatile memory, such as hard disk, RAM, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other non-volatile solid-state storage device.
[0054] At least one processor 102 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Processor 102 may be a microprocessor or any conventional processor. Processor 102 is the control center of electronic device 100, connecting various parts of electronic device 100 via various interfaces and lines.
[0055] The memory 101 in the electronic device 100 stores multiple instructions to implement a computing center load prediction method, and the processor 102 can execute multiple instructions to achieve the following: Collect relevant data from the computing center; The relevant data is processed into relevant data; The relevant data is input into a pre-built load forecasting model, which outputs a load forecast value. The load forecasting model adopts a GRU-MHSA-GRU three-tier architecture, including: The first-layer gated recurrent unit (GRU) is used to perform primary temporal encoding on the relevant data, capture local temporal features, and obtain feature vectors. The intermediate-layer global multi-head self-attention (MHSA) is used to parse the global dependencies between the feature vectors output by the first-layer GRU, and obtain global features and local temporal features. The final gated recurrent unit (GRU) is used to fuse the global features and local temporal features and output the final hidden state. The fully connected layer is used to perform linear regression on the final hidden state to generate load forecasts.
[0056] Example 4 If the modules / units integrated in the electronic device 100 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, and read-only memory (ROM).
[0057] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0058] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0059] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0060] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0061] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.
Claims
1. A method for predicting the load of a computing center, characterized in that, The specific steps include the following: Collect relevant data from the computing center; The relevant data is input into a pre-built load forecasting model, which outputs a load forecast value. The load forecasting model adopts a GRU-MHSA-GRU three-tier architecture, including: The first-layer gated recurrent unit (GRU) is used to perform primary temporal encoding on the relevant data, capture local temporal features, and obtain feature vectors. The intermediate-layer global multi-head self-attention (MHSA) is used to parse the global dependencies between the feature vectors output by the first-layer GRU, and obtain global features and local temporal features. The final gated recurrent unit (GRU) is used to fuse the global features and local temporal features and output the final hidden state. The fully connected layer is used to perform linear regression on the final hidden state to generate load forecasts.
2. The computing center load prediction method as described in claim 1, characterized in that, In the step of collecting relevant data from the computing center, the relevant data includes historical load data, temperature data, irradiance data, humidity data, working day data, and PUE data.
3. The computing center load prediction method as described in claim 2, characterized in that, In the step of processing the relevant data into relevant data, the historical load data, temperature, irradiance, humidity and PUE data of the computing center are standardized, and the weekday type data is one-hot encoded to obtain the raw data, which is then input into the load prediction model.
4. The computing center load prediction method as described in claim 1, characterized in that, The first-layer gated recurrent unit (GRU) is used to perform primary temporal encoding on relevant data, capture local temporal features, and obtain feature vectors. The steps include: The relevant data is input into the first-level gated recurrent unit (GRU), and the first-level gated recurrent unit (GRU) outputs the feature vector. The feature vector is mapped to a three-dimensional space through a linear transformation to generate a query vector Q, a key vector K, and a value vector V. Calculate the attention score between the query vector Q and the key vector K; scale the attention score to obtain the scaled dot product attention score. The scaled dot product attention score is calculated using the Softmax function and then converted into a probability distribution. The final attention is obtained by weighting the probability distribution using attention weights. The final attention is concatenated and subjected to a linear transformation to obtain the final output.
5. The computing center load prediction method as described in claim 4, characterized in that, The calculation query vector Q and key vector K Attention score between; The steps to scale the attention score to obtain the scaled dot product attention score include: Calculate the query vector Q and key vector K Attention score between: ； in, Represents the query vector Q and key vector K Perform dot product calculation between them; Represents the key vector K Transpose of; Divide the attention score by a constant. The scaled dot product attention score is obtained as follows: ； in, It is the dimension of the key vector.
6. The computing center load prediction method as described in claim 5, characterized in that, The step of calculating the scaled dot product attention score using the Softmax function and converting the scaled dot product attention score into a probability distribution includes: 。 7. The computing center load prediction method as described in claim 6, characterized in that, The step of using attention weights to perform a weighted summation of the probability distribution to obtain the final attention step includes: ； The final attention is concatenated and subjected to a linear transformation to obtain the final output. : ； ； in, , , This represents the appropriate parameters learned by the model during training. It is the final output linear transformation matrix. h For the number of attention heads.
8. A load prediction device for a computing center, characterized in that, include: The data acquisition module is used to collect relevant data from the computing center. The prediction module is used to input the relevant data into a pre-built load prediction model, and the load prediction model outputs load prediction values. The load forecasting model adopts a GRU-MHSA-GRU three-tier architecture, including: The first-layer gated recurrent unit (GRU) is used to perform primary temporal encoding on the relevant data, capture local temporal features, and obtain feature vectors. The intermediate-layer global multi-head self-attention (MHSA) is used to parse the global dependencies between the feature vectors output by the first-layer GRU, and obtain global features and local temporal features. The final gated recurrent unit (GRU) is used to fuse the global features and local temporal features and output the final hidden state. The fully connected layer is used to perform linear regression on the final hidden state to generate load forecasts.
9. An electronic device, characterized in that, It includes a processor and a memory, the processor being used to execute a computer program stored in the memory to implement the computing center load prediction method as described in any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one instruction, which, when executed by a processor, implements the computing center load prediction method as described in any one of claims 1 to 7.