A multi-model fusion computing power center load prediction method and related device

Abstract

The application discloses a kind of multi-model fusion's computing power center load prediction method and related device, belong to electric power system load prediction technical field.Method includes: real-time acquisition computing power center current time's load data, temperature data and workday type data;Data is standardized and encoded processing, constructs input feature vector;Input feature vector is respectively input to pre-trained long short-term memory network prediction model, time convolution neural network prediction model and multi-head self-attention-gate recurrent unit prediction model, obtains the load prediction value of future multiple time that three model outputs;Three prediction values are fused and calculated using Yager fusion rule, to obtain final load prediction value, for the energy consumption management or power grid scheduling of computing power center.The application significantly improves the accuracy and robustness of computing power center load prediction by fusing three models with different mechanisms and processing prediction conflicts using Yager rule.

Description

Technical Field

[0001] This invention belongs to the field of power system load forecasting technology, specifically relating to a multi-model fusion computing center load forecasting method and related devices. Background Technology

[0002] With the increasing popularity of cloud services, the iterative advancement of next-generation artificial intelligence technologies, and the rapid growth in data storage demand, the computing center industry has entered a stage of rapid development. Over the past decade, major global technology companies such as Google and Oracle have continuously increased their investment in the construction and expansion of computing center facilities worldwide, driving the continuous expansion of the scale and number of computing centers.

[0003] The rapid development of the computing center industry has been accompanied by a significant increase in its electricity demand. This continuously growing demand poses numerous challenges to the existing power grid system, raising widespread concerns about its ability to stably support such new high-power-density loads. At the same time, increasingly stringent global environmental governance requirements are driving the computing center industry towards modular design and the application of renewable energy.

[0004] Computing centers are prone to large and transient power fluctuations during cold starts, shutdowns, load transfers, or sudden interruptions. Since large computing centers typically consume tens to hundreds of megawatts of power, and their power fluctuations can reach several megawatts within seconds, these fluctuations directly impact the grid's frequency control, thus placing higher demands on the grid's frequency regulation response speed. Furthermore, computing centers operate with numerous power electronic devices, such as uninterruptible power supplies, control switches, and rectifiers. These nonlinear components introduce harmonics into the grid during transient events, leading to a decline in power quality and further threatening the stability and reliability of grid operation. Therefore, accurate prediction of computing center energy consumption has become an indispensable prerequisite for effectively improving grid reliability and achieving effective management of the computing center's power load.

[0005] Power consumption prediction for computing centers is a complex and critical technical task. The accuracy of the prediction results depends on a variety of factors, including the specifications of the equipment installed in the computing center, the type of workload being run, and the actual operating conditions. The inherent diversity and volatility of power consumption data from computing centers, including numerous sudden power drops and peaks, further increase the difficulty and complexity of power consumption prediction.

[0006] Currently, existing power consumption prediction technologies for computing centers have many shortcomings and deficiencies, making it difficult to meet practical application needs. On the one hand, existing research either fails to fully consider external factors affecting changes in the power load of computing centers, resulting in low prediction accuracy; or it simply analyzes the correlation between the overall load of the computing center and external factors without deeply exploring the inherent characteristics of the load composition of the computing center. In fact, the load composition of computing centers is unique, making the factors affecting their power load extremely complex. As the main source of energy consumption in computing centers, IT load is directly affected by internal operating parameters such as CPU utilization, memory utilization, disk throughput, and network throughput. High-performance servers in operation generate a large amount of heat, causing equipment temperatures to rise. Therefore, most computing centers are equipped with cooling systems to maintain normal equipment operation. The power consumption of cooling load is not only affected by IT workload but also significantly influenced by external environmental factors such as ambient temperature. These various influencing factors interact and are coupled with each other, posing a significant challenge to the construction of input features for the prediction model.

[0007] On the other hand, among existing forecasting methods, traditional time series forecasting models such as the autoregressive integral moving average model are difficult to effectively handle complex nonlinear patterns in high-dimensional data and cannot fully adapt to the high volatility and strong mutation characteristics of power consumption data in computing centers. As a result, the accuracy of current computing center load forecasting is generally not high, which makes it difficult to meet the actual engineering needs of stable power grid operation and refined energy consumption management of computing centers. Summary of the Invention

[0008] To address the problems of low accuracy of single prediction models and insufficient prediction accuracy due to the lack of effective integration of multi-source heterogeneous influencing factors in existing computing center load prediction methods, this invention provides a multi-model fusion computing center load prediction method and related apparatus, aiming to improve the accuracy and robustness of prediction.

[0009] This invention is achieved through the following technical solution: This invention discloses a multi-model fusion method for predicting the load of computing centers, comprising the following steps: The load impact data of the computing center at the current moment can be obtained in real time or at a preset period. The load impact data includes at least the load data, temperature data and workday type data corresponding to the current date at the current moment. The obtained load impact data is standardized, including: normalizing the load data and temperature data using maximum and minimum values, and performing one-hot encoding on the workday type data to obtain the standardized input feature vector. The input feature vector is input into a pre-trained long short-term memory network prediction model, a temporal convolutional neural network prediction model, and a multi-head self-attention-gated recurrent unit prediction model, respectively, to obtain the load prediction values ​​for multiple consecutive future time periods output by the three prediction models; The load forecast values ​​output by the three forecast models are fused and calculated using the Yager fusion rule to obtain the final load forecast value, which is used for energy consumption management or power grid dispatching in the computing center.

[0010] Furthermore, the load data at the current moment is the actual total power consumption of the computing center at the current sampling moment; the temperature data is the ambient temperature of the environment where the computing center is located or the air intake temperature of the server rack; the workday type data includes the category identifiers from Monday to Sunday and holiday identifiers.

[0011] Furthermore, the output calculation process of the Long Short-Term Memory Network prediction model is as follows: the cell state is updated through the forget gate, input gate, and output gate, and the hidden state of the last time step is mapped to the load prediction value through a fully connected layer.

[0012] Furthermore, the output calculation process of the temporal convolutional neural network prediction model is as follows: temporal features are extracted through dilated convolution and residual structure, and then the load prediction value is output through a fully connected layer.

[0013] Furthermore, the multi-head self-attention-gated recurrent unit prediction model includes: The input layer is used to receive the input feature vector; A gated recurrent unit layer, with its input connected to the input layer, is used to extract temporal dependent features from the input feature vector at each time step and output the hidden state at each time step. The attention weighting module, whose input is connected to the output of the gated recurrent unit layer, is used to calculate the attention score of the hidden state at each time step, obtain the attention weight by normalization through the Softmax function, and use the attention weight to perform weighted summation on the hidden state to output a comprehensive representation vector. A fully connected layer, with its input connected to the output of the attention weighting module, is used to map the comprehensive representation vector into a load prediction value; The attention weighting module contains multiple parallel attention heads, forming a multi-head self-attention mechanism. Each attention head independently calculates a weighting vector, and the outputs of all attention heads are concatenated or summed before being input into the fully connected layer.

[0014] Furthermore, the attention weighting module calculates the attention score for the hidden state at each time step, expressed as: ; In the formula, As a learnable parameter, W and b These are learnable weights and biases; Attention score; Hide the current state; The attention weights are used to perform a weighted summation of the hidden states, outputting a comprehensive representation vector; the specific expression is: ; In the formula, c This is the comprehensive representation vector after being focused by the attention mechanism; Attention weights; It is in a hidden state.

[0015] Furthermore, let the predicted load values ​​for the Long Short-Term Memory Network (LSTM) prediction model, the Temporal Convolutional Neural Network (TCNN) prediction model, and the Multi-Head Self-Attention Gated Recurrent Unit (MLUG) prediction model at the same time be respectively... P 1. P 2. P 3, the corresponding prediction confidence levels are respectively w 1. w 2. w 3, and w 1+ w 2+ w 3 = 1; The load forecast values ​​output by the three forecast models are fused using the Yager fusion rule to obtain the final load forecast value, specifically: The Yager fusion rule first calculates a weighted average as the initial fusion result, and then adjusts the fusion weights based on the consistency between the models: if the prediction result of a certain model deviates from the results of the other two models by more than a preset threshold, the prediction confidence of that model is reduced; the final load forecast is calculated as follows:

[0016] in, This is the adjusted prediction confidence level; This is the final load forecast value; For the first j Load forecast values ​​for each model.

[0017] This invention also discloses a multi-model fusion computing center load prediction system, comprising: The data acquisition module is used to acquire the load impact data of the computing center at the current moment in real time or at a preset period. The load impact data includes at least the load data, temperature data and workday type data corresponding to the current date at the current moment. The data preprocessing module performs standardization processing on the acquired load impact data. The standardization processing includes: normalizing the load data and temperature data using maximum and minimum values, and performing one-hot encoding on the workday type data to obtain the standardized input feature vector. The multi-model prediction module is internally deployed with a pre-trained long short-term memory network prediction model, a temporal convolutional neural network prediction model, and a multi-head self-attention-gated recurrent unit prediction model. It is used to input the input feature vector into the pre-trained long short-term memory network prediction model, the temporal convolutional neural network prediction model, and the multi-head self-attention-gated recurrent unit prediction model respectively, and obtain the load prediction values ​​of multiple consecutive future time periods output by the three prediction models respectively. The fusion module uses the Yager fusion rule to fuse the load forecast values ​​output by the three prediction models to obtain the final load forecast value. The output module is used to output the final load prediction value for energy management or power grid dispatching in the computing center.

[0018] The present invention also discloses a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the multi-model fusion computing center load prediction method described above.

[0019] The present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the multi-model fusion computing center load prediction method.

[0020] Compared with the prior art, the present invention has the following beneficial technical effects: This invention provides a multi-model fusion method for predicting the load of computing centers. Firstly, the data selection includes not only load data but also data such as temperature and working days. Secondly, it constructs an LSTM prediction model adept at mining time-series features, a TCN prediction model integrating joint time and spatial features, and an MHSA-GRU prediction model adept at extracting long- and short-range dependencies. The TCN and MHSA-GRU prediction models are the first to achieve computing center load prediction. Regarding the fusion of multi-model prediction results, the Yager fusion rule is used to fuse multiple prediction results, avoiding the shortcomings of single-model prediction and effectively improving the accuracy of load prediction. By fusing three models with different mechanisms, it can better handle the highly volatile and multi-coupled characteristics of computing center loads than using a single model. The Yager rule is not a simple averaging; it automatically reduces the weight of unreliable models when model prediction results conflict, resulting in higher prediction accuracy and stability after fusion. The input includes external factors such as temperature and working days, making it more closely aligned with actual operating scenarios than predicting only load.

[0021] Furthermore, the system specifies that the load data at the current moment is the actual total power consumption, the temperature is either ambient temperature or rack intake air temperature, and the workday type includes Monday to Sunday and public holidays. This specification has a very practical effect: it clarifies the specific form of the data source and avoids ambiguous interpretations. The actual power value directly reflects the actual power consumption of the computing center, rack intake air temperature has a greater impact on cooling load than ambient temperature, and the public holiday label can capture load fluctuations caused by changes in working modes.

[0022] Furthermore, the internal architecture of the MHSA-GRU model is described in detail: input layer, GRU layer, attention weighting module, and fully connected layer. The attention weighting module is constrained to consist of multiple parallel heads forming a multi-head self-attention mechanism. The GRU first extracts the hidden states at each time step, and then the attention mechanism reweights these states, allowing the model to focus more on the time points that are truly important for prediction, rather than mechanically averaging all time steps. The multi-head mechanism can also capture different dependencies from different subspaces, finally synthesizing them into a single vector for the fully connected output.

[0023] Furthermore, the attention weighting module is limited to calculating the attention score. This non-linear transformation, combined with tanh activation, allows the model to adaptively learn the importance weights of each hidden state, rather than manually setting them. The technical effect is that the attention score is entirely data-driven; when load characteristics change across different computing centers or seasons, the model can automatically adjust its focus without manual parameter tuning.

[0024] Furthermore, the entire hidden state sequence output by the GRU is compressed into a fixed-length vector. However, this vector is not simply averaged, but aggregated according to the learned weights. Therefore, information from key time steps is preserved, while secondary information is suppressed. This results in higher quality data input to the fully connected layer, leading to more accurate predictions. Attached Figure Description

[0025] Figure 1 To build a multi-model fusion computing center load prediction framework; Figure 2 This is a flowchart of a multi-model fusion computing center load prediction method according to the present invention; Figure 3 This is a schematic diagram of a multi-model fusion computing center load prediction system according to the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the present invention clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention; that is, the described embodiments are only a part of the embodiments of the present invention, and not all of them.

[0027] The detailed description of the embodiments of the present invention provided in the following drawings is not intended to limit the scope of the claimed invention, but merely to illustrate one selected embodiment of the invention. All other embodiments obtained by those skilled in the art based on the drawings and embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0028] The features and performance of the present invention will be further described in detail below with reference to embodiments.

[0029] like Figure 1 As shown, this invention first introduces the computing center load prediction framework from a design perspective, mainly including the selection and standardization of computing center load-related data, the construction and training of multiple load prediction models, model verification, testing, analysis of multi-model prediction results, and Yager fusion. The specific details are as follows: 1. Obtain relevant data on the computing center load. The data related to the computing center load involves multiple dimensions, covering key indicators such as historical load, temperature, power consumption, and equipment utilization. Considering the feasibility of obtaining actual data, we focus on historical load, temperature, and date type as the key analysis objects.

[0030] 2. Data standardization processing Considering the different attributes and magnitudes of historical load data, temperature data, and workday type data from the computing center, in order to ensure the reliability and stability of subsequent load prediction model training, it is necessary to standardize various types of data to obtain standardized input feature vectors, which are then integrated to obtain the dataset.

[0031] For historical load data, a maximum-minimum method is used for standardization, as follows: (1) In the formula, x i The first in the load sample vector i One element; x For the load sample vector; max( x ) and min ( x ) represent the load sample vectors respectively. x The maximum and minimum values ​​of the elements in the middle; The standardized load value is located between [0,1] and is used as the input feature vector.

[0032] The temperature data from the computing center is also standardized using the maximum-minimum method, and the calculation formula is as follows: (2) In the formula, y i The first element in the temperature sample vector i One element, y Given a temperature sample vector, max( y ) and min ( y ) represent taking temperature sample vectors respectively y The maximum and minimum values ​​of the elements in the middle. The standardized temperature value ranges from [0,1].

[0033] Finally, the weekday data is processed using one-hot encoding, as shown in Table 1 below.

[0034] Table 1. Workday Type Data Encoding

[0035] 3. Dataset partitioning The dataset is divided into training, validation, and test sets according to a certain ratio. The training set is used for training subsequent multi-prediction models, the validation set is used for fine-tuning the hyperparameters of the trained models, and the test set is used for the final testing of the model's performance.

[0036] 4. Construction and training of multiple prediction models 4.1 Construction and Training of the Long Short-Term Memory Network Prediction Model (LSTM Prediction Model) For input sample data d t =[x t ,y t ,z t ], x t Indicates the current t Historical load data at any time, y t Indicates the current t Temperature data at any time, z t Indicates the current t The time and date are encoded, and the forget gate is calculated first. The forget gate determines the state of cell C. t-1 How much information is lost, i.e.: (3) In the formula, f t It is the output of the forget gate. For activation functions such as the Sigmoid function, W f and b f These are the weight matrix and bias terms of the forget gate. This is the hidden state from the previous moment. d t It is the input at the current moment.

[0037] The input gate determines which information is added to the cell state, and its calculation formula is as follows: (4) In the formula, i t It is the output of the input gate. W i and b i These are the weights and biases of the input gate. A candidate value vector is then constructed based on these. The calculation formula is as follows: (5) Based on the forget gate and the input gate, the cell state is updated as follows: (6) In the formula, C t This represents the current state of the cell. C t-1 for t Cell state at time -1 This indicates element-wise multiplication.

[0038] Furthermore, the output based on the current cell state is represented as: (7) In the formula, This indicates the output of the output gate. W o and b o These are the weights and biases of the output gate.

[0039] Based on the output of the output gate and the current cell state, the current hidden state is calculated as follows: (8) Current hidden state Typically used as the output of an LSTM layer, it can be directly used as the input to the next hidden layer or ultimately mapped through a fully connected layer to obtain the predicted output.

[0040] During training, the LSTM prediction model is trained using the training set data. Let n be the number of neurons in the final network output (representing the number of consecutive prediction time steps). The last fully connected layer uses a linear regression function as the activation function to normalize the output of each neuron. The likelihood variable Y=[Y1,…,Y] is normalized. n ] represents the actual output of each neuron, i.e.: (9) In the formula, h f It is the output feature of the fully connected layer.

[0041] Using the classification cross-entropy function as the loss function, its form can be expressed as: (10) In the formula, Y k This is the actual output of the neuron. Predict the output for the target. n This represents the number of consecutive prediction time points. The network's key parameters, such as weights and biases, are updated using the backpropagation algorithm. Simultaneously, the network parameters are fine-tuned using the validation set, and early stopping techniques are employed to prevent overfitting.

[0042] 4.2 Construction and Training of the Temporal Convolutional Neural Network Prediction Model (TCN Prediction Model) For input sample data, dilated convolution in a temporal convolutional neural network allows the input of the convolution kernel to be interval-sampled, with the interval size determined by the dilation coefficient. Its convolution formula is: (11) In the formula, m The kernel size; w The coefficient of thermal expansion; ds-wi For the sample data used in the calculation, s This is the current step size position. wi For expansion intervals; f ( ) represents a filter; This refers to the feature mapping quantity.

[0043] If it is the first layer, the sample data used for calculation is the input sample data; then the output of the previous layer is used as the sample data used for calculation in the next layer.

[0044] The output of a temporal convolutional neural network containing residual structures can be calculated as follows: (12) In the formula, For activation function, Z k This is the output of a time-convolutional neural network containing residual structures. k=1,2....n; b k For bias terms; W 1 represents the weight matrix of the convolutional layer; W 2 is the weight matrix of the residual layer; d This is the input feature vector.

[0045] During training, the network output uses a fully connected layer, and its loss function can be designed with reference to equations (9)-(10). The network's key parameters, such as kernel size, weights, and bias values, are optimized by combining the backpropagation algorithm. At the same time, the network is fine-tuned using the validation set.

[0046] 4.3 Construction and Training of the Multi-Head Self-Attention Gated Recurrent Unit Prediction Model (MHSA-GRU Prediction Model) The multi-head self-attention-gated recurrent unit prediction model includes: The input layer is used to receive the standardized input feature vector; The gated recurrent unit (GRU) layer, with its input connected to the input layer, is used to extract temporal dependent features from the input feature vector at each time step and output the hidden state at each time step. The attention weighting module, whose input is connected to the output of the gated recurrent unit layer, is used to calculate the attention score of the hidden state at each time step, obtain the attention weight by normalization through the Softmax function, and use the attention weight to perform weighted summation on the hidden state to output a comprehensive representation vector. A fully connected layer, with its input connected to the output of the attention weighting module, is used to map the comprehensive representation vector into a load prediction value; The attention weighting module contains multiple parallel attention heads, forming a multi-head self-attention mechanism. Each attention head independently calculates a weighting vector, and the outputs of all attention heads are concatenated or summed before being input into the fully connected layer.

[0047] The MHSA-GRU prediction model combines the information-focusing ability of multi-head self-attention with the modeling ability of GRU for time-series data. Its core lies in using the attention mechanism to assign different weights to the hidden states output by the GRU at each time step, thus allowing the model to focus more on the parts of the sequence that are more relevant to the current task. The specific calculation process is as follows: (1) Input sample data is fed into the GRU layer step by step in time step, and processed step by step in time order. At each time step t Generate a hidden state at the current moment. .

[0048] (2) For each time step in the hidden state t Calculate an attention score a t It measures the hidden state at the current moment. h t The calculation process for the importance of the entire feature vector time series is as follows: (13) In the formula, As a learnable parameter, W and b These are the learnable weights and biases.

[0049] (3) Attention scores for all time steps Attention weights are obtained by normalization using the Softmax function. The sum of all of them is 1, that is: (14) (4) Utilizing attention weights We perform a weighted summation of the hidden states of the GRU across all time steps, i.e.: (15) In the formula, c This is the comprehensive representation vector after focusing through the attention mechanism.

[0050] (5) The integrated representation vector after focusing through the attention mechanism c The input is fed into a fully connected layer, and the final output is the load forecast values ​​for multiple consecutive future time periods. The training process of the network model can be found in the LSTM training section. The network parameters are updated by training the model using the training set, and the parameters are fine-tuned using the validation set.

[0051] 5. Model Testing and Performance Evaluation For each load forecasting model, its prediction accuracy on the test set is evaluated. The evaluation metrics mainly 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: (16) In the formula, N is the number of data samples in the test set. for i The actual load value of the computing center at any given time. for i The computing center predicts the load value at any given time.

[0052] 6. Fusion of prediction results from multiple models Based on the prediction results of LSTM, TCN and MHSA-GRU load forecasting models, Yager fusion rules are used to fuse the multiple load forecasting results to achieve comprehensive load forecasting for computing centers.

[0053] Let the load forecasts of the LSTM, TCN, and MHSA-GRU models at the same time be respectively... P 1. P 2. P 3. The corresponding initial prediction confidence levels (which can be set based on the historical prediction accuracy on the validation set) are as follows: w 1. w 2. w 3, and w 1+ w 2+ w 3 = 1.

[0054] Yager's fusion rule first calculates a weighted average as the initial fusion result, and then adjusts the fusion weights based on the consistency between the models: if the prediction result of one model deviates from the results of the other two models by more than a preset threshold, the prediction confidence of that model is reduced. The final fusion result is:

[0055] in, This is the adjusted prediction confidence level; This is the final load forecast value; For the first j The load forecast values ​​for each model. The adjusted forecast confidence level remains 1.

[0056] The evaluation metrics for each model are shown in Table 2: Table 2 Evaluation metrics for each model

[0057] As can be seen from Table 2, the prediction results after fusion using the Yager fusion rule are significantly better than any single model in terms of RMSE, MAE, and MAPE, thus verifying the effectiveness of this scheme.

[0058] When applying, such as Figure 2 As shown, the multi-model fusion method for predicting the load of a computing center according to the present invention includes the following steps: The load impact data of the computing center at the current moment can be obtained in real time or at a preset period. The load impact data includes at least the load data, temperature data and workday type data corresponding to the current date at the current moment. The obtained load impact data is standardized, including: normalizing the load data and temperature data using maximum and minimum values, and performing one-hot encoding on the workday type data to obtain the standardized input feature vector. The input feature vector is input into a pre-trained long short-term memory network prediction model, a temporal convolutional neural network prediction model, and a multi-head self-attention-gated recurrent unit prediction model, respectively, to obtain the load prediction values ​​for multiple consecutive future time periods output by the three prediction models; The load forecast values ​​output by the three forecast models are fused and calculated using the Yager fusion rule to obtain the final load forecast value, which is used for energy consumption management or power grid dispatching in the computing center.

[0059] like Figure 3 As shown, this invention also discloses a multi-model fusion computing center load prediction system, comprising: The data acquisition module is used to acquire the load impact data of the computing center at the current moment in real time or at a preset period. The load impact data includes at least the load data, temperature data and workday type data corresponding to the current date at the current moment. The data preprocessing module performs standardization processing on the acquired load impact data. The standardization processing includes: normalizing the load data and temperature data using maximum and minimum values, and performing one-hot encoding on the workday type data to obtain the standardized input feature vector. The multi-model prediction module is internally deployed with a pre-trained long short-term memory network prediction model, a temporal convolutional neural network prediction model, and a multi-head self-attention-gated recurrent unit prediction model. It is used to input the input feature vector into the pre-trained long short-term memory network prediction model, the temporal convolutional neural network prediction model, and the multi-head self-attention-gated recurrent unit prediction model respectively, and obtain the load prediction values ​​of multiple consecutive future time periods output by the three prediction models respectively. The fusion module uses the Yager fusion rule to fuse the load forecast values ​​output by the three prediction models to obtain the final load forecast value. The output module is used to output the final load prediction value for energy management or power grid dispatching in the computing center.

[0060] This invention also discloses a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the multi-model fusion computing center load prediction method. The memory may include main memory, such as high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device. The processor, network interface, and memory are interconnected via an internal bus, which may be an industry-standard architecture bus, a peripheral component interconnection standard bus, an extended industry-standard architecture bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. The memory stores the program; specifically, the program may include program code, which includes computer operation instructions. The memory may include main memory and non-volatile memory, and provides instructions and data to the processor.

[0061] This invention also discloses a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the steps of the multi-model fusion computing center load prediction method. Specifically, the computer-readable storage medium includes, but is not limited to, volatile memory and / or non-volatile memory. The volatile memory may include random access memory and / or cache memory, etc. The non-volatile memory may include read-only memory, hard disk, flash memory, optical disk, magnetic disk, etc.

[0062] 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, optical storage, etc.) containing computer-usable program code.

[0063] 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.

[0064] 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.

[0065] 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.

[0066] 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 protection scope of the present invention.

Claims

1. A multi-model fusion method for predicting the load of a computing center, characterized in that, Includes the following steps: The load impact data of the computing center at the current moment can be obtained in real time or at a preset period. The load impact data includes at least the load data, temperature data and workday type data corresponding to the current date at the current moment. The obtained load impact data is standardized, including: normalizing the load data and temperature data using maximum and minimum values, and performing one-hot encoding on the workday type data to obtain the standardized input feature vector. The input feature vector is input into a pre-trained long short-term memory network prediction model, a temporal convolutional neural network prediction model, and a multi-head self-attention-gated recurrent unit prediction model, respectively, to obtain the load prediction values ​​for multiple consecutive future time periods output by the three prediction models; The load forecast values ​​output by the three forecast models are fused and calculated using the Yager fusion rule to obtain the final load forecast value, which is used for energy consumption management or power grid dispatching in the computing center.

2. The multi-model fusion method for predicting computing center load according to claim 1, characterized in that, The load data at the current moment is the actual total power consumption of the computing center at the current sampling time; the temperature data is the ambient temperature of the environment where the computing center is located or the air intake temperature of the server rack; the workday type data includes the category identifiers from Monday to Sunday and holiday identifiers.

3. The multi-model fusion method for predicting computing center load according to claim 1, characterized in that, The output calculation process of the Long Short-Term Memory Network prediction model is as follows: the cell state is updated through the forget gate, input gate and output gate, and the hidden state of the last time step is mapped to the load prediction value through a fully connected layer.

4. The multi-model fusion method for predicting computing center load according to claim 1, characterized in that, The output calculation process of the temporal convolutional neural network prediction model is as follows: extract temporal features through dilated convolution and residual structure, and then output load prediction values ​​through a fully connected layer.

5. The multi-model fusion method for predicting computing center load according to claim 1, characterized in that, The multi-head self-attention-gated recurrent unit prediction model includes: The input layer is used to receive the input feature vector; A gated recurrent unit layer, with its input connected to the input layer, is used to extract temporal dependent features from the input feature vector at each time step and output the hidden state at each time step. The attention weighting module, whose input is connected to the output of the gated recurrent unit layer, is used to calculate the attention score of the hidden state at each time step, obtain the attention weight by normalization through the Softmax function, and use the attention weight to perform weighted summation on the hidden state to output a comprehensive representation vector. A fully connected layer, with its input connected to the output of the attention weighting module, is used to map the comprehensive representation vector into a load prediction value; The attention weighting module contains multiple parallel attention heads, forming a multi-head self-attention mechanism. Each attention head independently calculates a weighting vector, and the outputs of all attention heads are concatenated or summed before being input into the fully connected layer.

6. The multi-model fusion method for predicting computing center load according to claim 5, characterized in that, The attention weighting module calculates the attention score for the hidden state at each time step, expressed as: ； In the formula, As a learnable parameter, W and b These are learnable weights and biases; Attention score; Hide the current state; The attention weights are used to perform a weighted summation of the hidden states, outputting a comprehensive representation vector; the specific expression is: ； In the formula, c This is the comprehensive representation vector after being focused by the attention mechanism; Attention weights; It is in a hidden state.

7. The multi-model fusion method for predicting the load of a computing center according to claim 1, characterized in that, Let the predicted load values ​​at the same time be as follows: Long Short-Term Memory Network prediction model, Temporal Convolutional Neural Network prediction model, and Multi-Head Self-Attention Gated Recurrent Unit prediction model. P 1. P 2. P 3, the corresponding prediction confidence levels are respectively w 1. w 2. w 3, and w 1+ w 2+ w 3 = 1; The load forecast values ​​output by the three forecast models are fused using the Yager fusion rule to obtain the final load forecast value, specifically: The Yager fusion rule first calculates a weighted average as the initial fusion result, and then adjusts the fusion weights based on the consistency between the models: if the prediction result of a certain model deviates from the results of the other two models by more than a preset threshold, the prediction confidence of that model is reduced; the final load forecast is calculated as follows: in, This is the adjusted prediction confidence level; This is the final load forecast value; For the first j Load forecast values ​​for each model.

8. A multi-model fusion computing center load forecasting system, characterized in that, The system is used to perform the method of any one of claims 1 to 7, comprising: The data acquisition module is used to acquire the load impact data of the computing center at the current moment in real time or at a preset period. The load impact data includes at least the load data, temperature data and workday type data corresponding to the current date at the current moment. The data preprocessing module performs standardization processing on the acquired load impact data. The standardization processing includes: normalizing the load data and temperature data using maximum and minimum values, and performing one-hot encoding on the workday type data to obtain the standardized input feature vector. The multi-model prediction module is internally deployed with a pre-trained long short-term memory network prediction model, a temporal convolutional neural network prediction model, and a multi-head self-attention-gated recurrent unit prediction model. It is used to input the input feature vector into the pre-trained long short-term memory network prediction model, the temporal convolutional neural network prediction model, and the multi-head self-attention-gated recurrent unit prediction model respectively, and obtain the load prediction values ​​of multiple consecutive future time periods output by the three prediction models respectively. The fusion module uses the Yager fusion rule to fuse the load forecast values ​​output by the three prediction models to obtain the final load forecast value. The output module is used to output the final load prediction value for energy management or power grid dispatching in the computing center.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the multi-model fusion computing center load prediction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the multi-model fusion computing center load prediction method as described in any one of claims 1 to 7.

Images