Demand prediction method and system based on computing power operation load of data center

Abstract

The invention provides a demand prediction method and system based on computing power operation load of a data center, and the method comprises the steps: carrying out the time sequence coding of historical load data of each computing power node through employing a long-short term memory (LSTM) network, and obtaining a time sequence feature vector; taking each computing power node as a node of an interaction graph, taking a dynamic adjacency matrix calculated through a graph learning layer as an edge of the interaction graph, and updating node features through graph convolution operation; a multi-head self-attention mechanism is introduced to carry out global information integration on node features output by a GCN, cross-regional global computing power interaction information is captured by calculating attention weights of all computing power node features, the features output by the multi-head self-attention mechanism are input into a residual decoder, and after residual connection processing is carried out, global computing power interaction information is obtained. And outputting a future computing power operation load prediction value and a prediction confidence interval of the data center. According to the method, the long-period trend and the short-term abrupt change characteristics of the computing power operation load of the data center are captured, and the accuracy of demand prediction of the computing power operation load is improved.

Description

Technical Field

[0001] This invention belongs to the field of data center computing power management and load forecasting technology, and particularly relates to a method and system for demand forecasting based on data center computing power operation load. Background Technology

[0002] As the main carrier of computing infrastructure, data centers integrate diverse computing nodes such as servers, racks, computing clusters, and AI computing pools. Short-term demand forecasting of their computing power operation load is a core link in achieving optimized scheduling of computing resources, improving computing power utilization efficiency, and energy consumption management. With the rapid development of cloud computing, big data, and artificial intelligence technologies, the coupling relationship between computing nodes in data centers is becoming increasingly close. The computing load is not only affected by the periodicity of business hours and task scheduling, but is also prone to non-periodic sudden changes due to factors such as sudden business requests, changes in computing scheduling strategies, and hardware failures.

[0003] Currently, most computing load prediction technologies are based on single time-series models or simple spatial correlation modeling, which are insufficient in capturing the temporal characteristics of computing load and cannot effectively integrate the correlation between the operating status of computing equipment and the load sequence. This results in limited accuracy in extracting long-term computing load trends and short-term abrupt changes. At the same time, they lack the ability to capture global correlation information such as cross-regional interactions and task scheduling dependencies between computing nodes, making it difficult to meet the reliability requirements of fine-grained computing load scheduling in data centers. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for demand forecasting based on data center computing power operation load, in order to solve the technical problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution.

[0006] According to an embodiment of the present invention, a demand forecasting method based on data center computing power operation load is provided, comprising the following steps: Collect multi-source data from the data center, including load data of computing nodes, equipment operating parameters, spatial location information, and node operating status parameters; The historical load data of each computing node is temporally encoded using a Long Short-Term Memory (LSTM) network to obtain a temporal feature vector that characterizes the time-series dependence of computing load. Each computing power node is used as a node in the interaction graph, and the dynamic adjacency matrix calculated through the graph learning layer is used as the edge of the interaction graph. The temporal feature vector and interaction graph are used as inputs to the Graph Convolutional Neural Network (GCN), and node features are updated through graph convolution operations. The local interaction layer aggregates the local direct interaction features of each computing node, while the global fusion layer fuses global indirect association features through residual connections. A multi-head self-attention mechanism is introduced to integrate global information from the node features output by the GCN. By calculating the attention weights of each computing node's features, global computing power interaction information across regions is captured. The features output by the multi-head self-attention mechanism are then input into the residual decoder, which includes a fully connected layer and a normalization layer. After processing through residual connections, the predicted future computing power operation load of the data center and the prediction confidence interval are output.

[0007] Furthermore, the step of using a Long Short-Time Memory (LSTM) network to perform time-series encoding of the historical load data of each computing node includes: Acquire historical load data and corresponding time period equipment operation status parameters for each computing node. The equipment operation status parameters include server start / stop status, real-time computing power conversion efficiency, and task scheduling strategy parameters. The computing load data and equipment operating status parameters are concatenated into a multi-source sequence; An LSTM network with integrated computing load change awareness is constructed. The preprocessed multi-source sequences are used as the input of the LSTM network, and the output is a temporal feature vector.

[0008] Furthermore, in constructing an LSTM network that integrates computing load abrupt change awareness: The bottom layer is a network structure that uses gated recurrent units (GRU) and LSTM in parallel. GRU is used to capture short-term computing load fluctuations, while LSTM is used to capture long-term computing power trends. The preprocessed multi-source sequences are used as the input to the bottom layer, and the output is a mixed time series feature. A computing load mutation detection module is introduced in the middle layer. The computing load mutation detection module dynamically weights the mixed time-series features to obtain a weighted mixed feature sequence. The top layer is a periodic attention layer, which performs matrix multiplication of the mixed feature sequence with a preset periodic mask, calculates the attention weight of each time step through attention, and performs weighted aggregation of the feature sequence; the aggregated features output by the periodic attention layer are reduced in dimensionality through a fully connected layer to obtain a temporal feature vector, which integrates long-term computing power trends, short-term mutation features and device status correlation information.

[0009] Furthermore, the step of dynamically weighting the mixed time-series features in the computing load mutation detection module includes: Construct a causal graph of sudden changes in computing power task status and computing power load, and mine the temporal correlation rules between sudden tasks, scheduling strategy changes and sudden changes in computing power load based on historical data, and quantify the rules into a causal coefficient matrix; The quantiles of historical computing load fluctuations during the same period are statistically analyzed using a sliding window and used as the adaptive threshold at time t. The mutation index is calculated based on causal mapping and adaptive thresholding, where the basic mutation index is expressed as: In the formula, This represents the adaptive threshold at time t. This represents the normalized computing load value at time t. This represents the normalized computing load value at time t-1. It represents the absolute change in computing load between adjacent time steps; A causal correction index is introduced to strengthen the causal relationship between computing task status and sudden changes in computing load, denoted as: in, This represents the normalized value of the state parameter of the k-th type of computing power device at time t. Represents the elements of the causal coefficient matrix; A nonlinear weighting function is constructed, and the nonlinear weights are multiplied element-wise with the mixed temporal features. At the same time, enhancement coefficients are applied to the feature dimensions with correlation strength exceeding the threshold in the mutation causal graph based on the causal coefficient matrix, resulting in a weighted mixed feature sequence.

[0010] Furthermore, the step of using the dynamic adjacency matrix computed through the graph learning layer as the edges of the interaction graph includes: In the graph learning layer, the two fully connected layers of the graph learning layer are used to map the features of the computing nodes into embedding vectors, and the asymmetric dependencies of the embedding vectors are calculated to generate the initial adjacency matrix. A state-driven mechanism is introduced to adjust the initial adjacency matrix based on the computing power node's running state parameters. The weighted sum of the average computing power utilization rate and the task execution efficiency coordination coefficient is used as the edge weight, while retaining the top-k strong dependencies and filtering weak correlation noise.

[0011] Furthermore, in the graph learning layer: Extract multidimensional features from each computing node, including the mean and standard deviation of computing load data, rated computing power of equipment, real-time operating efficiency, and computing network topology coordinates. Calculate the Euclidean distance between node i and node j. Multidimensional features are concatenated into node feature vectors of dimension d; Construct a first fully connected network, which uses the ReLU activation function to map the node feature vectors from d dimensions to 64-dimensional hidden feature vectors; A second fully connected network is constructed, which uses the LeakyReLU activation function to map the 64-dimensional hidden feature vector to a 32-dimensional embedding vector. Calculate the asymmetric dependencies of the embedded vectors using the parameter matrix. Perform a linear transformation on the embedding vector and calculate the dependency weights between node i and node j, expressed as: In the formula, Let represent the elements of the initial adjacency matrix, and W represents the parameter matrix. Represents the embedding vector Perform a transpose operation; Represents the embedding vector of node j; For the initial adjacency matrix Normalization is performed to obtain the normalized initial adjacency matrix. The diagonal elements of the degree matrix D are represented as follows: The normalized representation is as follows: . Furthermore, in the introduced state-driven mechanism: Calculate the real-time computing power utilization rate of each computing power node, and calculate the average computing power utilization rate of node i and node j through a sliding window; The efficiency deviation of node task execution is calculated. Based on the efficiency deviation sequence over a preset time period, the Pearson correlation coefficient is used to calculate the efficiency deviation coordination coefficient between node i and node j, expressed as: in, Positive values ​​indicate that the deviation trend is consistent; This represents the deviation in task execution efficiency for node i. This represents the deviation in task execution efficiency for node j. Introducing a data transmission delay attenuation coefficient, expressed as: in, This represents the average transmission latency of the data center's computing nodes. This represents the latency of computing power data transmission between node i and node j; The state-driven weights are calculated based on the average computing power utilization rate, efficiency deviation coordination coefficient, and data transmission delay attenuation coefficient, and are expressed as follows: In the formula, Indicates the weighting coefficient. This indicates that only positive collaborative relationships are retained. This represents the average computing power utilization rate of node i and node j; The state-driven weights are compared with the normalized initial adjacency matrix. Element-wise multiplication yields the adjusted adjacency matrix. ; Sort the elements of each row of the adjusted adjacency matrix in descending order, retain the top k largest weight values, filter out weak correlation noise, and obtain the final dynamic adjacency matrix.

[0012] Furthermore, the step of using the temporal feature vector and interaction graph as input to the graph convolutional neural network (GCN) and updating node features through graph convolution operations includes: The temporal feature vectors are dimension-matched with the dynamic adjacency matrix, and the matrix is ​​concatenated to form the GCN input feature matrix. ; In the local interaction layer, local direct interaction features are aggregated through graph convolution, as shown below: In the formula, Represents a dynamic adjacency matrix. This represents the weight matrix of the local interaction layer, where b represents the bias term. Represents the local feature matrix, and ReLU represents the activation function. Represents the input feature matrix; In the global fusion layer, residual connections are introduced to fuse global indirect correlation features, represented as: In the formula, Represents the second power of the dynamic adjacency matrix. This represents the weight matrix of the global interaction layer. Represents the bias term, connected via residuals. Preserve local features and output globally updated features. ,matrix It also includes the local direct interaction features of each computing node and the cross-regional global association features.

[0013] This invention introduces a multi-head self-attention mechanism to perform global information integration on the node features output by GCN, including: Update the global feature matrix output by GCN. Local contextual features are fused using 1×1 convolution, and then a linear transformation is performed to generate a query vector Q, a key vector K, and a value vector V; where: In the formula, , , Both represent learnable linear transformation matrices; This represents a 1×1 convolution operation, used to aggregate the local neighborhood features of each node; This indicates a global update of the feature matrix; Multiple attention heads are set up, and Q, K, and V are reshaped from the spatial priority dimension to the channel priority dimension by transpose. The attention weight of the h-th head is calculated as follows: In the formula, , , The transpose of the h-th attention head is represented by the vector obtained by splitting Q, K, and V according to the attention head. , This represents the matrix dot product of the key vector and the query vector in the h-th attention head, used to characterize the similarity between channels; represents the scaling factor, and softmax represents the normalization function used to convert the dot product result into attention weights, the sum of which is 1; The output of each attention head is corrected based on the corrected weights, which are expressed as follows: In the formula, This represents the correction factor. Indicates attention head features, This represents the global average value used to characterize the typical computing load characteristics of the entire data center during the current time period. This represents the global standard deviation used to reflect the overall dispersion of the data center's computing load characteristics; c is a constant used to avoid the denominator being zero. The corrected outputs of multiple attention heads are transposed from the channel-priority dimension back to the space-priority dimension and concatenated into a global interactive feature matrix. The features are then integrated through a fully connected layer, and the global features are enhanced through residual connections and layer normalization to output the globally integrated features.

[0014] According to another embodiment of the present invention, a demand forecasting system based on data center computing power operation load is provided. The forecasting system includes the following modules: The data acquisition module is used to collect multi-source data from the data center, including load data of computing nodes, equipment operating parameters, spatial location information, and node operating status parameters. The data processing module is used to perform time-series encoding on the historical load data of each computing node using a Long Short-Term Memory (LSTM) network to obtain a time-series feature vector that characterizes the time-series dependence of computing load. The interaction graph construction module is used to use each computing power node as a node of the interaction graph and the dynamic adjacency matrix calculated by the graph learning layer as the edge of the interaction graph. The load forecasting module takes time-series feature vectors and interaction graphs as inputs to a graph convolutional neural network (GCN) and updates node features through graph convolution operations. The local interaction layer aggregates the local direct interaction features of each computing node, while the global fusion layer fuses global indirect association features through residual connections. A multi-head self-attention mechanism is introduced to integrate global information from the node features output by the GCN. By calculating the attention weights of each computing node's features, it captures global computing power interaction information across regions. The features output by the multi-head self-attention mechanism are then input into a residual decoder, which includes a fully connected layer and a normalization layer. After processing through residual connections, the module outputs the predicted future computing power operation load of the data center and its prediction confidence interval. Compared with existing technologies, the demand forecasting method and system based on data center computing power operation load in this application has the following technical advantages: This application combines a GRU and LSTM parallel network with a computing load mutation detection module to separate long-term computing power trends from short-term mutation characteristics, solving the problem that a single model cannot take into account computing power characteristics at different time scales. At the same time, it introduces a computing power task and load mutation causal graph and an adaptive threshold to integrate the mutation index with data fluctuations and business physical correlation, which greatly reduces the prediction error in computing power mutation scenarios. This application adopts a dual mechanism of state-driven dynamic adjacency matrix and data transmission delay constraint, which breaks through the limitations of fixed graph structure; it adjusts edge weights based on the average computing power utilization rate and the coordination coefficient of task execution efficiency, so that the node association strength changes dynamically with the computing power operation status; it incorporates a data transmission delay attenuation coefficient to quantify the impact of computing power data transmission loss on association strength, making spatial interaction modeling more in line with the computing power layout characteristics of data centers. The GCN local interaction layer of this application aggregates the direct associations of adjacent computing power nodes, and the global fusion layer captures the indirect associations across clusters / data centers through the second power of the adjacency matrix, and avoids feature loss by combining residual connections; the multi-head self-attention mechanism adopts channel-priority dimension calculation, focuses on the channel associations of different computing power types, and strengthens the key node features that deviate from the global benchmark by parameterless weight correction, which significantly improves the accuracy of global interaction capture across regions and computing power types. Attached Figure Description

[0015] 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 1A flowchart illustrating the implementation of the demand forecasting method based on data center computing power operation load provided in this application embodiment; Figure 2 This is a schematic diagram of the prediction model architecture for a demand forecasting method based on data center computing power operation load, as described in an embodiment of this application. Figure 3 This is a time-series coding sub-flowchart of the demand forecasting method based on data center computing power operation load in an embodiment of this application; Figure 4 This is a sub-flowchart of the computing load mutation detection method based on the data center computing power operation load demand forecasting method in this application embodiment; Figure 5 A structural block diagram of a demand forecasting system based on data center computing power operation load provided in the embodiments of this application; Figure 6 This is a hardware structure block diagram of a computer terminal based on the demand forecasting method for data center computing power operation load according to this application. Detailed Implementation

[0016] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0017] According to the embodiments of this application, a method embodiment of a demand forecasting method based on data center computing power operation load is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Also, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0018] like Figure 1 and Figure 2 As shown, in one implementation of the present invention, a demand forecasting method based on data center computing power operation load is provided, including the following steps: Step S101: Collect multi-source data from the data center; The multi-source data provided by this invention includes: Load data for computing nodes, such as CPU utilization, number of computing tasks, and throughput; Equipment operating parameters, such as server start / stop status, heat dissipation power, and power supply stability; Spatial location information, such as data center / rack layout and computing network topology; Node running status parameters, such as computing power utilization and task execution efficiency deviation; The computing nodes in this embodiment of the invention include server nodes, rack computing nodes, computing cluster nodes, AI computing pool nodes, and auxiliary equipment nodes such as heat dissipation and power supply. Furthermore, step S101 of the present invention also includes preprocessing the collected data such as timestamp alignment and outlier removal, using Z-score normalization to process the computing load data, and using min-max normalization to process the device status parameters. Step S102: Use a Long Short-Term Memory (LSTM) network to perform time-series encoding on the historical load data of each computing node to obtain a time-series feature vector that characterizes the time-series dependence of computing load. In step S102 of the present invention, historical load data of each computing power node and equipment operation status parameters of the corresponding time period are obtained, the computing power load data and equipment operation status parameters are concatenated into a multi-source sequence, an LSTM network that integrates computing power load change perception is constructed, the preprocessed multi-source sequence is input into the network, and a time-series feature vector is output. The LSTM network consists of a parallel structure of GRU and LSTM at the bottom layer. GRU is used to capture short-term computing power fluctuations, while LSTM is used to capture long-term computing power trends. The middle layer dynamically weights the hybrid temporal features through a computing power load mutation detection module. The top layer aggregates features and reduces dimensionality through a periodic attention layer to obtain a temporal feature vector that integrates multi-dimensional information.

[0019] Specifically, such as Figure 3 As shown, in step S102, the step of using a Long Short-Time Memory (LSTM) network to perform time-series encoding of the historical load data of each computing node includes: Step S1021: Obtain historical load data and corresponding equipment operating status parameters for each computing node; The device operation status parameters provided by this invention include server start / stop status, real-time computing power conversion efficiency, and task scheduling strategy parameters. Step S1022: Concatenate the computing load data and equipment operating status parameters into a multi-source sequence; Among them, computing load data is time-series data reflecting the computing power consumption of the data center, such as CPU utilization, memory usage, number of computing tasks executed, and task processing latency; equipment operating status parameters are real-time status data of computing node hardware, such as server start / stop status, cooling system power, power supply stability, and data transmission bandwidth utilization. During the concatenation process, the computing load data and equipment operating status parameters are first timestamped and aligned. Then, multiple feature vectors are constructed according to time steps. This includes: for each time point (e.g., time t), the computing load data and equipment operating status parameters at that time are concatenated along the feature dimension to form a multi-source feature vector for that time step. Combining the feature vectors from all time steps yields the final multi-source sequence. By concatenating the two types of data to form a multi-source sequence, the subsequent time-series coding module can simultaneously capture the temporal patterns of computing power consumption and the correlation between hardware status and computing load, thus solving the problem of one-sided feature representation caused by insufficient information from a single data dimension. Please continue to refer to... Figure 3 In step S102, the step of using a Long Short-Time Memory (LSTM) network to perform time-series encoding of the historical load data of each computing node further includes: Step S1023: Construct an LSTM network that integrates computing load change awareness, using the preprocessed multi-source sequences as input to the LSTM network, and outputting a temporal feature vector; This invention constructs an LSTM network for sensing sudden changes in computing load, comprising a bottom-layer network, a middle-layer network, and a top-layer network. The bottom-layer network is a parallel structure of Gated Recurrent Units (GRUs) and an LSTM. The GRUs are used to capture short-term computing load fluctuations, while the LSTM captures long-term computing load trends. Preprocessed multi-source sequences are used as input to the bottom-layer network, outputting mixed temporal features. In the middle-layer network, a computing load change detection module is introduced, which dynamically weights the mixed temporal features to obtain a weighted mixed feature sequence. The top-layer network is a periodic attention layer that performs matrix operations with a preset periodic mask on the mixed feature sequence. Multiplication is performed, and attention weights at each time step are calculated through attention to weighted aggregation of the feature sequences. The aggregated features output by the periodic attention layer are then dimensionality-reduced through a fully connected layer to obtain a temporal feature vector. This temporal feature vector integrates long-term computing power trends, short-term mutation features, and equipment status correlation information. This application achieves the separation of long-term computing power trends and short-term mutation features by combining a GRU and LSTM parallel network with a computing power load mutation detection module, solving the problem that a single model cannot take into account computing power features at different time scales. At the same time, a causal graph of computing power tasks and load mutations and an adaptive threshold are introduced to integrate the mutation index with data fluctuations and business physical correlations, significantly reducing prediction errors in computing power mutation scenarios.

[0020] like Figure 4 As shown, in one implementation of the present invention, the step of dynamically weighting the mixed time-series features by the computing load mutation detection module includes: Step S10231: Construct a causal graph of sudden changes in computing power task status and computing power load, and mine the time correlation rules of sudden tasks, scheduling strategy changes and sudden changes in computing power load based on historical data, and quantify the rules into a causal coefficient matrix; Computing task status includes task type, task start / end time, and task size; scheduling policy changes include policy adjustment type and change execution time; computing load includes the time point of load mutation, mutation direction, and mutation magnitude. This invention integrates three types of data into an event load mutation dataset based on timestamps; and uses a time series association analysis algorithm to mine the time association rules between task and strategy changes and computing load mutations. The discovered association rules are further transformed into a causal coefficient matrix. The matrix's behavior represents the computing power task / scheduling strategy type, the matrix's columns represent the computing power load mutation type, and the matrix element values ​​are the association strength coefficients. Step S10232: The quantiles of historical computing load fluctuations during the same period are statistically analyzed using a sliding window and used as the adaptive threshold at time t. This invention calculates the absolute change in load between adjacent time steps based on historical computing load data during the same period within the sliding window to obtain the computing load fluctuation. By using an adaptive threshold, it solves the problem that fixed thresholds can easily lead to normal fluctuations during peak periods being misjudged as sudden changes and true sudden changes during trough periods being missed. Based on the adaptive threshold of historical data, this invention can match the computing load fluctuation patterns of different time periods and reduce the false detection rate and false negative rate of sudden change judgment. Step S10233: Calculate the mutation index based on the causal map and adaptive threshold; The basic mutation index provided by this invention is expressed as follows: In the formula, This represents the adaptive threshold at time t. This represents the normalized computing load value at time t. This represents the normalized computing load value at time t-1. It represents the absolute change in computing load between adjacent time steps; Step S10234: Introduce a causal correction index to strengthen the causal relationship between computing task status and sudden changes in computing load; The causal correction index provided by this invention is expressed as follows: in, This represents the normalized value of the state parameter of the k-th type of computing power device at time t. Represents the elements of the causal coefficient matrix; Step S10235: Construct a nonlinear weighting function, multiply the nonlinear weights element-wise with the mixed time-series features, and simultaneously apply enhancement coefficients to the feature dimensions with correlation strength exceeding the threshold in the mutation causal graph based on the causal coefficient matrix to obtain the weighted mixed feature sequence; The nonlinear weighting function used in this invention dynamically adjusts the time step weights based on the causal correction index, so that the features at the mutation moment are retained as the focus, while the redundant features in the normal period are weakened.

[0021] like Figure 1 and Figure 2 As shown, in one implementation of the present invention, a demand forecasting method based on data center computing power operation load is provided, including the following steps: Step S103: Treat each computing power node as a node in the interaction graph, and treat the dynamic adjacency matrix calculated by the graph learning layer as an edge in the interaction graph; In step S103, the present invention uses computing power nodes as vertices and calculates dynamic adjacency matrices as edges through a graph learning layer; the graph learning layer extracts multidimensional features of computing power nodes and maps them into embedding vectors, and calculates asymmetric dependencies to generate an initial adjacency matrix; In one implementation of the present invention, traditional computing node interaction graphs often use fixed adjacency matrices, such as setting association relationships based on physical topology. Therefore, fixed adjacency matrices cannot reflect changes in the real-time operating status of nodes. The state-driven mechanism of the present invention adjusts edge weights through computing power utilization and efficiency coordination coefficients, so that the association relationships of the interaction graph are dynamically updated with the operating status of computing nodes, which is more in line with the actual scenario of data center computing power operation. Specifically, the present invention uses the dynamic adjacency matrix calculated by the graph learning layer as the edge of the interaction graph. The steps include: in the graph learning layer, using the two fully connected layers of the graph learning layer to map the features of the computing power nodes into embedding vectors, calculating the asymmetric dependencies of the embedding vectors to generate the initial adjacency matrix; introducing a state-driven mechanism to adjust the initial adjacency matrix based on the computing power node running state parameters, using the weighted sum of the average computing power utilization rate and the task execution efficiency coordination coefficient as the edge weight, and retaining the top-k strong dependencies to filter weak correlation noise. The top-k filtering can filter weak connections that are physically connected but have no correlation in running state, avoiding noise interference with the feature extraction accuracy of subsequent graph convolution operations. Furthermore, in one implementation of the present invention, in the graph learning layer, multidimensional features of each computing node are extracted. These multidimensional features include the mean and standard deviation of the computing load data, the rated computing power of the equipment, real-time operating efficiency, and the computing network topology coordinates. Calculate the Euclidean distance between node i and node j. The process involves concatenating multidimensional features into a node feature vector of dimension d; constructing a first fully connected network layer using the ReLU activation function to map the node feature vector from d dimensions to a 64-dimensional hidden feature vector; constructing a second fully connected network layer using the LeakyReLU activation function to map the 64-dimensional hidden feature vector to a 32-dimensional embedding vector; and calculating the asymmetric dependencies of the embedding vectors using the parameter matrix. Perform a linear transformation on the embedding vector and calculate the dependency weights between node i and node j, expressed as: In the formula, Let represent the elements of the initial adjacency matrix, and W represents the parameter matrix. Represents the embedding vector Perform a transpose operation; Represents the embedding vector of node j; For the initial adjacency matrix Normalization is performed to obtain the normalized initial adjacency matrix. The diagonal elements of the degree matrix D are represented as follows: The normalized representation is as follows: .

[0022] Furthermore, in step S103 of this invention, by introducing a state-driven mechanism, the initial adjacency matrix is ​​adjusted by combining the computing power utilization rate, the task execution efficiency coordination coefficient, and the data transmission delay attenuation coefficient, while retaining the top-k strong dependency relationship and filtering noise, to obtain the final dynamic adjacency matrix, and a computing power node interaction graph is constructed based on this. Specifically, in the state-driven mechanism introduced in this invention, the real-time computing power utilization rate of each computing node is calculated, and the average computing power utilization rate of node i and node j is calculated through a sliding window; the node task execution efficiency deviation value is calculated, and based on the efficiency deviation sequence over a preset time period, the Pearson correlation coefficient is used to calculate the efficiency deviation coordination coefficient between node i and node j, expressed as: in, Positive values ​​indicate that the deviation trend is consistent; This represents the deviation in task execution efficiency for node i. This represents the deviation in task execution efficiency for node j. This invention introduces a data transmission delay attenuation coefficient, expressed as: in, This represents the average transmission latency of the data center's computing nodes. This represents the latency of computing power data transmission between node i and node j; This invention also calculates the state-driven weights based on the average computing power utilization rate, efficiency deviation coordination coefficient, and data transmission delay attenuation coefficient, expressed as: In the formula, Indicates the weighting coefficient. This indicates that only positive collaborative relationships are retained. This represents the average computing power utilization rate of node i and node j; The state-driven weights are compared with the normalized initial adjacency matrix. Element-wise multiplication yields the adjusted adjacency matrix. Sort the elements of each row of the adjusted adjacency matrix in descending order, retain the top k largest weight values, filter out weak correlation noise, and obtain the final dynamic adjacency matrix. This application employs a dual mechanism of state-driven dynamic adjacency matrix and data transmission delay constraint, overcoming the limitations of fixed graph structures. Furthermore, it adjusts edge weights based on the average computing power utilization rate and the coordination coefficient of task execution efficiency, causing the node association strength to dynamically change with the computing power operation status. By incorporating a dynamic adjacency matrix with a decay coefficient, this invention can filter out false association node pairs with similar operating states but extremely high transmission delays. This allows the Graph Convolutional Neural Network (GCN) to focus only on nodes with feasible physical transmission and close actual interactions when aggregating node features, avoiding interference from invalid associations on feature extraction and improving the accuracy of computing power load prediction.

[0023] like Figure 1 and Figure 2 As shown, in one implementation of the present invention, a demand forecasting method based on data center computing power operation load is provided, including the following steps: Step S104: The temporal feature vector and interaction graph are used as inputs to the Graph Convolutional Neural Network (GCN), and the node features are updated through graph convolution operations. The local interaction layer aggregates the local direct interaction features of each computing node, and the global fusion layer fuses the global indirect association features through residual connections. A multi-head self-attention mechanism is introduced to integrate the global information of the node features output by the GCN. The global computing power interaction information across regions is captured by calculating the attention weights of the features of each computing node. The features output by the multi-head self-attention mechanism are input into the residual decoder. The residual decoder includes a fully connected layer and a normalization layer. After processing through residual connections, the predicted value of the future computing power operation load of the data center and the prediction confidence interval are output.

[0024] In step S104 of the present invention, the temporal feature vector and the interaction graph of computing power nodes are input into the GCN, and the temporal pattern and spatial correlation are combined through graph convolution operation so that the prediction result fits the temporal change trend of computing power load. In step S104 of the present invention, the temporal feature vector is dimension-matched with the dynamic adjacency matrix to form the GCN input feature matrix, which is then input into the GCN for graph convolution operation. The GCN includes a local interaction layer and a global fusion layer. The local interaction layer aggregates the local direct interaction features of the first-order neighborhood. The global fusion layer captures the second-order indirect associations through the second power of the adjacency matrix and retains local features by combining residual connections. Furthermore, this invention integrates the node features output by GCN with global information by introducing a multi-head self-attention mechanism, calculates the attention weight of each computing node to capture cross-regional computing power interaction, and then inputs the integrated features into the residual decoder to output the predicted value of the future computing power operation load of the data center and the 95% prediction confidence interval, so as to quantify the uncertainty of the prediction.

[0025] In one implementation of the present invention, the step of updating node features by using temporal feature vectors and interaction graphs as inputs to a graph convolutional neural network (GCN) and performing graph convolution operations includes: The temporal feature vectors are dimension-matched with the dynamic adjacency matrix, and the matrix is ​​concatenated to form the GCN input feature matrix. ; In the local interaction layer, local direct interaction features are aggregated through graph convolution, as shown below: In the formula, Represents a dynamic adjacency matrix. This represents the weight matrix of the local interaction layer, where b represents the bias term. Represents the local feature matrix, and ReLU represents the activation function. Represents the input feature matrix; In the global fusion layer, residual connections are introduced to fuse global indirect correlation features, represented as: In the formula, Represents the second power of the dynamic adjacency matrix. This represents the weight matrix of the global interaction layer. Represents the bias term, connected via residuals. Preserve local features and output globally updated features. ,matrix It also includes the local direct interaction features of each computing node and the cross-regional global association features.

[0026] In one implementation of the present invention, the step of introducing a multi-head self-attention mechanism to perform global information integration of node features output by the GCN includes: Update the global feature matrix output by GCN. Local contextual features are fused using 1×1 convolution, and then a linear transformation is performed to generate a query vector Q, a key vector K, and a value vector V; where: In the formula, , , Both represent learnable linear transformation matrices; This represents a 1×1 convolution operation, used to aggregate the local neighborhood features of each node; This indicates a global update of the feature matrix; Multiple attention heads are set up, and Q, K, and V are reshaped from the spatial priority dimension to the channel priority dimension by transpose. The attention weight of the h-th head is calculated as follows: In the formula, , , The transpose of the h-th attention head is represented by the vector obtained by splitting Q, K, and V according to the attention head. , This represents the matrix dot product of the key vector and the query vector in the h-th attention head, used to characterize the similarity between channels; represents the scaling factor, and softmax represents the normalization function used to convert the dot product result into attention weights, the sum of which is 1; The output of each attention head is corrected based on the corrected weights, which are expressed as follows: In the formula, This represents the correction factor. Indicating attentional characteristics, This represents the global average value used to characterize the typical computing load characteristics of the entire data center during the current time period. This represents the global standard deviation used to reflect the overall dispersion of the data center's computing load characteristics; c is a constant used to avoid the denominator being zero. This invention transposes the corrected outputs of multiple attention heads from the channel-priority dimension back to the space-priority dimension and concatenates them into a global interactive feature matrix. The features are then integrated through a fully connected layer, and the global features are enhanced through residual connections and layer normalization to output the globally integrated features.

[0027] As can be seen, in step S104 of the present invention, the GCN local interaction layer aggregates the direct associations of adjacent computing power nodes, the global fusion layer captures the indirect associations across clusters / data centers through the adjacency matrix quadratic power, and avoids feature loss by combining residual connections; the multi-head self-attention mechanism adopts channel priority dimension calculation, focuses on the channel associations of different computing power types, and strengthens the key node features that deviate from the global benchmark through parameterless correction weights, which significantly improves the global interaction capture accuracy across regions and computing power types.

[0028] like Figure 5 As shown, in another embodiment of the present invention, a demand forecasting system based on data center computing power operation load is provided. The forecasting system includes the following modules: The data acquisition module 201 is used to collect multi-source data from the data center, including load data of computing nodes, equipment operating parameters, spatial location information and node operating status parameters. Data processing module 202 is used to perform time-series encoding on the historical load data of each computing node using a Long Short-Term Memory (LSTM) network to obtain a time-series feature vector that characterizes the time-series dependence of computing load. The interaction graph construction module 203 is used to use each computing power node as a node of the interaction graph and the dynamic adjacency matrix calculated by the graph learning layer as the edge of the interaction graph. The load prediction module 204 is used to take the time-series feature vector and interaction graph as input to the graph convolutional neural network (GCN) and update the node features through graph convolution operations. Among them, the local interaction layer aggregates the local direct interaction features of each computing node, and the global fusion layer fuses the global indirect correlation features through residual connections. A multi-head self-attention mechanism is introduced to integrate the global information of the node features output by the GCN. By calculating the attention weight of the features of each computing node, the global computing power interaction information across regions is captured, and the features output by the multi-head self-attention mechanism are input into the residual decoder. The residual decoder includes a fully connected layer and a normalization layer. After processing through residual connections, the predicted value of the future computing power operation load of the data center and the prediction confidence interval are output.

[0029] like Figure 6 As shown, according to another embodiment of this application, a computer device is provided, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the demand forecasting method based on data center computing power operation load as described in any of the above embodiments.

[0030] The computer equipment can be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal equipment may include, but is not limited to, processors and memory.

[0031] The processor can be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, off-the-shelf programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting various parts of the terminal device via various interfaces and lines.

[0032] The memory can be used to store the computer program. The processor implements various functions of the terminal device by running or executing the computer program stored in the memory and calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function, etc.; the data storage area may store data created based on the use of the mobile phone, etc.

[0033] In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart memory card, secure digital card, flash memory card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0034] Another preferred embodiment of the present invention provides a storage medium, which is a computer-readable storage medium, and a computer program is stored in the computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of the demand forecasting method based on data center computing power operation load of the above embodiments.

[0035] The computer program includes computer program code, which may be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory, random access memory, electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0036] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for demand prediction based on operation load of data center computing power, characterized in that, The method comprises the following steps: Collecting multi-source data of the data center, the multi-source data comprising load data, equipment operation parameters, spatial position information and node operation state parameters of the computing power nodes; Using a long short-term memory network (LSTM) to perform time series coding on historical load data of each computing power node to obtain a time series feature vector for representing time series dependence features of the computing power load; Taking each computing power node as a node of an interaction graph and taking a dynamic adjacency matrix calculated by a graph learning layer as an edge of the interaction graph; Taking the time series feature vector and the interaction graph as inputs of a graph convolutional neural network (GCN) and updating node features by graph convolution operation; wherein, a local interaction layer aggregates local direct interaction features of each computing power node, and a global fusion layer fuses global indirect correlation features through a residual connection; Introducing a multi-head self-attention mechanism to integrate global information of node features output by the GCN, capturing global computing power interaction information across regions by calculating attention weights of features of each computing power node, and inputting features output by the multi-head self-attention mechanism into a residual decoder, the residual decoder comprising a fully connected layer and a normalization layer, and outputting a future computing power operation load prediction value and a prediction confidence interval of the data center after processing by the residual connection. 2.The method of claim 1, wherein, The step of using a long short-term memory network (LSTM) to perform time series coding on historical load data of each computing power node comprises: Obtaining historical load data of each computing power node and equipment operation state parameters corresponding to a time period; Concatenating the computing power load data and the equipment operation state parameters into multi-source sequences; Constructing an LSTM network fusing computing power load mutation awareness, taking the preprocessed multi-source sequences as inputs of the LSTM network, and outputting a time series feature vector. 3.The method of claim 2, wherein, In the construction of the LSTM network fusing computing power load mutation awareness: The bottom layer is a network structure in which a gated recurrent unit (GRU) and the LSTM are parallel; The middle layer introduces a computing power load mutation detection module; The top layer is a periodic attention layer, which performs matrix multiplication on a mixed feature sequence and a preset periodic mask, calculates attention weights of each time step by attention, and aggregates the feature sequence by weighting; The aggregated features output by the periodic attention layer are reduced in dimension by a fully connected layer to obtain a time series feature vector.

4. The method of claim 3, wherein, The step of dynamically weighting the mixed time series features by the computing power load mutation detection module comprises: Constructing a mutation causal graph of computing power task states and computing power loads, mining time correlation rules of burst tasks, scheduling strategy changes and computing power load mutations based on historical data, and quantizing the rules into a causal coefficient matrix; Using a sliding window to statistically calculate quantiles of historical same-period computing power load fluctuations as adaptive thresholds; Calculating a mutation index based on the causal graph and the adaptive thresholds; Introducing a causal correction index for strengthening causal correlation between computing power task states and computing power load mutations; Constructing a nonlinear weight function, multiplying nonlinear weights and mixed time series features element by element, and applying an enhancement coefficient to feature dimensions with correlation strengths exceeding thresholds in the mutation causal graph based on the causal coefficient matrix to obtain a weighted mixed feature sequence.

5. The method of claim 4, wherein, The step of taking a dynamic adjacency matrix calculated by a graph learning layer as an edge of an interaction graph comprises: In the graph learning layer, the computing power node features are mapped into embedding vectors by a two-layer fully connected network of the graph learning layer, and an initial adjacency matrix is generated based on the asymmetric dependency of the embedding vectors; A state-driven mechanism is introduced to adjust the initial adjacency matrix based on the running state parameters of the computing power nodes, and the weighted sum of the average computing power utilization rate and the task execution efficiency synergy coefficient is used as the edge weight, and the top-k strong dependency relationship is retained to filter weakly associated noise.

6. The method of claim 5, wherein, In the graph learning layer: Extract the multi-dimensional features of each computing power node, calculate the Euclidean distance between node i and node j Splice the multi-dimensional features into a node feature vector with dimension d; A first layer fully connected network is constructed, which uses a ReLU activation function to map the node feature vector from d dimensions to a 64-dimensional hidden feature vector; A second layer fully connected network is constructed, which uses a LeakyReLU activation function to map the 64-dimensional hidden feature vector to a 32-dimensional embedding vector; Asymmetric dependency of the computed embedding vectors, through a parameter matrix Linear transformation of the embedding vectors, to compute the dependency weight of node i and node j, denoted as: wherein, denotes an element of the initial adjacency matrix, and W denotes a parameter matrix, denotes the transpose operation on the embedding vector of node j; denotes the embedding vector of node j; to the initial adjacency matrix is normalized to obtain a normalized initial adjacency matrix where the diagonal elements of the degree matrix D are expressed as: The normalization is expressed as: .

7. The method of claim 6, wherein, In the introduced state-driven mechanism: The real-time computing power utilization rate of each computing power node is calculated, and the computing power utilization rate of node i and node j is calculated by a sliding window; The node task execution efficiency bias value is calculated, and based on the efficiency bias sequence of the preset time period, the Pearson correlation coefficient is used to calculate the efficiency bias synergy coefficient of node i and node j; A data transmission delay attenuation coefficient is introduced; The state-driven weight is calculated based on the average computing power utilization rate, the efficiency bias synergy coefficient and the data transmission delay attenuation coefficient; state-driven weights to a normalized initial adjacency matrix element-wise multiplication, resulting in an adjusted adjacency matrix; The elements of each row of the adjusted adjacency matrix are sorted in descending order, and the top-k largest weight values are retained to filter weakly associated noise, and the final dynamic adjacency matrix is obtained. 8.The method of claim 7, wherein, The time series feature vector and the interaction graph are used as the input of the graph convolutional neural network GCN, and the steps of updating the node features by graph convolution operation include: The time sequence feature vector is dimensionally matched with the dynamic adjacency matrix, and a GCN input feature matrix is formed through matrix splicing ; In the local interaction layer, the local direct interaction features are aggregated by graph convolution, represented as: wherein, denotes a dynamic adjacency matrix, denotes a weight matrix of the local interaction layer, b denotes a bias term, denotes a local feature matrix, ReLU denotes an activation function, denotes an input feature matrix; In the global fusion layer, residual connection is introduced to fuse global indirect association features, represented as: In the formula, denotes the square of the dynamic adjacency matrix, denotes the weight matrix of the global interaction layer, denotes the bias term, and the residual connection retains local features and outputs global updated features , the matrix simultaneously contains local direct interaction features and cross-region global correlation features of each computing node. 9.The method of claim 8, wherein, The steps of introducing multi-head self-attention mechanism to integrate global information of node features output by GCN include: Update the global feature matrix output by GCN. Local contextual features are fused using 1×1 convolution, and then a linear transformation is performed to generate a query vector Q, a key vector K, and a value vector V; where: wherein, , , all represent learnable linear transformation matrices; represents a 1x1 convolution operation to aggregate local neighborhood features of each node; represents a global updated feature matrix; Multiple attention heads are set, Q, K, and V are reshaped from spatial priority dimension to channel priority dimension by transposition, and the output of each attention head is modified based on the modified weight; The modified outputs of multiple attention heads are transposed from channel priority dimension to spatial priority dimension, and are concatenated into a global interaction feature matrix, which is integrated by a fully connected layer, and then the global features are enhanced by residual connection and layer normalization, and the global integrated features are output.

10. A prediction system for implementing the method of demand prediction based on data center computing power operational load according to any one of claims 1 to 9, characterized in that, The prediction system includes the following modules: A data acquisition module is used to collect multi-source data of the data center, including load data, device running parameters, spatial location information and node running state parameters of the computing power nodes; A data processing module is used to encode the historical load data of each computing power node by a long short-term memory network LSTM to obtain a time series feature vector representing the time series dependency features of the computing power load; An interaction graph construction module is used to construct an interaction graph by taking each computing power node as a node and taking the dynamic adjacency matrix calculated by the graph learning layer as an edge. The load prediction module is configured to take the time sequence feature vector and the interaction graph as inputs of a graph convolutional neural network (GCN), and update node features through graph convolution operation; wherein, the local interaction layer aggregates local direct interaction features of each computing power node, and the global fusion layer fuses global indirect correlation features through residual connection; A multi-head self-attention mechanism is introduced to integrate global information of the node features output by the GCN. The multi-head self-attention mechanism captures global computing power interaction information across regions by calculating attention weights of the computing power node features, and inputs the features output by the multi-head self-attention mechanism into a residual decoder. The residual decoder includes a full connection layer and a normalization layer. After processing through the residual connection, a future computing power operation load prediction value and a prediction confidence interval of the data center are output.

Images